Datasets:

datajuicer

/

redpajama-arxiv-refined-by-data-juicer

like 1

Follow

Data-Juicer 10

\section{Introduction} With the escalating growth of mobile data traffic caused by the proliferation of smart mobile devices, and the growing online services (i.e., Video on Demand (VoD) streaming, Facebook, Twitter,...), corresponding energy consumption is increasing considerably \cite{cisco}. Furthermore, video streaming and social applications require high bandwidth and strict delay constraint, which in turn has further adverse effects on user experience quality and Operational Expenses (OpEx). In addition, It has been reported that the network data traffic will continue growing in the future years \cite{cisco}. Hence, efficient energy utilization is of paramount importance in the design of future wireless networks. In this paper, our objective is to exploit the cache-enabled user devices to reduce the energy consumption by focusing particularly on the wireless transmission cost The massive mobile users generate issues to be addressed, such as, extremely high throughput and stringent Quality of Service (QoS) requirements, and excessive energy consumption. One way to deal with such issues is to deploy larger number of small cells and fog style access points \cite{6171992,6476878}. However, this approach does not address the energy efficiency and still large undesirable latency and network congestion could be induced during peak-traffic hours. Hence, content cashing at the edge of wireless networks has attracted much attention from both academia and industry \cite{cachinginwnetworks,latency,fundamental_caching_madah,maria16,contpushing_with_rdi_17}. Moving contents to the edge of network emerged as a prospective technique, that can significantly reduce transmission latency\cite{cachinginwnetworks,latency}, control traffic load \cite{fundamental_caching_madah}, and reduce energy consumption \cite{maria16}, or can be employed to increase overall system throughput\cite{contpushing_with_rdi_17}. Content caching utilizes the recent advances in the field of context awareness and leverages on the low price memory storage to enhance the system performance. Content caching enables proactive transmission (i.e., \textit{pre-serve}) that allows the user to download content files over a longer period, hence, reducing energy consumption. It can also mitigate severe network condition (i.e.,peak-traffic hours) by pushing at favorable transmission times. For instance, by exploiting the user demand information the base station (BS) can push the desired content files prior to the playback time (i.e., in video streaming). Hence, saving a significant amount of energy by avoiding the unfavorable channel conditions. It also provides another advantage by shifting traffic load (or equivalently reducing traffic variability) for better system utilization. In another line of research, the objective is to increase the throughput of content-centric wireless networks to better support the exponential growth of wireless data traffic. In \cite{contpushing_with_rdi_17} the available cache memory at the user is exploited via a joint pushing and caching method to increase the system throughput under non-causal, statistical, and causal knowledge of user request delay information (RDI). In this paper, we consider an optimization problem of a proactive caching wireless communications channel, with limited data buffer capacity available at the receiver end. Specifically, we formulate the joint pushing and caching (JPC) problem as an infinite horizon average cost Markov decision process (MDP) and devise a randomized policy to minimize the average energy consumption over time. In each timeslot the amount of data to transmit is a random variable, whose distribution is decided by the optimal policy, taking both the buffer occupancy and user requests into account. Numerically solving the optimization problem is computationally demanding due to the two-dimensional state space and randomized policy, known as the curse of dimensionality. Therefore, we develop a novel approach to decouple the factors of buffer occupancy and user requests by introducing the degenerated state space, and hence breakdown the optimization into three more tractable sub-problems. Furthermore, we found a structural property that gives a non-iterative way to design the optimal policy under certain constraints in the degenerated space. We name the property as \textit{generalized monotonicity}, which brings significant improvement to the computational complexity. The model in this work performs well in comparison against two theoretical bounds, which were derived from assuming non-causal knowledge of user request delay information (RDI)\cite{Zafer2009} and unlimited cache capacity, respectively. \section{System Model} Consider a wireless communication system between a server and a user equipped with a limited buffer as depicted in Fig. \ref{System_Model}. The system operates over an infinite time horizon in a discrete time fashion, with timeslots $t=0,1,2,...$. At the beginning of each timeslot, the user requests for a certain amount of data, which must be fulfilled by the end of the timeslot. The data may be transmitted in advance, stored and read from the buffer (proactively), or be transmitted on-demand (reactively), or combined. Also, we consider the scenario where data stream has been temporally ordered, i.e., the server knows what to be requested in the near future, while not knowing how soon the requests will be made. \begin{figure}[htbp] \begin{center} \vspace{-0.5cm} \includegraphics[width=0.5\textwidth]{system_model2} \vspace{-1cm} \caption[]{System model. As explained in section (\ref{sec2a}), fetching data from the buffer always has higher priority over via on-demand transmission, based on the assumption of temporal order. This observation gives the data flow on each edge.} \label{System_Model} \end{center} \end{figure} \subsection{Request model and buffer state}\label{sec2a} We assume that data is requested in content items of identical size. Thus, we may denote the number of content items requested in timeslot t as $x_t \in\mathcal{X}\triangleq\{0,1,\dots,X\}$, where $X$ is the maximum possible value of $x_t$. We further assume that $x_t, t=0,1,\dots$ are \textit{i.i.d.} integer random variables bounded by $X$, yielding probability mass function (p.m.f) $f_X(\cdot)$. One can always limit $x_t$ to integer values by choosing a small enough size for content items. Let a $(X+1)$-dimensional vector $\bm{p}$ denotes the p.m.f. of $x_t$, i.e., $p_x=f_X(x), x\in\mathcal{X}$. Let buffer state $b_t$ denote the number of content items at the beginning of timeslot $t$, $b_t\in\mathcal{B}\triangleq\{0, 1, \dots,B\}$.This implies the buffer storage capacity being $B$ content items. Since data stream is well-ordered, for any two content items, we always know which one would be requested first, and thus which to proactively push first. Therefore, as long long as the buffer is not empty, there is no reason to transmit data on-demand. In other words, on-demand transmission happens if and only if $x_t>b_t$. \subsection{System State Model} The system state space is denoted by $\mathcal{S}$ which consists of $b$ and $x$ elements and has the cardinality of $(B+1)\times(X+1)$. The pair $(b_t,x_t)$ constitutes each state $s_t\in \mathcal{S}$. Furthermore, we define degenerated state $\mathfrak{b}_b=\{s: s=(b,x), x\in\mathcal{X}\}$, and degenerated state space $\mathfrak{B}=\{\mathfrak{b}_b: b\in\mathcal{B}\}$. When it does not cause confusion, we use $\mathfrak{b}_b$ and $b$ interchangeably. Let $y_t$ denote the number of content items downloaded within timeslot $t$. Under a given policy, $y_t$ is a random variable with distribution being contingent on buffer occupancy $b_t$ and user requests $x_t$, i.e., $s_t$. Buffer occupancy evolves as a time-homogeneous Markov chain which can be described as \begin{equation} b_{t+1}=b_t+y_t-x_t, \label{bt_evolve} \end{equation} where \begin{equation}\begin{aligned} 0\leq b_t\leq B, && \forall t. \end{aligned} \label{bt_constraint} \end{equation} the first inequality in (\ref{bt_constraint}) comes from the model setting that requests must be fulfilled within the requested timeslot $t$, and thus $b_t+y_t\geq x_t$. The Fig. \ref{cache_state} depicts the dynamical relation between $y_t$, $b_t$ and $x_t$ in each timeslot. A policy $\pi$ would determine the distribution of $y_t$ from $b_t$ and $x_t$. \begin{figure}[htbp] \begin{center} \vspace{-0.6cm} \includegraphics[width=0.4\textwidth]{Drawing42_b_ocpnsy3} \vspace{-0.6cm} \caption[]{Buffer state evolution. An illustration of the relationship between the buffer occupancy $b_t$, user requests $x_t$, and data transmission $y_t$ in a timeslot.} \label{cache_state} \end{center} \end{figure} We consider a causal knowledge of user RDI, that the user requests for a future timeslot stays unknown until the beginning of the timeslot. The transmission action in each timeslot incurs a certain energy cost, whose corresponding power is typically a convex and continuous function of data rate. Conventionally we assume this function to be exponential\cite{gungurgunduz}. By \cite{Zafer2009}, constant transmission rate should be used within a timeslot to minimize energy consumption. Absorbing all constants and normalizing the scalar in our model, we give the energy consumption in timeslot $t$ by: \begin{equation} \rho_t=\rho(y_t)=\eta^{y_t}-1, \label{eq3} \end{equation} where $\rho_t$ is the energy consumption and $\eta>1$ is a constant. Since, $y_t$ takes finitely many values, it can be shown that $\mathbf{E}\rho_t<\infty$. \section{A Randomized MDP problem Formulation} In a randomized policy, a transmission action $y_t$ yields a distribution which we design contingent on $s_t$, i.e., $y_t \sim f_{Y|s_t}(y)$. The optimal transmission policy selects a distribution for $y_t$. However, the range where random variable $y_t$ takes non-zero probability changes given different $s_t$. By (\ref{bt_evolve}), $\sigma(y_t\mid s_t) = \sigma(b_{t+1}|s_t)$. Therefore, we study the conditional probability of $b_{t+1}$ instead of $y_t$. The system evolves as depicted in Fig. \ref{policy1}. Let a $(B+1)$-dimensional vector $\bm{d}^{s,\pi}$ denote a randomized decision, i.e., p.m.f., conditioning on state $s$ under policy $\pi$, where $d^{s,\pi}_{b^+}\triangleq\Pr\left(b_{t+1}={b^+}\mid s_t=s,\pi\right)$. All $(\bm{d}^{s,\pi})_{s\in\mathcal{S}}$ give the transition probabilities in $\mathcal{S}$ and $\mathfrak{B}$. Denote the transition probability matrix in $\mathfrak{B}$ under policy $\pi$ by $\bm{A}^\pi\triangleq{[\bm{a}^{0,\pi},\bm{a}^{1,\pi},...,\bm{a}^{B,\pi}]}^\intercal$, where $\bm{a}^{b,\pi}=\sum_{x}\bm{d}^{(b,x),\pi}p_x$. \begin{figure}[h] \begin{center} \vspace{-.5cm} \includegraphics[width=0.5\textwidth]{Drawing42_policy11} \vspace{-1cm} \caption[]{At each timeslot $t$ the controller, i.e., the server, observes the current system state $s_t$ and applies a control $y_t\sim f_{Y\mid s_t}(\cdot)$ (and equivalently, $b_{t+1}\sim\bm{d}^{s_t,\pi}$) contingent on the state. \label{policy1 \end{center \end{figure Since every state in $\mathcal{S}$ has to be assigned with a randomized decision, the corresponding policy space is $(B+1)^2\times(X+1)$- dimensional. Denote the probability that the system occupies state $(b,x)\in\mathcal{S}$ in timeslot $t$ under policy $\pi$ as $q_{b,x}^{t,\pi}\triangleq\Pr\left(s_t=(b,x)\mid\pi\right)$, and define matrix $\bm{Q}^{t,\pi}\triangleq{(q^{t,\pi}_{b,x})}_{b\in\mathcal{B},x\in\mathcal{X}}$, which contains the probabilities for all states. Similarly, we denote the state probability of $\mathfrak{b}_b\in\mathfrak{B}$ in timeslot $t$ under policy $\pi$ as $r_b^{t,\pi}\triangleq\Pr\{b_t=b\mid\pi\}$, and define vector $\bm{r}^{t,\pi}\triangleq(r_b^{t,\pi})_{b\in\mathcal{B}}$. Since, $x_t\sim f_X(\cdot)$ is i.i.d., we have \begin{equation} \bm{Q}^{t,\pi}=\bm{r}^{t,\pi}{\bm{p}}^\intercal,\hspace{0.5cm} \forall t>0, \pi. \end{equation} Under a given policy $\pi$, the expected energy cost in timeslot $t$ solely depends on state $s_t$. We define the expected energy cost for state $s=(b,x)$ as $\omega_s^\pi\triangleq \mathbf{E}[\rho_t|s_t=s,\pi]$. Let $\bm{\Phi}^B\triangleq{[1,\eta,\eta^2,\dots,\eta^B]}^\intercal$ and $\bm{\Phi}^X\triangleq{[1,\eta,\eta^2,\dots,\eta^X]}^\intercal$, we have \begin{equation} \omega_s^\pi=\eta^{x-b}\sum_{m=0}^{B}\eta^{m}d_m^{s,\pi}-1=\eta^{x-b}{\bm{\Phi}^B}^\intercal\bm{d}^{s,\pi}-1. \end{equation} denote with matrix $\bm{\Omega}^\pi\triangleq{(\omega^\pi_{b,x})}_{b\in\mathcal{B},x\in\mathcal{X}}$ the expected energy costs for all states. Under the average reward criterion, we formally state the optimization problem as \begin{subequations} \begin{align} & \underset{\pi}{\text{minimize}} & & \mathcal{L}(\pi)={\bm{r}^{\infty,\pi}}^\intercal\bm{\Omega}^\pi\bm{p}\label{prb6a}\\ & \text{subject to} & & {\bm{A}^\pi}^\intercal\bm{r}^{\infty,\pi}=\bm{r}^{\infty,\pi},\label{prb6b}\\ & & & \bm{d}^{s,\pi}\geq 0,& \forall s,\label{prb6c}\\ & & & \bm{1}^\intercal\bm{d}^{s,\pi}=1,& \forall s,\label{prb6d}\\ & & & d_m^{(b,x),\pi}=0,& \forall m <b-x.\label{prb6e} \end{align} \label{opt_1 \end{subequations} where constraint (\ref{prb6b}) implies stationary distribution and (\ref{prb6c}) and (\ref{prb6d}) ensure that $\bm{d}^{s,\pi}$ corresponds to a p.m.f.. The last constraint rules out the possibility of $y_t$ being negative, since the number of content item transmitted must be positive or $0$. As the policy space is $(B+1)^2\times(X+1)$-dimensional, problem (\ref{opt_1}) is an optimization with high dimensionality, whose convexity is hard to determine. Value iteration being a popular algorithm regarding solving MDP problems, overwhelming computational complexity is yet a problem if to run value iteration in $\mathcal{S}$. We found a novel approach to decouple the influence from $b_t$ and $x_t$ and hence turn (\ref{opt_1}) to three sub-problems, which are low-dimensional and more tractable. By solving the three sub-problems specified below, we equivalently solves (\ref{opt_1}). First, we assume a transition matrix in $\mathfrak{B}$ is given as $\bm{A}$ and find the optimal policy, i.e., $\{\bm{d}^{s,\pi}: s\in\mathcal{S}\}$, among all policies that results in $\bm{A}$. Secondly, we find the optimal $\bm{A}$ whose corresponding optimal policy minimizes the expected energy consumption in one iteration step. Thirdly, we carry out value iteration in $\mathfrak{B}$, each step searching for the optimal $\bm{A}$ iteratively, until the transition matrix converges to $\bm{A}^*$. By \cite{puterman2014markov}, $\bm{A}^*$ is the minimizer for average energy cost among all possible $\bm{A}$. We argue that we have also found the optimal $\pi^*$, among all policy space, that minimize the average energy cost as \begin{equation} \underset{\bm{A}}{\min}\left(\underset{\pi}{\min}\mathcal{L}(\pi)\mid\bm{A}\right)=\underset{\pi}{\min}\mathcal{L}(\pi). \end{equation} We next elaborate each of the three sub-problems. \subsection{Optimal Decisions with Given Transition Matrix in $\mathfrak{B}$} \label{sectionIIIb} Denote with $\omega_b^\pi$ the expected energy cost under policy $\pi$ in a timeslot belonging to state $\mathfrak{b}_b$. We have \begin{equation} \omega_b^\pi\triangleq\mathbf{E}[\rho_t|b_t=b,\pi]=\sum_{x=0}^{X}p_x\omega^\pi_{(b,x)}=\bm{\Omega}^\pi_{(b,:)}\bm{p}, \label{exp_enrg_cost} \end{equation} where $\bm{\Omega}_{(b,:)}$ denotes the row vector from the $b^{th}$ row of $\Omega$. Recall that $\bm{A}^\pi={\left[\begin{array}{cccc} \bm{a}^{0,\pi}&\bm{a}^{1,\pi}&\dots&\bm{a}^{B,\pi} \end{array}\right]}^\intercal$ and $\bm{a}^{b,\pi}=\sum_{x}\bm{d}^{(b,x),\pi}$, i.e., the $b^{th}$ row of $\bm{A}$ is solely determined by decisions from $\mathfrak{b}_b$. We formulate a matrix that represents decisions from $\mathfrak{b}_b$ as $\bm{D}^{b,\pi}\triangleq{\left[\begin{array}{cccc} \bm{d}^{(b,0),\pi}&\bm{d}^{(b,1),\pi}&\dots&\bm{d}^{(b,X),\pi} \end{array}\right]}^\intercal$. The first sub-problem, finding the optimal decisions under the constraint of a fixed $\bm{A}$, is formally stated as: \begin{subequations} \renewcommand{\theequation}{\theparentequation.\alph{equation}} \begin{align} & \underset{\bm{D}^{b,\pi}}{\text{minimize}} & & \omega_b^\pi=\eta^{-b}{\bm{\Phi}^X}^\intercal\mathrm{diag}(\bm{p})\bm{D}^{b,\pi}\bm{\Phi}^B-1\label{prb5a}\\ & \text{subject to} & & {\bm{D}^{b,\pi}}^\intercal\bm{p}=\textbf{\textit{a}}^\pi_b,\label{prb5b}\\ & & & \bm{D}^{b,\pi}\bm{1}=\bm{1},\label{prb5c}\\ & & & D^{b,\pi}_{m,n}\geq 0,\hspace{1cm} \forall m,n\label{prb5d}\\ & & & D^{b,\pi}_{m,n}=0, \hspace{1cm} \forall m+n<b\label{prb5e}, \end{align \label{opt_2 \end{subequations for every $b\in\mathcal{B}$. We define a function $h: \mathbb{R}^{X+1}\rightarrow \mathbb{R}$ as: \begin{equation} h(\bm{a}^{b,\pi})=\min_\pi\left( \omega^\pi_b|{\bm{a}^{b,\pi}}\right), \end{equation} in light of the optimization problem (\ref{opt_2}). In section \ref{sectionIV}, we propose an efficient algorithm to find $h(\bm{a}^\pi_b)$ without iteration, exploiting a certain structural property we name as \textit{generalized monotonicity}. \begin{subsection}{Optimal Transition Probability Matrix in $\mathfrak{B}$} \label{subprob2} In a finite MDP with $N$ timeslots, define $\upsilon^t_b$ as the expected total cost over period $\{N-t+1, N-t+2, ..., N\}$, starting from $b_{N-t+1}=b$. Denote vector $\bm{\upsilon}^t\triangleq(\upsilon^t_b)_{b\in\mathcal{B}}$. The iteration from $t-1$ to $t$ can be written as \begin{subequations} \renewcommand{\theequation}{\theparentequation.\alph{equation}} \begin{align} & \underset{\bm{a}^{b,\pi}}{\text{minimize}} & & \upsilon^t_b=h(\bm{a}^{b,\pi})+{\bm{a}^{b,\pi}}^\intercal\bm{\upsilon}^{t-1}\label{eq.13a}\\ & \text{subject to} & & \bm{a}^{b,\pi}\geq0,\label{eq.13b}\\ & & & \bm{1}^\intercal\bm{a}^{b,\pi}=1,\label{eq.13c}\\ & & & a^{b,\pi}_m\leq\sum_{x=b-m}^{X}p_x,\quad \forall m=0,1,...,b.\label{eq.13d} \end{align \label{opt_3 \end{subequations (\ref{eq.13b}) and (\ref{eq.13c}) are the nature of p.m.f., and (\ref{eq.13d}) is derived from (\ref{prb5b}) and (\ref{prb5e}). Now we prove that optimization problem (\ref{opt_3}) is convex, and hence can be easily solved by conventional optimization tools, say, the interior-point method. \begin{IEEEproof} Let $\bm{a}^{b,\pi_1}$ and $\bm{a}^{b,\pi_2}$ be two feasible points of (\ref{opt_3}), whose corresponding solutions in (\ref{opt_2}) are $\bm{D}^{b,\pi_1}$ and $\bm{D}^{b,\pi_2}$, respectively. For all $\lambda\in\{\lambda\in\mathbb{R}|0<\lambda<1\}$, let \begin{equation} \bm{a}^{b,\pi'}=(1-\lambda)\bm{a}^{b,\pi_1}+\lambda\bm{a}^{b,\pi_2}, \end{equation} and, let \begin{equation} \bm{D}^{b,\pi'}=(1-\lambda)\bm{D}^{b,\pi_1}+\lambda\bm{D}^{b,\pi_2}. \end{equation} it can be directly shown that $\bm{a}^{b,\pi'}$ and $D^{b,\pi'}$ satisfy constraint sets (\ref{eq.13b})-(\ref{eq.13d}) and (\ref{prb5c})-(\ref{prb5e}) respectively. Also we have \begin{equation} \begin{aligned} {\bm{D}^{b,\pi'}}^\intercal\bm{p}=& ((1-\lambda)\bm{D}^{b,\pi_1}+\lambda\bm{D}^{b,\pi_2})^\intercal\bm{p}\\ =& (1-\lambda)\bm{a}^{b,\pi_1}+\lambda\bm{a}^{b,\pi_2}\\ =& \bm{a}^{b,\pi'}, \end{aligned} \end{equation i.e., $\bm{D}^{b,\pi'}$ satisfies (\ref{prb5b}) with $\bm{a}^{b,\pi'}$. Thus, $\bm{a}^{b,\pi'}$ and $\bm{D}^{b,\pi'}$ are feasible points of (\ref{opt_3}) and (\ref{opt_2}) respectively. Now denote with $\omega_b^{\pi'}$ the expected average energy cost carried out by $\bm{D}^{b,\pi'}$. Because (\ref{prb5a}) is linear on $\bm{D}^{b,\pi}$, $\omega_b^{\pi'}=(1-\lambda)h(\bm{a}^{b,\pi_1})+\lambda h(\bm{a}^{b,\pi_2})$, which gives a feasible value of (\ref{opt_2}) under $\bm{a}^{b,\pi'}$. By definition, \begin{equation} \begin{aligned} h(\bm{a}^{b,\pi'})=&\min_\pi\left(\omega^\pi_b\mid\bm{a}^{b,\pi'}\right)\\ \leq& \omega^{\pi'}_b\\ =& (1-\lambda)h(\bm{a}^{b,\pi_1})+\lambda h(\bm{a}^{b,\pi_2}). \end{aligned} \end{equation} thus, $h$ is convex. Since that ${\bm{a}^{b,\pi}}^\intercal\bm{\upsilon}^{t-1}$ is linear, optimization problem (\ref{opt_3}) is convex. \end{IEEEproof} \end{subsection} \begin{subsection}{Value Iteration in degenerated state space $\mathfrak{B}$} By \ref{sectionIIIb} and \ref{subprob2}, we find the optimal transition matrix in $\mathfrak{B}$ and its corresponding policy, in light of one-step iteration. Now we start from:\begin{equation} \bm{\upsilon}^0=\bm{0}, \end{equation} and carry out the optimality equations given by \begin{equation} \begin{aligned} \upsilon^t_b=& \min_{\bm{a}^{b,\pi}} h(\bm{a}^{b,\pi})+\sum_{b^+=0}^Ba^{b,\pi}_{b^+}\upsilon^{t-1}_{b^+}\\ =& \min_{\bm{a}^{b,\pi}} h(\bm{a}^{b,\pi})+{\bm{a}^{b,\pi}}^\intercal\bm{\upsilon}^{t-1}, \end{aligned} \label{optimality_equs} \end{equation} for $\forall b\in\mathcal{B}$ and $t=1,2,\dots$. The transition matrix $A^\pi$ in $\mathfrak{B}$ and its corresponding policy $\pi$ is updated on each iteration step. By \cite{puterman2014markov}, the iteration would eventually converge to satisfy the $\varepsilon$-optimal stopping criterion: \begin{equation} \label{convergence} \left|\max_i(\upsilon^t_i)-\min_j(\upsilon^t_j)\right|<\varepsilon. \end{equation} when (\ref{convergence}) holds, the transition matrix $A^{\pi*}$ in $\mathfrak{B}$ and its corresponding $\pi^*$ give a global $\varepsilon$-optimal policy. By taking limit $\varepsilon\rightarrow0$, the global optimum of (\ref{opt_1}) is attained. \end{subsection} \section{Generalized Monotonic Structure} \label{sectionIV} As mentioned in subsection \ref{sectionIIIb}, an efficient algorithm that solves problem (\ref{opt_2}) without iteration is proposed in this section, and its corresponding time complexity analysis is given in section \ref{sectionV}. A special pattern of the optimal decision matrix $D^{b,\pi}$ is revealed by theorem 1, which we name \textit{generalized monotonicity}. \subsection{Solution of Optimization without Iteration} Intuitively the \textit{generalized monotonicity} captures the feature that the optimal decision matrix has all its non-zero entries lying in a stripe expanding from the top-right corner to the bottom-left. If a entry is non-zero, then the block adjacent to its bottom-right corner shold be all-zero. We formally describe and prove it as the following theorem: \begin{theorem} \label{theorem1} If $\bm{D}^{b,\pi^*}$ is a solution of (\ref{opt_2}), and there exist $ x^-, b^-$ s.t. $D^{b,\pi^*}_{x^-,b^-}>0$, then $D^{b,\pi^*}_{x^+,b^+}=0$ for $\forall x^+\in\{x\in\mathbb{N}: x^-<x\leq X\}, b^+\in\{b\in\mathbb{N}: b^-<b\leq B\}$. \end{theorem} \begin{IEEEproof} Suppose there exist $x^-, b^-$ s.t. $0\leq x^-<X, 0\leq b^-< B$ and $D^{b,\pi^*}_{x^-,b^-}>0$ (if not, the case is trivial and Theorem \ref{theorem1} still holds). We prove by contradiction, starting by assuming that $\exists x^+, b^+\in\mathbb{N}$ s.t. $x^-<x^+\leq X$, $b^-<b^+\leq B$ and $D^{b,\pi^*}_{x^+,b^+}>0$. Now we pick any positive number $\delta$ that satisfies $p_{b^+}\delta<D^{b,\pi^*}_{x^-,b^-}$ and $p_{b^-}\delta<D^{b,\pi^*}_{x^+,b^+}$. We formulate another decision matrix $D^{b,\pi'}$ which is identical to $D^{b,\pi^*}$ except the following elements \begin{subequations} \renewcommand{\theequation}{\theparentequation.\alph{equation}} \begin{align} &D^{b,\pi'}_{x^-,b^-}=D^{b,\pi^*}_{x^-,b^-}-p_{b^+}\delta\\ &D^{b,\pi'}_{x^-,b^+}=D^{b,\pi^*}_{x^-,b^+}+p_{b^+}\delta\\ &D^{b,\pi'}_{x^+,b^+}=D^{b,\pi^*}_{x^+,b^+}-p_{b^-}\delta\\ &D^{b,\pi'}_{x^+,b^-}=D^{b,\pi^*}_{x^+,b^-}+p_{b^-}\delta \end{align} \end{subequations} because $\bm{D}^{b,\pi^*}$ is a solution of (\ref{opt_2}), it is easy to verify that $\bm{D}^{b,\pi'}$ satisfies (\ref{prb5b}) and (\ref{prb5c}). Also, (\ref{prb5d}) and (\ref{prb5e}) are naturally satisfied by the choice of $\delta$. Hence, $D^{b,\pi'}$ is a feasible point of (\ref{opt_2}). Further we have \begin{equation} \begin{aligned} &\quad\omega^{\pi^*}_b-\omega^{\pi'}_b\\ &=\eta^{-b}{\bm{\Phi}^X}^\intercal\mathrm{diag}(\bm{p})(\bm{D}^{b,\pi^*}-\bm{D}^{b,\pi'})\bm{\Phi}^B\\ &=\eta^{-b+x^-+b^-}p_{x^+}p_{x^-}(\eta^{x^+-x^-}-1)(\eta^{b^+-b^-}-1)>0. \end{aligned} \end{equation} $\bm{D}^{b,\pi'}$ results in a smaller objective value in (\ref{prb5a}) than $D^{b,\pi^*}$, which contradicts the prerequisite that $D^{b,\pi^*}$ is a solution. Therefore, the original assumption must be false, proving the theorem. \end{IEEEproof} Now we formally state the Fast Assignment of State Transition (FAST) algorithm that gives a solution to the problem in (\ref{opt_2}) without iteration. \begin{algorithm} \caption{The FAST Algorithm} \begin{algorithmic}[1] \State Initialization: $m=0, n=B, u_0=p_0, w_B=a^{b,\pi}_B, \bm{D}^{b,\pi}=\bm{0}$; \If{ $u_m<w_n$, }\label{goto2} \State $D^{b,\pi}_{m,n}\triangleq\frac{u_m}{p_m}, w_n\triangleq w_n-u_m, m\triangleq m+1, u_m\triangleq p_m$, go to \ref{goto9}; \EndIf \If{$u_m>w_n$,} \State $D^{b,\pi}_{m,n}\triangleq\frac{w_n}{p_m}, u_m\triangleq u_m-w_n, n\triangleq n-1, w_n\triangleq a^{b,\pi}_n$, go to \ref{goto9}; \EndIf \State $D^{b,\pi}_{m,n}\triangleq\frac{w_n}{p_m}, m\triangleq m+1, n\triangleq n-1, u_m\triangleq p_m, w_n\triangleq a^{b,\pi}_n$; \If{$m\leq X$ and $n\geq 0$,}\label{goto9} \State go to \ref{goto2}; \EndIf \State $\bm{D}^{b,\pi^*}\triangleq \bm{D}^{b,\pi}$; \end{algorithmic} \end{algorithm} Validity of the FAST algorithm is given by the following Theorem \ref{theorem2}. \begin{theorem} \label{theorem2} Let $\bm{D}^{b,\pi^*}$ be a solution of (\ref{opt_2}). $\forall m,n\in\mathbb{N}$ s.t. $0\leq m\leq X, 0 \leq n\leq B$, if denote \begin{subequations} \renewcommand{\theequation}{\theparentequation.\alph{equation}} \begin{align} &u_m=(a^{b,\pi}_n-\sum_{i=0}^{m-1}D^{b,\pi^*}_{i,n}p_i)/p_m\\ &w_n=1-\sum_{i=n+1}^BD^{b,\pi^*}_{m,i}, \end{align} \end{subequations} then $D^{b,\pi^*}_{m,n}=\min(u_m, w_n)$. \end{theorem} \begin{IEEEproof} Let $m,n\in\mathbb{N}$ s.t. $0\leq m\leq X, 0 \leq n\leq B$. By (\ref{prb5b}) to (\ref{prb5d}), we have \begin{subequations} \renewcommand{\theequation}{\theparentequation.\alph{equation}} \begin{align} &\sum_{i=0}^{m}D^{b,\pi^*}_{i,n}p_i\leq a^{b,\pi}_n \label{eq22a}\\ &\sum_{i=n}^BD^{b,\pi^*}_{m,i}\leq 1 \label{eq22b} \end{align} \end{subequations} the proof is given by contradiction. Assume that neither equation in (\ref{eq22a}) and (\ref{eq22b}) holds. By (\ref{prb5b}) $\sum_{i=0}^{X}D^{b,\pi^*}_{i,n}p_i=a^{b,\pi}_n,$ and also we have $ \sum_{i=0}^{m}D^{b\pi^*}_{i,n}p_i<a^{b,\pi}_n,$ Therefore \begin{equation*} \exists m^+>m \hspace{1cm}\text{s.t.}\hspace{1cm} D^{b,\pi^*}_{m^+,n}>0, \end{equation*} by (\ref{prb5c}) and the assumption that the equation in (\ref{eq22b}) does not hold, we have $\exists n^-<n$ s.t. $D^{b,\pi^*}_{m,n^-}>0$. Therefore, we have $D^{b,\pi^*}_{m,n^-}>0$ and $D^{b,\pi^*}_{m^+,n}>0$, contradicting Theorem \ref{theorem1}. Thus, the original assumption is false, i.e., at least one of the equations in (\ref{eq22a})(\ref{eq22b}) holds. Rewrite (\ref{eq22a}) and (\ref{eq22b}) as \begin{subequations} \renewcommand{\theequation}{\theparentequation.\alph{equation}} \begin{align} &D^{b,\pi^*}_{m,n}\leq (a^{b,\pi}_n-\sum_{i=0}^{m-1}D^{b,\pi^*}_{i,n}p_i)/p_m\\ &D^{b,\pi^*}_{m,n}\leq 1-\sum_{i=n+1}^BD^{b,\pi^*}_{m,i}, \end{align} \end{subequations} it is clear that the equation with smaller value on the right side should hold, hence proving theorem \ref{theorem2}. \end{IEEEproof} \subsection{Complexity Analysis} \label{sectionV} Suppose the required precision of value iteration is $\epsilon_v$, and interior-point method is applied for all convex optimizations. By\cite{puterman2014markov}, the number of iteration steps is bounded by $-C_1\log\epsilon_v$, where $C_1$ is a constant. $B+1$ convex optimizations are solved with interior-point method in each iteration step. If precision $\epsilon_i$ is required for interior-point method, the number of times Newton's Method is called in each interior-point method is bounded by $-\sqrt{2(B+1)}\log\epsilon_i$ \cite{boyd2004convex}. In Newton's Method the FAST algorithm, whose time complexity is $O(X+B)$, is called $B+1$ times. Therefore, the time cost of value iteration in $\mathfrak{B}$ with the FAST algorithm is bounded by $O((X+B)B^{2.5}\log\epsilon_i\log\epsilon_v)$. With similar analysis, we have the time complexity of value iteration in $\mathcal{S}$ which is $O(X^2B^{3.5}\log\epsilon_i\log\epsilon_v)$. An interesting insight into the two complexities is that when $X\ll B$ the difference between value iterations in $\mathfrak{B}$ and $\mathcal{S}$ fades out, which is the case that buffer size is much larger than expected data request in a timeslot. Since there is nearly unlimited caching space, an optimal policy becomes meaningless. However, in most cases, value iteration in $\mathfrak{B}$ with the FAST algorithm brings significant improvement to time efficiency. \section{Simulation Results} We first demonstrate the reasonably small compromise of the causal MDP method proposed in this paper, in comparison with the non-causal \textit{tightest string method} which attains the absolute optimum \cite{Zafer2009}. Then we compare the time consumption of value iteration in $\mathfrak{B}$, as proposed in this paper, and the conventional MDP method of value iteration in $\mathcal{S}$. \begin{figure}[htbp] \centering \vspace{-0.4cm} \includegraphics[width=3in]{Fig_1} \vspace{-0.25cm} \caption{Average energy cost decreases as buffer size grows, and the causal MDP method attain similar performance to the non-causal optimum. Here, $x\sim U^d[0,20]$ and $\eta=1.4$.} \label{sim1} \end{figure} \begin{figure}[htbp] \begin{center} \vspace{-0.5cm} \includegraphics[width=3in]{Fig_5} \vspace{-0.25cm} \caption[Comparision of time complexity]{Average energy cost increases rapidly as data request grows. Similar to the non-causal optimum, causal MDP policy brings significant improvement. Here, buffer size $B=8$ in content items and $\eta=1.4$.} \label{sim3 \end{center \end{figure \begin{figure}[htbp] \begin{center} \vspace{-0.8cm} \includegraphics[width=3in]{Fig_9} \vspace{-0.25cm} \caption[Comparision of time complexity]{Value iteration in degenerated space $\mathfrak{B}$ with the FAST algorithm significantly saves time compared with that in $\mathcal{S}$. Here, the user request is of size $X=1.5B$ where $B$ denotes the buffer size with $\eta=1.4$} \label{sim6} \end{center} \end{figure} We first assume that data request yields discrete uniform distribution on all integers in $[0,20]$, and buffer size varies. $\eta$ takes typical values $1.4$ and $2$ respectively. Fig. \ref{sim1} shows that the causal policy designed by MDP compromises little compared with the non-causal optimum with full RDI. The curve \textit{No buffer} corresponds to real-time transmission, and \textit{Infinite buffer} shows the energy cost for stationary transmission at the average rate. The two curves together give the upper and lower bounds of energy cost. We then fix the buffer size and study the tendency of energy cost as data request pattern changes. Assume that data request always yields discrete uniform distribution on all integers in $[0,X]$. As $X$ grows, Fig. \ref{sim3} shows that energy cost grows exponentially. The optimal MDP policy still brings significant improvement, and its performance synchronically grows with the non-causal full-RDI method. As the simulations are done with identical parameter settings, where $X=1.5B$ always holds and $B$ takes $\{2,4,...,16\}$, Fig. \ref{sim6} shows that value iteration in full state space $\mathcal{S}$ consumes much more time than the method proposed in this paper, which is value iteration in $\mathfrak{B}$ with the FAST algorithm, though the two methods end with the same policy. \section{Conclusion} In this paper, we introduced Markov decision processes to the point-to-point proactive caching problem in wireless communications. Though it is possible to directly apply conventional MDP algorithms to design optimal JPC policies, we largely adapted MDP model for the specified problem setting. We revealed and exploited benefits from a special structure, which we proposed as generalized monotonicity. The algorithm designed based on the very structure has significantly accelerated the conventional MDP algorithm in this problem. Additionally, the energy performance of the attained causal policy is comparatively satisfactory, with regard to the globally optimal non-causal policy. Future works may continue to analytically study the improvement of energy consumption, and generalize conclusions in this paper to the scenario of multicasting. \bibliographystyle{IEEEtran}

\section{Introduction} Over the last decades the properties of hadrons in nuclei have been matter of intense investigation. On the one hand, it is believed that hadrons in nuclei is an excellent scenario to test certain symmetries of the theory of the strong interaction, Quantum Chromodynamics (QCD), such as the chiral symmetry in the low-energy regime or heavy-quark symmetries as hadrons with charm or beauty content are produced in the laboratory. On the other hand, it is of crucial importance to understand the excitation mechanisms in the nucleus as well as the nature of certain excited hadronic states, whose structure could be studied in a hot and/or dense nuclear medium. In order to address hadrons in nuclei one can resort to theoretical and/or experimental analyses \cite{}. From the theoretical side, there is an extensive variety of models that aim at understanding the properties of hadrons in nuclei, ranging from relativistic-mean field models (RMF), Nambu-Jona-Lasinio schemes (NJL), quark-meson coupling models (QMC), QCD sum-rule studies to unitarized approaches based on effective theories or meson-exchange models. Experimentally, photon-, electron-, neutrino- and hadron-induced reactions as well as heavy-ion collisions (HiCs) offer a gateway to the properties of hadrons in nuclei. The measurement of transparency ratios is extremely useful for the analysis of the imaginary part of the hadron-nucleon interaction in matter, whereas the excitation functions and the meson-momentum distributions are of fundamental importance to understand the real part of the interaction \cite{}. Moreover, it is crucial to connect the theoretical predictions to the experimental results, using transport model calculations or collision models based on nuclear spectral functions, to fully understand the dynamics of hadrons in nuclei. In this paper we concentrate on the analysis of the properties of mesons with strangeness and charm content that interact with nuclei. Nowadays, strange and charmed hadrons are being produced in nuclear and particle facilities, such as GSI, CERN or RHIC, while they are the subject of future experimental programs, such as FAIR, NICA or J-PARC. \section{Strangeness in nuclei} Strangeness in nuclei has received a lot attention in connection with the study of neutron stars interior \cite{}, the properties of exotic atoms \cite{}, and strangeness production in heavy-ion collisions (HICs) \cite{}. In particular, the dynamics of strange mesons, such as $\bar K$, in vacuum and in the nuclear medium is still a challenge for theory and experiments. In this section the $\bar K N$ interaction is studied, paying a special attention to the role of the $\Lambda(1405)$ and the formation of bound states, such as $\bar KNN$. Moreover, the production and propagation of strangeness in heavy-ion collisions (HICs) is investigated, in view of the present and forthcoming experimental programs on strangeness. \subsection{$\bar K N$ interaction: the $\Lambda(1405)$} The $\bar K N$ scattering in the $I=0$ channel is governed by the presence of the $\Lambda(1405)$, located only 27 MeV below the $\bar K N$ threshold. The dynamical origin of the $\Lambda(1405)$ dates back more than 50 years ago to the work of Dalitz and Tuan \cite{}. Recently, it has been revisited by means of unitarized theories using meson-exchange models \cite{} or chiral Lagrangians \cite{, , }, these latter ones analyzing the effects of including a complete basis of meson-baryon channels, studying the differences in the regularization of the equations, including s- and u-channel Born terms in the Lagrangian, implementing next-to-leading (NLO) contributions,... . All these recent efforts have culminated in establishing the $\Lambda(1405)$ as a superposition of two poles of the scattering matrix \cite{}, that are generated dynamically from the unitarized coupled-channel scheme. A renewed interest in the $\bar K N$ interaction has been developed in the past years after the availability of a more precise measurement of the energy shift and width of the $1s$ state in kaonic hydrogen by the SIDDHARTA Collaboration at DA$\Phi$NE \cite{}, that has helped to clarify the discrepancies between the KEK \cite{} and the DEAR \cite{} measurements. The obtained value of the energy shift is $\Delta E = 283 \pm 36 \pm 6$ eV with a width of $\Gamma=541 \pm 89 \pm 22$ eV, in good agreement with KEK results. Furthermore, the SIDDHARTA measurement has provided new constraints on the theoretical predictions reported in \cite{}. \begin{figure}[t] \includegraphics[width=.5\textwidth]{nagae.eps} \includegraphics[width=.5\textwidth]{Kbar.eps} \caption{Left: Comparison of the binding energy and width of the $K^-pp$ between experiments and theoretical predictions, taken from \cite{}. Right: $\bar K$ spectral function for different densities, temperatures and momenta, taken from \cite{}.} \label{fig1} \end{figure} \subsection{$\bar K NN$ state} The dynamical generation of the two-pole structure of the $\Lambda(1405)$ indicates that the $\bar K N$ interaction might be attractive enough to produce bound states. Indeed, it has been suggested that $\bar K$-nuclear clusters may form, such as the $\bar K NN$ in isospin $I=1/2$. The $I=1/2$ $\bar K NN$ state has been extensively studied, both theoretically and experimentally, as shown in the left-hand side of {\bf Fig.~\ref{fig1}} (see Ref.~\cite{} and references therein). This state was initially seen by the FINUDA \cite{}, DISTO \cite{} and OBELIX \cite{} Collaborations, but could be explained by means of conventional processes \cite{} or not be reproduced \cite{}. Experiments performed by the Spring8/LEPS \cite{}, J-PARC E15 \cite{}, AMADEUS \cite{} Collaborations do not find any state, or, if found \cite{}, may have other interpretation, such as as a possible $I=3/2$, $J^{\pi}=2^+$ resonance near the $\pi \Sigma N$ threshold \cite{}. More recently, the J-PARC E15 experiment has found a structure near the $\bar K NN$ threshold \cite{}, that has been interpreted as a $\bar K NN$ bound state with a binding energy of $\sim 20-40$ MeV \cite{}. \subsection{$\bar K N$ in matter} Antikaonic atoms \cite{} give us information on the antikaon interaction with nucleons. The antikaon-nucleus potential has been extracted from best-fit analysis of antikaonic-atom data and some solutions agree with a very strongly attractive potential of the order of -200 MeV at normal saturation density $\rho_0$. However, some criticism has been raised because the antikaonic-atom data only tests matter at the surface of the nucleus. Recent analysis on $K^- N$ scattering amplitudes from chiral SU(3) effective field theories supplemented with phenomenological terms for $K^-$ multinucleon interactions indicate that antikaonic atoms are insensitive to densities above $\rho_0$ \cite{}. Early works based on relativistic mean-field models \cite{} also obtained very deep potentials of a few hundreds of MeVs at $\rho_0$. However, later approaches on unitarized theories in coupled channels based on the chiral effective theory \cite{} or on meson-exchange potentials \cite{} obtain a potential much less attractive. In fact, in the unitarized coupled-channels models, the attraction is a consequence of the modified $s$-wave $\Lambda(1405)$ resonance in the medium due to Pauli blocking \cite{}, together with the self-consistent inclusion of the $\bar K$ self-energy \cite{}, and the implementation of self-energies of the mesons and baryons in the intermediate states \cite{}. As a result, the $\bar K$ spectral function can be obtained, as shown in the right-hand side of Fig.~\ref{fig1}. The $\bar K$ spectral function shows that $\bar K$ in matter feel a slight attraction while acquiring a remarkable width. Moreover, the knowledge of higher-partial waves beyond $s$-wave \cite{} becomes essential for analyzing the results of HiCs at beam energies below 2GeV per nucleon \cite{}. \subsection{Strangeness production in HICs} The production of $K$ and $\bar K$ close to threshold has been extensively investigated in low-energy HICs by the KaoS \cite{}, FOPI \cite{} and HADES Collaborations \cite{}. The analysis of experimental data together with microscopic transport approaches have permitted drawing several conclusions regarding the production mechanisms and the freeze-out conditions of strange mesons. However, the role of the in-medium properties of strange hadrons in their production and production in HiCs is still an open question. Recent results from HADES and FOPI indicate that, while the $K^+$ shows a repulsive interaction in matter, the $\Phi$ decay into $K^-$ washes out the effects of the $K^-$ potential in the spectra and flow \cite{leifels}. Therefore, more systematic and high statistic data on $K^-$ production are necessary, while further information in elementary reactions is required. \section{Open charm in nuclei} The interest in the properties of open and hidden charmed mesons was triggered in HiCs due to the possible charmonium suppression as a probe for the formation of quark-gluon plasma (QGP). Nowadays, the nature of newly observed baryon and meson states with the charm degree of freedom is a matter of high interest in connection with many on-going experiments, such as BESIII, BelleII, ALICE, LHCb, amongst others, as well as with planned facilities, e.g. FAIR, NICA and the J-PARC upgrade. The goal is to understand whether these states can be accommodated within the quark model picture and/or qualify better as being dynamically generated via hadron-hadron scattering processes. To this end, a large part of the experimental program in hadronic physics at PANDA (FAIR) will be devoted to charmonium spectroscopy. Also, the CBM (FAIR) experiment will extend the GSI program for in-medium modification of hadrons in the light quark sector and provide the first insight into the charm-nucleus interaction. Indeed, the influence of medium modifications in the charmonium production at finite baryon densities would affect the formation of the QGP phase of QCD at high densities. \subsection{ $D N$ interaction: the $\Lambda_c(2595)$} Given the success of unitarized coupled-channel approaches in the description of some of the existing experimental data in the light-quark sector, the charm degree of freedom has been recently incorporated in these models and several experimental states have been described as dynamically-generated baryon molecules (see Ref.~\cite{} and references therein). This is the case, for example, of the $\Lambda_c(2595)$, which is the charmed counterpart of the $\Lambda(1405)$. Whereas a separable potential for the bare meson-baryon interaction with no strange degree of freedom was assumed in \cite{}, later on unitarized approaches were based on a bare meson-baryon interaction saturated with the $t$-channel exchange of vector mesons between pseudoscalar mesons and baryons in the zero-range approximation \cite{} or using the full $t$-dependence \cite{}. Other approaches have made use of the J\"ulich meson-exchange model \cite{}, while some others have relied on the hidden gauge formalism \cite{, }. More recent schemes incorporate heavy-quark symmetry constraints explicitly, such as those based on a pion-exchange model \cite{} or on an extended Weinberg-Tomozawa interaction for four flavors that includes pseudoscalar and vector mesons together with $1/2^+$ and $3/2^+$ baryons \cite{, , }. In all these unitarized coupled-channel models, the $\Lambda_c(2595)$ is obtained dynamically, some of them \cite{, , } even obtaining a double-pole structure, in a similar manner as found for the $\Lambda(1405)$. \subsection{ $\bar D N$ interaction} The $C=-1$ has also been investigated within unitarized coupled-channel models, pion-exchange schemes with heavy-quark symmetry constraints or chiral quark models \cite{}. Interestingly, some of the models find a $J=1/2$ state close to the $\bar DN$ threshold \cite{}. In Ref.~\cite{} this state was generated by the $ \bar D N$ and $\bar D^* N$ coupled channel dynamics, and it appears to be a consequence of treating heavy pseudoscalars and heavy vector mesons on an equal footing, because no resonance would be generated unless $\bar D^* N$ channel is considered. \subsection{ $D NN$ and $\bar D NN$ states} Given the fact that the $DN$ and $\bar DN$ interactions are so attractive that allow for the formation of bound states, the question arises whether $D$ or $\bar D$-nuclear clusters may form. In \cite{} a $I=1/2$, $J=0^-$ $DNN$ state was found with mass $~3500$ MeV and width $\sim 20-40$ MeV, being interpreted as a quasibound state of the $\Lambda_c(2595)$ and a nucleon. Moreover, in \cite{} a state with $I=1/2$, $J=0^-$ and 5.2 MeV binding was found together with a $I=1/2$, $J=1^-$ state at 111.2 MeV above threshold. \begin{figure}[t] \begin{center} \includegraphics[width=.3\textwidth,angle =-90]{D0.eps} \end{center} \caption{$D^0$-nucleus bound states, taken from \cite{}. } \label{figd0} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=.45\textwidth]{Dmenos.eps} \includegraphics[width=.45\textwidth]{D0bar.eps} \end{center} \caption{$D^-$ and $\bar D^0$- nucleus bound states, taken from \cite{}.} \label{figdm} \end{figure} \subsection{Open charm in matter} The properties of open-charm mesons in matter have been object of theoretical interest due to the consequences for charmonium suppression, as observed at SPS energies by the NA50 collaboration. The change of the properties of D mesons in matter would modify the $J/\Psi$ absorption in a hot and dense nuclear medium and can provide an explanation for $J/\Psi$ suppression. Furthermore, there have been speculations about the existence of $D$-meson bound states in nuclei \cite{}. Several theoretical works have addressed the properties of open-charm mesons in dense baryonic matter: QMC schemes \cite{}, QCD sum-rule approaches \cite{, }, NJL models \cite{}, chiral effective models in hot and dense matter \cite{} or pion-exchange approaches with heavy-quark symmetry constraints \cite{}. The full spectral features (mass and width) of the open-charm mesons in dense nuclear matter have been obtained in self-consistent unitarized coupled-channel schemes, where the intermediate meson-baryon propagators contain different sources of density dependence \cite{}. \subsection{D-mesic nuclei} A possible experimental scenario for the detection of the changes in matter of the properties of open-charm mesons would be the formation of D-mesic nuclei, where a D-meson binds in nuclear orbits. In fact, $D$ and $\bar D$-meson bound states in $^{208}$Pb were predicted in Ref.~\cite{}, relying upon an attractive $D$ and $\bar D$ -meson potential in the nuclear medium, obtained within a QMC model. Within the unitarized coupled-channel model of Ref.~\cite{}, it is found that $D^0$-nucleus states are weakly bound (see Fig.~\ref{figd0}), in contrast to previous results using the QMC model \cite{}, while having significant widths. The best chances for observation of bound states are in the region of $^{24}\mbox{Mg}$, provided an orbital angular momentum separation can be done, where there is only one $s-$ bound state and its half width is about a factor of two smaller than the binding energy (see Fig.~\ref{figd0}). With regards to $D^+$-nuclear states, the Coulomb interaction prevents the formation of observable bound states. As for $\bar D$-mesic nuclei, not only $D^-$ but also $\bar{D}^0$ bind in nuclei as seen in Fig.~\ref{figdm}. The spectrum contains states of atomic and of nuclear types for all nuclei for $D^-$ while, as expected, only nuclear states are present for $\bar{D}^0$ in nuclei. Compared to the pure Coulomb levels, the atomic states are less bound. The nuclear ones are more bound and may present a sizable width, existing only for low angular momenta \cite{}. This is in contrast to \cite{} for $^{208}$Pb, but close to results in \cite{}. The experimental detection of $D$ and $\bar D$-meson bound states is, though, a difficult task \cite{}. The formation of D-mesic nuclei with antiprotons beams at PANDA (FAIR) might be possible if ($\bar p$, D+N) or ($\bar p$, D+2N) reactions with small or even zero momentum transfer are produced, although the formation cross sections could be suppressed because of the complexity of the reaction mechanisms. Other competing mechanisms could involve the emission of pions by intermediate $D^*$ or $\bar D^*$ with subsequent trapping of the pseudoscalar charmed mesons by the final nucleus \cite{}. \section*{Acknowledgements} L.T. acknowledges support from the Heisenberg Programme of the Deutsche Forschungsgemeinschaft under the Project Nr. 383452331, the Ram\'on y Cajal research programme, FPA2013-43425-P and FPA2016-81114-P Grants from MINECO, and THOR COST Action CA15213. \bibliographystyle{JHEP}

\section{Introduction} Task-oriented dialogue, different from chit-chat type of conversation, requires the system to produce responses by accessing information from knowledge bases and planning over multiple dialogue turns. Conventional task-oriented dialogue systems have a complex pipeline \cite{raux2005let,young2013pomdp} consisting of independently developed and modularly connected components for natural language understanding (NLU) \cite{mesnil2015using,Liu2016}, dialogue state tracking (DST) \cite{henderson2014word,mrkvsic2016neural}, and dialogue policy \cite{gasic2014gaussian,su2016line}. A limitation with such pipelined design is that errors made in upper stream modules may propagate to downstream components, making it hard to identify and track the source of errors. Moreover, each component in the pipeline is ideally re-trained as preceding components are updated, so that we have inputs similar to the training examples at run-time. This domino effect causes several issues in practice. To ameliorate these limitations with the conventional pipeline dialogue systems, recent efforts have been made in designing neural network based end-to-end learning solutions. Such end-to-end systems aim to optimize directly towards final system objectives (e.g. response generation, task success rate) instead of performing component-wise optimization. Many of the recently proposed end-to-end models are trained in supervised manner \cite{wenN2N16,bordes2017,eric2017copy,Liu2017} by learning from human-human or human-machine dialogue corpora. Deep reinforcement learning (RL) based systems \cite{li2017end,liu2017iterative,williams2017hybrid,dhingra2017towards} that learns by interacting with human user or user simulator have also been studied in the literature. Comparing to supervised training models, systems trained with deep RL showed improved task success rate and model robustness towards diverse dialogue scenarios. In this work, we present a neural network based task-oriented dialogue system that can be optimized end-to-end with deep RL. The system is built with neural network components for natural language encoding, dialogue state tracking, and dialogue policy learning. Each system component takes in underlying component's outputs in a continuous from which is fully differentiable with respect to the system optimization target, and thus the entire system can be trained end-to-end. In the experiments on a movie booking domain, we show that our system trained with deep RL leads to significant improvement on dialogue task success rate comparing to supervised training systems. We further illustrate the benefit of performing end-to-end optimization comparing to only updating the policy network during online policy learning as in many previous work \cite{gasic2014gaussian,su2016line}. \section{Related Work} Traditional task-oriented dialogue systems typically require a large number of handcrafted features, making it hard to extend a system to new application domains. Recent approaches to task-oriented dialogue treat the task as a partially observable Markov Decision Process (POMDP) \cite{young2013pomdp} and use RL for online policy optimization by interacting with users \cite{gavsic2013line}. The dialogue state and action space have to be carefully designed in order to make the reinforcement policy learning tractable \cite{young2013pomdp}. With the success of end-to-end trainable neural network models in modeling non-task-oriented chit-chat dialogues \cite{serban2015building,li2016deep}, efforts have been made in carrying over the good performance of end-to-end models to task-oriented dialogues. Bordes and Weston \cite{bordes2017} proposed modeling task-oriented dialogues with a machine reading approach using end-to-end memory networks. Their model removes the dialogue state tracking module and selects the final system response directly from candidate responses. Comparing to this approach, our model explicitly tracks user's goal in dialogue state over the sequence of turns, as robust dialogue state tracking has been shown \cite{jurvcivcek2012reinforcement,dhingra2017towards} to be useful for interfacing with a knowledge base (KB) and improving task success rate. Wen et al.~\cite{wenN2N16} proposed an end-to-end trainable neural network model with modularly connected system components. This system is trained in a supervised manner, and thus may not be robust enough to handle diverse dialogue situations due to the limited varieties in the training dialogue samples. Our system is trained by a combination of SL and deep RL methods, as it is shown that RL training may effectively improve the system robustness and dialogue success rate \cite{li2017end,williams2017hybrid}. Dhingra et al.~\cite{dhingra2017towards} proposed an end-to-end RL dialogue agent for information access. Their model focuses on bringing differentiability to the KB query operation by introducing a "soft" retrieval process in selecting the KB entries. Such soft-KB lookup may be prone to information updates in the KB, which is common in real world information systems. In our model, we use symbolic query and leave the selection of KB entities to external services (e.g. a recommender system), as entity ranking in real world systems can be made with much richer feature sets (e.g. user profiles, location and time context, etc.). Quality of the generated query is directly related to the performance of our dialog state tracking module, which can be optimized during user interactions in the proposed end-to-end reinforcement learning model. \section{Proposed Method} \subsection{System Architecture} \label{} Figure \ref{} shows the overall system architecture of the proposed end-to-end task-oriented dialogue model. A continuous form dialogue state over a sequence of turns is maintained in the state $s_k$ of a dialogue-level LSTM. At each dialogue turn $k$, this dialogue-level LSTM takes in the encoding of the user utterance $U_k$ and the encoding of the previous system action $A_{k-1}$, and produces a probability distribution $P(l^{m}_k)$ over candidate values for each of the tracked goal slots: \begin{align} & s_k = \operatorname{LSTM}(s_{k-1}, \hspace{1mm} [U_k, \hspace{1mm} A_{k-1}]) \\ & P(l^{m}_k \hspace{1mm} | \hspace{1mm} \mathbf{U}_{\le k}, \hspace{1mm} \mathbf{A}_{< k}) = \operatorname{SlotDist}_{m}(s_k) \end{align} where $\operatorname{SlotDist}_{m}$ is a single hidden layer MLP with $\operatorname{softmax}$ activation function over slot type $m \in M$. In encoding natural language user utterance to a continuous vector $U_k$, we use a bidirectional LSTM (i.e. an utterance-level LSTM) reader by concatenating the last forward and backward LSTM states. \begin{figure*}[t] \centering \includegraphics[width=220pt]{e2e_dialogue_nn.pdf} \vspace*{-2ex} \caption{{ Proposed end-to-end task-oriented dialogue model architecture. }} \label{} \vspace*{-2ex} \end{figure*} Based on slot-value pair outputs from dialogue state tracking, a query command is formulated by filling a query template with candidate values that have the highest probability for each tracked goal slot. Alternatively, an n-best list of queries can be generated with the most probable candidate values. The query is sent to a KB to retrieve user requested information. Finally, a system action is emitted in response to the user's input based on the current dialogue state and the information retrieved from the knowledge base: \begin{align} P(a_{k} \hspace{1mm} | \hspace{1mm} U_{\le k}, \hspace{1mm} A_{< k}, \hspace{1mm} E_{\le k}) = \operatorname{PolicyNet}(s_{k}, v_{k}, E_{k}) \end{align} where $v_{k}$ represents the concatenated log probabilities of candidate values for each goal slot. $E_{k}$ is the encoding of the retrieved result from the knowledge base (e.g. item availability and number of matched items). $\operatorname{PolicyNet}$ is an MLP with $\operatorname{softmax}$ activation function over all system actions. The emitted system action is then translated to a system response in natural language format by combining the state tracking outputs and the query results. We use a template based natural language generator (NLG) in this work. \subsection{Model Training} We first train the system in a supervised manner using task-oriented dialogue corpora. Based on system inputs with past user utterances, system actions, and KB results, the model tracks the user's goal slot values and predict the next system action. We optimize the model to minimize the linear interpolation of cross-entropy losses for dialogue state tracking and system action prediction: \begin{equation} \begin{split} \min_{\theta} \sum_{k=1}^{K} -\Big[ \sum_{m=1}^{M} &\lambda _{l^{m}} \log P({l^{m}_k}^{*} | \mathbf{U}_{\le k}, \mathbf{A}_{< k}, \mathbf{E}_{< k}; \theta) \\ + &\lambda _a \log P(a_k^{*} | \mathbf{U}_{\le k}, \mathbf{A}_{< k}, \mathbf{E}_{\le k}; \theta) \hspace{1mm} \Big] \\ \end{split} \end{equation} where $\lambda$s are the linear interpolation weights for the cost of each system output. ${l^{m}_k}^{*}$ and $a_k^{*}$ are the ground truth labels for goal slots and system action the $k$th turn. After the supervised training stage, we further optimize the system with RL by letting the agent to interact with users and collecting user feedback. We apply REINFORCE algorithm \cite{williams1992simple} in optimizing the network parameters. We use softmax policy during RL training to encourage the agent to explore the dialogue action space. Feedback is only collected at the end of a dialogue. A positive reward is assigned for success tasks, and a zero reward is assigned for failure tasks. A small step penalty is applied to each dialogue turn to encourage the agent to complete the task in fewer steps. We use policy gradient method for dialogue policy learning. With likelihood ratio gradient estimator, the gradient of the objective function $J_k(\theta)$ can be derived as: \begin{equation} \begin{split} \nabla _{\theta} J_k(\theta) = \nabla _{\theta} \mathbb E_{\theta}\left[ R_k \right] = \mathbb E_{\theta_a}\left[ \nabla _{\theta} \log \pi _{\theta}(a_{k} | s_{k}) R_{k} \right] \end{split} \end{equation} This last expression above gives us an unbiased gradient estimator. We sample the agent action based on the currently learned policy at each dialogue turn and compute the gradient. \section{Experiments} \subsection{Datasets} We evaluate the proposed method on DSTC2 \cite{henderson2014second} dataset in restaurant search domain and an internally collected dialogue corpus in movie booking domain. The movie booking corpus is generated with rule based dialogue agent and user simulator. The same user simulator is used to interact with our end-to-end learning agent during RL training. We use an extended set of NLG templates during model testing to evaluate the end-to-end model's generalization capability in handling diverse natural language inputs. \subsection{Training Settings} We set state size of the dialogue-level and utterance-level LSTM as 200 and 150 respectively. Hidden layer size of the policy network is set as 100. We used randomly initialized word embedding of size 300. Adam optimization method \cite{kingma2014adam} with initial learning rate of 1e-3 is used for mini-batch training. Dropout rate of 0.5 is applied during training to prevent the model from over-fitting. In dialogue simulation, we take a task-oriented dialogue as successful if the goal slot values estimated by the state tracker fully match to the user's true goal values, and the system is able to offer an entity which is finally accepted by the user. Maximum allowed number of dialogue turn is set as 15. A positive reward of +15.0 is given to the agent at the end of a success dialogue, and a zero reward is given in a failure case. We apply a step penalty of -1.0 for each turn to encourage shorter dialogue in completing the task. \subsection{Results and Analysis} Table \ref{} and Table \ref{} show the supervised training model performance on DSTC2 and the movie booking dialogue dataset. The model is evaluated on dialogue state tracking accuracy. On DSTC2 dataset, our end-to-end model achieves near-state-of-the-art state tracking performance comparing to the recent published results using RNN \cite{henderson2014robust} and NBT \cite{mrkvsic2016neural}. On the movie booking dataset, our model also achieves promising performance on individual slot tracking and joint slot tracking accuracy. \begin{table}[th] \caption{Belief tracking results on DSTC2 corpus (with ASR hypothesis as input)} \label{} \centering \begin{tabular}{l c c c c} \hline \textbf{Model} & \textbf{Area} & \textbf{Food} & \textbf{Price} & \textbf{Joint} \\ \hline RNN \cite{henderson2014robust} & 92 & 86 & 86 & 69 \\ NBT \cite{mrkvsic2016neural} & 90 & 84 & 94 & 72 \\ Our end-to-end model & 90 & 84 & 92 & 72 \\ \hline \end{tabular} \end{table} \begin{table}[th] \caption{Belief tracking results on movie booking dataset} \label{} \centering \begin{tabular}{l c c c c c c} \hline \textbf{Model} & \textbf{Num\_ticket} & \textbf{Movie} & \textbf{Theater} & \textbf{Date} & \textbf{Time} & \textbf{Joint} \\ \hline Our end-to-end model & 98.22 & 91.86 & 97.33 & 99.31 & 97.71 & 84.57 \\ \hline \vspace*{-1ex} \end{tabular} \end{table} Figure \ref{} shows the RL curves of the proposed model on dialogue task success rate and average dialogue turn size. Evaluation is based on dialogue simulations between our proposed end-to-end dialogue agent and the rule based user simulator. This is different from the evaluations based on fixed dialogue corpora as in Table \ref{} and \ref{}. The policy gradient based RL training is performed on top of the supervised training model. We compare models with two RL training settings, the end-to-end training and the policy-only training, to the baseline supervised learning (SL) model. As shown in Figure \ref{, the SL model performs poorly during user interaction, indicating the limited generalization capability of the SL model to unseen dialogue state. Any mistake made by the agent during user interaction may lead to deviation of the dialogue from the training dialogue trajectories and states. The SL agent does not know how to recover from an unknown state, which leads to final task failure. RL model training, under both end-to-end learning and policy-only learning settings, continuously improves the task success rate with the growing number of user interactions. We see clear advantage of performing end-to-end model update in achieving higher dialogue task success rate comparing to only updating the policy network during interactive learning. Figure \ref{ shows the learning curves for the average number of turns in successful dialogues. We observe decreasing number of dialogue turns along the growing number of interactive learning episodes. This shows that the dialogue agent learns better strategies to successfully complete the task in fewer numbers of turns. Similar to the results for task success rate, the end-to-end training model outperforms the model with policy-only optimization during RL training, achieving lower average number of dialogue turns in successfully completing a task. \begin{figure}[t] \centering \includegraphics[width=\linewidth]{e2e_training_curves_short.pdf} \vspace*{-4ex} \caption{RL curves on (a) dialogue task success rate and (b) average dialogue turn size.} \label{} \end{figure} \subsection{Human Evaluations} We further evaluate our proposed method with human judges recruited via Amazon Mechanical Turk. Each judge is asked to read a dialogue between our model and the user simulator and rate each system turn on a scale of 1 (frustrating) to 5 (optimal way to help the user). Each turn is rated by 3 different judges. We rate the three models with 100 dialogues each: (i) the SL model, (ii) SL with policy-only RL model, and (iii) SL with end-to-end RL model. Table \ref{} lists the mean and standard deviation of human evaluation scores over all system turns: end-to-end optimization with RL clearly improves the quality of the model according to human judges. \begin{table}[th] \caption{Human evaluation results with mean and standard deviation of crowd worker scores.} \label{} \centering \begin{tabular}{l|c|c|c} \hline \textbf{Model} & SL & SL + policy-only RL & SL + end-to-end RL \\ \hline \textbf{Score} & 3.987 $\pm$ 0.086 & 4.261 $\pm$ 0.089 & 4.394 $\pm$ 0.087 \\ \hline \end{tabular} \end{table} \vspace*{-0.5ex} \section{Conclusions} In this work, we propose a neural network based task-oriented dialogue system that can be trained end-to-end with supervised learning and deep reinforcement learning. We first bootstrap a dialogue agent with supervised training by learning directly from task-oriented dialogue corpora, and further optimize it with deep RL during its interaction with users. We show in the experiments that deep RL optimization on top of the supervised training model leads to significant improvement on task success rate and reduction in dialogue length comparing to supervised training baseline model. The simulation and human evaluation results further illustrate benefits of performing end-to-end model training with deep RL comparing to component-wise optimization.

\subsection{Empirical convergence rates for LLG} \label{example1}% We aim to illustrate the accuracy and the computational effort of Algorithm~\ref{}. We compare the following three strategies for the integration of the lower-order terms: \begin{itemize} \item {\tt TPS2}: fully implicit approach~\eqref{} analyzed in~\cite{akst2014}; \item {\tt TPS2+AB}: Adams--Bashforth approach~\eqref{} proposed and analyzed in this work; \item {\tt TPS2+EE}: explicit Euler approach, where $\boldsymbol{\pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1}) := \boldsymbol{\pi}_h^i(\boldsymbol{m}_h^{i})$. \end{itemize} Besides Figure~\ref{}, in which we compare the effect of different choices for the stabilization function $\rho$, we always employ the canonical choices~\eqref{} for $M$ and $\rho$. In addition, we consider the original (first-order) tangent plane scheme of~\cite{alouges2008a} with $\theta = 1/2$ (i.e., formal Crank--Nicolson-type integrator, which can be seen as a special case of Algorithm~\ref{} with $W_{M(k)} \equiv \alpha$ and $\rho \equiv 0$) together with the following two strategies for the integration of the lower-order terms: \begin{itemize} \item {\tt TPS1+AB}: Adams--Bashforth approach~\eqref{}; \item {\tt TPS1+EE}: explicit Euler approach analyzed in~\cite{akt2012,bffgpprs2014}. \end{itemize} For the first time-step, {\tt TPS2} is used to preserve a possible second-order convergence. To test the schemes, we use the model problem proposed in~\cite{prs2016}: We consider the initial boundary value problem~\eqref{} with $\omega = (0,1)^3$, $\boldsymbol{m}^0 \equiv (1,0,0)$, $\alpha=1$, and $T=5$. For the effective field~\eqref{}, we choose $\ell_{\mathrm{ex}}=1$, a constant field $\boldsymbol{f} \equiv (-2,-0.5,0)$, as well as an operator $\boldsymbol{\pi}$ which only involves the stray field, i.e., $\boldsymbol{\pi}(\boldsymbol{m}) = -\nabla u$; see~\eqref{}. We consider a fixed triangulation $\mathcal T_h$ of $\omega$ generated by Netgen, which consists of \SI{3904}{} elements and 898 nodes (mesh size $h = 1/8$). The exact solution is unknown. To compute the empirical convergence rates, we consider a reference solution $\boldsymbol{m}_{hk_{\rm ref}}$ computed with {\tt TPS2}, using the above mesh and the time-step size $k_{\rm ref} = 5\cdot10^{-5}$. Table~\ref{} gives the average computational time per time-step for each of the considered five integrators. In Figure~\ref{}, we plot the cumulative computational costs for the integration up to the final time $T$. We observe a vast improvement if the lower-order terms (i.e., the stray field) are integrated explicitly in time. The extended (first-order) tangent plane scheme from~\cite{alouges2008a} leads to the cheapest costs, since Algorithm~\ref{} involves additional computations in each time-step for $\lambda_h^i$ as well as the weighted mass matrix corresponding to $\prod{W_{M(k)}(\lambda_h^i)\,\cdot}{\cdot}_\omega$. Figure~\ref{} visualizes the experimental convergence order of the five integrators. As expected, {\tt TPS2} and {\tt TPS+AB} lead to second-order convergence in time. Essentially, both integrators even lead quantitatively to the same accuracy of the numerical solution. {\tt TPS2+EE} as well as {\tt TPS1+EE} yield first-order convergence, since the explicit Euler integration of the stray field is only first-order accurate. Differently from the classical $\theta$-method for linear second-order parabolic PDEs, due to the presence of the nodal projection, the original tangent plane scheme~\cite{alouges2008a} with $\theta=1/2$ (Crank--Nicolson-type {\tt TPS1+AB}) does not lead to any improvement of the convergence order in time (from first-order to second-order). Overall, the proposed {\tt TPS2+AB} integrator appears to be the method of choice with respect to both computational time and empirical accuracy. Figure~\ref{} shows the accuracy of the {\tt TPS2+AB} integrator for different choices of the stabilization $\rho(k) = k^{\delta}$ for $\delta \in \{0,0.1,0.2,0.3,0.4,0.5\}$, where $M(k) = |k \log k|^{-1}$ for all computations. As expected from Proposition~\ref{}, we observe an increase in convergence order for bigger values of $\delta$. In our example, however, the increase is stronger than the analysis predicts, e.g., for $\delta = 0.5$, Proposition~\ref{} predicts convergence order $3/2$, yet, Figure~\ref{} shows order $2$. \begin{figure}[ht] \centering \input{plots/rho_rates.tex} \caption{ Experiment of Section~\ref{example1}: Reference error $\max_j \norm{\boldsymbol{m}_{hk_{\textrm{ref}}}(t_j) - \boldsymbol{m}_{hk}(t_j)}{\LL{2}{\Omega}}$ for $k = 2^\ell \, k_{\rm ref}$ with $\ell \in \{1,2,3,4,5\}$ and $k_{\rm ref} = 5\cdot10^{-5}$. } \label{} \end{figure} \subsection{Spintronic extensions of LLG} \label{} We consider two spintronic extensions of LLG, where the energy-based effective field is supplemented by terms which model the effect of the so-called spin transfer torque (STT)~\cite{slonczewski1996,berger1996}. We show that these extended forms of LLG are covered by the abstract operator $\boldsymbol{\Pi}$ of Section~\ref{} and perform physically relevant numerical experiments with the integrator {\tt TPS2+AB}; cf.~Section~\ref{example1}. \subsubsection{Slonczewski model} \label{} Since the discovery of the GMR effect~\cite{bbfvnpecfc1988,bgsz1989}, magnetic multilayers, i.e., systems consisting of alternating ferromagnetic and nonmagnetic sublayers, have become the subject of intense research in metal spintronics. A phenomenological model to include the spin transfer torque in magnetic multilayer with current-perpendicular-to-plane injection geometry was proposed in~\cite{slonczewski1996}. This model is covered by our framework for LLG~\eqref{} by considering \begin{subequations} \label{} \begin{equation} \label{} \boldsymbol{\Pi}: \LL{\infty}{\omega} \rightarrow \LL{2}{\omega}, \quad \boldsymbol{\Pi}(\boldsymbol{\varphi}) := G(\boldsymbol{\varphi}\cdot\boldsymbol{p}) \, \boldsymbol{\varphi} \times \boldsymbol{p}, \end{equation} where $\boldsymbol{p} \in \mathbb{R}^3$ with $\lvert \boldsymbol{p} \rvert = 1$ is constant, while the function $G: [-1,1] \to \mathbb{R}$ belongs to $C^1([-1,1])$. Using the chain rule and the product rule, we obtain that \begin{equation*} \partial_t \boldsymbol{\Pi}(\boldsymbol{\varphi}) = G'(\boldsymbol{\varphi}\cdot\boldsymbol{p}) (\partial_t \boldsymbol{\varphi} \cdot \boldsymbol{p}) \, \boldsymbol{\varphi} \times \boldsymbol{p} + G(\boldsymbol{\varphi}\cdot\boldsymbol{p}) \, \partial_t \boldsymbol{\varphi} \times \boldsymbol{p}. \end{equation*} Hence, we recover the framework of Section~\ref{} if we consider the operator \begin{equation} \DPi{\boldsymbol{\varphi}}{\boldsymbol{\psi}} = G'(\boldsymbol{\varphi}\cdot\boldsymbol{p}) (\boldsymbol{\psi} \cdot \boldsymbol{p}) \, \boldsymbol{\varphi} \times \boldsymbol{p} + G(\boldsymbol{\varphi}\cdot\boldsymbol{p}) \, \boldsymbol{\psi} \times \boldsymbol{p} \end{equation} as well as the ``discrete'' operators \begin{equation} \boldsymbol{\Pi}_h(\boldsymbol{\varphi}) = \boldsymbol{\Pi}(\boldsymbol{\varphi}) \quad \text{and} \quad \DPih{\boldsymbol{\varphi}}{\boldsymbol{\psi}} = \DPi{\boldsymbol{\varphi}}{\boldsymbol{\psi}}; \end{equation} \end{subequations} see Remark~\ref{}. We show that the resulting approximate operators $\boldsymbol{\Pi}_h^i$, defined in the implicit case by~\eqref{} and in the explicit case by~\eqref{}, and the piecewise constant time reconstruction $\boldsymbol{\Pi}_{hk}^-$ satisfy the assumptions~\eqref{} of Theorem~\ref{}. Let $i=0,\dots,N-1$ and consider arbitrary $\boldsymbol{\psi}_h, \widetilde{\boldsymbol{\psi}}_h \in \boldsymbol{V}_{\!\!h}$ and $\boldsymbol{\varphi}_{h},\widetilde{\boldsymbol{\varphi}}_{h} \in \boldsymbol{\mathcal{M}}_h$. In the implicit case~\eqref{}, the Lipschitz-type continuity~\eqref{} follows from the estimate \begin{equation*} \begin{split} & \norm{\Pih{i}{\boldsymbol{\psi}_h}{\boldsymbol{\varphi}_h}{\widetilde{\boldsymbol{\varphi}}_h} - \Pih{i}{\widetilde{\boldsymbol{\psi}}_h}{\boldsymbol{\varphi}_h}{\widetilde{\boldsymbol{\varphi}}_h}}{\omega} \stackrel{\eqref{} \frac{k}{2} \norm{\DPih{\boldsymbol{\varphi}_h}{\boldsymbol{\psi}_h} - \DPih{\boldsymbol{\varphi}_h}{\widetilde{\boldsymbol{\psi}}_h}}{\omega} \\ & \quad \stackrel{\eqref{} \frac{k}{2} \norm{ G'(\boldsymbol{\varphi}_h\cdot\boldsymbol{p}) [(\boldsymbol{\psi}_h-\widetilde{\boldsymbol{\psi}}_h) \cdot \boldsymbol{p}] \, \boldsymbol{\varphi}_h \times \boldsymbol{p} + G(\boldsymbol{\varphi}_h\cdot\boldsymbol{p}) \, (\boldsymbol{\psi}_h-\widetilde{\boldsymbol{\psi}}_h) \times \boldsymbol{p} }{\omega} \lesssim k \norm{\boldsymbol{\psi}_h-\widetilde{\boldsymbol{\psi}}_h}{\omega}, \end{split} \end{equation*} where the hidden constant in the last estimate depends only on $\norm{G}{C^1([-1,1])}$. In the explicit case~\eqref{}, \eqref{} is trivially satisfied. We prove the stability estimate~\eqref{}: In the implicit case~\eqref{}, it holds that \begin{equation*} \begin{split} & \prod{\Pih{i}{\boldsymbol{\psi}_h}{\boldsymbol{\varphi}_{h}}{\widetilde{\boldsymbol{\varphi}}_{h}}}{\boldsymbol{\psi}_h}_{\omega} \stackrel{\eqref{} \prod{\boldsymbol{\Pi}_h(\boldsymbol{\varphi}_{h}) + \frac{k}{2} \DPih{\boldsymbol{\varphi}_{h}}{\boldsymbol{\psi}_h}}{\boldsymbol{\psi}_h}_{\omega} \\&\quad \stackrel{\eqref{} \prod{G(\boldsymbol{\varphi}_{h}\cdot\boldsymbol{p}) \boldsymbol{\varphi}_{h} \times \boldsymbol{p}}{\boldsymbol{\psi}_h}_{\omega} + \frac{k}{2} \prod{G'(\boldsymbol{\varphi}_{h}\cdot\boldsymbol{p})(\boldsymbol{\psi}_h \cdot \boldsymbol{p}) \boldsymbol{\varphi}_{h} \times \boldsymbol{p}}{\boldsymbol{\psi}_h}_{\omega} \\ & \qquad + \frac{k}{2} \prod{G(\boldsymbol{\varphi}_{h}\cdot\boldsymbol{p})(\boldsymbol{\psi}_h \times \boldsymbol{p})}{\boldsymbol{\psi}_h}_{\omega} \\&\quad \stackrel{\phantom{\eqref{} \norm{\boldsymbol{\psi}_h}{\omega} \left( \norm{\boldsymbol{\varphi}_{h}}{\omega} + k \norm{\boldsymbol{\psi}_h}{\omega} \right). \end{split} \end{equation*} In the explicit case~\eqref{}, it holds that \begin{equation*} \begin{split} & \prod{\Pih{i}{\boldsymbol{\psi}_h}{\boldsymbol{\varphi}_{h}}{\widetilde{\boldsymbol{\varphi}}_{h}}}{\boldsymbol{\psi}_h}_{\omega} \stackrel{\eqref{} \prod{\boldsymbol{\Pi}_h(\boldsymbol{\varphi}_{h}) + \frac{1}{2} \DPih{\boldsymbol{\varphi}_{h}}{\boldsymbol{\varphi}_{h}} - \frac{1}{2} \DPih{\boldsymbol{\varphi}_{h}}{\widetilde{\boldsymbol{\varphi}}_{h}}}{\boldsymbol{\psi}_h}_{\omega} \notag \\ & \stackrel{\eqref{} \prod{G(\boldsymbol{\varphi}_{h}\cdot\boldsymbol{p}) \boldsymbol{\varphi}_{h} \times \boldsymbol{p}}{\boldsymbol{\psi}_h}_{\omega} + \frac{1}{2} \prod{G'(\boldsymbol{\varphi}_{h}\cdot\boldsymbol{p})(\boldsymbol{\varphi}_{h} \cdot \boldsymbol{p}) \boldsymbol{\varphi}_{h} \times \boldsymbol{p}}{\boldsymbol{\psi}_{h}}_{\omega} \\ & \quad + \frac{1}{2} \prod{G(\boldsymbol{\varphi}_{h}\cdot\boldsymbol{p})(\boldsymbol{\varphi}_{h} \times \boldsymbol{p})}{\boldsymbol{\psi}_{h}}_{\omega} - \frac{1}{2} \prod{G'(\boldsymbol{\varphi}_{h}\cdot\boldsymbol{p})(\widetilde{\boldsymbol{\varphi}}_{h} \cdot \boldsymbol{p}) \boldsymbol{\varphi}_{h} \times \boldsymbol{p}}{\boldsymbol{\psi}_{h}}_{\omega} \\ & \qquad - \frac{1}{2} \prod{G(\boldsymbol{\varphi}_{h}\cdot\boldsymbol{p})(\widetilde{\boldsymbol{\varphi}}_h \times \boldsymbol{p})}{\boldsymbol{\psi}_{h}}_{\omega} \\ & \stackrel{\phantom{\eqref{} \norm{\boldsymbol{\psi}_h}{\omega} \left( \norm{\boldsymbol{\varphi}_{h}}{\omega}+ \norm{\widetilde{\boldsymbol{\varphi}}_{h}}{\omega} \right). \end{split} \end{equation*} Finally, we prove the consistency property~\eqref{} with strong convergence: To that end, let $\boldsymbol{\psi}_{hk}$ in $\LL{2}{0,T;\boldsymbol{V}_{\!\!h}}$ and $\boldsymbol{\varphi}_{hk}$, $\widetilde{\boldsymbol{\varphi}}_{hk}$ in $\LL{2}{0,T;\boldsymbol{\mathcal{M}}_h}$ with $\boldsymbol{\psi}_{hk} \rightharpoonup \boldsymbol{\psi}$ and $\boldsymbol{\varphi}_{hk}, \widetilde{\boldsymbol{\varphi}}_{hk} \rightarrow \boldsymbol{\varphi}$ in $\LL{2}{\omega_T}$. As in~\cite{bffgpprs2014}, the Lebesgue dominated convergence theorem proves that \begin{align*} \boldsymbol{\Pi}_h(\boldsymbol{\varphi}_{hk}) \rightarrow \boldsymbol{\Pi}(\boldsymbol{\varphi}) \quad \textrm{in } \LL{2}{\omega_T} \quad \textrm{as } h,k \rightarrow 0. \end{align*} Moreover, we infer from $G \in C^1([-1,1])$ and $\norm{\boldsymbol{\varphi}_{hk}}{\LL{\infty}{\omega_T}} = 1$, that \begin{align*} & \norm{ \DPih{ \boldsymbol{\varphi}_{hk} }{ \boldsymbol{\varphi}_{hk} - \widetilde{\boldsymbol{\varphi}}_{hk} } }{\LL{2}{\omega}} + \norm{ \DPih{ \boldsymbol{\varphi}_{hk} }{ k \boldsymbol{\psi}_{hk} } }{\LL{2}{\omega}} \notag \\ & \quad \stackrel{\eqref{} \norm{ \boldsymbol{\varphi}_{hk} - \widetilde{\boldsymbol{\varphi}}_{hk} }{\LL{2}{\omega}} + k \norm{ \boldsymbol{\psi}_{hk} }{\LL{2}{\omega}} \rightarrow 0 \quad \textrm{as } h,k \rightarrow 0. \end{align*} Thus, for both, the implicit case~\eqref{} and the explicit case~\eqref{}, we get consistency~\eqref{} with strong convergence. \begin{figure}[ht] \centering \begin{subfigure}{0.46\textwidth} \centering \includegraphics[width=0.9\textwidth]{./pics/sublayer.png} \caption{Computational domain.} \label{} \end{subfigure} \begin{subfigure}{0.49\textwidth} \centering \begin{tikzpicture} \pgfplotstableread{plots/switching.dat}{\switching} \pgfplotstableread{plots/energy.dat}{\energy} \begin{axis}[ axis y line*=left, xlabel={Time [\si{\nano\second}]}, ylabel={$\langle m_x \rangle$}, ylabel shift=-0.5cm, width = 0.9\textwidth, xmin=0, xmax=3, ymin=-1.1, ymax=1.1, ] \addplot[red,ultra thick] table[x=t, y=mx]{\switching}; \label{plot:mx} \end{axis} \begin{axis}[ axis y line*=right, axis x line=none, ylabel={Energy [\SI{e-18}{\joule}]}, ylabel shift=-0.2cm, width=0.9\textwidth, xmin=0, xmax=3, ymin=1, ymax=11, ] \addplot[blue,dashed,ultra thick] table[x=t, y=total]{\energy}; \end{axis} \end{tikzpicture} \caption{Time evolutions of $\langle m_x \rangle$ (red solid line, left scale) and of the total energy (blue dashed line, right scale) during the switching process.} \label{} \end{subfigure} \caption{ Experiment of Section~\ref{}: Current-induced switching of a ferromagnetic film.} \end{figure} To test the effectivity of our algorithm, we simulate the writing process of an STT random access memory~\cite{hyybhyyshfnk2005} and reproduce the switching of a ferromagnetic film without external field. The computational domain is an elliptic cylinder of height $d=$ \SI{10}{\nano\meter} (aligned with the $z$-direction) and elliptic cross section with semiaxis lengths $a=$ \SI{60}{\nano\meter} and $b=$ \SI{35}{\nano\meter} (parallel to the $xy$-plane); see Figure~\ref{}. The film is supposed to be the free layer of a magnetic trilayer, which also includes a second ferromagnetic layer (the so-called fixed layer) with constant uniform magnetization $\boldsymbol{p} = (1,0,0)$ and a conducting nonmagnetic spacer. For the material parameters, we choose the values of permalloy: damping parameter $\alpha=$~\num{0.1}, saturation magnetization $M_s=$ \SI{8.0e5}{\ampere\per\meter}, exchange stiffness constant $A=$ \SI{1.3e-11}{\joule\per\meter}, uniaxial anisotropy with constant $K=$ \SI{5.0e2}{\joule\per\meter\cubed} and easy exis $\boldsymbol{a} = (1,0,0)$. With these choices, the effective field takes the form~\eqref{} with \begin{equation*} \ell_{\mathrm{ex}} = \sqrt{\frac{2A}{\mu_0 M_s^2}} \approx \SI{5.7}{\milli\meter}, \quad \boldsymbol{\pi}(\boldsymbol{m})= \frac{2K}{\mu_0 M_s^2} (\boldsymbol{a}\cdot\boldsymbol{m})\boldsymbol{a} - \nabla u, \quad \text{and} \quad \boldsymbol{f} \equiv \boldsymbol{0}, \end{equation*} where $\mu_0 =$ \SI{4 \pi e-7}{\newton\per\ampere\squared} is the vacuum permeability and $u$ is the magnetostatic potential, solution of the transmission problem~\eqref{}. Starting from the uniform initial configuration $\boldsymbol{m}^0 \equiv (-1,0,0)$, we solve~\eqref{} with $\boldsymbol{\Pi} \equiv \boldsymbol{0}$ for \SI{1}{\nano\second} to reach a relaxed state. Then, we inject a spin-polarized electric current with intensity $J_e =$ \SI{1e11}{\ampere\per\meter\squared} for \SI{1}{\nano\second}. The resulting spin transfer torque is modeled by the operator~\eqref{}, where the function $G$ takes the phenomenological expression \begin{equation*} G(x) = \frac{\hbar J_e}{e \mu_0 M_s^2 d} \left[ \frac{(1+P)^3 (3+x)}{4 P^{3/2}} -4 \right]^{-1} \quad \text{for all } x \in [-1,1]; \end{equation*} see~\cite{slonczewski1996}. Here, $\hbar =$ \SI{1.054571800e-34}{\joule\second} is the reduced Planck constant, $e =$ \SI{1.602176621e-19}{\coulomb} is the elementary charge, while $P = 0.8$ is the polarization parameter. We solve~\eqref{} with $\boldsymbol{\Pi} \equiv \boldsymbol{0}$ for \SI{1}{\nano\second} to relax the system to the new equilibrium. For the spatial discretization, we consider a tetrahedral triangulation with mesh size \SI{3}{\nano\meter} (\num{15 885} elements, \num{3 821} nodes) generated by Netgen. For the time discretization, we consider a constant time-step size of~\SI{0.1}{\pico\second} (\num{30 000} time-steps). The time evolutions of the average $x$-component $\langle m_x \rangle$ of the magnetization in the free layer and of the total energy~\eqref{} are depicted in Figure~\ref{}. Since the size of the film is well under the single-domain limit (see, e.g., \cite{adm2015}), the initial uniform magnetization configuration is preserved by the first relaxation process. By applying a perpendicular spin-polarized current, the uniform magnetization of the free layer can be switched from $(-1,0,0)$ to $\boldsymbol{p}=(1,0,0)$. The fundamental physics underlying this phenomenon is understood as the mutual transfer of spin angular momentum between the $\boldsymbol{p}$-polarized conduction electrons and the magnetization of the film. During the switching process, the classical energy dissipation modulated by the damping parameter $\alpha$ is lost as an effect of the Slonczewski contribution; cf.\ the fourth term on the left-hand side of~\eqref{}. The new state is also stable and is preserved by the final relaxation process. \subsubsection{Zhang--Li model}\label{} In~\cite{zl2004}, the authors derived an extended form of LLG to model the effect of an electric current flow on the magnetization dynamics in single-phase samples characterized by a current-in-plane injection geometry. A similar equation was obtained in a phenomenological way in~\cite{tnms2005} for the description of the current-driven motion of domain walls in patterned nanowires. Here, the operator takes the form \begin{subequations} \label{} \begin{equation} \label{} \boldsymbol{\Pi}: \boldsymbol{H}{1}{\omega} \cap \LL{\infty}{\omega} \rightarrow \LL{2}{\omega}, \quad \boldsymbol{\Pi}(\boldsymbol{\varphi}) := \boldsymbol{\varphi} \times ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi} + \beta ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi}, \end{equation} where $\boldsymbol{u} \in \LL{\infty}{\omega}$ and $\beta>0$ is constant. The product rule yields that \begin{equation*} \partial_t \boldsymbol{\Pi}(\boldsymbol{\varphi}) = \partial_t \boldsymbol{\varphi} \times ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi} + \boldsymbol{\varphi} \times ( \boldsymbol{u} \cdot \nabla ) \partial_t \boldsymbol{\varphi} + \beta ( \boldsymbol{u} \cdot \nabla ) \partial_t \boldsymbol{\varphi}. \end{equation*} Hence, we recover the framework of Section~\ref{} if we consider the operators \begin{gather} \DPi{\boldsymbol{\varphi}}{\boldsymbol{\psi}} = \boldsymbol{\psi} \times ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi} + \boldsymbol{\varphi} \times ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\psi} + \beta ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\psi}, \\ \boldsymbol{\Pi}_h(\boldsymbol{\varphi}) = \boldsymbol{\Pi}(\boldsymbol{\varphi}), \quad \text{and} \quad \DPih{\boldsymbol{\varphi}}{\boldsymbol{\psi}} = \DPi{\boldsymbol{\varphi}}{\boldsymbol{\psi}}. \end{gather} \end{subequations} Under the additional assumption \begin{equation} \label{} \sup_{h>0} \norm{\nabla\boldsymbol{m}_h^0}{\LL{\infty}{\omega}} \leq C, \end{equation} we show that Theorem~\ref{ (a)}-{\rm (b)} still holds for the Adams--Bashforth approach from Remark~\ref{ (ii)}. To see this, recall Remark~\ref{ (iv)}. First, note that the Lipschitz-type continuity~\eqref{} is trivially satisfied in the explicit case ($i=1,\dots,N-1$). In the implicit case ($i=0$), it is not fulfilled for all $\boldsymbol{\psi}_h, \widetilde{\boldsymbol{\psi}}_h \in \boldsymbol{V}_{\!\!h}$ and $\boldsymbol{\varphi}_{h},\widetilde{\boldsymbol{\varphi}}_{h} \in \boldsymbol{\mathcal{M}}_h$. However, thanks to~\eqref{}, the desired inequality is satisfied in the specific situation of Remark~\ref{ (iv)}. Indeed, for arbitrary $\boldsymbol{\psi}_h, \widetilde{\boldsymbol{\psi}}_h \in \boldsymbol{V}_{\!\!h}$, it holds that \begin{equation}\label{} \begin{split} & \norm{\Pih{0}{\boldsymbol{\psi}_h}{\boldsymbol{m}_h^0}{\boldsymbol{m}_h^{-1}} - \Pih{0}{\widetilde{\boldsymbol{\psi}}_h}{\boldsymbol{m}_h^0}{\boldsymbol{m}_h^{-1}}}{\omega} \stackrel{\eqref{} \frac{k}{2} \norm{\DPih{\boldsymbol{m}_h^0}{\boldsymbol{\psi}_h} - \DPih{\boldsymbol{m}_h^0}{\widetilde{\boldsymbol{\psi}}_h}}{\omega} \\ & \stackrel{\eqref{} \frac{k}{2} \norm{(\boldsymbol{\psi}_h-\widetilde{\boldsymbol{\psi}}_h) \times ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{m}_h^0 + \boldsymbol{m}_h^0 \times ( \boldsymbol{u} \cdot \nabla ) (\boldsymbol{\psi}_h-\widetilde{\boldsymbol{\psi}}_h) + \beta ( \boldsymbol{u} \cdot \nabla ) (\boldsymbol{\psi}_h-\widetilde{\boldsymbol{\psi}}_h)}{\omega} \\ & \stackrel{\phantom{\eqref{} k \, \norm{\boldsymbol{\psi}_h-\widetilde{\boldsymbol{\psi}}_h}{\boldsymbol{H}{1}{\omega}}, \end{split} \end{equation} where the hidden constant depends on $C$ and $\norm{\boldsymbol{u}}{\LL{\infty}{\omega}}$. Next, we prove the stability estimate~\eqref{}: In the implicit case~\eqref{}, it holds that \begin{equation*} \begin{split} & \prod{\Pih{i}{\boldsymbol{\psi}_h}{\boldsymbol{\varphi}_{h}}{\widetilde{\boldsymbol{\varphi}}_{h}}}{\boldsymbol{\psi}_h}_{\omega} \stackrel{\eqref{} \prod{\boldsymbol{\Pi}_h(\boldsymbol{\varphi}_{h}) + \frac{k}{2} \DPih{\boldsymbol{\varphi}_{h}}{\boldsymbol{\psi}_h}}{\boldsymbol{\psi}_h}_{\omega} \\ & \quad \stackrel{\eqref{} \prod{\boldsymbol{\varphi}_{h} \times ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi}_{h}}{\boldsymbol{\psi}_h}_{\omega} + \beta \prod{( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi}_{h}}{\boldsymbol{\psi}_h}_{\omega} \\ & \qquad + \frac{k}{2} \prod{\boldsymbol{\varphi}_{h} \times ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\psi}_h}{\boldsymbol{\psi}_h}_{\omega} + \frac{\beta}{2} \, k \, \prod{( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\psi}_h}{\boldsymbol{\psi}_h}_{\omega} \\&\quad \stackrel{\phantom{\eqref{} \norm{\boldsymbol{\psi}_h}{\omega} \big( \norm{\nabla \boldsymbol{\varphi}_{h}}{\omega} + k \norm{\nabla \boldsymbol{\psi}_h}{\omega} \big). \end{split} \end{equation*} In the explicit case~\eqref{}, it holds that \begin{equation*} \begin{split} & \prod{\Pih{i}{\boldsymbol{\psi}_h}{\boldsymbol{\varphi}_{h}}{\widetilde{\boldsymbol{\varphi}}_{h}}}{\boldsymbol{\psi}_h}_{\omega} \stackrel{\eqref{} \prod{\boldsymbol{\Pi}_h(\boldsymbol{\varphi}_{h}) + \frac{1}{2} \DPih{\boldsymbol{\varphi}_{h}}{\boldsymbol{\varphi}_{h}} - \frac{1}{2} \DPih{\boldsymbol{\varphi}_{h}}{\widetilde{\boldsymbol{\varphi}}_{h}}}{\boldsymbol{\psi}_h}_{\omega} \notag \\ & \stackrel{\eqref{} \prod{\boldsymbol{\varphi}_{h} \times ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi}_{h}}{\boldsymbol{\psi}_h}_{\omega} + \beta \prod{( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi}_{h}}{\boldsymbol{\psi}_h}_{\omega} + \prod{\boldsymbol{\varphi}_{h} \times ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi}_{h}}{\boldsymbol{\psi}_h}_{\omega} \\ & \quad + \frac{\beta}{2} \, \prod{( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi}_{h}}{\boldsymbol{\psi}_h}_{\omega} - \frac{1}{2} \, \prod{\widetilde{\boldsymbol{\varphi}}_{h} \times ( \boldsymbol{u} \cdot \nabla ) \boldsymbol{\varphi}_{h}}{\boldsymbol{\psi}_h}_{\omega} \\ & \quad - \frac{1}{2} \, \prod{\boldsymbol{\varphi}_{h} \times ( \boldsymbol{u} \cdot \nabla ) \widetilde{\boldsymbol{\varphi}}_{h}}{\boldsymbol{\psi}_h}_{\omega} -\frac{\beta}{2} \, \prod{( \boldsymbol{u} \cdot \nabla ) \widetilde{\boldsymbol{\varphi}}_{h}}{\boldsymbol{\psi}_h}_{\omega} \\ & \stackrel{\phantom{\eqref{} \norm{\boldsymbol{\psi}_h}{\omega} \big( \norm{\nabla \boldsymbol{\varphi}_{h}}{\omega}+ \norm{\nabla \widetilde{\boldsymbol{\varphi}}_{h}}{\omega} \big). \end{split} \end{equation*} To see~\eqref{}, let $\boldsymbol{\zeta} \in \LL{2}{\omega_T}$. We show that \begin{align}\label{} \Int{0}{T} \prod{ \boldsymbol{\Pi}_{hk}^-(\boldsymbol{v}_{hk}^- ; \boldsymbol{m}_{hk}^- , \boldsymbol{m}_{hk}^= ) }{ \boldsymbol{\zeta} }_{\omega} \d{t} \rightarrow \Int{0}{T} \prod{ \boldsymbol{\Pi}(\boldsymbol{m}) }{ \boldsymbol{\zeta} }_{\omega} \d{t} \quad \textrm{as } h,k \rightarrow 0. \end{align} Unwrapping the Adams--Bashforth approach from Remark~\ref{ (ii)}, we see that \begin{align*} & \Int{0}{T} \prod{ \boldsymbol{\Pi}_{hk}^-(\boldsymbol{v}_{hk}^- ; \boldsymbol{m}_{hk}^- , \boldsymbol{m}_{hk}^= ) }{ \boldsymbol{\zeta} }_{\omega} \d{t} \\ & = \Int{0}{k} \prod{ \boldsymbol{\Pi}_{hk}(\boldsymbol{m}_{hk}^-) }{ \boldsymbol{\zeta} }_{\omega} \d{t} + \frac{k}{2} \Int{0}{k} \prod{ \DPih{\boldsymbol{v}_{hk}^-}{\boldsymbol{m}_{hk}^-} }{ \boldsymbol{\zeta} }_{\omega} \d{t} \\ & \qquad + \Int{k}{T} \prod{ \boldsymbol{\Pi}_{hk}(\boldsymbol{m}_{hk}^-) }{ \boldsymbol{\zeta} }_{\omega} \d{t} + \frac12 \Int{k}{T} \prod{ \DPih{\boldsymbol{m}_{hk}^-}{\boldsymbol{m}_{hk}^-}}{ \boldsymbol{\zeta} }_{\omega} \d{t} - \frac12 \Int{k}{T} \prod{ \DPih{\boldsymbol{m}_{hk}^-}{\boldsymbol{m}_{hk}^=}}{ \boldsymbol{\zeta} }_{\omega} \d{t} \\ & =: T_{hk}^1 + \dots + T_{hk}^5. \end{align*} With the definitions from~\eqref{}, the convergence properties from Lemma~\ref{} prove that \begin{align*} T_{hk}^1 \rightarrow 0 \quad \textrm{and} \quad T_{hk}^3 + T_{hk}^4 + T_{hk}^5 \rightarrow \Int{0}{T} \prod{\boldsymbol{\Pi}(\boldsymbol{m})}{\boldsymbol{\zeta}}_{\Omega} \d{t} \quad \textrm{as } h,k \rightarrow 0. \end{align*} It remains to show that $T_{hk}^{2} \rightarrow 0$ as $h,k \rightarrow 0$. Note that \begin{align*} T_{hk}^2 & \stackrel{\eqref{} \frac{k}{2} \Int{0}{k} \prod{\boldsymbol{v}_{hk}^- \times (\boldsymbol{u} \cdot \nabla) \boldsymbol{m}_{hk}^-}{\boldsymbol{\zeta}}_{\omega} \d{t} \notag \\ & \quad + \frac{k}{2} \Int{0}{k} \prod{\boldsymbol{m}_{hk}^- \times (\boldsymbol{u} \cdot \nabla) \boldsymbol{v}_{hk}^-}{\boldsymbol{\zeta}}_{\omega} \d{t} + \frac{\beta k}{2} \Int{0}{k} \prod{(\boldsymbol{u} \cdot \nabla) \boldsymbol{v}_{hk}^-}{\boldsymbol{\zeta}}_{\omega} \d{t}. \end{align*} With the convergence properties from Lemma~\ref{}, the last two terms of $T_{hk}^2$ go to $0$ as $h,k \rightarrow 0$. To deal with the first term, we derive from $\boldsymbol{m}_{hk}^-(t)|_{[0,k)} = \boldsymbol{m}_h^0$ that \begin{align*} &\frac{k}{2} \Int{0}{k} \prod{\boldsymbol{v}_{hk}^- \times (\boldsymbol{u} \cdot \nabla) \boldsymbol{m}_{hk}^-}{\boldsymbol{\zeta}}_{\omega} \d{t} = \frac{k}{2} \Int{0}{k} \prod{\boldsymbol{v}_{hk}^- \times (\boldsymbol{u} \cdot \nabla) \boldsymbol{m}_{h}^0}{\boldsymbol{\zeta}}_{\omega} \d{t} \\ & \qquad \lesssim k \, \norm{\boldsymbol{u}}{\LL{\infty}{\omega}} \norm{\nabla \boldsymbol{m}_{h}^0}{\LL{\infty}{\omega_T}} \norm{\boldsymbol{v}_{hk}^-}{\omega_T} \norm{\boldsymbol{\zeta}}{\omega_T} \stackrel{\eqref{} 0 \quad \textrm{as } h,k \rightarrow 0. \end{align*} This verifies~\eqref{}. \begin{figure}[ht] \begin{subfigure}{0.48\textwidth} \includegraphics[width=0.9\textwidth]{./pics/mumag5_initial.png} \caption{Initial vortex state.} \label{} \end{subfigure}\hspace*{\fill} \begin{subfigure}{0.48\textwidth} \includegraphics[width=0.9\textwidth]{./pics/mumag5_final.png} \caption{Final vortex state.} \label{} \end{subfigure} \caption{Experiment of Section~\ref{}: Vortex states in $\mu$MAG standard problem \#5. The color scale refers to the $x$-component of the magnetization.} \end{figure} \begin{figure}[ht] \centering \begin{tikzpicture} \pgfplotstableread{plots/sp5-oommf.dat}{\oommf} \pgfplotstableread{plots/sp5-tps2ab.dat}{\tpsab} \pgfplotstableread{plots/sp5-mpsab.dat}{\mpsab} \begin{axis}[ width = 100mm, xlabel={Time [\si{\nano\second}]}, xmin=0, xmax=8, ymin=-0.4, ymax=0.4, legend columns=2, legend style={/tikz/column 2/.style={column sep=5pt}}, ] \addplot[blue,ultra thick] table[x=t, y=mx]{\tpsab}; \addplot[cyan,ultra thick] table[x=t, y=my]{\tpsab}; \addplot[purple,dashed,ultra thick] table[x=t, y=mx]{\oommf}; \addplot[red,dashed,ultra thick] table[x=t, y=my]{\oommf}; \addplot[brown,densely dotted,ultra thick] table[x=t, y=mx]{\mpsab}; \addplot[olive,densely dotted,ultra thick] table[x=t, y=my]{\mpsab}; \legend{ \texttt{TPS2+AB} $\langle m_x \rangle$, \texttt{TPS2+AB} $\langle m_y \rangle$, \texttt{OOMMF} $\langle m_x \rangle$, \texttt{OOMMF} $\langle m_y \rangle$, \texttt{MPS} $\langle m_x \rangle$, \texttt{MPS} $\langle m_y \rangle$ } \end{axis} \end{tikzpicture} \caption{Experiment of Section~\ref{}: Comparison of the results obtained for $\mu$MAG standard problem \#5 with \texttt{TPS2+AB}, \texttt{OOMMF}, and \texttt{MPS}.} \label{} \end{figure} The extended LLG forms of~\cite{zl2004,tnms2005} are the subject of the $\mu$MAG standard problem \#5, proposed by the Micromagnetic Modeling Activity Group of the National Institute of Standards and Technology (NIST)~\cite{MUMAG}. The sample under consideration is a permalloy film with dimensions \SI[scientific-notation=false]{100}{\nano\meter} $\times$ \SI[scientific-notation=false]{100}{\nano\meter} $\times$ \SI[scientific-notation=false]{10}{\nano\meter}, aligned with the $x$, $y$, and $z$ axes of a Cartesian coordinate system, with origin at the center of the film. For the material parameters, we consider the same values as in Section~\ref{}, except for the magnetocrystalline anisotropy, which is neglected, i.e., $K=0$. The initial state is obtained by solving~\eqref{} for the initial condition $\boldsymbol{m}^0(x,y,z) = (-y,x,10) / \sqrt{x^2 + y^2 + 100}$ and $\boldsymbol{\Pi} \equiv \boldsymbol{0}$ for a sufficiently long time, until the system relaxes to equilibrium; see Figure~\ref{}. Given the spin velocity vector $\boldsymbol{u}_T = (-72.17,0,0)$ \si{\meter\per\second} and the gyromagnetic ratio $\gamma_0 =$ \SI{2.21e5}{\meter\per\ampere\per\second}, we define $\boldsymbol{\Pi}$ by~\eqref{} for $\boldsymbol{u}=-\boldsymbol{u}_T/(\gamma_0 M_s)$ and $\beta = 0.05$. With the relaxed magnetization configuration as initial condition, we solve~\eqref{} for \SI{8}{\nano\second}, which turns out to be a sufficiently long time to reach the new equilibrium; see Figure~\ref{}. For the simulation, we consider a tetrahedral triangulation of the domain into \num{25 666} elements with maximal diameter \SI{3}{\nano\meter} (\num{5 915} nodes) generated by Netgen. For the time discretization, we consider a constant time-step size of~\SI{0.1}{\pico\second} (\num{80 000} \emph{uniform} time-steps). We compare our results with those obtained with the finite difference code \texttt{OOMMF}~\cite{dp1999} and with the implicit-explicit midpoint scheme of~\cite{prs2016} (\texttt{MPS}). For~\texttt{OOMMF}, we consider the data downloadable from the $\mu$MAG homepage~\cite{MUMAG}, which refer to a uniform partition of the computational domain into \num{12 500} cubes with \SI{2}{\nano\meter} edge and \num{42350} \emph{adaptive} time-steps. In~\texttt{OOMMF}, the solution of~\eqref{} for the stray field computation is based on a fast Fourier transform algorithm. The results for~\texttt{MPS} refer to the same triangulation used for \texttt{TPS2+AB}, but are obtained with a \num{20} times smaller time-step size (\SI{5}{\femto\second}, i.e., \num{1600000} time-steps), which is necessary to ensure the well-posedness of the fixpoint iteration which solves the nonlinear system~\cite{prs2016}. Figure~\ref{} shows the evolution of the averages $\langle m_x\rangle$ and $\langle m_y\rangle$ of the $x$- and $y$-component of $\boldsymbol{m}$, respectively. Despite the different nature of the methods, the results are in full qualitative agreement. \subsection{Empirical convergence rates for ELLG}\label{example4} We aim to illustrate the accuracy and the computational effort of Algorithm~\ref{}. We neglect $\boldsymbol{m}$-dependent lower-order terms and dissipative effects, i.e., $\boldsymbol{\pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1}) = \boldsymbol{0} = \boldsymbol{\Pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1})$, and compare the following four strategies: \begin{itemize} \item {\tt FC}: The fully-coupled approach $\boldsymbol{h}_h^{i,\boldsymbol{\Theta}} = \midh{\boldsymbol{h}}{i}$ from~\eqref{} for all $i = 0,\dots, N-1$; \item {\tt DC-2}:~The decoupled second-order Adams--Bashforth approach from Remark~\ref{ (ii)}; \item {\tt DC-1}:~The decoupled explicit Euler approach $\boldsymbol{h}_h^{i,\boldsymbol{\Theta}} = \boldsymbol{h}_h^i$ for all $i = 0,\dots, N-1$; \item {\tt SF}: The simplified decoupled Adams--Bashforth approach from Remark~\ref{ (iv)}. \end{itemize} We always employ the canonical choices~\eqref{} for $M$ and $\rho$. For the first time-step, we always use {\tt FC} to preserve a possible second-order convergence in time. For all implicit time-steps, the system~\eqref{} of Algorithm~\ref{} is solved by the fixpoint iteration from the proof of Theorem~\ref{} which is stopped if \begin{align*} \norm{\boldsymbol{\eta}_h^\ell - \boldsymbol{\eta}_h^{\ell-1}}{\omega} + \norm{\boldsymbol{\nu}_h^{\ell}-\boldsymbol{\nu}_h^{\ell-1}}{\Omega} \le 10^{-10}. \end{align*} To test the schemes, we use a setting similar to that of Section~\ref{example1}: We choose $\omega = (-0.125,0.125)^3$, $\Omega = (0,1)^3$, and $T=7$. For the LLG-part~\eqref{}, we choose $\alpha = 1$, $\ell_{\rm ex}^2 = 1$, and set $\boldsymbol{f} = (f_1,0,0)^T \in \boldsymbol{C}^1([0,T])$, where \begin{align*} f_1(t) := \begin{cases} 15 t^2 & \textrm{for } 0 \leq t \leq 1, \\ 30 - 15(t - 2)^2 & \textrm{for } 1 < t \leq 2, \\ 30 & \textrm{for } 2 < t \leq 4, \\ 30 - 15(t - 4)^2 & \textrm{for } 4 < t \leq 5, \\ 15 (t-6)^2 & \textrm{for } 5 < t \leq 6, \\ 0 & \textrm{for } 6 < t \leq 7. \end{cases} \end{align*} For the eddy current part~\eqref{}, we choose $\mu_0 = 1$, and $\sigma \in \L{\infty}{\Omega}$ with $\sigma|_{\omega} = 100$ and $\sigma|_{\Omega \setminus \overline{\omega}} = 1$. We use a fixed triangulation $\mathcal{T}_h$ of $\Omega$ generated by Netgen, which resolves $\omega$ and satisfies \begin{itemize} \item $\max_{K \in \mathcal{T}_h|_{\omega}} h_K = 0.03$ on the sub-mesh $\mathcal{T}_h|_{\omega}$, \item $\max_{K \in \mathcal{T}_h|_{\Omega \setminus \omega}} h_K = 0.125$ on the outer mesh $\mathcal{T}_h \setminus \mathcal{T}_h|_{\omega}$. \end{itemize} This yields \num{2 388} elements and \num{665} nodes for $\mathcal{T}_h|_{\omega}$ as well as \num{22 381} elements and \num{4 383} nodes for the overall mesh $\mathcal{T}_h$. In all cases, the initial values $\boldsymbol{m}_h^0$ and $\boldsymbol{h}_h^0$ are obtained with setting $\boldsymbol{f} = 0$ and relaxing with {\tt FC} the nodal interpolation of \begin{align*} \boldsymbol{m}^0 = (-1,-1,-1) / \sqrt{3} \qquad \textrm{and} \qquad \boldsymbol{h}^0(\boldsymbol{x}) = \begin{cases} - \boldsymbol{m}^0(\boldsymbol{x}) & \textrm{for } \boldsymbol{x} \in \overline{\omega}, \\ 0 & \textrm{for } \boldsymbol{x} \in \overline{\Omega} \setminus \overline{\omega}. \end{cases} \end{align*} The exact solution is unknown. To compute the empirical convergence rates, we consider a reference solution $\boldsymbol{m}_{hk_{\rm ref}}$ computed with {\tt DC-2}, using the above mesh and the time-step size $k_{\textrm{ref}} = 2^{-14}$. Figure~\ref{} visualizes the experimental convergence orders of the four integrators. As expected, the fully-coupled approach {\tt FC} and the decoupled approach {\tt DC-2} lead to second-order convergence in time. As mentioned in Remark~\ref{ (iv)}, the simplified and fully linear approach {\tt SF} as well as {\tt DC-1} only lead to first-order convergence in time. Moreover, from the second time-step on the decoupled approach {\tt DC-2} is computationally as expensive as {\tt DC-1}. In contrast to that, the fully-coupled approach {\tt FC} employs a fixed-point iteration at each time-step and thus comes at the highest computational cost. Overall, the decoupled approach {\tt DC-2} appears to be the method of choice with respect to both, computational time and empirical accuracy. \begin{figure}[ht] \centering \begin{subfigure}{0.48\textwidth} \scalebox{0.7}{ \input{plots/H1_rates_ELLG.tex} } \end{subfigure} \begin{subfigure}{0.48\textwidth} \scalebox{0.7}{ \input{plots/Hcurl_rates_ELLG.tex} } \end{subfigure} \caption{Experiment of Section~\ref{example4}: Reference error $\max_j ( \norm{\boldsymbol{m}_{hk_{\rm ref}}(t_j) - \boldsymbol{m}_{hk}(t_j)}{\boldsymbol{H}{1}{\omega}})$ (left) and $\max_j ( \norm{\boldsymbol{h}_{hk_{\rm ref}}(t_j) - \boldsymbol{h}_{hk}(t_j)}{\Hcurl{\Omega}} )$ (right) for $k = 2^\ell \, k_{\rm ref}$ with $\ell \in \{1,2,3,4,5\}$ and $k_{\rm ref} = 5\cdot10^{-5}$.} \label{} \end{figure} \section{Introduction} \subsection{State of the art} Nowadays, the study of magnetization processes in magnetic materials and the development of fast and reliable tools to perform large-scale micromagnetic simulations are the focus of considerable research as they play a fundamental role in the design of many technological devices. Applications to magnetic recording, in which the external field can change fast so that the hysteresis properties are not accurately described by a static approach, require accurate numerical methods to study the dynamics of the magnetization distribution. In this context, a well-established model to describe the time evolution of the magnetization in ferromagnetic materials is the Landau--Lifshitz--Gilbert equation (LLG) introduced in~\cite{ll1935,gilbert1955}. Recently, the numerical integration of LLG has been the subject of many mathematical studies; see, e.g., the review articles~\cite{kp2006,garciaCervera2007,cimrak2008b}, the monograph~\cite{prohl2001} and the references therein. The main challenges concern the strong nonlinearity of the problem, a nonconvex unit-length constraint which enforces the solutions to take their pointwise values in the unit sphere, an intrinsic energy law, which combines conservative and dissipative effects and should be preserved by the numerical scheme, as well as the presence of nonlocal field contributions, which prescribe the (possibly nonlinear) coupling with other partial differential equations (PDEs). One important aspect of the research is related to the development of unconditionally convergent methods, for which the numerical analysis does not require to impose any CFL-type condition on the relation of the spatial and temporal discretization parameters. The seminal works~\cite{bp2006,alouges2008a} propose numerical time-marching schemes, based on lowest-order finite elements in space, which are proven to be unconditionally convergent towards a weak solution of LLG in the sense of~\cite{as1992}. The implicit midpoint scheme of~\cite{bp2006} is formally of second order in time. It inherently preserves some of the fundamental properties of LLG, such as the pointwise constraint (at the nodes of the mesh) and the energy law. However, it requires the solution of a nonlinear system of equations per time-step. The tangent plane scheme of~\cite{alouges2008a} is based on an equivalent reformulation of LLG, which gives rise to a variational formulation in the tangent space of the current magnetization state. It requires only the solution of one linear system per time-step and employs the nodal projection at each time-step to enforce the pointwise constraint at a discrete level. For an implicit-explicit approach for the full effective field, we refer to~\cite{akt2012,bffgpprs2014}. Moreover, extensions for the discretization of the coupling of LLG with the full Maxwell equations, the eddy current equation, a balance law for the linear momentum (magnetostriction), and a spin diffusion equation for the spin accumulation have been considered in~\cite{lt2013,bppr2013,ahpprs2014,lppt2015,bpp2015,ft2017}. Formally, the tangent plane scheme and its aforementioned extensions are of first order in time. The nodal projection step can be omitted and this leads to an additional consistency error which is also (formally) first-order in time~\cite{ahpprs2014}. For this projection-free variant, the recent work~\cite{ft2017} derives rigorous \emph{a~priori} error estimates which are of first order in time and space. A tangent plane scheme with an improved convergence order in time is introduced and analyzed in~\cite{akst2014}. Like~\cite{alouges2008a}, the proposed method is based on a predictor-corrector approach which combines a linear reformulation of the equation with the use of the nodal projection for the numerical treatment of the pointwise constraint. However, the variational formulation for the linear update is designed in such a way that the scheme has a consistency error of order $2-\varepsilon$ for any $0<\varepsilon<1$. For this reason, this method is named \emph{(almost) second-order tangent plane scheme} in~\cite{akst2014}. \subsection{Contributions of the present work} In this work, we propose a threefold extension of the improved tangent plane scheme from~\cite{akst2014}: {\large$\bullet$} During the design and the implementation of a micromagnetic code, one of the main issues concerns the computation of the nonlocal magnetostatic interactions. In many situations, it turns out to be the most time-consuming part of micromagnetic simulations~\cite{aesds2013}. To cope with this problem, we follow the approach of~\cite{prs2016} and propose an implicit-explicit treatment for the lower-order effective field contributions. Then, only one expensive stray field computation per time-step needs to be carried out. Nevertheless, our time-stepping preserves the (almost) second-order convergence in time of the scheme as well as the unconditional convergence result. {\large$\bullet$} The discovery of the giant magnetoresistance (GMR) effect in~\cite{bbfvnpecfc1988,bgsz1989} determined a breakthrough in magnetic hard disk storage capacity and encouraged several extensions of the micromagnetic model, which aim to describe the effect of spin-polarized currents on magnetic materials. The most used approaches involve extended forms of LLG, in which the classical energy-based effective field is augmented by additional terms in order to take into account the spin transfer torque effect; see~\cite{slonczewski1996,zl2004,tnms2005}. In this work, we extend the abstract setting, the proposed algorithm, and the convergence result so that the aforementioned extended forms of LLG are covered by our analysis. {\large$\bullet$} For the treatment of systems in which LLG is bidirectionally coupled with another time-dependent PDE, the works~\cite{bppr2013,ahpprs2014,lppt2015,bpp2015} propose integrators which completely decouple the time integration of LLG and the coupled equation. This treatment is very attractive in terms of computational cost and applicability of the scheme, since existing implementations, including solvers and preconditioning strategies for the building blocks of the system, can easily be reused. We show how such an approach can also be adopted for the improved tangent plane scheme. Combined with a second-order method for the coupled equation, this leads to algorithms of global (almost) second order, for which the convergence result can be generalized. As an illustrative example, we analyze the coupling of LLG with eddy currents (ELLG), which is of relevant interest in several concrete applications~\cite{bmms2002,scmw2004,hseskdf2005}. \subsection{Outline} The remainder of this work is organized as follows: We conclude this section by recalling the notation used throughout the paper. In Section~\ref{} and in Section~\ref{}, we present the problem formulation and introduce the numerical algorithm. Then, we state the convergence result for pure LLG and ELLG, respectively. In Section~\ref{}, for the convenience of the reader, we reformulate the argument of~\cite{akst2014} and propose a formal derivation of the (almost) second-order algorithm. In Section~\ref{} and Section~\ref{}, we prove the main results for pure LLG (Theorem~\ref{}) and ELLG (Theorem~\ref{}), respectively. Finally, Section~\ref{} is devoted to numerical experiments. \subsection{General notation} Throughout this work, we use the standard notations for Lebesgue, Sobolev, and Bochner spaces and norms. For any domain $U$, we denote the scalar product in $L^2(U)$ by $\prod{\cdot}{\cdot}_U$ and the corresponding norm by $\norm{\cdot}{U}$. Vector-valued functions (as well as the corresponding function spaces) are indicated by bold letters. To abbreviate notation in proofs, we write $A \lesssim B$ when $A \le c B$ for some generic constant $c > 0$, which is clear from the context and always independent of the discretization parameters. Moreover, $A \simeq B$ abbreviates $A \lesssim B \lesssim A$. \section{Landau--Lifshitz--Gilbert equation} \label{} \subsection{LLG and weak solutions}% \label{} For a bounded Lipschitz domain $\omega\subset\mathbb{R}^3$, the Gilbert form of LLG reads \begin{subequations}\label{} \begin{align} \label{} \partial_t \boldsymbol{m} &= -\boldsymbol{m}\times \big( \boldsymbol{h}_{\operatorname{eff}}(\boldsymbol{m}) \, + \, \boldsymbol{\Pi}(\boldsymbol{m}) \big) + \alpha\,\boldsymbol{m}\times\partial_t \boldsymbol{m} &&\text{ in } \omega_T := (0,T)\times\omega \,, \\ \label{} \partial_{\boldsymbol{n}}\boldsymbol{m} &= \boldsymbol{0} &&\text{ on } (0,T)\times\partial\omega \,, \\ \label{} \boldsymbol{m}(0) &= \boldsymbol{m}^0 &&\text{ in } \omega, \end{align} where $T>0$ is the final time, $\alpha \in (0,1]$ is the Gilbert damping constant, and $\boldsymbol{m}^0\in\boldsymbol{H}{1}{\omega}$ is the initial data with $|\boldsymbol{m}^0|=1$ a.e.\ in $\omega$. The so-called effective field reads \begin{align}\label{} \boldsymbol{h}_{\operatorname{eff}}(\boldsymbol{m}) = \ell_{\rm ex}^2\Delta\boldsymbol{m} + \boldsymbol{\pi}(\boldsymbol{m}) + \boldsymbol{f}, \end{align} \end{subequations} where $\ell_{\rm ex}>0$ is the exchange length, while the linear, self-adjoint, and bounded operator $\boldsymbol{\pi}: \LL{2}{\omega} \rightarrow \LL{2}{\omega}$ collects the lower-order terms such as stray field and uniaxial anisotropy, and $\boldsymbol{f} \in C^1([0,T],\LL{2}{\omega})$ is the applied field. We note that \begin{align}\label{} \boldsymbol{h}_{\operatorname{eff}}(\boldsymbol{m}) = -\frac{\delta \mathcal{E}_{\textrm{LLG}}(\boldsymbol{m})}{\delta \boldsymbol{m}}, \text{ where }~ \mathcal{E}_{\textrm{LLG}}(\boldsymbol{m}) = \frac{\ell_{\rm ex}^2}{2} \norm{\nabla \boldsymbol{m}}{\omega}^2 \!-\! \frac12 \prod{\boldsymbol{\pi}(\boldsymbol{m})}{\boldsymbol{m}}_{\omega} \!-\! \prod{\boldsymbol{f}}{\boldsymbol{m}}_{\omega} \end{align} is the micromagnetic energy functional. Further dissipative (e.g., spintronic) effects such as the Slonczewski contribution~\cite{slonczewski1996} or the Zhang--Li contribution~\cite{zl2004,tnms2005} are collected in the (not necessarily linear) operator $\boldsymbol{\Pi}: \boldsymbol{H}{1}{\omega} \cap \LL{\infty}{\omega} \rightarrow \LL{2}{\omega}$. Note that solutions of LLG~\eqref{} formally satisfy $0 = \partial_t \boldsymbol{m}\cdot\boldsymbol{m} = \frac12\,\partial_t|\boldsymbol{m}|^2$ and hence $|\boldsymbol{m}|=1$ a.e.\ in $\omega_T$. As in~\cite{as1992}, we use the following notion of weak solutions of~\eqref{}: \begin{definition}\label{def::weaksolution} With the foregoing notation, a function $\boldsymbol{m}:\omega_T\to\mathbb{R}^3$ is called a weak solution of LLG~\eqref{}, if the following properties \ref{} are satisfied: \begin{enumerate}[label=\rm{(\roman*)}] \item\label{} $\boldsymbol{m}\in\boldsymbol{H}{1}{\omega_T} \, \cap \, \L{\infty}{0,T;\boldsymbol{H}{1}{\omega}}$ with $|\boldsymbol{m}|=1$ a.e.\ in $\omega_T$; \item\label{} $\boldsymbol{m}(0)=\boldsymbol{m}^0$ in the sense of traces; \item\label{} for all $\boldsymbol{\varphi}\in\boldsymbol{H}{1}{\omega_T}$, it holds that \begin{align}\label{} \int_0^T\prod{\partial_t \boldsymbol{m}}{\boldsymbol{\varphi}}_{\omega}\,\d{t} &= \ell_{\rm ex}^2\int_0^T\prod{\boldsymbol{m}\times\nabla\boldsymbol{m}}{\nabla\boldsymbol{\varphi}}_{\omega}\,\d{t} - \int_0^T\prod{\boldsymbol{m}\times\boldsymbol{f}}{\boldsymbol{\varphi}}_{\omega}\,\d{t} \notag \\ &\quad - \int_0^T\prod{\boldsymbol{m}\times\boldsymbol{\pi}(\boldsymbol{m})}{\boldsymbol{\varphi}}_{\omega}\,\d{t} - \int_0^T\prod{\boldsymbol{m}\times\boldsymbol{\Pi}(\boldsymbol{m})}{\boldsymbol{\varphi}}_{\omega}\,\d{t} \notag \\ &\quad + \alpha \, \int_0^T\prod{\boldsymbol{m}\times\partial_t \boldsymbol{m}}{\boldsymbol{\varphi}}_{\omega}\,\d{t}; \end{align} \item\label{} for almost all $\tau \in (0,T)$, it holds that \begin{align}\label{} \!\!\! \mathcal{E}_{\rm{LLG}}(\boldsymbol{m}(\tau)) + \alpha\!\int_0^{\tau} \! \norm{\partial_t \boldsymbol{m}}{\omega}^2\,\d{t} +\! \int_0^{\tau} \!\! \prod{\partial_t \boldsymbol{f}}{\boldsymbol{m}}_{\omega} \d{t} -\! \int_0^{\tau} \!\! \prod{\boldsymbol{\Pi}(\boldsymbol{m})}{\partial_t \boldsymbol{m}}_{\omega} \d{t} \le \mathcal{E}_{\rm{LLG}}(\boldsymbol{m}^0). \end{align} \end{enumerate} \end{definition} \subsection{Time discretization}% \label{} Let $N \in \mathbb{N}$ and $k:= T/N$. Consider the uniform time-steps $t_i := ik$ for $i=0,\dots,N$ and the corresponding midpoints $t_{i+1/2} := (t_{i+1}+t_{i})/2$. For a Banach space $\boldsymbol{B}$, e.g., $\boldsymbol{B} \in \{ \LL{2}{\omega}, \boldsymbol{H}{1}{\omega} \}$, and a sequence $\left( \boldsymbol{\varphi}^i \right)_{i=-1}^N \in \boldsymbol{B}$, define the mean value and the discrete time derivative by \begin{subequations} \begin{align}\label{} \mid{\boldsymbol{\varphi}}{i} := \frac{\boldsymbol{\varphi}^{i+1} + \boldsymbol{\varphi}^i}{2} \quad \textrm{and} \quad \dt{\boldsymbol{\varphi}}{i} := \frac{\boldsymbol{\varphi}^{i+1} - \boldsymbol{\varphi}^i}{k} \quad \textrm{for} \quad i = 0,\dots,N-1. \end{align} For $\boldsymbol{\Theta}^k \in \mathbb{R}^{N \times 3}$, define the two-step approach by \begin{align}\label{} \gentime{\boldsymbol{\varphi}}{i} := \boldsymbol{\Theta}_{i3}^k \boldsymbol{\varphi}^{i+1} + \boldsymbol{\Theta}_{i2}^k \boldsymbol{\varphi}^{i} + \boldsymbol{\Theta}_{i1}^k \boldsymbol{\varphi}^{i-1} \quad \textrm{for} \quad i = 0,\dots,N-1. \end{align} \end{subequations} For $i=0,\dots,N-1$ and $t \in [t_i,t_{i+1})$, define \begin{align}\label{} \begin{split} & \boldsymbol{\varphi}_{k}^{=} \!\left(t\right) := \boldsymbol{\varphi}^{i-1}, \quad \boldsymbol{\varphi}_{k}^-\!\left(t\right) := \boldsymbol{\varphi}^i, \quad \boldsymbol{\varphi}_{k}^+\!\left(t\right) := \boldsymbol{\varphi}^{i+1}, \quad \overline{\boldsymbol{\varphi}}_{k}\!\left(t\right) := \mid{\boldsymbol{\varphi}}{i}, \quad \\ &\boldsymbol{\varphi}_{k}^{\boldsymbol{\Theta}} \!\left(t\right) := \gentime{\boldsymbol{\varphi}}{i}, \quad \textrm{and} \quad \boldsymbol{\varphi}_{k}\!\left(t\right) := \boldsymbol{\varphi}^{i+1} \frac{t-t_i}{t_{i+1}-t_i} + \boldsymbol{\varphi}^i \frac{t_{i+1}-t}{t_{i+1}-t_i}. \end{split} \end{align} Note that $\boldsymbol{\varphi}_{k}^{=}, \boldsymbol{\varphi}_{k}^{-}, \boldsymbol{\varphi}_{k}^{+}, \overline{\boldsymbol{\varphi}}_{k}, \boldsymbol{\varphi}_{k}^{\boldsymbol{\Theta}} \in \L{2}{0,T;\boldsymbol{B}}$ and $\boldsymbol{\varphi}_k \in \boldsymbol{H}{1}{0,T;\boldsymbol{B}}$ with $\partial_t \boldsymbol{\varphi}_k (t) = \dt{\boldsymbol{\varphi}}{i}$ for $t \in (t_i,t_{i+1})$. \subsection{Space discretization}% \label{} Let $\mathcal{T}_h$ be a conforming triangulation of $\omega$ into compact tetrahedra $K \in \mathcal{T}_h$ which is $\kappa$-quasi-uniform, i.e., the global mesh-size $h>0$ assures that \begin{align}\label{} \kappa^{-1} \, h \leq \abs{ K }^{1/3} \leq \diam{K} \leq \kappa \, h \quad \textrm{for all } K \in \mathcal{T}_h. \end{align} We suppose that $\mathcal{T}_h$ is weakly acute, i.e., the dihedral angles of all elements are $\leq \pi / 2$. Define the space of $\mathcal{T}_h$-piecewise affine and globally continuous functions by \begin{align*} \boldsymbol{V}_{\!\!h} := (V_h)^3 \subset \boldsymbol{H}{1}{\omega} \quad \text{with} \quad V_h := \{ v_h \in C \lr{\omega} : v_h|_{K} \in \mathcal{P}^1 \lr{K} \textrm{ for all } K \in \mathcal{T}_h \}. \end{align*} Let $\mathcal N_h$ be the set of nodes of $\mathcal{T}_h$. Recall that $|\boldsymbol{m}|=1$ and $\boldsymbol{m} \cdot\partial_t \boldsymbol{m} =0$. To mimic the latter properties, we define the set of discrete admissible magnetizations on $\omega$ by \begin{align} \boldsymbol{\mathcal{M}}_h := \set{\boldsymbol{\psi}_h\in\boldsymbol{V}_{\!\!h}}{|\boldsymbol{\psi}_h(\boldsymbol{z}_h)|=1\text{ for all } \boldsymbol{z}_h\in\mathcal N_h} \end{align} as well as the discrete tangent space \begin{align}\label{eq::defKpsih} \Kh{\boldsymbol{\psi}_h} := \set{\boldsymbol{\varphi}_h\in\boldsymbol{V}_{\!\!h}}{\boldsymbol{\psi}_h(\boldsymbol{z}_h)\cdot\boldsymbol{\varphi}_h(\boldsymbol{z}_h)=0\text{ for all } \boldsymbol{z}_h\in\mathcal N_h} \, \quad \text{for all } \boldsymbol{\psi}_h\in\boldsymbol{\mathcal{M}}_h. \end{align} \subsection{Almost second-order tangent plane scheme}% \label{} In this section, we formulate our numerical integrator which is analyzed below. For each time-step $t_i$, it approximates $\boldsymbol{m}(t_i)$ by $\boldsymbol{m}_h^i \in \boldsymbol{\mathcal{M}}_h$ and $\partial_t \boldsymbol{m}(t_i)$ by $\boldsymbol{v}_h^i \in \Kh{\boldsymbol{m}_h^i}$. For a practical formulation, suppose (computable) operators \begin{align*} \boldsymbol{\pi}_h,\boldsymbol{\Pi}_h : \boldsymbol{\mathcal{M}}_h \to \LL{2}{\omega} \quad\text{and}\quad \boldsymbol{\pi}_h^i, \boldsymbol{\Pi}_h^i : \boldsymbol{V}_{\!\!h} \times \boldsymbol{\mathcal{M}}_h \times \boldsymbol{\mathcal{M}}_h \rightarrow \LL{2}{\omega} \end{align*} such that $\boldsymbol{\pi}_h(\boldsymbol{m}_h^i) \approx \boldsymbol{\pi}(\boldsymbol{m}_h^i)$ and $\boldsymbol{\Pi}_h(\boldsymbol{m}_h^i) \approx \boldsymbol{\Pi}(\boldsymbol{m}_h^i)$ as well as $\boldsymbol{\pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1}) \approx \boldsymbol{\pi}(\boldsymbol{m}_h^{i+1/2})$ and $\boldsymbol{\Pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1}) \approx \boldsymbol{\Pi}(\boldsymbol{m}_h^{i+1/2})$. We suppose that the operators $\boldsymbol{\pi}_h^i$ and $\boldsymbol{\Pi}_h^i$ are affine in $\boldsymbol{v}_h^i$ (cf.~Remark~\ref{}). We extend the construction in~\eqref{} to the operator sequences $(\boldsymbol{\pi}_h^i)_{i=0}^{N-1}$ and $(\boldsymbol{\Pi}_h^i)_{i=0}^{N-1}$ via \begin{align}\label{} \boldsymbol{\pi}_{hk}^-(t) := \boldsymbol{\pi}_{h}^i \quad \textrm{and} \quad \boldsymbol{\Pi}_{hk}^-(t) := \boldsymbol{\Pi}_{h}^i \quad \textrm{for all } t \in [t_i,t_{i+1}). \end{align} Finally, we require two stabilizations $M: \mathbb{R}_{>0} \to \mathbb{R}_{>0}$ and $\rho: \mathbb{R}_{>0} \to \mathbb{R}_{>0}$ such that \begin{subequations}\label{} \begin{align}\label{} \lim_{k\to0}M(k) = \infty \quad &\text{with} \quad \lim_{k\to0}M(k)k = 0, \\ \label{} \lim_{k\to0} \rho(k) = 0 \quad &\text{with} \quad \lim_{k\to0} k^{-1}\rho(k) = \infty. \end{align} The canonical choices are \begin{align}\label{} \rho(k) := |k\,\log k| \quad\text{and}\quad M(k) := |k\,\log k|^{-1}, \end{align} \end{subequations} We proceed as in \cite[p.415]{akst2014} and define the weight function \begin{align}\label{} W_{M(k)}(s) := \begin{cases} \displaystyle\alpha + \frac{k}{2} \, \min\{s,M(k)\} \quad&\text{for } s\ge 0 \,,\\ \displaystyle\frac{\alpha}{1+\frac{k}{2\alpha}\min\{-s,M(k)\}} \quad&\text{for } s<0 \,. \end{cases} \end{align} With these preparations, our numerical integrator reads as follows: \begin{algorithm}[Implicit-explicit tangent plane scheme]\label{} {\bf Input:} Approximation $\boldsymbol{m}_h^{-1} := \boldsymbol{m}_h^0\in\boldsymbol{\mathcal{M}}_h$ of initial magnetization.\\ {\bf Loop:} For $i=0,\dots,N-1$, iterate the following steps {\rm(a)}--{\rm(c)}:\\ {\rm (a)} Compute the discrete function \begin{align} \lambda_h^i := - \ell_{\rm ex}^2 \, |\nabla \boldsymbol{m}_h^i |^2 + \big( \boldsymbol{f}(t_i) + \boldsymbol{\pi}_h(\boldsymbol{m}_h^i) + \boldsymbol{\Pi}_h(\boldsymbol{m}_h^i) \big) \cdot \boldsymbol{m}_h^i. \end{align} {\rm(b)} Find $\boldsymbol{v}_h^{i}\in\Kh{\boldsymbol{m}_h^i}$ such that, for all $\boldsymbol{\varphi}_h\in\Kh{\boldsymbol{m}_h^i}$, it holds that \begin{align}\label{} \begin{split} &\prod{W_{M(k)}(\lambda_h^i) \, \boldsymbol{v}_h^i}{\boldsymbol{\varphi}_h}_{\omega} + \prod{\boldsymbol{m}_h^i\times\boldsymbol{v}_h^i}{\boldsymbol{\varphi}_h}_{\omega} + \frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k)) \, \prod{\nabla\boldsymbol{v}_h^i}{\nabla\boldsymbol{\varphi}_h}_{\omega} \\ &= - \ell_{\rm ex}^2 \, \prod{\nabla\boldsymbol{m}_h^i}{\nabla\boldsymbol{\varphi}_h}_{\omega} + \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{\varphi}_h}_{\omega} + \prod{\pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{\varphi}_{h}}_{\omega} \\& \qquad + \prod{\Pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{\varphi}_h}_{\omega} \,. \end{split} \end{align} {\rm(c)} Define $\boldsymbol{m}_h^{i+1}\in\boldsymbol{\mathcal{M}}_h$ by \begin{align}\label{} \boldsymbol{m}_h^{i+1}(\boldsymbol{z}_h) := \frac{\boldsymbol{m}_h^i(\boldsymbol{z}_h)+k\boldsymbol{v}_h^i(\boldsymbol{z}_h)}{|\boldsymbol{m}_h^i(\boldsymbol{z}_h)+k\boldsymbol{v}_h^i(\boldsymbol{z}_h)|} \quad\text{ for all } \boldsymbol{z}_h\in\mathcal N_h. \end{align} {\bf Output:} Approximations $\boldsymbol{m}_h^i\approx\boldsymbol{m}(t_i)$ for all $i=0,\dots,N$. \qed \end{algorithm} \begin{remark}\label{} {\rm(i)} The variational formulation~\eqref{} gives rise to a linear system for $\boldsymbol{v}_h^i$. However, in particular for stray field computations, the part of the resulting system matrix which corresponds to the $\boldsymbol{\pi}_h^i$-contribution may be fully populated (resp.\ not explicitly available for hybrid FEM-BEM methods~\cite{fk1990}). To deal with this issue, we can, on the one hand, employ the fixpoint iteration~\eqref{} and, on the other hand, choose $\boldsymbol{\pi}_h^i$ and $\boldsymbol{\Pi}_h^i$ independent of $\boldsymbol{v}_h^i$. For the latter, see {\rm (ii)} below for a choice that does not impair the convergence order. {\rm (ii)} Natural choices for the operators $\boldsymbol{\pi}_h^i$ and $\boldsymbol{\Pi}_h^i$ are the following: The chain rule gives rise to an operator $\DPi{\boldsymbol{m}}{\partial_t \boldsymbol{m}} := \partial_t[\boldsymbol{\Pi}(\boldsymbol{m})]$ which is linear in the second argument. Suppose that $\boldsymbol{D}_h$ is a discretization of $\boldsymbol{D}$. For $i=0$, we follow~\cite[Algorithm 2]{akst2014} and choose \begin{subequations}\label{} \begin{align}\label{ \pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}} &:= \boldsymbol{\pi}_h(\boldsymbol{m}_h^i) + \frac{k}{2} \boldsymbol{\pi}_h(\boldsymbol{v}_h^{i}),\\\label{} \Pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}} &:= \boldsymbol{\Pi}_h(\boldsymbol{m}_h^i) + \frac{k}{2} \DPih{\boldsymbol{m}_h^i}{\boldsymbol{v}_h^i}; \end{align} \end{subequations} cf.\ Proposition~\ref{ (i)}. For the subsequent time-steps $i\geq1$ and unlike~\cite{akst2014}, we set \begin{subequations}\label{} \begin{align}\label{} \pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}} &:= \frac{3}{2} \boldsymbol{\pi}_h(\boldsymbol{m}_h^i) - \frac{1}{2} \boldsymbol{\pi}_h(\boldsymbol{m}_h^{i-1}), \\\label{} \Pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}} &:= \boldsymbol{\Pi}_h(\boldsymbol{m}_h^i) + \frac{1}{2} \DPih{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^i} - \frac{1}{2} \DPih{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}; \end{align} \end{subequations} cf.\ Proposition~\ref{ (ii)}. In this way, the right-hand side of~\eqref{} is independent of $\boldsymbol{v}_h^i$. {\rm(iii)} The choice of $\rho$ in~\eqref{} leads to formally almost second-order in time convergence of Algorithm~\ref{}. In principle, it suffices to choose a sufficiently large constant $M(k)\equiv M > 0$; for details see Proposition~\ref{}. {\rm(iv)} Theorem~\ref{} states well-posedness of Algorithm~\ref{} as well as unconditional convergence towards a weak solution of LLG in the sense of Definition~\ref{def::weaksolution}. Moreover, the proof of Theorem~\ref{ (i)} provides a convergent fixpoint solver for the first time-step $i=0$. {\rm (v)} Unlike the stabilized scheme, we note that the non-stabilized scheme $\rho(k) \equiv 0$ requires a CFL-type coupling $k = \mathbf{o}(h)$ for the convergence proof; see also Remark~\ref{}. \end{remark} \subsection{Main theorem for LLG integration}% \label{} To formulate the main result which generalizes~\cite[Theorem 2]{akst2014}, we require the following assumptions: {\large$\bullet$} Let $\mathcal T_h$ satisfy the assumptions of Section~\ref{}. {\large$\bullet$} Let $\boldsymbol{m}_h^0\in\boldsymbol{\mathcal{M}}_h$ satisfy \begin{align}\label{} \boldsymbol{m}_h^0 \rightharpoonup \boldsymbol{m}^0 \quad \textrm{in } \boldsymbol{H}{1}{\omega} \quad\text{as } h \rightarrow 0. \end{align} {\large$\bullet$} For all $i \in \{0,\dots,N-1\}$, let $\boldsymbol{\pi}_h^i$ satisfy, for all $\boldsymbol{\psi}_h, \widetilde{\boldsymbol{\psi}}_h \in \boldsymbol{V}_{\!\!h}$ and all $\boldsymbol{\varphi}_{h},\widetilde{\boldsymbol{\varphi}}_{h} \in \boldsymbol{\mathcal{M}}_h$, \begin{subequations}\label{} the Lipschitz-type continuity \begin{align}\label{} \norm{\pih{i}{\boldsymbol{\psi}_h}{\boldsymbol{\varphi}_h}{\widetilde{\boldsymbol{\varphi}}_h} - \pih{i}{\widetilde{\boldsymbol{\psi}}_h}{\boldsymbol{\varphi}_h}{\widetilde{\boldsymbol{\varphi}}_h}}{\omega} \le C_{\mathrm{\bf \ppi}} \, k \, \norm{\boldsymbol{\psi}_h\!-\!\widetilde{\boldsymbol{\psi}}_h}{\omega} \end{align} and the stability-type estimate \begin{align}\label{} \norm{\pih{i}{\boldsymbol{\psi}_h}{\boldsymbol{\varphi}_h}{\widetilde{\boldsymbol{\varphi}}_h}}{\omega} \le C_{\mathrm{\bf \ppi}} \, \big( k \, \norm{\boldsymbol{\psi}_h}{\omega} + \norm{\boldsymbol{\varphi}_h}{\omega} + \norm{\widetilde{\boldsymbol{\varphi}}_h}{\omega} \big). \end{align} Moreover, for all sequences $\boldsymbol{\psi}_{hk}$ in $\L{2}{0,T;\boldsymbol{V}_{\!\!h}}$ and $\boldsymbol{\varphi}_{hk}$, $\widetilde{\boldsymbol{\varphi}}_{hk}$ in $\L{2}{0,T;\boldsymbol{\mathcal{M}}_h}$ with $\boldsymbol{\psi}_{hk} \rightharpoonup \boldsymbol{\psi}$ and $\boldsymbol{\varphi}_{hk}, \widetilde{\boldsymbol{\varphi}}_{hk} \rightarrow \boldsymbol{\varphi}$ in $\LL{2}{\omega_T}$, we suppose consistency \begin{align}\label{} \boldsymbol{\pi}_{hk}^-(\boldsymbol{\psi}_{hk};\boldsymbol{\varphi}_{hk},\widetilde{\boldsymbol{\varphi}}_{hk}) \rightharpoonup \boldsymbol{\pi}(\boldsymbol{\varphi}) \quad \textrm{in } \LL{2}{\omega_{T}} \quad\text{as } h,k \rightarrow 0. \end{align} \end{subequations} {\large$\bullet$} For all $i \in \{0,\dots,N-1\}$, let $\boldsymbol{\Pi}_h^i$ satisfy, for all $\boldsymbol{\psi}_h, \widetilde{\boldsymbol{\psi}}_h \in \boldsymbol{V}_{\!\!h}$ and all $\boldsymbol{\varphi}_{h},\widetilde{\boldsymbol{\varphi}}_{h} \in \boldsymbol{\mathcal{M}}_h$, \begin{subequations}\label{} the Lipschitz-type continuity \begin{align}\label{} \norm{\Pih{i}{\boldsymbol{\psi}_h}{\boldsymbol{\varphi}_h}{\widetilde{\boldsymbol{\varphi}}_h} - \Pih{i}{\widetilde{\boldsymbol{\psi}}_h}{\boldsymbol{\varphi}_h}{\widetilde{\boldsymbol{\varphi}}_h}}{\omega} \le C_{\mathrm{\bf \Ppi}} \, k \, \norm{\boldsymbol{\psi}_h-\widetilde{\boldsymbol{\psi}}_h}{\boldsymbol{H}{1}{\omega}} \end{align} and the stability-type estimate \begin{align}\label{} \prod{\Pih{i}{\boldsymbol{\psi}_h}{\boldsymbol{\varphi}_h}{\widetilde{\boldsymbol{\varphi}}_h}}{ \boldsymbol{\psi}_h }_{\omega} \le C_{\mathrm{\bf \Ppi}} \, \norm{\boldsymbol{\psi}_h}{\omega} \big( k \, \norm{\boldsymbol{\psi}_h}{\boldsymbol{H}{1}{\omega}} + \norm{\boldsymbol{\varphi}_h}{\boldsymbol{H}{1}{\omega}} + \norm{\widetilde{\boldsymbol{\varphi}}_h}{\boldsymbol{H}{1}{\omega}} \big). \end{align} Moreover, for all sequences $\boldsymbol{\psi}_{hk}$ in $\L{2}{0,T;\boldsymbol{V}_{\!\!h}}$ and $\boldsymbol{\varphi}_{hk}$, $\widetilde{\boldsymbol{\varphi}}_{hk}$ in $\L{2}{0,T;\boldsymbol{\mathcal{M}}_h}$ with $\boldsymbol{\psi}_{hk} \rightharpoonup \boldsymbol{\psi}$ and $\boldsymbol{\varphi}_{hk}, \widetilde{\boldsymbol{\varphi}}_{hk} \rightarrow \boldsymbol{\varphi}$ in $\LL{2}{\omega_T}$, we suppose consistency \begin{align} \label{} \boldsymbol{\Pi}_{hk}^-(\boldsymbol{\psi}_{hk};\boldsymbol{\varphi}_{hk},\widetilde{\boldsymbol{\varphi}}_{hk}) & \rightharpoonup \boldsymbol{\Pi}(\boldsymbol{\varphi}) \quad \textrm{in } \LL{2}{\omega_T} \quad \text{as } h,k \rightarrow 0. \end{align} \end{subequations} \begin{theorem}\label{} {\rm(i)} If the Lipschitz-type estimates~\eqref{} and~\eqref{} are satisfied, then there exists $k_0>0$ such that for all $0<k<k_0$ the discrete variational formulation~\eqref{} admits a unique solution $\boldsymbol{v}_h^i \in \Kh{\boldsymbol{m}_h^i}$. In particular, Algorithm~\ref{} is well-defined. {\rm(ii)} Suppose that all preceding assumptions~\eqref{} are satisfied. Let $\boldsymbol{m}_{hk}$ be the postprocessed output~\eqref{} of Algorithm~\ref{}. Then, there exists $\boldsymbol{m} \in \boldsymbol{H}{1}{\omega_T} \cap \L{\infty}{0,T;\boldsymbol{H}{1}{\omega}}$ which satisfies Definition~\ref{def::weaksolution}\ref{}, and a subsequence of $\boldsymbol{m}_{hk}$ converges weakly in $\boldsymbol{H}{1}{\omega_T}$ towards $\boldsymbol{m}$ as $h,k\to0$. {\rm(iii)} Under the assumptions of {\rm(ii)}, suppose that the convergence properties~\eqref{} as well as~\eqref{} and~\eqref{} hold with strong convergence. Moreover, suppose additionally that $\boldsymbol{f} \in C^1([0,T],\LL{2}{\omega}) \cap C([0,T],\LL{3}{\omega})$ and that $\boldsymbol{\pi}$ is $\boldsymbol{L}^3$-stable, i.e., \begin{align}\label{} \norm{\boldsymbol{\pi}(\boldsymbol{\varphi})}{\LL{3}{\omega}} \leq C_{\boldsymbol{\pi}}' \, \norm{\boldsymbol{\varphi}}{\LL{3}{\omega}} \quad \textrm{for all } \boldsymbol{\varphi} \in \boldsymbol{H}{1}{\omega}. \end{align} Then, $\boldsymbol{m}$ from {\rm(ii)} is a weak solution of LLG in the sense of Definition~\ref{def::weaksolution}\ref{}. \end{theorem} \newpage \begin{remark}\label{} {\rm (i)} If the solution $\boldsymbol{m} \in \boldsymbol{H}{1}{\omega_T} \cap \L{\infty}{0,T;\boldsymbol{H}{1}{\omega}}$ of Definition~\ref{def::weaksolution} is unique, then the full sequence $\boldsymbol{m}_{hk}$ converges weakly in $\boldsymbol{H}{1}{\omega_T}$ towards $\boldsymbol{m}$. {\rm(ii)} The assumptions~\eqref{} on $\boldsymbol{\pi}_h$ (even with strong convergence) as well as~\eqref{} on $\boldsymbol{\pi}$ are satisfied, e.g., for uniaxial anisotropy and stray field. Possible stray field discretizations include hybrid FEM-BEM approaches~\cite{fk1990}; see~\cite{bffgpprs2014,prs2016}. {\rm(iii)} The Lipschitz-type estimates~\eqref{} and~\eqref{} are only used to prove that Algorithm~\ref{} is well-defined for sufficiently small $k>0$. The stability estimates~\eqref{} and~\eqref{} are then used to prove some discrete energy estimate (Lemma~\ref{}). Finally, the consistency assumptions~\eqref{} and~\eqref{} are used to show that the existing limit satisfies the variational formulation of Definition~\ref{def::weaksolution}\ref{}. {\rm (iv)} The assumptions on $\boldsymbol{\pi}_h$ and $\boldsymbol{\Pi}_h$ from~\eqref{} are only exploited for $\boldsymbol{\varphi}_{h} = \boldsymbol{m}_{h}^i$, $\widetilde{\boldsymbol{\varphi}}_{h} = \boldsymbol{m}_{h}^{i-1}$, and $\boldsymbol{\varphi}_{hk} = \boldsymbol{m}_{hk}^-$, $\widetilde{\boldsymbol{\varphi}}_{hk} = \boldsymbol{m}_{hk}^=$, $\boldsymbol{\psi}_{hk} = \boldsymbol{v}_{hk}^-$. Moreover, onl ~\eqref{} and~\eqref{} require general $\boldsymbol{\psi}_{h}, \widetilde{\boldsymbol{\psi}}_h$, whil ~\eqref{} and~\eqref{} are only exploited for $\boldsymbol{\psi}_h = \boldsymbol{v}_h^i$. \end{remark} The proof of Theorem~\ref{} is postponed to Section~\ref{}. Unique solvability of~\eqref{} will be proved with the Banach fixpoint theorem. Theorem~\ref{ will be proved through an energy argument which consists of the following three steps: \begin{itemize} \item bound the discrete energy (Lemma~\ref{}); \item extract weakly convergent subsequences having a common limit (Lemma~\ref{}); \item verify that this limit is a weak solution in the sense of Definition~\ref{def::weaksolution}. \end{itemize} \section{Landau--Lifshitz--Gilbert equation with eddy currents} \label{} \subsection{ELLG and weak solutions}% \label{} Let $\Omega \subset \mathbb{R}^3$ be a bounded Lipschitz domain with $\Omega \supseteq \omega$ that represents a conducting body with its ferromagnetic part $\omega$. ELLG reads \begin{subequations}\label{} \begin{align} \partial_t \boldsymbol{m} &= - \boldsymbol{m} \times \big(\boldsymbol{h}_{\operatorname{eff}}(\boldsymbol{m}) + \boldsymbol{\Pi}(\boldsymbol{m}) + \boldsymbol{h} \big) + \alpha \, \boldsymbol{m} \times \partial_t \boldsymbol{m} &&\hspace*{-3mm} \textrm{ in } \omega_T := (0,T) \times \omega, \label{} \\ - \mu_0 \, \partial_t \boldsymbol{m} &= \mu_0 \, \partial_t \boldsymbol{h} + \nabla \times (\sigma^{-1} \nabla \times \boldsymbol{h} ) &&\hspace*{-3mm} \textrm{ in } \Omega_T := (0,T) \times \Omega, \label{} \\ \partial_{\mathbf{n}} \boldsymbol{m} &= \boldsymbol{0} &&\hspace*{-3mm} \textrm{ on } \left(0,T\right) \times \partial \omega, \label{} \\ \hspace*{-1.5mm}( \nabla \times \boldsymbol{h} ) \times \mathbf{n} &= \boldsymbol{0} &&\hspace*{-3mm} \textrm{ on } \left(0,T\right) \times \partial \Omega, \label{} \\ (\boldsymbol{m}, \boldsymbol{h}) \lr{0} &= (\boldsymbol{m}^0,\boldsymbol{h}^0) &&\hspace*{-3mm} \textrm{ in } \omega \times \Omega. \label{} \end{align} \end{subequations} Here, $\sigma \in \L{\infty}{\Omega}$ with $\sigma \ge \sigma_0 > 0$ is the conductivity of $\Omega$, and $\mu_0 > 0$ is the vacuum permeability. The initial condition $\boldsymbol{h}^0 \in \Hcurl{\Omega}$ satisfies the compatibility conditions \begin{align}\label{} \nabla \cdot ( \boldsymbol{h}^0 + \chi_{\omega} \boldsymbol{m}^0 ) = 0 \textrm{ in } \Omega \quad \textrm{and} \quad (\boldsymbol{h}^0 + \chi_{\omega} \boldsymbol{m}^0 ) \cdot \mathbf{n} = 0 \textrm{ on } \partial \Omega. \end{align} Unlike \cite[Definition 2.1]{lt2013}, we define the energy functional \begin{align}\label{} \mathcal{E}_{\textrm{ELLG}}(\boldsymbol{m},\boldsymbol{h}) := \frac{\ell_{\rm ex}^2}{2} \, \norm{\nabla \boldsymbol{m}}{\omega}^2 - \frac12 \prod{\boldsymbol{\pi}(\boldsymbol{m})}{\boldsymbol{m}}_{\omega} - \prod{\boldsymbol{f}}{\boldsymbol{m}}_{\omega} + \frac12 \norm{\boldsymbol{h}}{\Omega}^2. \end{align} As in~\cite{as1992,lt2013}, we use the following notion of weak solutions of ELLG~\eqref{}: \begin{definition}\label{} With the foregoing notation, the pair $(\boldsymbol{m},\boldsymbol{h})$ is called a weak solution of ELLG~\eqref{}, if the following conditions \ref{} are satisfied: \begin{enumerate}[label=\rm{(\roman*)}] \item \label{} $\boldsymbol{m} \in \boldsymbol{H}{1}{\omega_T} \, \cap \, \L{\infty}{0,T;\boldsymbol{H}{1}{\omega}}$ with $\left|\boldsymbol{m} \right| = 1$ a.e. in $\omega_T$; \item \label{} $\boldsymbol{h} \in \boldsymbol{H}{1}{0,T;\LL{2}{\Omega}} \cap \L{\infty}{0,T;\boldsymbol{H}({\rm curl};\Omega)}$. \item \label{} $\boldsymbol{m} \!\left(0\right) = \boldsymbol{m}^0$ and $\boldsymbol{h} \!\left(0\right) = \boldsymbol{h}^0$ in the sense of traces; \item \label{} for all $\boldsymbol{\varphi}\in\boldsymbol{H}{1}{\omega_T}$ and for all $\boldsymbol{\zeta} \in \L{2}{0,T;\boldsymbol{H}({\rm curl},\Omega)}$, it holds that \begin{subequations}\label{} \begin{align}\label{} \Int{0}{T} \prod{\partial_t \boldsymbol{m}}{\boldsymbol{\varphi}}_{\omega} \d{t} &=\ell_{\rm ex}^2 \Int{0}{T} \prod{\boldsymbol{m} \times \nabla \boldsymbol{m}}{\nabla \boldsymbol{\varphi}}_{\omega} \d{t} - \Int{0}{T} \prod{\boldsymbol{m} \times \boldsymbol{f}}{\boldsymbol{\varphi}}_{\omega} \d{t}\notag \\ &\quad- \Int{0}{T} \prod{\boldsymbol{m} \times \boldsymbol{\pi}(\boldsymbol{m})}{\boldsymbol{\varphi}}_{\omega} \d{t} - \int_0^T\prod{\boldsymbol{m}\times\boldsymbol{\Pi}(\boldsymbol{m})}{\boldsymbol{\varphi}}_{\omega}\,\d{t} \notag \\ &\quad -\Int{0}{T} \prod{\boldsymbol{m} \times \boldsymbol{h}}{\boldsymbol{\varphi}}_{\omega} \d{t} + \alpha \Int{0}{T} \prod{\boldsymbol{m} \times \partial_t \boldsymbol{m}}{\boldsymbol{\varphi}}_{\omega} \d{t}, \\ \label{} -\mu_0 \Int{0}{T} \prod{\partial_t \boldsymbol{m}}{\boldsymbol{\zeta}}_{\omega} \d{t} &= \mu_0 \Int{0}{T} \prod{\partial_t \boldsymbol{h}}{\boldsymbol{\zeta}}_{\Omega} \d{t} + \Int{0}{T} \prod{\sigma^{-1} \nabla \times \boldsymbol{h}}{\nabla \times \boldsymbol{\zeta}}_{\Omega} \d{t}; \end{align} \end{subequations} \item\label{} for almost all $\tau \in (0,T)$, it holds that \begin{align}\label{} & \mathcal{E}_{{\rm ELLG}}(\boldsymbol{m}(\tau),\boldsymbol{h}(\tau)) + \alpha \Int{0}{\tau} \norm{\partial_t \boldsymbol{m}}{\omega}^2 \d{t} + \int_0^{\tau} \prod{\partial_t \boldsymbol{f}}{\boldsymbol{m}}_{\omega} \d{t} \notag \\ & \quad + \frac{1}{\mu_0} \Int{0}{\tau} \norm{\sigma^{-1/2}\nabla \times \boldsymbol{h}}{\Omega}^2 \d{t} - \int_0^{\tau} \prod{\boldsymbol{\Pi}(\boldsymbol{m})}{\partial_t \boldsymbol{m}}_{\omega} \d{t} \leq \mathcal{E}_{{\rm ELLG}}(\boldsymbol{m}^0,\boldsymbol{h}^0). \end{align} \end{enumerate} \end{definition} \subsection{Discretization}% \label{} We adopt the notation of Section~\ref{}. Let $\mathcal{T}_h$ be a conforming triangulation of $\Omega$ into compact tetrahedra $K \in \mathcal{T}_h$. Suppose that $\mathcal{T}_h$ is $\kappa$-quasi-uniform (cf.~\eqref{}). We suppose that $\mathcal{T}_h$ resolves $\omega$, i.e., \begin{align} \overline\omega = \bigcup_{K\in\mathcal{T}_h|_\omega}K, \quad\text{where}\quad \mathcal{T}_h|_\omega := \set{K\in\mathcal{T}_h}{K \subseteq \overline\omega}. \end{align} Let $\mathcal N_h$ (resp.\ $\mathcal N_h|_\omega$) be the set of nodes of $\mathcal{T}_h$ (resp.\ $\mathcal{T}_h|_\omega$). We suppose that $\mathcal{T}_h|_\omega$ is weakly acute and define $\boldsymbol{V}_{\!\!h}$, $\boldsymbol{\mathcal{M}}_h$, and $\Kh{\cdot}$ with respect to $\mathcal{T}_h|_\omega$ as in Section~\ref{}. Finally, the space of N\'{e}d\'{e}lec edge elements of second type~\cite{nedelec1986} on $\Omega$ reads \begin{align*} \mathcal{X}_h := \{ \boldsymbol{\zeta}_h \in \Hcurl{\Omega} : \boldsymbol{\zeta}_h|_{K} \in \mathcal{P}^1 \lr{K} \textrm{ for all } K \in \mathcal{T}_h \} \subset \Hcurl{\Omega}. \end{align*} \subsection{Almost second-order tangent plane scheme}% \label{} In this section, we extend Algorithm~\ref{} to ELLG~\eqref{}. More precisely, we combine Algorithm~\ref{} for the LLG-part with an implicit midpoint scheme for the eddy current part. \begin{algorithm}[Decoupled ELLG algorithm]\label{} {\bf Input:} $\boldsymbol{\Theta}^k \in \mathbb{R}^{N \times 3}$, approximations $\boldsymbol{m}_h^{-1} := \boldsymbol{m}_h^0 \in \boldsymbol{\mathcal{M}}_h$ and $\boldsymbol{h}_h^{-1} := \boldsymbol{h}_h^0 \in \mathcal{X}_h$ of the initial values. \\ {\bf Loop:} For $i=0,\dots,N-1$, iterate the following steps {\rm(a)}--{\rm(d)}:\\ {\rm (a)} Compute the discrete function \begin{align} \lambda_h^i := - \ell_{\rm ex}^2 \, |\nabla \boldsymbol{m}_h^i |^2 + \big(\boldsymbol{f}(t_i) + \boldsymbol{\pi}_h(\boldsymbol{m}_h^i) + \boldsymbol{\Pi}_h(\boldsymbol{m}_h^i) + \boldsymbol{h}_h^i \big) \cdot \boldsymbol{m}_h^i. \end{align} {\rm (b)} Find $\boldsymbol{v}_h^{i} \in \Kh{\boldsymbol{m}_h^i}$ such that, for all $\boldsymbol{\varphi}_h \in \Kh{\boldsymbol{m}_h^i}$, it holds that \begin{subequations}\label{} \begin{align}\label{} &\prod{W_M(\lambda_h^i)\boldsymbol{v}_h^i}{\boldsymbol{\varphi}_h}_{\omega} + \prod{\boldsymbol{m}_h^i\times\boldsymbol{v}_h^i}{\boldsymbol{\varphi}_h}_{\omega} + \frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k)) \, \prod{\nabla\boldsymbol{v}_h^i}{\nabla\boldsymbol{\varphi}_h}_{\omega} \notag \\ &= -\ell_{\rm ex}^2\prod{\nabla\boldsymbol{m}_h^i}{\nabla\boldsymbol{\varphi}_h}_{\omega} + \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{\varphi}_h}_{\omega} + \prod{\pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{\varphi}_h}_{\omega} \notag \\ & \quad + \prod{\Pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{\varphi}_h}_{\omega} + \prod{\gentimeh{\boldsymbol{h}}{i}}{\boldsymbol{\varphi}_h}_{\omega}. \end{align} {\rm (c)} Define $\boldsymbol{m}_h^{i+1}\in\boldsymbol{\mathcal{M}}_h$ by \begin{align}\label{} \boldsymbol{m}_h^{i+1}(\boldsymbol{z}_h) := \frac{\boldsymbol{m}_h^i(\boldsymbol{z}_h)+k\boldsymbol{v}_h^i(\boldsymbol{z}_h)}{|\boldsymbol{m}_h^i(\boldsymbol{z}_h)+k\boldsymbol{v}_h^i(\boldsymbol{z}_h)|} \quad\text{ for all } \boldsymbol{z}_h\in\mathcal N_h. \end{align} {\rm(d)} Find $\boldsymbol{h}_h^{i+1} \in \mathcal{X}_h$ such that, for all $\boldsymbol{\zeta}_h \in \mathcal{X}_h$, it holds that \begin{align}\label{} - \mu_0 \prod{\dth{\boldsymbol{m}}{i}}{\boldsymbol{\zeta}_h}_{\omega} = \mu_0 \prod{\dth{\boldsymbol{h}}{i}}{\boldsymbol{\zeta}_h}_{\Omega} + \prod{\sigma^{-1} \nabla \times \midh{\boldsymbol{h}}{i}}{\nabla \times \boldsymbol{\zeta}_h}_{\Omega}. \end{align} \end{subequations} {\bf Output:} Approximations $\boldsymbol{m}_h^i \approx \boldsymbol{m}(t_i)$ and $\boldsymbol{h}_h^i \approx \boldsymbol{h}(t_i)$ for all $i=0,\dots,N$. \qed \end{algorithm} \begin{remark}\label{} {\rm(i)} It depends on the choice of $\boldsymbol{\Theta}^k$ (and hence of $\gentimeh{\boldsymbol{h}}{i}$ in~\eqref{}) whether step~{\rm(b)--(d)} of Algorithm~\ref{} have to be solved simultaneously or sequentially. {\rm(ii)} For the first time-step $i=0$, we choose $\boldsymbol{\Theta}^k_{03} : = \boldsymbol{\Theta}^k_{02} := 1/2$ and $\boldsymbol{\Theta}^k_{01} = 0$, i.e., \begin{align}\label{} \gentimeh{\boldsymbol{h}}{i} = \midh{\boldsymbol{h}}{i} = \boldsymbol{h}_h^{i} + \frac{k}{2} \dth{\boldsymbol{h}}{i}. \end{align} Then, step~{\rm(b)--(d)} of Algorithm~\ref{} have to be solved simultaneously. Moreover, because of step~{\rm(c)}, the overall system is nonlinear in $\boldsymbol{v}_h^i$. For the subsequent time-steps $i \geq 1$, we choose $\boldsymbol{\Theta}^k_{i3} : = 0$, $\boldsymbol{\Theta}^k_{i2} := 3/2$, and $\boldsymbol{\Theta}^k_{i1} = -1/2$, i.e., \begin{align}\label{} \gentimeh{\boldsymbol{h}}{i} = \frac32 \boldsymbol{h}_h^{i} - \frac12 \boldsymbol{h}_h^{i-1}. \end{align} Then, the numerical integrator decouples the time-stepping for LLG and eddy currents, and we need to solve only two linear systems per time-step. {\rm(iii)} The choice of $\boldsymbol{\Theta}^k_i$ for $i\ge1$ is motivated by the explicit Adams--Bashforth two-step method which is of second order in time. By choice~\eqref{} of $\rho$, Algorithm~\ref{} then is formally of almost second order in time. {\rm(iv)} In principle, we can replace $-\mu_0\prod{\dth{\boldsymbol{m}}{i}}{\boldsymbol{\zeta}_h}_{\omega}$ in equation~\eqref{} of Algorithm~\ref{} by $- \mu_0 \prod{\boldsymbol{v}_h^i}{\boldsymbol{\zeta}_h}_{\omega}$. Even in the implicit case~\eqref{}, the resulting scheme is then linear in $\boldsymbol{v}_h^i$ and $\midh{\boldsymbol{h}}{i}$. However, Lemma~\ref{lemma::vformula} predicts that $\partial_t \boldsymbol{m} = \boldsymbol{v} + \mathcal{O}(k)$, and thus we only expect first-order convergence in time. This is also confirmed numerically in Section~\ref{example4}. {\rm(v)} In practice, we solve~\eqref{} for the unknown $\boldsymbol{\nu}_h := \midh{\boldsymbol{h}}{i} \in \mathcal{X}_h$, i.e., we compute the unique $\boldsymbol{\nu}_h \in \mathcal{X}_h$ such that, for all $\boldsymbol{\zeta}_h \in \mathcal{X}_h$, \begin{align}\label{} \frac{2\mu_0}{k} \prod{\boldsymbol{\nu}_h}{\boldsymbol{\zeta}_h}_{\Omega} + \prod{\sigma^{-1} \nabla \times \boldsymbol{\nu}_h}{\nabla \times \boldsymbol{\zeta}_h}_{\Omega} = - \mu_0 \prod{\dth{\boldsymbol{m}}{i}}{\boldsymbol{\zeta}_h}_{\omega} + \frac{2\mu_0}{k} \prod{\boldsymbol{h}_h^i}{\boldsymbol{\zeta}_h}_{\Omega}. \end{align} Then, $\boldsymbol{h}_h^{i+1} := 2 \, \boldsymbol{\nu}_h - \boldsymbol{h}_h^i$ solves~\eqref{}. {\rm(vi)} Theorem~\ref{} states well-posedness of Algorithm~\ref{} as well as unconditional convergence towards a weak solution of ELLG in the sense of Definition~\ref{}. The proof of Theorem~\ref{ (i)} provides a convergent fixpoint solver for the first (nonlinear) time-step $i = 0$. \end{remark} \subsection{Main theorem for ELLG integration} This section states our convergence result for Algorithm~\ref{}. In~\cite{lt2013,lppt2015} and~\cite{bpp2015}, similar results are proved for a first-order tangent plane scheme for ELLG and for the Maxwell-LLG system, respectively. \begin{theorem}\label{} {\rm(i)} Suppose that the Lipschitz-type estimates~\eqref{} and~\eqref{} are satisfied. Suppose that the coupling parameter $\boldsymbol{\Theta}^k \in \mathbb{R}^{N \times 3}$ satisfies \begin{align}\label{} \max_{ij} \abs{\boldsymbol{\Theta}^k_{ij}} \leq C_{\boldsymbol{\Theta}} < \infty \quad \textrm{and} \quad \sum\limits_{j=1}^3 \boldsymbol{\Theta}^k_{ij} = 1 \quad \textrm{for all } i \in \{0,\dots,N-1\}. \end{align} Then, there exists $k_0>0$ such that for all $0<k<k_0$ the discrete variational formulation~\eqref{} admits a unique solution $(\boldsymbol{v}_h^i,\boldsymbol{h}_h^{i+1}) \in \Kh{\boldsymbol{m}_h^i} \times \mathcal{X}_h$. In particular, Algorithm~\ref{} is well-defined. {\rm(ii)} Suppose that~\eqref{} as well as all assumptions from Section~\ref{} are satisfied. Let $(\boldsymbol{m}_{hk}, \boldsymbol{h}_{hk})$ be the postprocessed output~\eqref{} of Algorithm~\ref{}. Then, there exist $\boldsymbol{m} \in \boldsymbol{H}{1}{\omega_T} \cap \L{\infty}{0,T;\boldsymbol{H}{1}{\omega}}$ and $\boldsymbol{h} \in \boldsymbol{H}{1}{0,T;\LL{2}{\Omega}} \cap \L{\infty}{0,T;\boldsymbol{H}({\rm curl},\Omega)}$ which satisfy Definition~\ref{}, and a subsequence of $(\boldsymbol{m}_{hk},\boldsymbol{h}_{hk})$ converges weakly in $\boldsymbol{H}{1}{\omega_T}\times\LL{2}{\Omega_T}$ towards $(\boldsymbol{m},\boldsymbol{h})$ as $h,k\to0$. {\rm(iii)} Under the assumptions of {\rm(ii)}, suppose that the convergence properties~\eqref{} as well as~\eqref{} and~\eqref{} hold with strong convergence. Moreover, suppose additionally that $\boldsymbol{f} \in C^1([0,T],\LL{2}{\omega}) \cap C([0,T],\LL{3}{\omega})$ and that $\boldsymbol{\pi}$ is $\boldsymbol{L}^3$-stable~\eqref{}. Finally, let the CFL condition $k = \mathbf{o}(h^2)$ be satisfied. Then, the pair $(\boldsymbol{m},\boldsymbol{h})$ from {\rm(ii)} is a weak solution of ELLG in the sense of Definition~\ref{}. \end{theorem} \begin{remark}\label{} {\rm(i)} If the solution $(\boldsymbol{m},\boldsymbol{h})$ of Definition~\ref{} is unique, then the full sequence $(\boldsymbol{m}_{hk},\boldsymbol{h}_{hk})$ converges weakly in $\boldsymbol{H}{1}{\omega_T}\times\LL{2}{\Omega_T}$ towards $(\boldsymbol{m},\boldsymbol{h})$. {\rm(ii)} If the CFL condition $k = \mathbf{o}(h^2)$ is satisfied (as required for Theorem~\ref{}), then Theorem~\ref{} holds also for vanishing stabilization $\rho(k)\equiv0$. {\rm(iii)} The compatibility condition~\eqref{} of the initial data is not exploited in the proof of Theorem~\ref{}. In particular, the discrete initial data do not have to satisfy any ``discrete compatibility condition''. {\rm(iv)} If $\prod{\dth{\boldsymbol{m}}{i}}{\boldsymbol{\zeta}_h}_{\omega}$ in~\eqref{} is replaced by $\prod{\boldsymbol{v}_h^i}{\boldsymbol{\zeta}_h}_{\omega}$, then Theorem~\ref{} holds without the CFL condition $k=\mathbf{o}(h^2)$ which is only exploited in~\eqref{} below. However, as mentioned in Remark~\ref{}, the overall integrator then appears to be only of first order in time. \end{remark} The proof of Theorem~\ref{} is postponed to Section~\ref{}. For the LLG-part of ELLG~\eqref{}, we follow our proof of Theorem~\ref{}. For the eddy current part, we adapt the techniques of~\cite{lt2013,lppt2015,bpp2015,prs2016} to the setting of Algorithm~\ref{}. \section{Derivation of second-order tangent plane scheme} \label{} In this section, we adapt \cite{akst2014,prs2016} in order to motivate Algorithm~\ref{} and to underpin that it is of (almost) second order in time. Since solutions $\boldsymbol{m}$ to LLG satisfy $\boldsymbol{m} \cdot \partial_t \boldsymbol{m} = 0$ a.e.\ in $\omega$, we define, for any $\boldsymbol{\psi}\in\Cinfty{\overline{\omega}}$ with $|\boldsymbol{\psi}|=1$ its tangent space \begin{align}\label{} \K{\boldsymbol{\psi}} := \set{\boldsymbol{\varphi}\in\Cinfty{\overline{\omega}}}{\boldsymbol{\psi}\cdot\boldsymbol{\varphi}=0\text{ in }\omega} \end{align} as well as the pointwise projection onto $\K{\boldsymbol{\psi}}$ by \begin{align}\label{eq::morthproj} \mathbb{P}_{\boldsymbol{\psi}}(\boldsymbol{u}) := \boldsymbol{u} - (\boldsymbol{u}\cdot\boldsymbol{\psi})\boldsymbol{\psi} \quad \text{for all }\boldsymbol{u} \in \Cinfty{\overline{\omega}}. \end{align} \begin{lemma}[{{\cite[p.~413]{akst2014}}}]\label{lemma::vformula} For $\boldsymbol{m}\in\Cinfty{\overline{\omega_T}}$ with $|\boldsymbol{m}|=1$, it holds that \begin{align}\label{eq::vformula} \frac{\boldsymbol{m}(t) + k\boldsymbol{v}(t)}{|\boldsymbol{m}(t) + k\boldsymbol{v}(t)|} = \boldsymbol{m}(t+k) + \mathcal{O}(k^3), \quad\text{where}\quad \boldsymbol{v}(t) := \partial_t \boldsymbol{m}(t) + \frac{k}{2} \, \mathbb{P}_{\boldsymbol{m}(t)}(\partial_{tt}\boldsymbol{m}(t)) \end{align} for all $t \in [0,T-k]$.\qed \end{lemma} Recall the stabilizations $M(\cdot)$ and $\rho(\cdot)$ from~\eqref{} as well as the weight function $W_{M(k)}(\cdot)$ from~\eqref{}. The following lemma is implicitly found in~\cite[p.415]{akst2014}. With~\eqref{}, it proves, in particular, that $W_{M(k)}(\cdot)\ge \alpha/2$, if $k>0$ is sufficiently small. \begin{lemma}\label{} The weight function $W_{M(k)}(\cdot)$ satisfies the following two assertions {\rm(i)}--{\rm(ii)}: \newline {\rm (i)} $|W_{M(k)}(s)-\alpha| \le \frac{M(k)k}{2}$ for all $s\in\mathbb{R}$. \newline {\rm (ii)} If $M(k)\ge B>0$ and $k<k_0:=\alpha/B$, then \begin{align}\label{eq::weightapprox} \Big| \, \alpha + \frac{k}{2}s - W_{M(k)}(s) \, \Big| \le \frac{B^2}{2 \alpha} \, k^2 \quad \textrm{for all } s\in[-B,B] . \end{align} \end{lemma} \begin{proof} For $s \geq 0$, {\rm (i)} follows immediately from the definition~\eqref{}. For $s < 0$, we get that \begin{align*} |W_{M(k)}(s)-\alpha| \stackrel{\eqref{} \frac{\frac{k}{2} \min\{-s,M(k)\}}{1+\frac{k}{2 \alpha}\min\{-s,M(k)\}} \leq \frac{k}{2} \min\{-s,M(k)\} \leq \frac{M(k)k}{2}. \end{align*} This proves {\rm (i)}. We come to the proof of {\rm (ii)}: If $0\le s\le B$, then $B\leq M$ yields that \begin{align*} \alpha + \frac{k}{2} s - W_{M(k)}(s) \stackrel{\eqref{} 0. \end{align*} If $-B\le s < 0$ and $j\in\mathbb{N}$, the $j$-th derivative of $W_{M(k)}(\cdot)$ reads \begin{align*} W_{M(k)}^{(j)}(s) =\frac{2\alpha^2k^jj!}{(2\alpha-ks)^{1+j}}\,. \end{align*} Since $k<k_0:=\alpha/B$, it holds that \begin{align* q := \Big| \frac{ks}{2 \alpha} \Big| \leq \Big| \frac{s}{2 B} \Big| \leq \frac12 \quad \textrm{and thus} \quad \sum\limits_{j=0}^{\infty} q^j = \frac{1}{1-q} \leq 2. \end{align*} The Taylor expansion at $s=0^-$ shows that \begin{align* \Big| \alpha + \frac{k}{2}s - W_{M(k)}(s)\Big| = \Big| \sum_{j=2}^\infty \frac{k^j s^j}{2^j\alpha^{j-1}} \Big| = \Big| \frac{k^2 s^2}{4\alpha}\sum_{j=0}^\infty \Big(\frac{ks}{2\alpha}\Big)^j \Big| \le 2 \, \frac{k^2 s^2}{4 \alpha} \leq \frac{B^2 k^2}{2 \alpha}. \end{align*} This concludes the proof. \end{proof} The following proposition is the main result of this section. It shows that the consistency error of Algorithm~\ref{} is of order $\mathcal{O}(k^2+k\rho(k))$. For $\boldsymbol{\Pi} = \boldsymbol{0}$ and $\boldsymbol{f}$ constant in time,~\eqref{} is implicitly contained in~\cite[Section 6]{akst2014}, while~\eqref{} adapts some further ideas from~\cite{prs2016}. Note that~\eqref{} corresponds to~\eqref{} in combination with~\eqref{} and requires the evaluation of $\boldsymbol{\pi}(\boldsymbol{v})$ and $\DPi{\boldsymbol{m}}{\boldsymbol{v}}$. In contrast to that,~\eqref{} corresponds to~\eqref{} in combination with~\eqref{} and avoids these evaluations at~$\boldsymbol{v}$. \begin{proposition}\label{} Let $\DPi{\cdot}{\cdot}$ be obtained from $\DPi{\boldsymbol{m}(t)}{\partial_t \boldsymbol{m}(t)} := \partial_t[\boldsymbol{\Pi}(\boldsymbol{m}(t))]$ and the chain rule. For $\boldsymbol{\psi} \in \Cinfty{\overline{\omega}}$ and $\boldsymbol{w}, \boldsymbol{\varphi} \in \boldsymbol{H}{1}{\omega}$, define the bilinear form \begin{subequations}\label{} \begin{align}\label{} \hspace*{-1.5mm} \mathbf{b} (\boldsymbol{\psi};\boldsymbol{w},\boldsymbol{\varphi}) &:= \langle W_{M(k)}\big(\lambda(\boldsymbol{\psi})\big) \boldsymbol{w},\boldsymbol{\varphi} \rangle_{\omega} + \prod{\boldsymbol{\psi}\times\boldsymbol{w}}{\boldsymbol{\varphi}}_{\omega} + \frac{\ell_{\rm ex}^2}{2} \, k \, (1 + \rho(k)) \, \prod{\nabla\boldsymbol{w}}{\nabla\boldsymbol{\varphi}}_{\omega}, \hspace*{-2mm} \\ \intertext{where} \label{} \lambda(\boldsymbol{\psi}) &:= \big( \boldsymbol{h}_{\operatorname{eff}}(\boldsymbol{\psi}) + \boldsymbol{\Pi}(\boldsymbol{\psi}) \big) \cdot \boldsymbol{\psi}. \end{align} \end{subequations} Let $\boldsymbol{m} \in \Cinfty{\overline{\omega_T}}$ be a strong solution of LLG~\eqref{} which satisfies \begin{align}\label{} B := \norm{ \lambda(\boldsymbol{m}) }{\L{\infty}{\omega_T}} \leq M(k) <\infty. \end{align} Then, $\boldsymbol{v}$ from~\eqref{eq::vformula} satisfies the following two assertions {\rm (i)}--{\rm (ii)}: \newline \begin{subequations} {\rm (i)} For $t \in [0,T-k]$, there exists $\boldsymbol{R}_1 = \mathcal{O}(k^2+ k \rho(k))$ such that, for all $\boldsymbol{\varphi}\in\K{\boldsymbol{m}}$, \begin{align}\label{} \begin{split} & \mathbf{b} (\boldsymbol{m};\boldsymbol{v}(t),\boldsymbol{\varphi}) - \frac{k}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{v}(t))}{\boldsymbol{\varphi}}_{\omega} - \frac{k}{2} \, \prod{\DPi{\boldsymbol{m}(t)}{\boldsymbol{v}(t)}}{\boldsymbol{\varphi}}_{\omega} \\ &\quad= -\ell_{\rm ex}^2\prod{\nabla\boldsymbol{m}(t)}{\nabla\boldsymbol{\varphi}}_{\omega} + \prod{\boldsymbol{f}\big( \, t + k/2 \, \big)}{\boldsymbol{\varphi}}_{\omega} + \prod{\boldsymbol{\pi}(\boldsymbol{m}(t))}{\boldsymbol{\varphi}}_{\omega} \\ & \quad\quad\quad + \prod{\boldsymbol{\Pi}(\boldsymbol{m}(t))}{\boldsymbol{\varphi}}_{\omega} \!+\! \prod{\boldsymbol{R}_1}{\boldsymbol{\varphi}}_{\omega}. \end{split} \end{align} \newline {\rm (ii)} For $t \in [k,T-k]$, there exists $\boldsymbol{R}_2 = \mathcal{O}(k^2+ k\rho(k))$ such that, for all $\boldsymbol{\varphi}\in\K{\boldsymbol{m}}$, \begin{align}\label{} \mathbf{b} (\boldsymbol{m};\boldsymbol{v}(t),\boldsymbol{\varphi}) &= -\ell_{\rm ex}^2\prod{\nabla\boldsymbol{m}(t)}{\nabla\boldsymbol{\varphi}}_{\omega} + \prod{\boldsymbol{f}\big( \, t + k/2 \, \big)}{\boldsymbol{\varphi}}_{\omega} + \frac{3}{2} \,\prod{\boldsymbol{\pi}(\boldsymbol{m}(t))}{\boldsymbol{\varphi}}_{\omega} \notag \\ & \quad - \frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}(t-k))}{\boldsymbol{\varphi}}_{\omega} + \prod{\boldsymbol{\Pi}(\boldsymbol{m}(t))}{\boldsymbol{\varphi}}_{\omega} + \frac{1}{2} \, \prod{\DPi{\boldsymbol{m}(t)}{\boldsymbol{m}(t)}}{\boldsymbol{\varphi}}_{\omega} \notag \\ & \quad - \frac{1}{2} \, \prod{\DPi{\boldsymbol{m}(t)}{\boldsymbol{m}(t-k)}}{\boldsymbol{\varphi}}_{\omega} + \prod{\boldsymbol{R}_2}{\boldsymbol{\varphi}}_{\omega}. \end{align} \end{subequations} Moreover, it holds that \begin{align}\label{} \lambda(\boldsymbol{m}) = - \ell_{\rm ex}^2 |\nabla \boldsymbol{m}|^2 + \big( \boldsymbol{f} + \boldsymbol{\pi}(\boldsymbol{m}) + \boldsymbol{\Pi}(\boldsymbol{m}) \big) \cdot \boldsymbol{m}. \end{align} \end{proposition} \begin{proof} The proof is split into the following three steps. {\bf Step~1}. For the proof of {\rm (i)}, we extend~\cite[Section 6]{akst2014} to our setting: Define \begin{subequations}\label{} \begin{align*} \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}) := \boldsymbol{h}_{\operatorname{eff}}(\boldsymbol{m}) + \boldsymbol{\Pi}(\boldsymbol{m}). \end{align*} The chain rule yields that \begin{align*} \partial_t \big[ \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}) \big] = \partial_t \big[ \boldsymbol{h}_{\operatorname{eff}}(\boldsymbol{m}) \big] + \partial_t \big[ \boldsymbol{\Pi}(\boldsymbol{m}) \big] = \ell_{\rm ex}^2 \Delta \partial_t \boldsymbol{m} + \partial_t \boldsymbol{f} \, + \boldsymbol{\pi}(\partial_t \boldsymbol{m}) + \DPi{\boldsymbol{m}}{\partial_t \boldsymbol{m}}. \end{align*} \end{subequations} Since $|\boldsymbol{m}|=1$ a.e.,~\eqref{} is equivalent to \begin{align}\label{} \alpha \, \partial_t \boldsymbol{m} + \boldsymbol{m} \times \partial_t \boldsymbol{m} = \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}) - \big( \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}) \cdot \boldsymbol{m} \big) \, \boldsymbol{m}; \end{align} see, e.g.,~\cite[Lemma 1.2.1]{goldenits2012}. Differentiating~\eqref{} with respect to time, we obtain that \begin{align* \begin{split} \alpha \, \partial_{tt}\boldsymbol{m} + \boldsymbol{m}\times\partial_{tt}\boldsymbol{m} &= \partial_t \big[ \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}) \big] - (\partial_t \big[ \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}) \big] \cdot\boldsymbol{m})\boldsymbol{m} \\ &\quad - (\boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m})\cdot\partial_t \boldsymbol{m})\boldsymbol{m} - (\boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m})\cdot\boldsymbol{m})\partial_t \boldsymbol{m} \,. \end{split} \end{align*} Since $\partial_t \boldsymbol{m} \cdot \boldsymbol{m} = 0$ in $\omega_T$, it holds that $\alpha \, \partial_t \boldsymbol{m} + \boldsymbol{m} \times \partial_t \boldsymbol{m} = \mathbb{P}_{\boldsymbol{m}}( \alpha \, \partial_t \boldsymbol{m} + \boldsymbol{m} \times \partial_t \boldsymbol{m} )$. Therefore, we further obtain that \begin{align}\label{} \begin{split} \alpha \, \boldsymbol{v} + \boldsymbol{m}\times\boldsymbol{v} &\stackrel{\eqref{eq::vformula}}{=} \mathbb{P}_{\boldsymbol{m}}\Big( \alpha \, \partial_t \boldsymbol{m} + \boldsymbol{m}\times\partial_t \boldsymbol{m} + \frac{k}{2}\alpha\,\partial_{tt}\boldsymbol{m} + \frac{k}{2}\,\boldsymbol{m}\times\partial_{tt}\boldsymbol{m} \Big) \\ &\stackrel{\eqref{} \mathbb{P}_{\boldsymbol{m}}\Big( \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}) + \frac{k}{2} \, \partial_t \big[ \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}) \big] - \frac{k}{2} \, (\boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m})\cdot\boldsymbol{m})\, \partial_t \boldsymbol{m} \Big). \end{split} \end{align} For $t\in[0,T-k]$, we multiply~\eqref{} by $\boldsymbol{\varphi}\in\K{\boldsymbol{m}}$ and integrate over $\omega$. We rearrange the terms and use that the definition~\eqref{eq::vformula} of $\boldsymbol{v}$ yields that $\partial_t \boldsymbol{m} = \boldsymbol{v} + \mathcal{O}(k)$. Then, \begin{align*} & \langle \alpha \, \boldsymbol{v}(t), \boldsymbol{\varphi} \rangle_{\omega} + \prod{\boldsymbol{m}(t)\times\boldsymbol{v}(t)}{\boldsymbol{\varphi}}_{\omega} + \frac{k}{2} \, \langle \mathbb{P}_{\boldsymbol{m}}\Big( \big( \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}(t))\cdot\boldsymbol{m}(t) \big) \boldsymbol{v}(t) \Big), \boldsymbol{\varphi} \rangle_{\omega} \\& \stackrel{\eqref{} \prod{\mathbb{P}_{\boldsymbol{m}}\big( \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}(t)) \big)}{\boldsymbol{\varphi}}_{\omega} + \frac{k}{2} \, \prod{\mathbb{P}_{\boldsymbol{m}} \big( \partial_t \big[ \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}(t)) \big] \big) }{\boldsymbol{\varphi}}_{\omega} + \mathcal{O}(k^2). \end{align*} Since any $\boldsymbol{\varphi} \in \K{\boldsymbol{m}}$ is invariant under the pointwise definition~\eqref{eq::morthproj}, the projection $\mathbb{P}_{\boldsymbol{m}}$ can be omitted in the latter equation. By definition~\eqref{} of $\lambda(\boldsymbol{m})$, it follows that \begin{align*} & \langle \big( \alpha + \frac{k}{2} \lambda(\boldsymbol{m}(t))\big) \boldsymbol{v}(t), \boldsymbol{\varphi} \rangle_{\omega} + \prod{\boldsymbol{m}(t)\times\boldsymbol{v}(t)}{\boldsymbol{\varphi}}_{\omega} \\ &\qquad = \prod{\boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}(t))}{\boldsymbol{\varphi}}_{\omega} + \frac{k}{2} \, \prod{\partial_t \big[ \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}(t)) \big] }{\boldsymbol{\varphi}}_{\omega} + \mathcal{O}(k^2). \end{align*} Note that $\partial_{\boldsymbol{n}} \boldsymbol{v} = \partial_t \partial_{\boldsymbol{n}} \boldsymbol{m} + \mathcal{O}(k) = \mathcal{O}(k)$. Therefore, integration by parts yields that \begin{align*} & \prod{\boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}(t))}{\boldsymbol{\varphi}}_{\omega} + \frac{k}{2} \prod{\partial_t \big[ \boldsymbol{h}_{\operatorname{eff},\boldsymbol{\Pi}}(\boldsymbol{m}(t)) \big] }{\boldsymbol{\varphi}}_{\omega} + \mathcal{O}(k^2) \notag \\& \stackrel{\eqref{} -\ell_{\rm ex}^2\prod{\nabla\boldsymbol{m}(t)}{\nabla\boldsymbol{\varphi}}_{\omega} + \prod{\boldsymbol{f}\big( t \big)}{\boldsymbol{\varphi}}_{\omega} + \prod{\boldsymbol{\pi}(\boldsymbol{m}(t))}{\boldsymbol{\varphi}}_{\omega} + \prod{\boldsymbol{\Pi}(\boldsymbol{m}(t))}{\boldsymbol{\varphi}}_{\omega} \notag \\&\quad - \frac{\ell_{\rm ex}^2}{2} \, k \, \prod{\nabla\boldsymbol{v}(t)}{\nabla\boldsymbol{\varphi}}_{\omega} + \frac{k}{2} \, \prod{\partial_t \boldsymbol{f}(t)}{\boldsymbol{\varphi}}_{\omega} + \frac{k}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{v}(t))}{\boldsymbol{\varphi}}_{\omega} + \frac{k}{2} \, \prod{ \DPi{\boldsymbol{m}(t)}{\boldsymbol{v}(t)}}{\boldsymbol{\varphi}}_{\omega}. \end{align*} Together with $\boldsymbol{f}(t) + \frac{k}{2} \, \partial_t \boldsymbol{f}(t) = \boldsymbol{f}(t+k/2) + \mathcal{O}(k^2)$, the latter two equations prove that \begin{align*} &\langle \big( \alpha + \frac{k}{2} \lambda(\boldsymbol{m}(t))\big) \boldsymbol{v}(t), \boldsymbol{\varphi} \rangle_{\omega} + \prod{\boldsymbol{m}(t)\times\boldsymbol{v}(t)}{\boldsymbol{\varphi}}_{\omega} \\ &\qquad = - \frac{\ell_{\rm ex}^2}{2} \, k \, \prod{\nabla\boldsymbol{v}(t)}{\nabla\boldsymbol{\varphi}}_{\omega} + \prod{\boldsymbol{f}(t+k/2)}{\boldsymbol{\varphi}}_\omega + \prod{\boldsymbol{\pi}(\boldsymbol{m}(t))}{\boldsymbol{\varphi}}_{\omega} + \prod{\boldsymbol{\Pi}(\boldsymbol{m}(t))}{\boldsymbol{\varphi}}_{\omega} \\&\qquad\qquad + \frac{k}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{v}(t))}{\boldsymbol{\varphi}}_{\omega} + \frac{k}{2} \, \prod{ \DPi{\boldsymbol{m}(t)}{\boldsymbol{v}(t)}}{\boldsymbol{\varphi}}_{\omega} + \mathcal{O}(k^2). \end{align*} It remains to replace $\big( \alpha + \frac{k}{2} \, \lambda(\boldsymbol{m}(t))\big)$ by $W_{M(k)}\big(\lambda(\boldsymbol{m}(t))\big)$, which yields an additional error of $\mathcal{O}(k^2)$ (see~\eqref{} and Lemma~\ref{), and to add the stabilization term $\frac{\ell_{\rm ex}^2}{2} \, k \, \rho(k) \, \prod{\nabla\boldsymbol{v}}{\nabla\boldsymbol{\varphi}}_{\omega}$, which yields an additional error of $\mathcal{O}(k\rho(k))$. This proves~(i). {\bf Step~2.} For the proof of~{\rm (ii)}, we first observe that an implicit Euler step satisfies that \begin{align* \boldsymbol{m}(t-k) + k \, \partial_t \boldsymbol{m}(t) = \boldsymbol{m}(t) + \mathcal{O}(k^2) \,. \end{align*} Together with Lemma~\ref{lemma::vformula}, we obtain that \begin{align* \boldsymbol{v}(t) \stackrel{\eqref{eq::vformula}}{=} \partial_t \boldsymbol{m}(t) + \mathcal{O}(k) = \frac{\boldsymbol{m}(t)-\boldsymbol{m}(t-k)}{k} + \mathcal{O}(k) \,. \end{align*} Since $\boldsymbol{\pi}$ is linear and bounded, it follows that \begin{eqnarray*}\label{eq::firstab} \frac{k}{2}\,\boldsymbol{\pi}(\boldsymbol{v}(t)) = \frac{1}{2}\,\boldsymbol{\pi}(\boldsymbol{m}(t)) - \frac{1}{2}\,\boldsymbol{\pi}(\boldsymbol{m}(t-k)) + \mathcal{O}(k^2) \,. \end{eqnarray*} Similarly, since $\boldsymbol{D}(\cdot,\cdot)$ is linear in the second argument, we also obtain that \begin{align*} \frac{k}{2} \, \DPi{\boldsymbol{m}(t)}{\boldsymbol{v}(t)} = \frac{1}{2} \, \DPi{\boldsymbol{m}(t)}{\boldsymbol{m}(t)} - \frac{1}{2} \, \DPi{\boldsymbol{m}(t)}{\boldsymbol{m}(t-k)} + \mathcal{O}(k^2). \end{align*} Combining the latter two equations with {\rm (i)}, we conclude the proof of {\rm (ii)}. {\bf Step~3.} Note that $|\boldsymbol{m}| = 1$ in $\omega$ yields that $\Delta \boldsymbol{m} \cdot \boldsymbol{m} = - | \nabla \boldsymbol{m} |^2$. In particular, $$ \lambda(\boldsymbol{m}) = \big( \boldsymbol{h}_{\operatorname{eff}}(\boldsymbol{m}) + \boldsymbol{\Pi}(\boldsymbol{m}) \big) \cdot \boldsymbol{m} = - \ell_{\rm ex}^2 |\nabla \boldsymbol{m}|^2 + \big( \boldsymbol{f} + \boldsymbol{\pi}(\boldsymbol{m}) + \boldsymbol{\Pi}(\boldsymbol{m}) \big) \cdot \boldsymbol{m}. $$ This proves~\eqref{} and concludes the proof. \end{proof} \begin{remark}\label{re::rhok} {\rm(i)} For $\rho(k) = \mathcal{O}(k)$, Proposition~\ref{} yields that $\boldsymbol{R}_1,\boldsymbol{R}_2 = \mathcal{O}(k^2)$ and therefore second-order accuracy of the consistency error. In this case, however, the convergence result of Theorem~\ref{} requires the CFL-type condition $k = \mathbf{o}(h)$. Instead, the choice $\rho(k) := k |\log(k)|$ requires no CFL-type condition and leads to $\rho(k)= \mathcal{O}(k^{1-\varepsilon})$ and hence to $\boldsymbol{R}_1,\boldsymbol{R}_2 = \mathcal{O}(k^{2-\varepsilon})$ for any $\varepsilon > 0$. For details, we refer to Remark~\ref{} below. {\rm(ii)} In the proof of Proposition~\ref{}, $\big( \alpha + \frac{k}{2} \, \lambda(\boldsymbol{m})\big)$ is replaced by $W_{M(k)}\big(\lambda(\boldsymbol{m})\big)$ to ensure ellipticity of the bilinear form $\mathbf{b}(\boldsymbol{m};\cdot,\cdot)$. \end{remark} \section{Proof of Theorem~\ref{} (Numerical integration of LLG)} \label{} \begin{proof}[Proof of Theorem~\ref{}] Together with the Lax--Milgram theorem, we employ a fixpoint iteration in order to solve~\eqref{} with $\boldsymbol{\eta}_{h}^{\ell} \approx \boldsymbol{v}_h^i$: Let $\boldsymbol{\eta}_h^0 := \boldsymbol{0}$. For $\ell\in\mathbb{N}_0$, let $\boldsymbol{\eta}_{h}^{\ell+1} \in \Kh{\boldsymbol{m}_h^i}$ solve, for all $\boldsymbol{\varphi}_h \in \Kh{\boldsymbol{m}_h^i}$, \begin{align}\label{} \begin{split} &\prod{W_{M(k)}(\lambda_h^i) \, \boldsymbol{\eta}_{h}^{\ell+1}}{\boldsymbol{\varphi}_h}_{\omega} + \prod{\boldsymbol{m}_h^i\times \boldsymbol{\eta}_{h}^{\ell+1}}{\boldsymbol{\varphi}_h}_{\omega} + \frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k))\prod{\nabla\boldsymbol{\eta}_{h}^{\ell+1}}{\nabla\boldsymbol{\varphi}_h}_{\omega} \\& \quad = -\ell_{\rm ex}^2\prod{\nabla\boldsymbol{m}_h^i}{\nabla\boldsymbol{\varphi}_h}_{\omega} + \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{\varphi}_h}_{\omega} + \prod{\pih{i}{\boldsymbol{\eta}_{h}^{\ell}}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{\varphi}_{h}}_{\omega} \\& \quad \quad \quad + \prod{\Pih{i}{\boldsymbol{\eta}_{h}^{\ell}}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{\varphi}_h}_{\omega} \,. \end{split} \end{align} We equip $\Kh{\boldsymbol{m}_h^i}$ with the norm $|\!|\!|\boldsymbol{\eta}_h|\!|\!|^2 := (\alpha/4) \, \norm{\boldsymbol{\eta}_h}{\omega}^2 + (\ell_{\rm ex}^2/2) \, k \, (1+\rho(k)) \, \norm{\nabla \boldsymbol{\eta}_h}{\omega}^2$. For sufficiently small $k$, Lemma~\ref{ and~\eqref{} prove that $W_{M(k)}(\cdot) \ge \alpha/2$. Since $\rho(k) \geq 0$, the bilinear form on the left-hand side of~\eqref{} is positive definite on $\Kh{\boldsymbol{m}_h^i}$, i.e., the fixpoint iteration is well-defined. We subtract~\eqref{} for $\boldsymbol{\eta}_h^{\ell+1}$ from~\eqref{} for $\boldsymbol{\eta}_h^{\ell}$ and test with $\boldsymbol{\varphi}_h := \boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell} \in \Kh{\boldsymbol{m}_h^i}$. With ~\eqref{} and~\eqref{}, we get that \begin{align}\label{} & \frac{\alpha}{2} \, \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega}^2 + \frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k)) \, \norm{\nabla \boldsymbol{\eta}_h^{\ell+1} - \nabla \boldsymbol{\eta}_h^{\ell}}{\omega}^2 \notag \\ & \quad \stackrel{\eqref{} \norm{\pih{i}{\boldsymbol{\eta}_h^{\ell}}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}} - \pih{i}{\boldsymbol{\eta}_h^{\ell-1}}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\omega} \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega} \notag \\ & \quad\quad \quad + \norm{\Pih{i}{\boldsymbol{\eta}_h^{\ell}}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}} - \Pih{i}{\boldsymbol{\eta}_h^{\ell-1}}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\omega} \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega} \notag \\ & \quad \,\,\leq\,\, (C_{\mathrm{\bf \ppi}} + C_{\mathrm{\bf \Ppi}}) \,k \,\norm{\boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1}}{\omega} \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega} \notag \\ & \quad\quad \quad + C_{\mathrm{\bf \Ppi}} \, k \, \norm{\nabla(\boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1})}{\omega} \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega}. \end{align} For arbitrary $\delta>0$, the Young inequality thus yields that \begin{align*} & \min\Big\{1 \, , \, 2 - \frac{1}{\alpha\delta} \, (C_{\mathrm{\bf \ppi}} + 2\,C_{\mathrm{\bf \Ppi}}) \, k \Big\} \, |\!|\!| \boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell} |\!|\!|^2 \notag \\ &\quad\le \Big[\frac{\alpha}{2} - \frac{1}{4\delta} \, (C_{\mathrm{\bf \ppi}} + 2\,C_{\mathrm{\bf \Ppi}}) \, k \Big] \, \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega}^2 + \frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k)) \, \norm{\nabla \boldsymbol{\eta}_h^{\ell+1} - \nabla \boldsymbol{\eta}_h^{\ell}}{\omega}^2 \notag \\ &\quad\le \delta \, (C_{\mathrm{\bf \ppi}} + C_{\mathrm{\bf \Ppi}}) \,k \,\norm{\boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1}}{\omega}^2 + \delta \, C_{\mathrm{\bf \Ppi}} \, k \, \norm{\nabla(\boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1})}{\omega}^2 \notag \\ &\quad\le \delta \, \max\Big\{\frac{4}{\alpha} \, (C_{\mathrm{\bf \ppi}} + C_{\mathrm{\bf \Ppi}}) \,k \, , \, 2 \, \frac{C_{\mathrm{\bf \Ppi}}}{\ell_{\rm ex}^2} \Big\} \, |\!|\!| \boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1} |\!|\!|^2 \quad \textrm{for all } \ell \in \mathbb{N}. \end{align*} For sufficiently small $\delta$ and $k$, the iteration is thus a contraction with respect to $|\!|\!|\cdot|\!|\!|$. The Banach fixpoint theorem yields the existence and uniqueness of the solution $\boldsymbol{v}_h^i \in \Kh{\boldsymbol{m}_h^i}$ to~\eqref{}. For all $\boldsymbol{z}_h \in \mathcal N_h$, it holds that $|\boldsymbol{m}_h^i(\boldsymbol{z}_h) + k \, \boldsymbol{v}_h^i(\boldsymbol{z}_h)|^2 = 1 + k^2 \, |\boldsymbol{v}_h^i(\boldsymbol{z}_h)|^2 \ge 1$ so that $\boldsymbol{m}_h^{i+1} \in \boldsymbol{\mathcal{M}}_h$ in~\eqref{} is well-defined. Altogether, Algorithm~\ref{} is thus well-posed. \end{proof} \begin{lemma}\label{} Under the assumptions of Theorem~\ref{}, the following assertions {\rm (i)}--{\rm (ii)} are satisfied, if $k>0$ is sufficiently small. {\rm (i)} For all $i = 0,\dots,N-1$, it holds that \begin{align}\label{} & \frac{\ell_{\rm ex}^2}{2} \mathop{\mathrm{d}_t} \norm{\nabla\boldsymbol{m}_h^{i+1}}{\omega}^2 + \prod{W_{M(k)}(\lambda_h^i)\boldsymbol{v}_h^i}{\boldsymbol{v}_h^i}_{\omega} + \frac{\ell_{\rm ex}^2}{2} \, k \, \rho(k) \, \norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 \notag \\& \quad \le \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{v}_h^i}_{\omega} + \prod{\pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega} + \prod{\Pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega} \end{align} {\rm (ii)} For all $j = 0,\dots,N$, it holds that \begin{align}\label{} \norm{\nabla \boldsymbol{m}_h^j}{\omega}^2 + k\sum_{i=0}^{j-1}\norm{\boldsymbol{v}_h^i}{\omega}^2 + k^2\rho(k)\sum_{i=0}^{j-1}\norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 \le C \,, \end{align} where $C>0$ only depends on $T$, $\omega$, $\alpha$, $\ell_{\rm ex}^2$, $\boldsymbol{f}$, $\boldsymbol{\pi}$, $\boldsymbol{\Pi}$, $\boldsymbol{m}^0$, and $\kappa$. \end{lemma} \begin{proof} Testing \eqref{} with $\boldsymbol{\varphi}_h := \boldsymbol{v}_h^i \in \Kh{\boldsymbol{m}_h^i}$, we see that \begin{align}\label{} &\prod{W_{M(k)}(\lambda_h^i)\boldsymbol{v}_h^i}{\boldsymbol{v}_h^i}_{\omega} + \frac{\ell_{\rm ex}^2}{2} \, k \, \norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 + \frac{\ell_{\rm ex}^2}{2} \, k \rho(k) \, \norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 \notag \\ &= -\ell_{\rm ex}^2\prod{\nabla\boldsymbol{m}_h^i}{\nabla\boldsymbol{v}_h^i}_{\omega} + \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{v}_h^i}_{\omega} \\& \qquad + \prod{\pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega} \notag + \prod{\Pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega}. \end{align} Since $\mathcal{T}_h$ is weakly acute, \cite[Lemma 3.2]{bartels2005} provides the estimate \begin{align*} \begin{split} \norm{\nabla\boldsymbol{m}_h^{i+1}}{\omega}^2 \le \norm{\nabla ( \boldsymbol{m}_h^i + k\boldsymbol{v}_h^i )}{\omega}^2 &= \norm{\nabla\boldsymbol{m}_h^i}{\omega}^2 + 2k \, \prod{\nabla\boldsymbol{m}_h^i}{\nabla\boldsymbol{v}_h^i}_{\omega} + k^2\norm{\nabla\boldsymbol{v}_h^i}{\omega}^2. \end{split} \end{align*} Rearranging this estimate and multiplying it by $\ell_{\rm ex}^2/(2k)$, we derive that \begin{align}\label{} \frac{\ell_{\rm ex}^2}{2} \mathop{\mathrm{d}_t} \norm{\nabla\boldsymbol{m}_h^{i+1}}{\omega}^2 - \frac{\ell_{\rm ex}^2}{2} \, k \, \norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 \le \ell_{\rm ex}^2\prod{\nabla\boldsymbol{m}_h^i}{\nabla\boldsymbol{v}_h^i}_{\omega}. \end{align} Adding \eqref{} and \eqref{}, we prove {\rm (i)}. To prove {\rm (ii)}, we sum~\eqref{} over $i=0,\dots,j-1$ and multiply by $k$. For $k$ sufficiently small, we have $W_{M(k)}(\cdot) \geq \alpha/2 > 0$ (cf. Lemma~\ref{ (i)}) and altogether get that \begin{align}\label{} \begin{split} {{\rm \chi}_j} &:= \frac{\ell_{\rm ex}^2}{2} \, \norm{ \nabla \boldsymbol{m}_h^j }{\omega}^2 + \frac{\alpha}{2} \, k \sum_{i=0}^{j-1} \norm{ \boldsymbol{v}_h^i }{\omega}^2 + \frac{\ell_{\rm ex}^2}{2} \, k^2\rho(k)\sum_{i=0}^{j-1}\norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 \\& \stackrel{\eqref{} \frac{\ell_{\rm ex}^2}{2} \, \norm{ \nabla \boldsymbol{m}_h^0 }{\omega}^2 + k \sum_{i=0}^{j-1} \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{v}_h^i}_{\omega} + k \sum_{i=0}^{j-1} \prod{\pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega} \\& \quad + k \sum_{i=0}^{j-1} \prod{\Pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^{i}}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega}. \end{split} \end{align} Let $\delta > 0$. The Young inequality proves that \begin{align}\label{} k^2 \sum_{i=0}^{j-1} \norm{ \nabla \boldsymbol{v}_h^i }{\omega} \norm{ \boldsymbol{v}_h^i }{\omega} \lesssim \delta k^2 \rho(k) \sum_{i=0}^{j-1} \norm{ \nabla \boldsymbol{v}_h^i }{\omega}^2 + \frac{k^2}{\delta \rho(k)} \, \sum_{i=0}^{j-1} \norm{\boldsymbol{v}_h^i }{\omega}^2. \end{align} Together with~\eqref{} and~\eqref{}, further applications of the Young inequality prove that \begin{align*} \chi_j &\stackrel{\eqref{ \norm{ \nabla \boldsymbol{m}_h^0 }{\omega}^2 + k \sum_{i=0}^{j-1} \norm{\boldsymbol{f}(t_{i+1/2)}}{\omega}\norm{ \boldsymbol{v}_h^i }{\omega} + k \sum_{i=0}^{j-1} \Big( k \, \norm{\boldsymbol{v}_h^i}{\omega} \!+\! \norm{\boldsymbol{m}_h^i}{\omega} \!+\! \norm{\boldsymbol{m}_h^{i-1}}{\omega} \Big) \norm{\boldsymbol{v}_h^i}{\omega} \\& \quad + k \sum_{i=0}^{j-1} \Big( k \, \norm{\boldsymbol{v}_h^i}{\boldsymbol{H}{1}{\omega}} + \norm{\boldsymbol{m}_h^i}{\boldsymbol{H}{1}{\omega}} + \norm{\boldsymbol{m}_h^{i-1}}{\boldsymbol{H}{1}{\omega}} \Big) \norm{\boldsymbol{v}_h^i}{\omega} \\& \stackrel{\eqref{} \norm{ \nabla \boldsymbol{m}_h^0 }{\omega}^2 + \frac{k}{\delta} \sum_{i=0}^{j-1} \norm{\boldsymbol{f}(t_{i+1/2})}{\omega}^2 + \frac{k}{\delta} \sum_{i=0}^{j-1} \norm{ \boldsymbol{m}_h^i }{\omega}^2 + \frac{k}{\delta} \sum_{i=0}^{j-1} \norm{ \nabla \boldsymbol{m}_h^i }{\omega}^2 \\& \quad + \Big(\delta + k + \frac{k}{\delta \rho(k)}\Big) \, k \, \sum_{i=0}^{j-1} \norm{ \boldsymbol{v}_h^i }{\omega}^2 + \delta \, k^2\rho(k) \, \sum_{i=0}^{j-1} \norm{ \nabla \boldsymbol{v}_h^i }{\omega}^2. \end{align*} Since $k/\rho(k)\to0$ as $k\to0$, we choose $\delta>0$ sufficiently small and can absorb the sums $k \sum_{i=0}^{j-1} \norm{ \boldsymbol{v}_h^i }{\omega}^2$ and $k^2 \rho(k) \sum_{i=0}^{j-1} \norm{ \nabla \boldsymbol{v}_h^i }{\omega}^{2}$ in $\chi_j$. Using that $\boldsymbol{f} \in C^1([0,T],\LL{2}{\omega})$, we altogether arrive at the estimate \begin{align} {{\rm \chi}_j} \lesssim \norm{ \nabla \boldsymbol{m}_h^0 }{\omega}^2 + k \sum_{i=0}^{j-1} \norm{\boldsymbol{f}(t_{i+1/2})}{\omega}^2 + k\sum_{i=0}^{j-1} \norm{ \boldsymbol{m}_h^i }{\omega}^2 + k \sum_{i=0}^{j-1} \norm{\nabla \boldsymbol{m}_h^i}{\omega}^2 \lesssim 1 + k \sum_{i=0}^{j-1} {{\rm \chi_i}} . \notag \end{align} This fits in the setting of the discrete Gronwall lemma (cf.\ \cite[Lemma 1.4.2]{qv1994}), i.e., \begin{align} {{\rm \chi}_j} \lesssim \alpha_0 + \beta \sum_{i=0}^{j-1} {{\rm \chi}_i} \notag \quad \textrm{with } \alpha_0 > 0 \textrm{ and } \beta \simeq k. \end{align} We obtain that \begin{align}\label{} {\rm \chi}_j \lesssim \alpha_0 \exp \left( \sum_{i=0}^{j-1} \beta \right) \lesssim \exp(T). \end{align} This concludes the proof. \end{proof} \begin{lemma}\label{} Under the assumptions of Theorem~\ref{}, consider the postprocessed output~\eqref{} of Algorithm~\ref{}. Then, there exists $\boldsymbol{m}\in\boldsymbol{H}{1}{\omega_T}\cap\L{\infty}{0,T;\boldsymbol{H}{1}{\omega}}$ as well as a subsequence of each $\boldsymbol{m}_{hk}^{\star} \in \{\boldsymbol{m}_{hk}, \boldsymbol{m}_{hk}^{=} , \boldsymbol{m}_{hk}^-, \boldsymbol{m}_{hk}^+\}$ and of $\boldsymbol{v}_{hk}^-$ such that the following convergence statements~\ref{} hold true simultaneously for the same subsequence as $h,k\to0$: \begin{enumerate}[label=\rm{(\roman*)}] \item \label{} $\boldsymbol{m}_{hk} \rightharpoonup \boldsymbol{m}$ in $\boldsymbol{H}{1}{\omega_T}$, \item \label{} $\boldsymbol{m}_{hk}^{\star} \stackrel{*}\rightharpoonup \boldsymbol{m}$ in $\L{\infty}{0,T;\boldsymbol{H}{1}{\omega}}$, \item \label{} $\boldsymbol{m}_{hk}^{\star} \rightharpoonup \boldsymbol{m}$ in $\L{2}{0,T;\boldsymbol{H}{1}{\omega}}$, \item \label{} $\boldsymbol{m}_{hk}^{\star} \to \boldsymbol{m}$ in $\LL{2}{\omega_T}$, \item \label{} $\boldsymbol{m}_{hk}^{\star} (t) \rightarrow \boldsymbol{m}(t)$ in $\LL{2}{\omega}$ for $t \in (0,T)$ a.e., \item \label{} $\boldsymbol{m}_{hk}^{\star} \rightarrow \boldsymbol{m}$ pointwise a.e.\ in $\omega_T$, \item \label{} $\boldsymbol{v}_{hk}^- \rightharpoonup \partial_t \boldsymbol{m}$ in $\LL{2}{\omega_T}$, \item \label{} $k \, \nabla \boldsymbol{v}_{hk}^- \rightarrow 0$ in $\LL{2}{\omega_T}$. \end{enumerate} \end{lemma} \begin{proof} Lemma~\ref{} yields uniform boundedness of $\boldsymbol{m}_{hk}$ in $\boldsymbol{H}{1}{\omega_T}$ and of $\boldsymbol{m}_{hk}^\star \in \L{\infty}{0,T;\linebreak\boldsymbol{H}{1}{\omega}}$. Therefore, \ref{} follow as in~\cite{alouges2008a,bffgpprs2014}. Lemma~\ref{ yields that \begin{align}\label{} k^2\norm{\nabla\boldsymbol{v}_{hk}^-}{\LL{2}{\omega_T}}^2 = k \rho(k)^{-1} \rho(k) k^2 \sum_{i=0}^{N-1} \norm{\nabla\boldsymbol{v}_{h}^i}{\omega}^2 \stackrel{\eqref{ k \rho(k)^{-1} \stackrel{\eqref{} 0 \quad\text{ as } h,k\to 0. \end{align} This proves~(viii) and concludes the proof. \end{proof} \begin{remark}\label{} Under the CFL-type condition $k = \mathbf{o}(h)$, one may choose $\rho(k)\equiv0$ and hence violate~\eqref{}. To see this, note that~\eqref{} is only used for the proof of~\eqref{} and~\eqref{}. An inverse inequality yields that \begin{align*} k^2 \sum_{i=0}^{j-1} \norm{ \nabla \boldsymbol{v}_h^i }{\omega} \norm{ \boldsymbol{v}_h^i }{\omega} \lesssim (kh^{-1}) \, k \sum_{i=0}^{j-1} \norm{ \boldsymbol{v}_h^i }{\omega}^2, \end{align*} where $k/h \to 0$ as $k\to0$. Similarly, \begin{align*} k^2 \, \norm{\nabla\boldsymbol{v}_{hk}^-}{\LL{2}{\omega_T}}^2 = k^3 \sum_{i=0}^{N-1} \norm{\nabla\boldsymbol{v}_{h}^i}{\omega}^2 \lesssim k^3 h^{-2} \sum_{i=0}^{N-1} \norm{\boldsymbol{v}_{h}^i}{\omega}^2 \stackrel{\eqref{ k^2h^{-2} \longrightarrow 0 \quad\text{ as } h,k\to 0. \end{align*} Therefore, Lemma~\ref{} as well as Lemma~\ref{} (and hence also Theorem~\ref{}) remain valid.\qed \end{remark} \begin{proof}[Proof of Theorem~\ref{}] We verify that $\boldsymbol{m}\in\boldsymbol{H}{1}{\omega_T}\cap\L{\infty}{0,T;\boldsymbol{H}{1}{\omega}}$ from Lemma~\ref{} satisfies Definition~\ref{def::weaksolution}\ref{}. The modulus constraint $|\boldsymbol{m}|=1$ a.e.\ in $\omega_T$ as well as $\boldsymbol{m}(0) = \boldsymbol{m}^0$ in the sense of traces follow as in~\cite{alouges2008a,bffgpprs2014}. Hence, $\boldsymbol{m}$ satisfies Definition~\ref{def::weaksolution}\ref{}. It remains to prove that $\boldsymbol{m}$ from Lemma~\ref{} satisfies the variational formulation~\eqref{} from Definition \ref{def::weaksolution}\ref{}: To that end, let $\boldsymbol{\varphi}\in\boldsymbol{C}^{\infty}(\overline{\omega_T})$. Let $\boldsymbol{\mathcal{I}}_{h}:\boldsymbol{C}(\overline{\omega}) \to \boldsymbol{V}_{\!\!h}$ be the (vector-valued) nodal interpolation operator. For $t \in [t_i,t_{i+1})$ and $i=0,\dots,N-1$, we test~\eqref{} with $\boldsymbol{\mathcal{I}}_{h}\big( \boldsymbol{m}_h^i \times \boldsymbol{\varphi}(t) \big) \in \Kh{\boldsymbol{m}_h^i}$ and integrate over time. With the definition of the postprocessed output~\eqref{} and $\overline{\boldsymbol{f}}_k(t) := \boldsymbol{f}(t_{i+1/2})$ for $t \in [t_i,t_{i+1})$, we obtain that \begin{align}\label{} & I_{hk}^1 + I_{hk}^2 + \frac{\ell_{\rm ex}^2}{2} \, I_{hk}^3 \notag \\& := \int_0^T\prod{W_{M(k)}(\lambda_{hk}^-)\boldsymbol{v}_{hk}^-}{\boldsymbol{\mathcal{I}}_{h}(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} + \int_0^T\prod{\boldsymbol{m}_{hk}^-\times\boldsymbol{v}_{hk}^-}{\boldsymbol{\mathcal{I}}_{h}(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} \notag \\& \quad + \frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k))\int_0^T\prod{\nabla\boldsymbol{v}_{hk}^-}{\nabla\boldsymbol{\mathcal{I}}_{h}(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} \notag \\& \stackrel{\eqref{} -\ell_{\rm ex}^2\int_0^T\prod{\nabla\boldsymbol{m}_{hk}^-}{\nabla\boldsymbol{\mathcal{I}}_{h}(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} + \int_0^T\prod{\boldsymbol{\pi}_{hk}^-(\boldsymbol{v}_{hk}^-;\boldsymbol{m}_{hk}^-,\boldsymbol{m}_{hk}^=)}{\boldsymbol{\mathcal{I}}_{h}(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} \notag \\& \quad + \int_0^T\prod{\overline{\boldsymbol{f}}_{k}}{\boldsymbol{\mathcal{I}}_{h}(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} + \int_0^T\prod{\boldsymbol{\Pi}_{hk}^-(\boldsymbol{v}_{hk}^-;\boldsymbol{m}_{hk}^-,\boldsymbol{m}_{hk}^=)}{\boldsymbol{\mathcal{I}}_{h}(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} \notag \\& =: - \ell_{\rm ex}^2 I_{hk}^4 + I_{hk}^5 + I_{hk}^6 + I_{hk}^7 . \end{align} In the following, we prove convergence of the integrals from~\eqref{} towards their continuous counterparts in the variational formulation~\eqref{}: To this end, recall the approximation properties of the nodal interpolation operator $\boldsymbol{\mathcal{I}}_{h}$ and note that $\overline{\boldsymbol{f}}_k\to\boldsymbol{f}$ in $C([0,T];\LL{2}{\Omega})$. With Lemma~\ref{}, we get as in~\cite{alouges2008a,bffgpprs2014} that \begin{align* I_{hk}^2 & \longrightarrow \int_0^T\prod{\boldsymbol{m}\times\partial_t \boldsymbol{m}}{\boldsymbol{m}\times\boldsymbol{\varphi}}_{\omega} \d{t} = \int_0^T\prod{\partial_t \boldsymbol{m}}{\boldsymbol{\varphi}}_{\omega} \d{t} &&\text{as $h,k\to0$}, \\ I_{hk}^4 & \longrightarrow - \int_0^T\prod{\boldsymbol{m}\times\nabla\boldsymbol{m}}{\nabla\boldsymbol{\varphi}}_{\omega} \d{t} &&\text{as $h,k\to0$}, \\ I_{hk}^6 & \longrightarrow \int_0^T\prod{\boldsymbol{f}}{\boldsymbol{m}\times\boldsymbol{\varphi}}_{\omega} \d{t} &&\text{as $h,k\to0$}. \end{align*} Since $M(k)k \to 0$, Lemma~\ref{ yields that $W_{M(k)}(\lambda_{hk}^-) \rightarrow \alpha$ in $\L{\infty}{\omega_T}$. Together with the assumption $\rho(k) \rightarrow 0$ as $k\to0$, we get as in~\cite{akst2014} that \begin{align} I_{hk}^1 & \longrightarrow \alpha \int_0^T\prod{\partial_t \boldsymbol{m}}{\boldsymbol{m}\times\boldsymbol{\varphi}}_{\omega} \d{t} \quad \textrm{and} \quad I_{hk}^3 \rightarrow 0 \quad\text{as $h,k\to 0$}. \end{align} With the assumptions~\eqref{} and~\eqref{} on $\boldsymbol{\pi}_{hk}^-$ and $\boldsymbol{\Pi}_{hk}^-$, respectively, we conclude that \begin{align} I_{hk}^5 &= \int_0^T\prod{\boldsymbol{\pi}_{hk}^-(\boldsymbol{v}_{hk}^-;\boldsymbol{m}_{hk}^-,\boldsymbol{m}_{hk}^=)}{\boldsymbol{\mathcal{I}}_{h}(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} \stackrel{\eqref{} \int_0^T\prod{\boldsymbol{\pi}(\boldsymbol{m})}{\boldsymbol{m}\times\boldsymbol{\varphi}}_{\omega} \d{t}, \notag \\ \label{} I_{hk}^7 &= \int_0^T\prod{\boldsymbol{\Pi}_{hk}^-(\boldsymbol{v}_{hk}^-;\boldsymbol{m}_{hk}^-,\boldsymbol{m}_{hk}^=)}{\boldsymbol{\mathcal{I}}_{h}(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} \stackrel{\eqref{} \int_0^T\prod{\boldsymbol{\Pi}(\boldsymbol{m})}{\boldsymbol{m}\times\boldsymbol{\varphi}}_{\omega} \d{t}, \end{align} as $h,k \rightarrow 0$. Altogether, $\boldsymbol{m}$ from Lemma~\ref{} satisfies the variational formulation~\eqref{}. \end{proof} \begin{proof}[Proof of Theorem~\ref{}] It remains to verify that $\boldsymbol{m}$ from Lemma~\ref{} satisfies the energy estimate of Definition~\ref{def::weaksolution}\ref{}: To that end, let $\tau\in(0,T)$ be arbitrary and $j \in \{1,\dots,N\}$ such that $\tau \in [t_{j-1},t_{j})$. Besides the shorthand notation $\boldsymbol{f}^j := \boldsymbol{f}(t_j)$, define the time reconstructions $\boldsymbol{f}_k$ and $\overline{\boldsymbol{f}}_k$ according to~\eqref{}. For any $i = 0,\dots, j-1$, Lemma~\ref{ (i)} shows that \begin{eqnarray}\label{} && \hspace*{-1.5cm} \mathcal{E}_{\textrm{LLG}}(\boldsymbol{m}_{hk}(t_{i+1})) - \mathcal{E}_{\textrm{LLG}}(\boldsymbol{m}_{hk}(t_{i})) \notag \\ &\stackrel{\eqref{} & \frac{\ell_{\rm ex}^2}{2} \, \norm{\nabla\boldsymbol{m}_h^{i+1}}{\omega}^2 - \frac{\ell_{\rm ex}^2}{2} \, \norm{\nabla\boldsymbol{m}_h^i}{\omega}^2 \notag \\ && \quad -\frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^{i+1})}{\boldsymbol{m}_h^{i+1}}_{\omega} + \frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^i)}{\boldsymbol{m}_h^i}_{\omega} - \prod{\boldsymbol{f}^{i+1}}{\boldsymbol{m}_h^{i+1}}_{\omega} + \prod{\boldsymbol{f}^i}{\boldsymbol{m}_h^i}_{\omega} \notag \\ &\stackrel{\eqref{ - k \, \prod{W_{M(k)}(\lambda_h^i)\boldsymbol{v}_h^i}{\boldsymbol{v}_h^i}_{\omega} - \frac{\ell_{\rm ex}^2}{2} \, k^2 \rho(k) \, \norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 + k \, \prod{\boldsymbol{\Pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1})}{\boldsymbol{v}_h^i}_{\omega} \notag \\ && \quad + \underbrace{ k \, \prod{\boldsymbol{\pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1})}{\boldsymbol{v}_h^i}_{\omega} -\frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^{i+1})}{\boldsymbol{m}_h^{i+1}}_{\omega} + \frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^i)}{\boldsymbol{m}_h^i}_{\omega}}_{=:T_{\boldsymbol{\pi}}} \notag \\ && \quad + \underbrace{ k \, \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{v}_h^i}_{\omega} - \prod{\boldsymbol{f}^{i+1}}{\boldsymbol{m}_h^{i+1}}_{\omega} + \prod{\boldsymbol{f}^i}{\boldsymbol{m}_h^i}_{\omega}}_{=:T_{\boldsymbol{f}}} \,. \end{eqnarray} Since $\boldsymbol{\pi}$ is linear and self-adjoint, we obtain that \begin{align}\label{} T_{\boldsymbol{\pi}} &= -\frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^{i+1})}{\boldsymbol{m}_h^{i+1}}_{\omega} + \frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^i)}{\boldsymbol{m}_h^i}_{\omega} + k \, \prod{\boldsymbol{\pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1})}{\boldsymbol{v}_h^i}_{\omega} \notag \\ &= k \, \prod{\boldsymbol{\pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1}) - \boldsymbol{\pi}(\boldsymbol{m}_h^i)}{\boldsymbol{v}_h^i}_{\omega} -\prod{\boldsymbol{\pi}(\boldsymbol{m}_h^i)}{\boldsymbol{m}_h^{i+1}-\boldsymbol{m}_h^i-k\boldsymbol{v}_h^i} \notag \\ &\quad -\frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^{i+1}-\boldsymbol{m}_h^i)}{\boldsymbol{m}_h^{i+1}-\boldsymbol{m}_h^i}_{\omega} . \end{align} Similarly, it holds that \begin{align*} T_{\boldsymbol{f}} &= k \, \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{v}_h^i}_{\omega} - k \, \prod{\mathop{\mathrm{d}_t} \boldsymbol{f}^{i+1}}{\boldsymbol{m}_h^i}_{\omega} - \prod{\boldsymbol{f}^{i+1}}{\boldsymbol{m}_h^{i+1}-\boldsymbol{m}_h^i}_{\omega} \\& = k \, \prod{\boldsymbol{f}(t_{i+1/2})-\boldsymbol{f}^{i+1}}{\boldsymbol{v}_h^i}_{\omega} - k \, \prod{\mathop{\mathrm{d}_t} \boldsymbol{f}^{i+1}}{\boldsymbol{m}_h^i}_{\omega} - \prod{\boldsymbol{f}^{i+1}}{\boldsymbol{m}_h^{i+1}-\boldsymbol{m}_h^i-k\boldsymbol{v}_h^i}_{\omega} \\& =: T_{\boldsymbol{f},1} - k \, \prod{\mathop{\mathrm{d}_t} \boldsymbol{f}^{i+1}}{\boldsymbol{m}_h^i}_{\omega} + T_{\boldsymbol{f},2}. \end{align*} Elementary calculations (see, e.g.,~\cite[Lemma~3.3.2]{goldenits2012}) show that \begin{align*} |\boldsymbol{m}_h^{i+1}(\boldsymbol{z}_h) - \boldsymbol{m}_h^i(\boldsymbol{z}_h)| \le k \, |\boldsymbol{v}_h^i(\boldsymbol{z}_h)| \quad \text{for all } \boldsymbol{z}_h \in \mathcal N_h. \end{align*} A scaling argument thus proves that \begin{align}\label{} k \, \norm{\mathop{\mathrm{d}_t} \boldsymbol{m}_h^{i+1}}{\LL{2}{\omega}} = \norm{\boldsymbol{m}_h^{i+1} - \boldsymbol{m}_h^i}{\LL{2}{\omega}} \lesssim k \, \norm{\boldsymbol{v}_h^i}{\LL{2}{\omega}}. \end{align} Elementary calculations (see, e.g.,~\cite[eq.~(22)]{aj2006} or~\cite[Lemma~3.3.3]{goldenits2012}) show that \begin{align*} |\boldsymbol{m}_h^{i+1}(\boldsymbol{z}_h) - \boldsymbol{m}_h^i(\boldsymbol{z}_h) - k \, \boldsymbol{v}_h^i(\boldsymbol{z}_h)| \le \frac{k^2}{2} \, |\boldsymbol{v}_h^i(\boldsymbol{z}_h)|^2 \quad \text{for all } \boldsymbol{z}_h \in \mathcal N_h. \end{align*} A scaling argument thus proves that \begin{align}\label{} k \, \norm{\mathop{\mathrm{d}_t} \boldsymbol{m}_h^{i+1} - \boldsymbol{v}_h^i}{\LL{p}{\omega}} = \norm{\boldsymbol{m}_h^{i+1} - \boldsymbol{m}_h^i - k \, \boldsymbol{v}_h^i}{\LL{p}{\omega}} \lesssim k^2 \, \norm{\boldsymbol{v}_h^i}{\LL{{2p}}{\omega}}^2 \text{ for } 1 \le p < \infty. \end{align}% The Sobolev embedding $\boldsymbol{H}{1}{\omega} \subset \LL{6}{\omega}$ yields that \begin{align*} \norm{\boldsymbol{v}_h^i}{\LL{3}{\omega}}^2 \le \norm{\boldsymbol{v}_h^i}{\LL{2}{\omega}} \, \norm{\boldsymbol{v}_h^i}{\LL{6}{\omega}} \lesssim \norm{\boldsymbol{v}_h^i}{\LL{2}{\omega}} \, \norm{\boldsymbol{v}_h^i}{\boldsymbol{H}{1}{\omega}} \le \norm{\boldsymbol{v}_h^i}{\omega}^2 + \norm{\boldsymbol{v}_h^i}{\omega} \, \norm{\nabla\boldsymbol{v}_h^i}{\omega}. \end{align*} Therefore, we obtain that \begin{align*} \norm{\boldsymbol{m}_h^{i+1} - \boldsymbol{m}_h^i - k \boldsymbol{v}_h^i}{\LL{{3/2}}{\omega}} \stackrel{\eqref{ k^2 \, \norm{\boldsymbol{v}_h^i}{\LL{3}{\omega}}^2 \lesssim k^2 \, \norm{\boldsymbol{v}_h^i}{\omega} \norm{\nabla \boldsymbol{v}_h^i}{\omega} + k^2 \, \norm{\boldsymbol{v}_h^i}{\omega}^2. \end{align*}% With the stronger boundedness~\eqref{} of $\boldsymbol{\pi}$ and the Hölder inequality, we derive that \begin{align}\label{} \begin{split} |T_{\boldsymbol{\pi}}| & \stackrel{\eqref{} k \, | \prod{\boldsymbol{\pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1}) - \boldsymbol{\pi}(\boldsymbol{m}_h^i)}{\boldsymbol{v}_h^i}_{\omega} | \\& \quad + \norm{\boldsymbol{\pi}(\boldsymbol{m}_h^i)}{\LL{3}{\omega}} \, \norm{\boldsymbol{m}_h^{i+1}-\boldsymbol{m}_h^i-k\boldsymbol{v}_h^i}{\LL{{3/2}}{\omega}} \\& \quad + \norm{\boldsymbol{\pi}(\boldsymbol{m}_h^{i+1} - \boldsymbol{m}_h^i)}{\omega} \, \norm{\boldsymbol{m}_h^{i+1} - \boldsymbol{m}_h^i}{\omega} \\& \,\lesssim\, k \, | \prod{\boldsymbol{\pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1}) - \boldsymbol{\pi}(\boldsymbol{m}_h^i)}{\boldsymbol{v}_h^i}_{\omega} | + k^2 \, \norm{\boldsymbol{v}_h^i}{\omega} \norm{\nabla \boldsymbol{v}_h^i}{\omega} + k^2 \, \norm{\boldsymbol{v}_h^i}{\omega}^2. \end{split} \end{align} Similarly, the additional assumption $\boldsymbol{f} \in C^1([0,T];\LL{2}{\omega}) \cap C([0,T];\LL{3}{\omega})$ yields that \begin{eqnarray}\label{} |T_{\boldsymbol{f},1} + T_{\boldsymbol{f},2}| & \lesssim & k |\prod{\boldsymbol{f}(t_{i+1/2})-\boldsymbol{f}^{i+1}}{\boldsymbol{v}_h^i}_{\omega}| + k^2 \, \norm{\boldsymbol{v}_h^i}{\omega} \norm{\nabla \boldsymbol{v}_h^i}{\omega} + k^2 \, \norm{\boldsymbol{v}_h^i}{\omega}^2. \end{eqnarray} The combination of~\eqref{} with~\eqref{} and summation over $i = 0,\dots, j-1$ yields that \begin{align}\label{eq::proofthmenergyeq1} \begin{split} &\mathcal{E}_{\textrm{LLG}}(\boldsymbol{m}_{hk}^+(\tau)) - \mathcal{E}_{\textrm{LLG}}(\boldsymbol{m}_h^0) + \int_0^{t_j}\prod{W_{M(k)}(\lambda_{hk}^-)\boldsymbol{v}_{hk}^-}{\boldsymbol{v}_{hk}^-}_{\omega} \,\d{t} \\ & \,\,\, + \int_0^{t_j} \prod{\partial_t \boldsymbol{f}_k}{\boldsymbol{m}_{hk}^-}_{\omega}\,\d{t} - \int_0^{t_j} \prod{\boldsymbol{\Pi}_{hk}(\boldsymbol{v}_{hk}^-;\boldsymbol{m}_{hk}^-,\boldsymbol{m}_{hk}^=)}{\boldsymbol{v}_{hk}^-}_{\omega}\,\d{t} \\ &\lesssim k\int_0^{t_j}\norm{\boldsymbol{v}_{hk}^-}{\omega}^2\,\d{t} + k\int_0^{t_j}\norm{\boldsymbol{v}_{hk}^-}{\omega} \norm{\nabla\boldsymbol{v}_{hk}^-}{\omega}\,\d{t} \\ &\,\,\, + \int_0^{t_j}| \prod{\boldsymbol{\pi}_{hk}(\boldsymbol{v}_{hk}^-;\boldsymbol{m}_{hk}^-,\boldsymbol{m}_{hk}^=) - \boldsymbol{\pi}(\boldsymbol{m}_{hk}^-)}{\boldsymbol{v}_h^j}_{\omega} |\,\d{t} + \int_0^{t_j}|\prod{\overline{\boldsymbol{f}}_{k}-\boldsymbol{f}_k^+}{\boldsymbol{v}_{hk}^-}_{\omega}|\,\d{t}. \end{split} \end{align} From strong convergence in~\eqref{}, it follows that $\mathcal{E}_{\textrm{LLG}}(\boldsymbol{m}_{h}^0) \rightarrow \mathcal{E}_{\textrm{LLG}}(\boldsymbol{m}^0)$ as $h,k \rightarrow 0$. The first and second term on the right-hand side vanish as $h,k \rightarrow 0$ due to Lemma~\ref{}. Thanks to~\eqref{} and~\eqref{} with strong convergence, the last two terms on the right-hand side of~\eqref{eq::proofthmenergyeq1} vanish as $h,k \rightarrow 0$. Standard lower semicontinuity arguments for the remaining terms in~\eqref{eq::proofthmenergyeq1} conclude the proof. \end{proof} \section{Proof of Theorem~\ref{} (Numerical integration of ELLG)} \label{} \begin{proof}[Proof of Theorem~\ref{}] Note that the right-hand side of~\eqref{} can depend non-linearly on $\boldsymbol{v}_h^i$. As in the proof of Theorem~\ref{, we employ a fixpoint iteration, where $\big(\boldsymbol{\eta}_h^\ell,\boldsymbol{\nu}_h^\ell\big) \approx \big(\boldsymbol{v}_h^i,\midh{\boldsymbol{h}}{i}\big) \in \Kh{\boldsymbol{m}_h^i} \times \mathcal{X}_h$. To this end, let $\boldsymbol{\eta}_h^0:=0$ and $\boldsymbol{\nu}_h^0:=\boldsymbol{h}_h^i$. For $\ell\in\mathbb{N}_0$, let $\boldsymbol{\eta}_h^{\ell+1}\in \Kh{\boldsymbol{m}_h^i}$ satisfy, for all $\boldsymbol{\varphi}_h \in \Kh{\boldsymbol{m}_h^i}$, \begin{subequations}\label{} \begin{align}\label{} \begin{split} &\!\! \prod{W_{M(k)}(\lambda_h^i)\boldsymbol{\eta}_h^{\ell+1}}{\boldsymbol{\varphi}_h}_{\omega} + \prod{\boldsymbol{m}_h^i\times\boldsymbol{\eta}_h^{\ell+1}}{\boldsymbol{\varphi}_h}_{\omega} + \frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k))\prod{\nabla\boldsymbol{\eta}_h^{\ell+1}}{\nabla\boldsymbol{\varphi}_h}_{\omega} \\ &\!\! = -\ell_{\rm ex}^2\prod{\nabla\boldsymbol{m}_h^i}{\nabla\boldsymbol{\varphi}_h}_{\omega} + \prod{\pih{i}{\boldsymbol{\eta}_h^{\ell}}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{\varphi}_h}_{\omega} + \prod{\midh{\boldsymbol{f}}{i}}{\boldsymbol{\varphi}_h}_{\omega} \\ &\!\! \quad + \prod{\Pih{i}{\boldsymbol{\eta}_h^{\ell}}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{\varphi}_h}_{\omega} + 2 \boldsymbol{\Theta}^k_{i3} \prod{\boldsymbol{\nu}_h^\ell}{\boldsymbol{\varphi}_h}_{\omega} + \prod{( \boldsymbol{\Theta}^k_{i2}\!-\!\boldsymbol{\Theta}^k_{i3}) \boldsymbol{h}_h^i \!+\! \boldsymbol{\Theta}^k_{i1} \boldsymbol{h}_h^{i-1} }{\boldsymbol{\varphi}_h}_{\omega}. \end{split} \end{align} Moreover, let $\boldsymbol{\nu}_h^{\ell+1} \in \mathcal{X}_h$ satisfy, for all $\boldsymbol{\zeta}_h \in \mathcal{X}_h$, \begin{align}\label{} \begin{split} & \frac{2\mu_0}{k} \prod{\boldsymbol{\nu}_h^{\ell+1}}{\boldsymbol{\zeta}_h}_{\Omega} + \prod{\sigma^{-1} \nabla \times \boldsymbol{\nu}_h^{\ell+1}}{\nabla \times \boldsymbol{\zeta}_h}_{\Omega} \\ & \quad = - \frac{\mu_0}{k} \langle \, \boldsymbol{\mathcal{I}}_h \bigg( \, \frac{\boldsymbol{m}_h^i + k \boldsymbol{\eta}_h^{\ell+1}}{|\boldsymbol{m}_h^i + k \boldsymbol{\eta}_h^{\ell+1}|} \, \bigg),\boldsymbol{\zeta}_h \rangle_{\omega} + \frac{\mu_0}{k} \prod{\boldsymbol{m}_h^i}{\boldsymbol{\zeta}_h}_{\omega} + \frac{2\mu_0}{k} \prod{\boldsymbol{h}_h^i}{\boldsymbol{\zeta}_h}_{\Omega}, \end{split} \end{align} where $\boldsymbol{\mathcal{I}}_h:\boldsymbol{C}(\overline\omega)\to\boldsymbol{V}_{\!\!h}$ denotes the nodal interpolation operator. \end{subequations} Since $\rho(k) \geq 0$ and $W_{M(k)}(\cdot) > 0$, the bilinear forms on the left-hand sides of~\eqref{} are elliptic on $(\Kh{\boldsymbol{m}_h^i},\norm{\cdot}{\omega})$ and $(\mathcal{X}_h,\norm{\cdot}{\Omega})$, respectively. Since $\boldsymbol{\eta}_h^{\ell+1}$ is known for the computation of $\boldsymbol{\nu}_h^{\ell+1}$, the fixpoint iteration is thus well-defined. We subtract~\eqref{} for $(\boldsymbol{\eta}_h^\ell,\boldsymbol{\nu}_h^\ell)$ from~\eqref{} for $(\boldsymbol{\eta}_h^{\ell+1},\boldsymbol{\nu}_h^{\ell+1})$ and test with $\boldsymbol{\varphi}_h := \boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell} \in \Kh{\boldsymbol{m}_h^i}$ and $\boldsymbol{\zeta}_h := \boldsymbol{\nu}_h^{\ell+1} - \boldsymbol{\nu}_h^{\ell} \in \mathcal{X}_h$. For sufficiently small $k$, Lemma~\ref{ and~\eqref{} prove that $W_{M(k)}(\cdot) \ge \alpha/2$. With $\sup_{k} \max_{ij} \abs{\boldsymbol{\Theta}^k_{ij}} \leq C_{\boldsymbol{\Theta}} < \infty$ as well as~\eqref{} and~\eqref{}, we get as in the proof of Theorem~\ref{ that \begin{align}\label{} &\frac{\alpha}{2} \, \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega}^2 + \frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k)) \, \norm{\nabla(\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell})}{\omega}^2 \notag \\&\quad \stackrel{\eqref{} \big( C_{\mathrm{\bf \ppi}} + C_{\mathrm{\bf \Ppi}} \big) \, k \, \norm{\boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1}}{\omega} \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega} + C_{\mathrm{\bf \Ppi}} \, k \, \norm{\nabla(\boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1})}{\omega} \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega} \notag \\&\qquad + 2 C_{\boldsymbol{\Theta}} \norm{\boldsymbol{\nu}_h^{\ell} - \boldsymbol{\nu}_h^{\ell-1}}{\Omega} \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega} \quad \textrm{for all } \ell \in \mathbb{N}. \end{align} For all $\boldsymbol{\varphi}_h \in \boldsymbol{V}_{\!\!h}$, it holds that $\norm{\boldsymbol{\mathcal{I}}_h\boldsymbol{\varphi}_h}{\omega} \leq \sqrt{5}\,\norm{\boldsymbol{\varphi}_h}{\omega}$; see, e.g.,~\cite[Lemma 2.2.3]{goldenits2012}. Moreover, for all $x,y \in \mathbb{R}$ with $|x|, |y| \geq 1$, it holds that $| x / |x| - y / |y| | \leq |x - y|$. Since $|\boldsymbol{m}_h^i(\boldsymbol{z}_h) + k \boldsymbol{\eta}_h^{\ell}(\boldsymbol{z}_h)| \geq 1$ for all $\ell \in \mathbb{N}_0$ and for all nodes $\boldsymbol{z}_h \in \mathcal N_h|_{\omega}$, we get that \begin{eqnarray*} & & \hspace*{-1.5cm} \norm{\boldsymbol{\nu}_h^{\ell+1} - \boldsymbol{\nu}_h^{\ell}}{\Omega}^2 + \frac{k}{2\mu_0} \norm{\sigma^{-1/2}\nabla \times \big( \boldsymbol{\nu}_h^{\ell+1} - \boldsymbol{\nu}_h^{\ell} )}{\Omega}^2 \\ &\stackrel{\eqref{} & - \frac12 \, \bigg\langle \boldsymbol{\mathcal{I}}_h \bigg(\, \frac{\boldsymbol{m}_h^i + k \boldsymbol{\eta}_h^{\ell+1}}{|\boldsymbol{m}_h^i + k \boldsymbol{\eta}_h^{\ell+1}|} \,\bigg) - \boldsymbol{\mathcal{I}}_h \bigg(\, \frac{\boldsymbol{m}_h^i + k \boldsymbol{\eta}_h^{\ell}}{|\boldsymbol{m}_h^i + k \boldsymbol{\eta}_h^{\ell}|} \,\bigg) , \boldsymbol{\nu}_h^{\ell+1} - \boldsymbol{\nu}_h^{\ell} \bigg\rangle_{\omega} \notag \\ & \leq & \frac{\sqrt{5}k}{2} \, k \, \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell} }{\omega} \norm{\boldsymbol{\nu}_h^{\ell+1} - \boldsymbol{\nu}_h^{\ell}}{\Omega} \quad \textrm{for all } \ell \in \mathbb{N}_0. \end{eqnarray*} The latter equation yields that \begin{align}\label{} \norm{\boldsymbol{\nu}_h^{\ell+1} - \boldsymbol{\nu}_h^{\ell}}{\Omega} \leq \frac{\sqrt{5}}{2} \, k \, \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell} }{\omega} \quad \textrm{for all } \ell \in \mathbb{N}_0. \end{align} We add~\eqref{} and obtain that \begin{align}\label{} \begin{split} &\frac{\alpha}{2} \, \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega}^2 + \frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k)) \, \norm{\nabla(\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell})}{\omega}^2 + \norm{\boldsymbol{\nu}_h^{\ell+1} - \boldsymbol{\nu}_h^{\ell}}{\Omega}^2 \\&\quad \leq \big( C_{\mathrm{\bf \ppi}} + C_{\mathrm{\bf \Ppi}} + \sqrt{5} \, C_{\boldsymbol{\Theta}}\big) \, k \, \norm{\boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1}}{\omega} \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega} \\&\qquad + C_{\mathrm{\bf \Ppi}} \, k \, \norm{\nabla(\boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1})}{\omega} \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega} + \frac{5}{4} \, k^2 \, \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell} }{\omega}^2 \quad \textrm{for all } \ell \in \mathbb{N}. \end{split} \end{align} We equip the product space $\Kh{\boldsymbol{m}_h^i} \times \mathcal{X}_h$ with the norm $|\!|\!|(\boldsymbol{\eta}_h,\boldsymbol{\nu}_h)|\!|\!|^2 := (\alpha/4) \, \norm{\boldsymbol{\eta}_h}{\omega}^2 + (\ell_{\rm ex}^2/2) \, k \, (1+\rho(k)) \, \norm{\nabla\boldsymbol{\eta}_h}{\omega}^2 + \norm{\boldsymbol{\nu}_h}{\Omega}^2$. For $\delta>0$, the Young inequality then proves that \begin{align*} & \min\Big\{1 \, , \, 2 - \frac{1}{\alpha\delta} \, (C_{\mathrm{\bf \ppi}} + 2\,C_{\mathrm{\bf \Ppi}} + \sqrt{5} \, C_{\boldsymbol{\Theta}}) \, k - \frac{5}{\alpha}\, k^2 \Big\} \, |\!|\!| (\boldsymbol{\eta}_h^{\ell+1},\boldsymbol{\nu}_h^{\ell+1}) - (\boldsymbol{\eta}_h^{\ell},\boldsymbol{\nu}_h^\ell) |\!|\!|^2 \notag \\ &\quad\le \Big[\frac{\alpha}{2} - \frac{1}{4\delta} \, (C_{\mathrm{\bf \ppi}} + 2\,C_{\mathrm{\bf \Ppi}} + \sqrt{5} \, C_{\boldsymbol{\Theta}}) \, k - \frac{5}{4}\, k^2 \Big] \, \norm{\boldsymbol{\eta}_h^{\ell+1} - \boldsymbol{\eta}_h^{\ell}}{\omega}^2 \notag \\&\quad\qquad + \frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k)) \, \norm{\nabla \boldsymbol{\eta}_h^{\ell+1} - \nabla \boldsymbol{\eta}_h^{\ell}}{\omega}^2 + \norm{\boldsymbol{\nu}_h^{\ell+1} - \boldsymbol{\nu}_h^{\ell}}{\Omega}^2 \notag \\ &\quad\le \delta \, (C_{\mathrm{\bf \ppi}} + C_{\mathrm{\bf \Ppi}} + \sqrt{5} \, C_{\boldsymbol{\Theta}}) \,k \,\norm{\boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1}}{\omega}^2 + \delta \, C_{\mathrm{\bf \Ppi}} \, k \, \norm{\nabla(\boldsymbol{\eta}_h^{\ell} - \boldsymbol{\eta}_h^{\ell-1})}{\omega}^2 \notag \\ &\quad\le \delta \, \max\Big\{\frac{4}{\alpha} \, (C_{\mathrm{\bf \ppi}} + C_{\mathrm{\bf \Ppi}}+ \sqrt{5} \, C_{\boldsymbol{\Theta}}) \,k \, , \, 2 \, \frac{C_{\mathrm{\bf \Ppi}}}{\ell_{\rm ex}^2} \Big\} \, |\!|\!| (\boldsymbol{\eta}_h^{\ell},\boldsymbol{\nu}_h^{\ell}) - (\boldsymbol{\eta}_h^{\ell-1},\boldsymbol{\nu}_h^{\ell-1}) |\!|\!|^2 \quad \textrm{for all } \ell \in \mathbb{N}. \end{align*} For sufficiently small $\delta$ and $k$, the iteration is thus a contraction with respect to $|\!|\!|\cdot|\!|\!|$. The Banach fixpoint theorem yields existence and uniqueness of $(\boldsymbol{v}_h^i,\midh{\boldsymbol{h}}{i}) \in \Kh{\boldsymbol{m}_h^i} \times \mathcal{X}_h$ of~\eqref{}. With $\boldsymbol{h}_h^{i+1} := 2\midh{\boldsymbol{h}}{i} - \boldsymbol{h}_h^i \in \mathcal{X}_h$, \ $(\boldsymbol{v}_h^i,\boldsymbol{h}_h^{i+1}) \in \Kh{\boldsymbol{m}_h^i} \times \mathcal{X}_h$ is the unique solution of~\eqref{}; cf.\ Remark~\ref{ (v)}. The remainder of the proof follows as for Theorem~\ref{. \end{proof} \begin{lemma}\label{} Under the assumptions of Theorem~\ref{}, the following assertions {\rm (i)}--{\rm (iii)} are satisfied, if $k>0$ is sufficiently small. {\rm (i)} For all $i = 0,\dots,N-1$ it holds that \begin{align}\label{} \begin{split} & \frac{\ell_{\rm ex}^2}{2} \mathop{\mathrm{d}_t} \norm{\nabla\boldsymbol{m}_h^{i+1}}{\omega}^2 + \prod{W_{M(k)}(\lambda_h^i)\boldsymbol{v}_h^i}{\boldsymbol{v}_h^i}_{\omega} + \frac{\ell_{\rm ex}^2 }{2} \, k\rho(k) \,\norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 \\ & \quad + \frac12 \mathop{\mathrm{d}_t} \norm{\boldsymbol{h}_h^{i+1}}{\Omega}^2 + \frac{1}{\mu_0} \norm{\sigma^{-1/2} \nabla \times \midh{\boldsymbol{h}}{i}}{\Omega}^2 \\ & \le \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{v}_h^i}_{\omega} + \prod{\pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega} + \prod{\Pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega} \\ & \quad + \prod{\boldsymbol{v}_h^i - \dth{\boldsymbol{m}}{i}}{\midh{\boldsymbol{h}}{i}}_{\omega} + \prod{\boldsymbol{v}_h^i}{\gentimeh{\boldsymbol{h}}{i} - \midh{\boldsymbol{h}}{i}}_{\omega}. \end{split} \end{align} {\rm (ii)} For all $i = 0,\dots,N-1$, it holds that \begin{align}\label{} \mu_0 \norm{\dth{\boldsymbol{h}}{i}}{\Omega}^2 + \mathop{\mathrm{d}_t} \norm{\sigma^{-1/2} \nabla \times \boldsymbol{h}_h^{i+1}}{\Omega}^2 &\leq \mu_0 \norm{\dth{\boldsymbol{m}}{i}}{\omega}^2. \end{align} {\rm (iii)} For all $j = 0,\dots,N-1$, it holds that \begin{align}\label{} \begin{split} & \norm{\nabla \boldsymbol{m}_h^j}{\omega}^2 + k\sum_{i=0}^{j-1}\norm{\boldsymbol{v}_h^i}{\omega}^2 + k^2\rho(k)\sum_{i=0}^{j-1}\norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 \\ & \quad + \norm{\boldsymbol{h}_h^{j}}{\Omega}^2 + \norm{\sigma^{-1/2}\nabla \times \boldsymbol{h}_h^{j}}{\Omega}^2 + k \sum_{i=0}^{j-1} \norm{\dth{\boldsymbol{h}}{i}}{\Omega}^2 \le C, \end{split} \end{align} where $C>0$ depends only on $T$, $\omega$, $\Omega$, $\alpha$, $\mu_0$, $\sigma_0$, $\ell_{\rm ex}^2$, $\boldsymbol{f}$, $\boldsymbol{\pi}$, $\boldsymbol{\Pi}$, $\boldsymbol{m}^0$, $\boldsymbol{h}^0$, $C_{\boldsymbol{\Theta}}$, and $\kappa$. \end{lemma} \begin{proof} For the LLG-part~\eqref{}, we argue as in the proof of Lemma~\ref{ to see that \begin{align}\label{} & \frac{\ell_{\rm ex}^2}{2} \mathop{\mathrm{d}_t} \norm{\nabla\boldsymbol{m}_h^{i+1}}{\omega}^2 + \prod{W_{M(k)}(\lambda_h^i)\boldsymbol{v}_h^i}{\boldsymbol{v}_h^i}_{\omega} + \frac{\ell_{\rm ex}^2}{2} \, k\rho(k) \, \norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 \\& \notag \,\,\, \le \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{v}_h^i}_{\omega} \!+\! \prod{\pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega} \!+\! \prod{\Pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega} \!+\! \prod{\gentimeh{\boldsymbol{h}}{i}}{\boldsymbol{v}_h^i}_{\omega}. \end{align} Testing~\eqref{} with $\boldsymbol{\zeta}_h := - (1 / \mu_0) \, \midh{\boldsymbol{h}}{i}$, we obtain that \begin{eqnarray}\label{} \prod{\dth{\boldsymbol{m}}{i}}{\midh{\boldsymbol{h}}{i}}_{\omega} & \stackrel{\eqref{} & - \prod{\dth{\boldsymbol{h}}{i}}{\midh{\boldsymbol{h}}{i}}_{\Omega} - \frac{1}{\mu_0} \norm{\sigma^{-1/2} \nabla \times \midh{\boldsymbol{h}}{i}}{\Omega}^2 \notag \\ & = & - \frac12 \mathop{\mathrm{d}_t} \norm{\boldsymbol{h}_h^{i+1}}{\Omega}^2 - \frac{1}{\mu_0} \norm{\sigma^{-1/2} \nabla \times \midh{\boldsymbol{h}}{i}}{\Omega}^2. \end{eqnarray} Inserting $\gentimeh{\boldsymbol{h}}{i}$ and $\boldsymbol{v}_h^i$ in~\eqref{}, we are led to \begin{align}\label{} \begin{split} \prod{\boldsymbol{v}_h^i}{\gentimeh{\boldsymbol{h}}{i}}_{\omega} &= \prod{\boldsymbol{v}_h^i}{\gentimeh{\boldsymbol{h}}{i} - \midh{\boldsymbol{h}}{i}}_{\omega} + \prod{\boldsymbol{v}_h^i - \dth{\boldsymbol{m}}{i}}{\midh{\boldsymbol{h}}{i}}_{\omega} \\ & \quad -\frac12 \mathop{\mathrm{d}_t} \norm{\boldsymbol{h}_h^{i+1}}{\Omega}^2 - \frac{1}{\mu_0} \norm{\sigma^{-1/2} \nabla \times \midh{\boldsymbol{h}}{i}}{\Omega}^2. \end{split} \end{align} Adding~\eqref{} and~\eqref{}, we prove {\rm (i)}. To prove {\rm (ii)}, we test~\eqref{} with $\boldsymbol{\zeta}_h := \dth{\boldsymbol{h}}{i}$. With the Young inequality, we obtain that \begin{align}\label{} \mu_0 \norm{\dth{\boldsymbol{h}}{i}}{\Omega}^2 + \frac{1}{2} \mathop{\mathrm{d}_t} \norm{\sigma^{-1/2} \nabla \times \boldsymbol{h}_h^{i+1}}{\Omega}^{2} &\stackrel{\eqref{} - \mu_0 \prod{\dth{\boldsymbol{m}}{i}}{\dth{\boldsymbol{h}}{i}}_{\omega} \notag \\ & \stackrel{\phantom{\eqref{} \frac{\mu_0}{2} \norm{\dth{\boldsymbol{m}}{i}}{\omega}^2 + \frac{\mu_0}{2} \norm{\dth{\boldsymbol{h}}{i}}{\Omega}^2. \end{align} This proves {\rm (ii)}. To prove~{\rm (iii)}, we sum~\eqref{} over $i = 0,\dots,j-1$ and multiply with $k$. With $W_{M(k)}(\cdot) \geq \alpha/2 > 0$ for $k$ being sufficiently small, we obtain that \begin{align*} {\rm \chi}_j & := \frac{\ell_{\rm ex}^2}{2} \, \norm{ \nabla \boldsymbol{m}_h^{j} }{\omega}^2 + \frac{\alpha}{2} \, k \sum_{i=0}^{j-1} \norm{ \boldsymbol{v}_h^i }{\omega}^2 + \frac{\ell_{\rm ex}^2}{2} \, k^2\rho(k) \, \sum_{i=0}^{j-1}\norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 \notag \\ & \quad +\frac12 \, \norm{\boldsymbol{h}_h^{j}}{\Omega}^2 + \frac{k}{\mu_0} \sum_{i=0}^{j-1} \norm{\sigma^{-1/2} \nabla \times \midh{\boldsymbol{h}}{i}}{\Omega}^2 \notag \\ & \stackrel{\eqref{} \frac{\ell_{\rm ex}^2}{2} \, \norm{ \nabla \boldsymbol{m}_h^0 }{\omega}^2 + k \sum_{i=0}^{j-1} \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{v}_h^i}_{\omega} + k \sum_{i=0}^{j-1} \prod{\pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega} \notag \\ & \quad + k \sum_{i=0}^{j-1} \prod{\Pih{i}{\boldsymbol{v}_h^i}{\boldsymbol{m}_h^i}{\boldsymbol{m}_h^{i-1}}}{\boldsymbol{v}_h^i}_{\omega} + \frac12 \, \norm{\boldsymbol{h}_h^{0}}{\Omega}^2 \notag \\ & \quad + k \sum_{i=0}^{j-1} \prod{\boldsymbol{v}_h^i - \dth{\boldsymbol{m}}{i}}{\midh{\boldsymbol{h}}{i}}_{\omega} + k \sum_{i=0}^{j-1} \prod{\boldsymbol{v}_h^i}{\gentimeh{\boldsymbol{h}}{i} - \midh{\boldsymbol{h}}{i}}_{\omega}. \end{align*} Recalling from~\eqref{} that $\norm{\dth{\boldsymbol{m}}{i}}{\omega} \lesssim \norm{\boldsymbol{v}_h^i}{\omega}$, we proceed as in the proof of Lemma~\ref{ (ii)}. Together with~\eqref{} and~\eqref{}, the Young inequality proves that \begin{align}\label{} \begin{split} {\rm \chi}_j &\lesssim \norm{\nabla \boldsymbol{m}_h^0}{\omega}^2 + \norm{\boldsymbol{h}_h^0}{\Omega}^2 + \frac{k}{\delta} \sum_{i=0}^{j-1} \norm{\boldsymbol{f}(t_{i+1/2})}{\omega}^2 + \frac{k}{\delta} \sum_{i=0}^{j-1} \norm{\boldsymbol{m}_h^i}{\omega}^2 + \frac{k}{\delta} \sum_{i=0}^{j-1} \norm{ \nabla \boldsymbol{m}_h^{i}}{\omega}^2 \\&\quad + \frac{k}{\delta} \sum_{i=0}^{j} \norm{\boldsymbol{h}_h^{i}}{\omega}^2 + \Big(\delta + k + \frac{k}{\delta\rho(k)}\Big) \, k \, \sum_{i=0}^{j-1} \norm{ \boldsymbol{v}_h^i }{\omega}^2 + \delta k^2 \rho(k) \, \sum_{i=0}^{j-1} \norm{ \nabla \boldsymbol{v}_h^i }{\omega}^2. \end{split} \end{align} Since $k/\rho(k) \to 0$ as $k\to0$, we choose $\delta$ sufficiently small and can absorb $k\delta^{-1} \, \norm{\boldsymbol{h}_h^{j}}{\omega}^2 + k \, \sum_{i=0}^{j-1} \norm{\boldsymbol{v}_h^i}{\omega}^2 + \delta k^2 \rho(k) \, \sum_{i=0}^{j-1} \norm{ \nabla \boldsymbol{v}_h^i }{\omega}^2$ in $\chi_j$. Altogether, we arrive at \begin{align} {\rm \chi}_j \lesssim 1 + k \sum_{i=0}^{j-1} \norm{ \nabla \boldsymbol{m}_h^{i}}{\omega}^2 + k\sum_{i=0}^{j-1} \norm{\boldsymbol{h}_{h}^i}{\Omega}^2 \lesssim 1 + k \Sum{i=0}{j-1} {\rm \chi}_i . \notag \end{align} Arguing as for Lemma~\ref{ (ii)}, we get that ${\rm \chi}_j $ is uniformly bounded for all $j=1,\dots,N$. In order to bound the remaining terms from~\eqref{}, we sum~\eqref{} for $i=0,\dots,j-1$ and multiply by $k$. Recall from~\eqref{} that $\norm{\dth{\boldsymbol{m}}{i}}{\omega} \lesssim \norm{\boldsymbol{v}_h^i}{\omega}$. This yields that \begin{align \mu_0 k \sum_{i=0}^{j-1} \norm{\dth{\boldsymbol{h}}{i}}{\Omega}^2 &+ \norm{\sigma^{-1/2} \nabla \times \boldsymbol{h}_h^{j}}{\Omega}^{2} \stackrel{\eqref{} \mu_0 k \sum_{i=0}^{j-1} \norm{\dth{\boldsymbol{m}}{i}}{\omega}^2 + \norm{\sigma^{-1/2} \nabla \times \boldsymbol{h}_h^{0}}{\Omega}^{2} \notag \\ &\lesssim k \sum_{i=0}^{j-1} \norm{\boldsymbol{v}_h^i}{\omega}^2 + \norm{\sigma^{-1/2} \nabla \times \boldsymbol{h}_h^{0}}{\Omega}^{2} \stackrel{\eqref{} {\rm \chi}_j + \norm{\sigma^{-1/2} \nabla \times \boldsymbol{h}_h^{0}}{\Omega}^{2} . \notag \end{align} Altogether, this proves {\rm (iii)} and concludes the proof. \end{proof} \begin{lemma}\label{} Under the assumptions of Theorem \ref{} consider the postprocessed output~\eqref{} of Algorithm \ref{}. Then, there exist $\boldsymbol{m} \in \boldsymbol{H}{1}{\omega_T} \cap \L{\infty}{0,T;\boldsymbol{H}{1}{\omega}}$ and $\boldsymbol{h} \in H^1(0,T;\LL{2}{\Omega}) \cap \L{\infty}{0,T;\boldsymbol{H}({\rm curl},\Omega)}$ and a subsequence of each $\boldsymbol{m}_{hk}^{\star} \in \{ \boldsymbol{m}_{hk}^{=}, \boldsymbol{m}_{hk}^-, \boldsymbol{m}_{hk}^+, \linebreak \boldsymbol{m}_{hk} \}$, $\boldsymbol{h}_{hk}^{\star} \in \{ \boldsymbol{h}_{hk}^{=}, \boldsymbol{h}_{hk}^-, \boldsymbol{h}_{hk}^+, \boldsymbol{h}_{hk}, \boldsymbol{h}_{hk}^{\boldsymbol{\Theta}} \}$, and of $\boldsymbol{v}_{hk}^-$ such that the following convergence statements~\ref{} hold true simultaneously for the same subsequence as $h,k\to0$: \begin{enumerate}[label=\rm{(\roman*)}] \item \label{} $\boldsymbol{m}_{hk} \rightharpoonup \boldsymbol{m}$ in $\boldsymbol{H}{1}{\omega_T}$, \item \label{} $\boldsymbol{m}_{hk}^{\star} \stackrel{*}\rightharpoonup \boldsymbol{m}$ in $\L{\infty}{0,T;\boldsymbol{H}{1}{\omega}}$, \item \label{} $\boldsymbol{m}_{hk}^{\star} \rightharpoonup \boldsymbol{m}$ in $\L{2}{0,T;\boldsymbol{H}{1}{\omega}}$, \item \label{} $\boldsymbol{m}_{hk}^{\star} \rightarrow \boldsymbol{m}$ in $\LL{2}{\omega_T}$, \item \label{} $\boldsymbol{m}_{hk}^{\star} (t) \rightarrow \boldsymbol{m}(t)$ in $\LL{2}{\omega}$\ for $t \in (0,T)$ a.e., \item \label{} $\boldsymbol{m}_{hk}^{\star} \rightarrow \boldsymbol{m}$ pointwise\ a.e. in $\omega_T$, \item \label{} $\boldsymbol{v}_{hk}^- \rightharpoonup \partial_t \boldsymbol{m}$ in $\LL{2}{\omega_T}$, \item \label{} $k \, \nabla \boldsymbol{v}_{hk}^- \rightarrow 0$ in $\LL{2}{\omega_T}$, \item \label{} $\boldsymbol{h}_{hk} \rightharpoonup \boldsymbol{h}$ in $\boldsymbol{H}{1}{0,T;\LL{2}{\Omega}}$, \item \label{} $\boldsymbol{h}_{hk}^{\star} \rightharpoonup \boldsymbol{h}$ in $\LL{2}{\Omega_T}$, \item \label{} $\boldsymbol{h}_{hk}^{\star} \stackrel{*}\rightharpoonup \boldsymbol{h}$ in $\L{\infty}{0,T;\boldsymbol{H}({\rm curl},\Omega)}$, \item \label{} $\boldsymbol{h}_{hk}^{\star} \rightharpoonup \boldsymbol{h}$ in $\L{2}{0,T;\boldsymbol{H}({\rm curl},\Omega)}$, \item \label{} $\boldsymbol{h}_{hk}^{\star} - \boldsymbol{h}_{hk} \rightarrow 0$ in $\LL{2}{\Omega_T}$. \end{enumerate} \end{lemma} \begin{proof} Lemma~\ref{ yields uniform boundedness of $\boldsymbol{m}_{hk} \in \boldsymbol{H}{1}{\omega_T}$, $\boldsymbol{m}_{hk}^\star \in \L{\infty}{0,T;\linebreak\boldsymbol{H}{1}{\omega}}$, $\boldsymbol{h}_{hk} \in H^1(0,T;\LL{2}{\Omega})$, and $\boldsymbol{h}_{hk}^\star \in \L{\infty}{0,T;\boldsymbol{H}({\rm curl},\Omega)}$. The proofs of~\ref{} and~\ref{} follow as in~\cite{alouges2008a,akst2014} resp.~\cite{lt2013,lppt2015,bpp2015}. Finally,~\eqref{} and Lemma~\ref{ and prove that \begin{align}\label{} \norm{\boldsymbol{h}_{hk} - \boldsymbol{h}_{hk}^{\star} }{\Omega_T}^2 \stackrel{\eqref{} k^2 \sum_{i=0}^{N-1} \norm{\dth{\boldsymbol{h}}{i}}{\Omega}^2 \stackrel{\eqref{} k \,\to\, 0 \quad \textrm{as } h,k \rightarrow 0. \end{align} This verifies~\ref{} and concludes the proof. \end{proof} \begin{proof}[Proof of Theorem~\ref{ (ii)}] We prove that $(\boldsymbol{m},\boldsymbol{h})$ satisfies Definition~\ref{}. Definition~\ref{} follows as for LLG. Definition~\ref{} is an immediate consequence of Lemma~\ref{}. Definition~\ref{} follows as in~\cite{lt2013,lppt2015,bpp2015} from Lemma~\ref{}} and~{\ref{}}. It remains to verify Definition~\ref{}: To that end, adopt the notation of the proof of Theorem~\ref{. In addition, let $\boldsymbol{\mathcal{J}}_h:\Hcurl{\Omega} \rightarrow \mathcal{X}_h$ denote the interpolation operator for first-order N\'ed\'elec elements of second type; see~\cite{nedelec1986}. Let $\boldsymbol{\varphi}\in \boldsymbol{C}^{\infty}(\overline{\omega_T})$, $\boldsymbol{\zeta} \in \boldsymbol{C}^{\infty}(\overline{\Omega_T})$ and $t\in[0,T]$. For $t \in [t_i,t_{i+1})$ and $i=0,\dots,N-1$, we test~\eqref{} with $\boldsymbol{\mathcal{I}}_h(\boldsymbol{m}_h^i \times \boldsymbol{\varphi}(t)) \in \Kh{\boldsymbol{m}_h^i}$ and $\boldsymbol{\mathcal{J}}_h(\boldsymbol{\zeta}(t)) \in \mathcal{X}_h$ and integrate over $(0,T)$. With the definition of the postprocessed output~\eqref{}, we obtain that \begin{subequations}\label{} \begin{align}\label{} & I_{hk}^1 + I_{hk}^2 + \frac{\ell_{\rm ex}^2}{2} \, I_{hk}^3 \notag \\ &=: \int_0^T\prod{W_{M(k)}(\lambda_{hk}^-)\boldsymbol{v}_{hk}^-}{\boldsymbol{\mathcal{I}}_h(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} + \int_0^T\prod{\boldsymbol{m}_{hk}^-\times\boldsymbol{v}_{hk}^-}{\boldsymbol{\mathcal{I}}_h(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} \notag \\ & \quad +\frac{\ell_{\rm ex}^2}{2} \, k \, (1+\rho(k))\int_0^T\prod{\nabla\boldsymbol{v}_{hk}^-}{\nabla\boldsymbol{\mathcal{I}}_h(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} \notag \\ & \stackrel{\eqref{} -\ell_{\rm ex}^2\int_0^T\prod{\nabla\boldsymbol{m}_{hk}^-}{\nabla\boldsymbol{\mathcal{I}}_h(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} + \int_0^T\prod{\boldsymbol{\pi}_{hk}^-(\boldsymbol{v}_{hk}^-;\boldsymbol{m}_{hk}^-,\boldsymbol{m}_{hk}^=)}{\boldsymbol{\mathcal{I}}_h(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} \notag \\ & \quad + \int_0^T\prod{\overline{\boldsymbol{f}}_{k}}{\boldsymbol{\mathcal{I}}_h(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} + \int_0^T\prod{\boldsymbol{\Pi}_{hk}^-(\boldsymbol{v}_{hk}^-;\boldsymbol{m}_{hk}^-,\boldsymbol{m}_{hk}^=)}{\boldsymbol{\mathcal{I}}_h(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} \notag \\ & \quad + \int_0^T\prod{\boldsymbol{h}_{hk}^{\boldsymbol{\Theta}}}{\boldsymbol{\mathcal{I}}_h(\boldsymbol{m}_{hk}^-\times\boldsymbol{\varphi})}_{\omega} \d{t} =: - \ell_{\rm ex}^2 I_{hk}^4 + I_{hk}^5 + I_{hk}^6 + I_{hk}^7 + I_{hk}^8, \end{align} and \begin{eqnarray}\label{} - \mu_0 I_{hk}^9 &:= & -\mu_0 \Int{0}{T} \prod{\partial_t \boldsymbol{m}_{hk}}{\boldsymbol{\mathcal{J}}_h\boldsymbol{\zeta}}_{\omega} \d{t} \notag \\ & \stackrel{\eqref{} & \mu_0 \Int{0}{T} \prod{\partial_t \boldsymbol{h}_{hk}}{\boldsymbol{\mathcal{J}}_h\boldsymbol{\zeta}}_{\Omega} \d{t} + \Int{0}{T} \prod{\sigma^{-1} \nabla \times \overline{\boldsymbol{h}}_{hk}}{\nabla \times (\boldsymbol{\mathcal{J}}_h\boldsymbol{\zeta})}_{\Omega} \d{t} \notag \\ & =: & \mu_0 I_{hk}^{10} + I_{hk}^{11}. \end{eqnarray} \end{subequations} With Lemma~\ref{}, convergence of the integrals $I_{hk}^1,\dots,I_{hk}^7$ in~\eqref{} towards their continuous counterparts in the variational formulation~\eqref{} follows the lines of the proof of Theorem~\ref{ (ii)}. Thus, we only consider the integrals $I_{hk}^8,\dots,I_{hk}^{11}$ from~\eqref{}: As in~\cite{lt2013,lppt2015}, we get from Lemma~\ref{} and the convergence properties of $\boldsymbol{\mathcal{I}}_h$ and $\boldsymbol{\mathcal{J}}_h$ that \begin{align*} I_{hk}^{10} & \to \Int{0}{T} \prod{\partial_t \boldsymbol{h}}{\boldsymbol{\zeta}}_{\Omega} \d{t} \quad \textrm{as } h,k \rightarrow 0. \end{align*} For the remaining terms, Lemma~\ref{} and the convergence properties of $\boldsymbol{\mathcal{J}}_h$ yield that \begin{align*} I_{hk}^9 \to \Int{0}{T} \! \prod{\partial_t \boldsymbol{m}}{\boldsymbol{\zeta}}_{\omega} \d{t}, \qua I_{hk}^8 \to \int_0^T \! \prod{\boldsymbol{h}}{\boldsymbol{m}\times\boldsymbol{\varphi}}_{\omega} \d{t}, \quad I_{hk}^{11} \to \Int{0}{T} \! \prod{\sigma^{-1}\nabla \times \boldsymbol{h}}{ \nabla \times \boldsymbol{\zeta}}_{\Omega} \d{t} \end{align*} as $h,k\to0$. This concludes the proof. \end{proof} \begin{proof}[Proof of Theorem~\ref{ (iii)}] It remains to verify that $(\boldsymbol{m},\boldsymbol{h})$ from Lemma~\ref{} satisfies the energy estimate of Definition~\ref{}. To that end, let $\tau\in(0,T)$ be arbitrary and $j \in \{1,\dots,N\}$ such that $\tau \in [t_{j-1},t_{j})$. Adopt the notation of the proof of Theorem~\ref{. For any $i = 0,\dots,j-1$, Lemma~\ref{ (i)} shows that \begin{align* &\mathcal{E}_{\textrm{ELLG}}(\boldsymbol{m}_{hk}(t_{i+1}),\boldsymbol{h}_{hk}(t_{i+1}))-\mathcal{E}_{\textrm{ELLG}}(\boldsymbol{m}_{hk}(t_{i}),\boldsymbol{h}_{hk}(t_{i})) \notag \\ &\stackrel{\eqref{} \frac{\ell_{\rm ex}^2}{2} \, \norm{\nabla\boldsymbol{m}_h^{i+1}}{\omega}^2 - \frac{\ell_{\rm ex}^2}{2} \, \norm{\nabla\boldsymbol{m}_h^i}{\Omega}^2 + \frac{1}{2} \, \norm{\boldsymbol{h}_h^{i+1}}{\omega}^2 - \frac{1}{2} \, \norm{\boldsymbol{h}_h^i}{\Omega}^2 \notag \\ &\quad -\frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^{i+1})}{\boldsymbol{m}_h^{i+1}}_{\omega} + \frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^i)}{\boldsymbol{m}_h^i}_{\omega} - \prod{\boldsymbol{f}^{i+1}}{\boldsymbol{m}_h^{i+1}}_{\omega} + \prod{\boldsymbol{f}^i}{\boldsymbol{m}_h^i}_{\omega} \notag \\ &\stackrel{\phantom{\eqref{} - k \, \prod{W_{M(k)}(\boldsymbol{m}_h^i)\boldsymbol{v}_h^i}{\boldsymbol{v}_h^i}_{\omega} - \frac{\ell_{\rm ex}^2 }{2} \, k^2\rho(k) \, \norm{\nabla\boldsymbol{v}_h^i}{\omega}^2 + k \, \prod{\boldsymbol{\Pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1})}{\boldsymbol{v}_h^i}_{\omega} \notag \\ &\quad + k \, \prod{\boldsymbol{\pi}_h^i(\boldsymbol{v}_h^i;\boldsymbol{m}_h^i,\boldsymbol{m}_h^{i-1})}{\boldsymbol{v}_h^i}_{\omega} -\frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^{i+1})}{\boldsymbol{m}_h^{j+i}}_{\omega} + \frac{1}{2} \, \prod{\boldsymbol{\pi}(\boldsymbol{m}_h^i)}{\boldsymbol{m}_h^i}_{\omega} \notag \\ &\quad + k \, \prod{\boldsymbol{f}(t_{i+1/2})}{\boldsymbol{v}_h^i}_{\omega} \!- \prod{\boldsymbol{f}^{i+1}}{\boldsymbol{m}_h^{i+1}}_{\omega} + \prod{\boldsymbol{f}^i}{\boldsymbol{m}_h^i}_{\omega} + k \, \prod{\boldsymbol{v}_h^i}{\gentimeh{\boldsymbol{h}}{i} \!- \midh{\boldsymbol{h}}{i}}_{\omega} \\& \quad + k \, \prod{\boldsymbol{v}_h^i - \dth{\boldsymbol{m}}{i}}{\midh{\boldsymbol{h}}{i}}_{\omega} . \end{align*} With the stronger assumptions on $\boldsymbol{\pi}$ and $\boldsymbol{f}$ from~\eqref{}, we can follow the lines of the proof of Theorem~\ref{}. This leads to \begin{align}\label{} &\mathcal{E}_{\textrm{ELLG}}(\boldsymbol{m}_{hk}^+(\tau),\boldsymbol{h}_{hk}^+(\tau)) - \mathcal{E}_{\textrm{ELLG}}(\boldsymbol{m}_h^0,\boldsymbol{h}_{h}^0) + \int_0^{t_j}\prod{W_{M(k)}(\boldsymbol{m}_{hk}^-)\boldsymbol{v}_{hk}^-}{\boldsymbol{v}_{hk}^-}_{\omega}\,\d{t} \notag \\ & \quad + \int_0^{t_j} \prod{\partial_t \boldsymbol{f}_k}{\boldsymbol{m}_{hk}^-}_{\omega}\,\d{t} - \int_0^{t_j} \prod{\boldsymbol{\Pi}_{hk}(\boldsymbol{v}_{hk}^-;\boldsymbol{m}_{hk}^-,\boldsymbol{m}_{hk}^=)}{\boldsymbol{v}_{hk}^-}_{\omega}\,\d{t} \notag \\ &\lesssim k\int_0^{t_j}\norm{\boldsymbol{v}_{hk}^-}{\Omega}^2\,\d{t} + k\int_0^{t_j}\norm{\boldsymbol{v}_{hk}^-}{\Omega}\norm{\nabla\boldsymbol{v}_{hk}^-}{\Omega}\,\d{t} \notag \\ &\quad + \int_0^{t_j}| \prod{\boldsymbol{\pi}_{hk}(\boldsymbol{v}_{hk}^-;\boldsymbol{m}_{hk}^-,\boldsymbol{m}_{hk}^=) - \boldsymbol{\pi}(\boldsymbol{m}_{hk}^-)}{\boldsymbol{v}_{hk}^-}_{\omega} |\,\d{t} + \int_0^{t_j}|\prod{\overline{\boldsymbol{f}_{k}}-\boldsymbol{f}}{\boldsymbol{v}_{hk}^-}_{\omega}|\,\d{t} \notag \\ & \quad + \int_0^{t_j}|\prod{\boldsymbol{h}_{hk}^{\boldsymbol{\Theta}}-\overline{\boldsymbol{h}}_{hk}}{\boldsymbol{v}_{hk}^-}_{\omega}|\,\d{t} + \int_0^{t_j}|\prod{\overline{\boldsymbol{h}}_{hk}}{\boldsymbol{v}_{hk}^- - \partial_t \boldsymbol{m}_{hk}}_{\omega}|\,\d{t} . \end{align} The only crucial term is the last one on the right-hand side of~\eqref{}: Recall the Sobolev embedding $\boldsymbol{H}{1}{\omega} \subset \LL{4}{\omega}$. Together with Lemma~\ref{ (iii)}, this yields that \begin{align}\label{} \begin{split} &\int_0^{t_j}|\prod{\overline{\boldsymbol{h}}_{hk}}{\boldsymbol{v}_{hk}^- - \partial_t \boldsymbol{m}_{hk}}_{\omega}|\,\d{t} \le \norm{\overline{\boldsymbol{h}}_{hk}}{\L{\infty}{0,T;\LL{2}{\Omega}}} \norm{\boldsymbol{v}_{hk}^- - \partial_t \boldsymbol{m}_{hk}}{\L{1}{0,T;\LL{2}{\omega}}} \\& \qquad \stackrel{\eqref{ k \norm{\boldsymbol{v}_{hk}^-}{\L{2}{0,T;\LL{4}{\omega}}}^2 \lesssim k \norm{\boldsymbol{v}_{hk}^-}{\L{2}{0,T;\boldsymbol{H}{1}{\omega}}}^2 \lesssim k(1+ h^{-2}) \, \norm{\boldsymbol{v}_{hk}^-}{\LL{2}{\omega_T}}^2 \longrightarrow 0 \end{split} \end{align} as $h,k \rightarrow 0$, where we have used an inverse inequality and the assumption $k = \mathbf{o}(h^2)$. Arguing by lower semicontinuity, we conclude the proof as for Theorem~\ref{}. \end{proof} \section{Numerical experiments} \label{} This section provides some numerical experiments for Algorithm~\ref{} and Algorithm~\ref{}. Our implementation is based on the C++/Python library Netgen/NGSolve~\cite{ngsolve}. The computation of the stray field $\boldsymbol{h}_{\mathrm{s}} = -\nabla u$ requires the approximation of the magnetostatic potential $u \in H^1(\mathbb{R}^3)$, which solves the full space transmission problem \begin{subequations} \label{} \begin{alignat}{2} -\Delta u &= -\operatorname{div} \boldsymbol{m} &\quad& \textrm{in } \omega, \\ -\Delta u &= 0 && \textrm{in } \mathbb{R}^3\setminus\overline{\omega}, \\ u^{\mathrm{ext}} - u^{\mathrm{int}} &= 0 && \textrm{on } \partial \omega, \\ (\nabla u^{\mathrm{ext}} - \nabla u^{\mathrm{int}})\cdot{\boldsymbol{n}} &= -\boldsymbol{m}\cdot\boldsymbol{n} && \textrm{on } \partial \omega, \\ u(\boldsymbol{x}) &= \mathcal{O}(\vert \boldsymbol{x} \vert^{-1}) && \textrm{as } \vert \boldsymbol{x} \vert \to \infty. \end{alignat} \end{subequations} Here, the superscript \emph{ext} (resp.\ \emph{int}) refers to the traces of $u$ on $\partial \omega$ with respect to the exterior domain $\mathbb{R}^3 \setminus \overline{\omega}$ (resp.\ the interior domain $\omega$), and $\boldsymbol{n}$ is the outer normal vector on $\partial \omega$. Recall from~\cite{praetorius2004} that $\boldsymbol{\pi}(\boldsymbol{m}):=-\nabla u$ gives rise to a self-adjoint operator $\boldsymbol{\pi}\in L(\L{2}{\omega},\L{2}{\mathbb{R}^d})$ which satisfies the stronger stability assumption~\eqref{}. To discretize~\eqref{}, we employ the hybrid FEM-BEM method from~\cite{fk1990}. We note that the latter satisfies~\eqref{} with strong convergence in~\eqref{}; see~\cite[Section~4.4.1]{bffgpprs2014} or~\cite[Section~4.1]{prs2016} for details. This part of the code builds upon the open-source Galerkin boundary element library BEM++~\cite{sbaps2015}. The arising linear systems are solved with GMRES (resp.\ with CG for the hybrid FEM-BEM approach) with tolerance $10^{-12}$. The implicit first time-step of Algorithm~\ref{} is solved by the fixpoint iteration used in the proof of Theorem~\ref{} which is stopped if $\norm{\boldsymbol{\eta}_h^\ell - \boldsymbol{\eta}_h^{\ell-1}}{\LL{2}{\omega}} \le 10^{-10}$. \input{example01.tex} \input{example02.tex} \input{example03.tex} \bibliographystyle{alpha}

\section{Introduction}\label{introduction} Frustrated quantum antiferromagnets are believed to harbor exotic states of matter known as quantum spin liquids. In their original guise, they were envisioned by Anderson \cite{Anderson73} to retain the full space group symmetry of some underlying lattice {\em as well as} global $SU(2)$ spin-rotational symmetry in their ground state. This gave rise to a picture where ground states are thought of in terms of fluctuating valence bond configurations, or resonating valence bonds (RVBs) (see \cite{PALEE} for a review). While originally it proved challenging to stabilize this scenario in models of local microscopic interactions, the assumption alone that this is possible has lead to deep insights regarding the role of topology in certain situations. Specifically, in the short-range RVB scenario, where all singlet bonds are limited in length (and which we will always refer to simply as ``RVB'' in the following), a four-fold ground state degeneracy has been predicted for toroidal topology, in addition to the presence of fractionalized excitations with non-trivial mutual statistics \cite{ReadChakraborty89, Kivelson89}. These are the hallmarks characterizing what was coined ``topological order'' by Wen \cite{WenTopo}, and necessitate the presence of ``long-range entanglement'' (see \cite{misguich_book} and \cite{sachdev15} for reviews of these concepts in the present context). To construct exactly solvable models whose ground states have the same essential topological features as the RVB-wave functions, Rokhsar and Kivelson introduced so called quantum dimer models (QDMs) \cite{Kivelson89}. Their construction principle is based on two assumptions that simplify the setting greatly: 1. The essential physics of quantum antiferromagnets in the targeted phase is well-captured by states containing only near (usually near{\em est}) neighbor bonds (``dimers''). 2. It is still captured when the inner product of the original spin-1/2 problem is changed so as to render certain loop dynamics Hermitian, and at the same time, render different dimer configurations orthogonal. Especially the latter assumption must be regarded as highly non-trivial, as any two of the original valence-bond configurations will have non-zero overlap (though they have been shown to be linearly independent on many lattices \cite{CCK, seidel09, wildeboer11}). Over time, however, these two assumptions have passed much scrutiny, as we will review in the following. The main purpose of this paper is to test thus far uncharted territory of this dimer-RVB correspondence. Irrespective of their relations with spin-1/2 RVB states, the study of ground states of QDMs at exactly solvable points has brought to light rich and interesting phase diagrams. At such ``Rokhsar-Kivelson'' (RK) points, the ground states of quantum dimer models tend to be equal amplitude superpositions of all dimer configurations (with possible restrictions to topological sectors). Counter to intuition, the original construction on the square lattice {\em failed} to stabilize a gapped liquid, but led to a critical point between many symmetry broken phases \cite{sachdev89, leung96}. Similar findings were made for a QDM on the honeycomb lattice \cite{MoessnerSondhiChandra01}. Moessner and Sondhi correctly attributed this failure to produce a gapped topological phase to the bipartiteness of these lattices, and demonstrated that such a phase exists in a QDM on the triangular lattice \cite{MS}. The dichotomy between the phase diagrams of QDMs on bipartite and non-bipartite lattices can be understood in terms of powerful mappings between QDMs and lattice gauge theories \cite{MoessnerSondhiFradkin, Fradkin_book} (see also \cite{Ng} for a recent review), which are exact in some cases \cite{misguich}. A non-trivial question now arises when considering the RVB-counterparts of RK-type wave functions, that is, equal amplitude superpositions of all nearest neighbor valence bond coverings in some lattice topology. One may ask what properties of RVB RK-states are faithfully captured by the well-understood quantum dimer RK-states. For the latter, correlation functions are often analytically known in terms of Pfaffians \cite{wangwu}. On the other hand, for RVB RK-states on some {\em bipartite} lattices, pertinent studies have a long tradition \cite{doucout, alet, tang}, following a loop-gas Monte-Carlo method due to Sutherland \cite{sutherland}. These studies have confirmed the critical behavior of RVB RK-states respecting a bipartite structure \cite{alet, tang} first inferred from QDM RK-states, albeit with somewhat different critical exponents. In the non-bipartite case, defining physical properties of RVB RK-states have only been calculated more recently, owing to a sign-problem of the loop-gas method in this case. \begin{figure}[tbp] \includegraphics[width=0.28\textwidth]{red_stuff}\quad \quad \quad \quad \includegraphics[width=0.28\textwidth]{kas}\quad \quad \quad \quad \includegraphics[width=0.27\textwidth]{my_ori} \caption{ a) A dimer covering on the frustrated square lattice. Note that one plaquette is internally dimerized with a pair of crossed valence bonds. This covering appears in \eqref{state} with a negative amplitude. b) The ``pre-Kastelyn'' orientation of the lattice discussed in the main text. This orientation is used to in the Pfaffian re-formulation of the wave function \eqref{state}. Note that this artificially doubles the unit cell, but this is without physical consequence (see c)). c) The link orientation used to define the sign of valence bonds in \Eq{state}. Note that this orientation manifestly preserves the full translational symmetry of the square lattice, thus ensuring that the wave function \Eq{state} has this symmetry. } \label{figure_1} \end{figure} In Ref.\cite{wildeboer12}, we developed a Pfaffian presentation of some RVB RK-states on the triangular and kagome lattice, which allows sign-problem free Monte Carlo studies. This method was later used to calculate entanglement properties, and also constrained the modular $\mathcal{S}$ and $\mathcal{U}$ matrices \cite{wildeboer15}. In all, these studied have strongly corroborated to hypothesis that these wave functions are in the same $\mathbb{Z}_2$ topological phase as their QDM counterparts \cite{MS, misguich}. Historically, the square lattice has played an important role in motivating the search for spin liquid physics, owing to its relative experimental abundance, notably in the high-T$_c$ parent compounds. Indeed, the frustrated geometries of non-bipartite lattices, which have produced a number of exciting experimental candidates for quantum spin liquids \cite{Shimizu, Itou, Helton, Okamoto, Feng}, are not expected to be necessary to access these physics, as interactions beyond nearest neighbors can introduce frustration in any geometry. In this context, an interesting prediction of QDMs is that a topological $\mathbb{Z}_2$ liquid phase {\em can} be accessed at the critical RK-point of the square lattice QDM as soon as non-zero amplitudes are introduced \cite{paul} for next-nearest neighbor dimers, residing on the diagonals of the lattice (Fig. \ref{figure_1}a). {\em The primary purpose of this paper is to consider RVB-wave functions with diagonal bonds on the square lattice, which admit a Pfaffian presentation and are thus amenable to Monte-Carlo studies of its correlations \underline{and} entanglement properties.} We demonstrate that such wave functions indeed lie in the expected $\mathbb{Z}_2$ topological phase. In closing this section, we briefly remark on the status of parent Hamiltonians for the $SU(2)$-invariant RVB RK-states. Indeed, these states do not easily lend themselves to the construction of local parent Hamiltonians, fueling controversy for some time. The earliest successful stabilizations of $\mathbb{Z}_2$ topological phases have indeed done away with $SU(2)$-invariance, as evidenced by the non-bipartite QDMs, as well as the celebrated Kitaev toric code \cite{toric}, all of which lie in the phase described by deconfined (weak coupling) Ising gauge theory in three dimensions \cite{wegner}. By now, however, various costruction principles for RVB RK-states have been discussed. In the limit of highly decorated lattices, the construction of RVB-stabilizing Hamiltonians has been argued to be perturbatively under control \cite{RamanMoessnerSondhi}. In the following, we will focus on {\em simple} lattices. For the bipartite case, (critical) RVB RK-states have first been stabilized by Fujimoto \cite{fujimoto}, with some simplifications (and, in one case, corrections) later given by Cano and Fendley \cite{cano}. The kagome (topological) RVB RK-state referenced above is stabilized by the parent Hamiltonian constructed by one of us \cite{seidel09}. Subsequent work has, moreover, proven the uniqueness of the ground states of this Hamiltonian, modulo four-fold topological degeneracy on the torus \cite{schuch, zhou}. The approach of Ref. \cite{cano} may also be applied to certain non-bipartite lattices, where, however, for the ground state uniqueness one relies (thus far) on the ability of Klein models \cite{klein} to gap out (at least on finite lattices) non-nearest-neighbor valence bond states. This property of Klein models appears to be well established only for some bipartite lattices \cite{CCK}, does not hold for the kagome or triangular lattice, but is at least {\em expected} to hold for sufficient lattice decoration \cite{RamanMoessnerSondhi}. The remainder of this paper is organized as follows: In Sec. \ref{RVBdef}, we briefly introduce the RVB wave function that will be investigated in this paper. in Sec. \ref{MCmethod}, we review the Pfaffian Monte-Carlo method of Refs. \cite{wildeboer12, wildeboer15} in some detail, and set the stage for its application to a situation with a non-planar lattice graph. To this end, the notion of a ``pre-Kasteleyn'' orientation is introduced, and various statements are proven. We then present the results of the application of these methods to the calculation of correlation functions and Renyi entropy, confirming the wave function's low energy effective theory to be the $\mathbb{Z}_2$ topological phase. In Sec. \ref{QDMSEC} we define a quantum dimer model related to our RVB wave function. We summarize and conclude in Sec. \ref{conclusion}. An appendix elaborates on some defining properties of Kasteleyn-like orientations. \section{The wave function for the nonbipartite square lattice \label{RVBdef}} In this work, we aim to analyze the correlation functions and the Renyi entanglement entropy of an RVB state defined on a square lattice topology including diagonal links, as shown in Fig. \ref{figure_1}a). We will refer to this topology as a nonbipartite and/or frustrated square lattice. The RVB wave function to be considered is defined via \begin{eqnarray} |{\sf RVB} \rangle = \sum_{D}(-1)^{n_{c}(D)}|D\rangle\;. \label{state} \end{eqnarray} Here, $D$ goes over all possible ``dimerizations'' or dimer coverings of the lattice, i.e., pairings of the lattice into neighboring pairs along the links of the non-bipartite square topology. A typical dimer covering is shown in Figure \ref{figure_1}(a). Each lattice site is equipped with a spin-$1/2$ degree of freedom. For each dimer covering $D$, $|D\rangle$ denotes a state where each dimer of the covering $D$ is realized by a singlet between the associated spins, and a sign convention is used that corresponds to an orientation of links as shown in Figure \ref{figure_1}(c). We point out that our lattice topology does not correspond to a planar graph. This precludes using methods of Kasteleyn \cite{kasteleyn} in their original form, which were quite essential in our Pfaffian presentation of RVB wave functions on the kagome and triangular lattices \cite{wildeboer12}. However, we find that a generalization of these methods is possible if we furnish the definition of the RVB-state with certain additional phases as shown in \Eq{state}. Here, the parameter $n_{c}(D)$ in \eqref{state} counts the number of crossings between two dimers that a particular dimer covering contains. Thus, the dimer coverings with an even number of crossed pair of dimers, $n_{c} = 0, 2, \ldots$ are equipped with a positive amplitude, whereas an odd number of crossing, $n_{c} = 1, 3, \ldots$, in a dimer covering leads to a negative amplitude. We emphasize that despite the need to introduce additional phases owing to the departure from planar graph topology, the rule that determines these phases is still {\em local}, and can, as we will show, in particular be obtained from a quantum dimer model with local interactions. We may therefore regard \Eq{state} as a natural variation of the RVB RK-state. We will now proceed to show in detail how the Pfaffian Monte-Carlo scheme of Refs. \cite{wildeboer12, wildeboer15} can successfully be applied to this wave function \Eq{state}. \section{Pfaffian Monte Carlo and Correlations and Entanglement Entropy\label{MCmethod}} In order to make this work self-contained, we will now briefly review the Pfaffian Monte Carlo technique. The standard method for calculating expectation values, e.g., for the product of two local operators ${\cal O}_1 {\cal O}_2$, for {\em bipartite} lattices is due to Sutherland \cite{sutherland}. It is based on the observation matrix elements of the form $\langle {\sf RVB}|{\cal O}_1 {\cal O}_2|{\sf RVB}\rangle$ will, quite generally, only depend on the configuration of close-packed non-intersecting loops associated to the overlap graph between $D$ and $D'$ (that is, basically, the union of the sets of all dimers in $D$ and $D'$, which can always be regarded as the disjoint union of close-packed non-intersecting loops). This maps the correlator \begin{eqnarray}\label{corr} \cfrac{\langle {\sf RVB}|{\cal O}_1 {\cal O}_2|{\sf RVB}\rangle}{\langle {\sf RVB}|{\sf RVB}\rangle} =\cfrac{\sum_{D,D'}\langle D|{\cal O}_1 {\cal O}_2|D'\rangle}{\sum_{D,D'}\langle D|D'\rangle}\,. \end{eqnarray} onto a {\em classical} loop gas problem, provided that all inner products $\langle D|D'\rangle$ are non-negative. In contrast, this is in general not the case for non-bipartite lattices: Typically, no sign convention for the states $\ket{D}$ exists that has this property. (Note that it is generally sufficient to identify three dimer covering on the lattice of interest for which the no such convention exists.) Any attempt to solve the problem by Monte-Carlo evaluation of \Eq{corr} will then be plagued by the sign problem. The only way to circumvent such a sign problem is abandon the description in term of the $\ket{D}$-states and instead move to a different expansion of the RVB-state in terms of states with non-negative mutual overlaps, while, {\em at the same time}, the coefficients in this expansion can be efficiently calculated. In particular, moving to any orthogonal basis precludes the appearance of negative overlaps, and moves all complexity to the problem of calculating coefficients. This is the route explored in Refs.\cite{wildeboer12,wildeboer15} for kagome and triangular lattice RVB RK-states, and will be established here for the ``non-planar'' square lattice RVB \Eq{state}. To this end, we consider the Ising-basis of local $S_z$-eigenstates $\ket{I}$ and write $\ket{{\sf RVB}}= \sum_I a_I \ket{I}$, the sum running over all Ising configurations. We may then write \begin{eqnarray} \label{corr2} \frac{\langle {\sf RVB}| {\cal O}_i {\cal O}_j \ket{\sf RVB}}{\bra{\sf RVB}{\sf RVB} \rangle} &=& \frac{\sum_I \sum_{I'} a_{I} a_{I'} \bra{I'} {\cal O}_i{\cal O}_j \ket{I}}{\sum_{I} |a_{I}|^2} \nonumber \\ &=& \frac{\sum_{I} |a_{I}|^2 \sum_{I'} \frac{a_{I'}}{a_{I}} \bra{I'} {\cal O}_i {\cal O}_j \ket{I} }{\sum_{I} |a_{I}|^2}\,. \nonumber \\ & & \end{eqnarray} This may now be interpreted as the classical expectation value of a quantity $f$: \begin{equation} \langle f \rangle = {\sum_{I} f_{I} e^{-E_I}}/{\sum_{I} e^{-E_I}}\;, \end{equation} where $e^{-E_I} = |a_I|^2$ and the value $f_I$ of the quantity $f$ in the Ising configuration $I$ is given by $f_{I} = \sum_{I'} \bra{I'}{\cal O}_i {\cal O}_j \ket{I} \frac{a_{I'}}{a_I}$. It turns out that for most local operators, the sum over $I'$ only contains a few non-zero terms. The problem is thus properly set up for Monte-Carlo evaluation, as long as the coefficients $a_I$ can be calculated. It turns out that for the choice of phases made in \Eq{state} this is the case. For reference purposes, it is a good idea to to first consider the ``plain vanilla'' RVB state, which, similar to the original Rokhsar-Kivelson quantum dimer states, has all amplitudes set equal to one: \begin{eqnarray} |{\sf RK} \rangle = \sum_{D}|D\rangle =\sum_I b_I\ket{I}\;. \label{RK} \end{eqnarray} Note that for any lattice topology {\em without} crossings, there is no difference between Eqs. \eqref{RK} and \eqref{state}. For any oriented lattice graph, the Ising-coefficients $b_I$ of $\ket{\sf RK}$ can naturally be written as a ``Hafnian'': \begin{eqnarray}\label{bI} b_{I} = \mbox{Haf}[M_{ij}(I)] \equiv\frac{1}{2^{N/2}(\frac{N}{2}!)} \sum_{\lambda \in S_N} M_{\lambda_1 \lambda_2}(I) M_{\lambda_3 \lambda_4}(I) \times \cdots \times M_{\lambda_{N-1} \lambda_N}(I) \,. \end{eqnarray} Here, $M$ is a symmetric matrix whose indices run over the $N$ lattice sites and which depends on the Ising configuration via $M_{ij}(I)= \Theta_{ij}( \delta_{\sigma_{i},\uparrow} \delta_{\sigma_{j},\downarrow} - \delta_{\sigma_{i},\downarrow} \delta_{\sigma_{j},\uparrow})$, where the $\sigma_{i}$ describe the Ising configuration $I$, $\Theta_{ij}$ encodes the orientation of links connecting nearest neighbors (NN) in the lattice topology via \begin{equation} \Theta_{ij}= \begin{cases} 1 & i<j \\ -1 & j<i \\ 0 & i,j\;\; \mbox{not NN} \end{cases} \end{equation} and ``$i<j$'', by abuse of notation, refers to a nearest neighbor pair $(i,j)$ ordered in accordance with the orientation. For the frustrated square lattice, we will consider the orientation indicated by the arrows in Figure \ref{figure_1}(c). $\lambda$ runs over all permutations of the $N$ sites. \begin{figure}[tbp] \includegraphics[width=0.16\textwidth]{bowtie_hor2}\quad \quad \quad \quad \includegraphics[width=0.16\textwidth]{bowtie_ver2}\quad \quad \quad \quad \includegraphics[width=0.16\textwidth]{bowtie_hor2_B}\quad \quad \quad \quad \includegraphics[width=0.16\textwidth]{bowtie_ver2_B} \caption{ The possible configurations for bowtie loops on the frustrated square lattice. Four configurations arise because of the doubling of the unit cell by the pre-Kasteleyn orientation (shown). One observes that traversal of any of these loops in any direction always leads to an even number of arrows parallel or anti-parallel to the traversal. } \label{bowtie} \end{figure} The correspondence between \Eq{RK} and \Eq{bI} is easy to see once one realizes that any $\lambda$ giving a non-zero contribution to \Eq{bI} naturally defines a dimerizarion $D$, as all $(\lambda_{2n-1},\lambda_{2n})$ must be nearest neighbor pairs, and the product of matrix elements in \Eq{bI} is then just the contribution of $\ket{D}$ to the Ising configuration $I$. The overall combinatorial factor in the equation compensates for a many-to-one correspondence between permutations and dimerizations. Note that the formal definition of the Hafnian is related to that of the Pfaffian through the absence/presence of the sign factor $(-1)^\lambda$, in a manner similar to how the permanent is related to the determinant. While the determinant or Pfaffian can be computed efficiently in polynomial time, it is not known how to do this for the permanent or Hafnian. \Eq{bI} therefore does not lend itself to a Monte-Carlo scheme as discussed above. To make progress, let's for the time being {\em define} an RVB-state of the form \begin{equation} \begin{split}\label{RVBp} \ket{\sf RVB}'&= \sum_I a_I' \ket{I}\\ a_I'&= \mbox{Pfaff} [M_{ij}(I)\tilde\Theta_{ij}]=\frac{1}{2^{N/2}(\frac{N}{2}!)}\sum_{\lambda \in S_N} (-1)^\lambda \,\tilde \Theta _{\lambda_1 \lambda_2} \times\dotsc\times \tilde \Theta _{\lambda_{N-1} \lambda_N}\, M_{\lambda_1 \lambda_2}(I) \times \cdots \times M_{\lambda_{N-1} \lambda_N}(I) \end{split} \end{equation} Here, $\tilde \Theta_{ij}$ is an orientation matrix describing yet another orientation of the lattice (usually different from $\Theta_{ij}$). Ordinarily, one requires $\tilde \Theta_{ij}$ to be a so-called ``Kasteleyn orientation'', which always exists for a planar lattice graph \cite{kasteleyn}. Presently, however, we will also need to admit somewhat more general orientations $\tilde \Theta_{ij}$, e.g., for the frustrated square lattice as shown in Fig. \ref{figure_1}b). The state \eqref{RVBp} is related to the $\ket{\sf RK}$-state in \Eq{RK} via the additional phases $q(\lambda)\equiv q(D)= (-1)^\lambda \,\tilde \Theta _{\lambda_1 \lambda_2} \times\dotsc\times \tilde \Theta _{\lambda_{N-1} \lambda_N}$, which, as on may easily see, only depend on the dimerization $D$ associated to $\lambda$. To better understand those phase, one may note that $q(D)$ and $q(D')$ are simply related whenever $D$ and $D'$ are related to each other by a single ``resonance move'' along a contractable closed loop of even length. That is, there exists such a loop, whose links alternatingly belong to $D$ and $D'$, respectively, while all dimers not along the loop are the same for $D$ and $D'$. Our definition of a loop also entails that each link may be traversed only once. \begin{figure}[tbp] \includegraphics[width=0.3\textwidth]{selfinta}\quad \quad \quad \quad \quad \quad \quad \quad \quad \includegraphics[width=0.3\textwidth]{selfintb} \caption{ Illustration of Lemma 3. The orientation $\tilde\Theta$ entering the Pfaffian formulation is shown. Only those dimers are shown that belong to a certain self-intersecting loop in the overlap of coverings $D$ (red) and $D'$ (blue). These dimers form a self-intersecting loop with both $D$ and $D'$ contributing to the cross. a) A cut with subsequent re-gluing is indicated (dashed line) that splits the single loop into two closed loops with the crossing eliminated. These two loops do not yet correspond to legitimate resonance moves, as the diagonal links of the cross are ``broken'' apart. b) A deformation of the two separate loops defined in a) that leads to valid resonance moves, which have the same effect as the single self-intersecting resonance move defined by the original dimer loop. Loops are indicated by pink and green lines, respectively. Note that these loops share one link (dashed, west of the cross), which is unoccupied for both $D$ and $D'$. When transforming $D$ into $D'$ (red into blue), the smaller, length 4 resonance move must be carried out first (green), followed by the other move (pink). } \label{selfcross} \end{figure} On any finite lattice, it makes sense to define a ``topological sector'' of dimerizations as an (equivalence) class of dimerizations related to each other by sequences of resonance moves along arbitrary contractible loops. We will then find that all phases $q(D)$ are simply related for dimerizations belonging to the same topological sector. The key to these results is the following property, somewhat weaker but more general than that of a true Kasteleyn orientation: {\underline{Definition:} \em We refer to a two-dimensional lattice endowed with an oriented link topology as ``pre-Kasteleyn'' if it has the following property: For any closed, non-intersecting, contractible loop along edges (links), the parity of the number of clockwise oriented edges is opposite to the parity of the number of sites enclosed by the loop. } Note that being pre-Kasteleyn is not an intrinsically ``graph theoretical'' property, since notions such as the ``enclosed sites'' will depend on the particular embedding of the graph into the plane or the torus\footnote{Even for a planar graph, these notions become meaningful only once a designated ``outer face'', or boundary, is chosen.}. (So do the notions of ``crossing'' links and ``contractible''.) That is why it is preferable in this context to think of the structure in question as a ``lattice with a topology'' rather than an abstract graph, which is in any case natural in the present physical situation: The term lattice will always imply that there is meaning to the positions of the lattice sites and links on some two-dimensional manifold (plane or torus). In this spirit, we will speak of a ``lattice with a planar topology'' if after the embedding of the lattice graph into the two-dimensional manifold, there are no crossings between different links. In particular, this does \underline{not} preclude toroidal topology. If, now, the lattice is equipped with a planar topology, the notion of pre-Kasteleyn defined above becomes equivalent to Kasteleyn, which we define as usual via {\underline{Definition:} \em A lattice with an oriented planar topology is ``Kasteleyn'' if the number of clockwise edges around each face is odd.} We establish the equivalence of these notions for planar lattice graphs in Appendix \ref{app}. The notion of ``pre-Kasteleyn'' as defined above is thus a very natural generalization of Kasteleyn-oriented lattices in the planar case. For all our purposes, the essential difference between pre-Kasteleyn and ``true'' Kasteleyn lies only in the fact that, on any planar lattice graph, any loop that can define a legitimate resonance move for any dimerization of the lattice necessarily encloses an even number of sites (all the enclosed sites form dimers with one another). This is not so in the general pre-Kasteleyn situation. The key observation for the remainder of this section is that the frustrated square lattice with the orientation $\tilde \Theta$ of Fig. \ref{figure_1}b) is pre-Kasteleyn. This, too, is proven in the Appendix, where we also discuss a simple general criterion. The following discussion will focus on the frustrated square lattice with the orientation of Fig. \ref{figure_1}b). Results simplify in obvious ways for the usual Kasteleyn-orientated planar lattice graphs, and are also expected to generalize readily to other situations with pre-Kasteleyn orientations. Obvious generalizations include the checkerboard lattice graph obtained by omission of diagonal links on every second plaquette. For the problem at hand, we now intend to show that for $D$, $D'$ in the same topological sector, \begin{equation} \label{qphase} \frac{q(D)}{q(D')} = (-1)^{n_c(D)-n_c(D')}\;. \end{equation} This is equivalent to saying that \Eq{RVBp} equals \Eq{state} up to a sign that may depend on the topological sector. We will not care about relative phases between different topological sectors, as local operators would be oblivious to such phases \cite{misguich} We demonstrate \Eq{qphase} via three small lemmas: {\em Lemma 1:} \Eq{qphase} holds if $D$ and $D'$ are related by a single resonance move along a loop that is not self-intersecting. Proof: For a single loop, we may choose $\lambda$ and $\lambda'$ associated to $D$, $D'$ such that $\lambda'=\mu \circ\lambda$, where $\mu$ is a cyclic permutation of lattice sites along the loop. Then it is easy to see that $ \frac{q(D)}{q(D')}=- (-1)^{N_\circlearrowright}$, where $N_\circlearrowright$ is the number of clockwise arrows encountered for the $\tilde \Theta$ orientation along the loop (or counter-clockwise, note that resonance loops are necessarily even in lengh). By the defining property of pre-Kasteleyn lattice graphs (cf. Appendix \ref{app}), $N_\circlearrowright=1+N_{\sf in} \mod 2$, where $N_{\sf in}$ is the number of sites enclosed by the loop. Moreover, $N_{\sf in}=N_{\sf x} \mod 2$, where $N_{\sf x}$ is the number of dimers in both $D$ and $D'$ that are crossing the loop. (Any such dimer is not part of the loop, which is free of self-intersections. Therefore, since no link of the frustrated square lattice crosses more than one other link, any such dimer has precisely one of its sites enclosed by the loop.) Finally, $N_{\sf x}=N_{\sf x}(D)+N_{\sf x}(D')$, where $N_{\sf x}(D)$ and $N_{\sf x}(D')$ are the crossed dimers {\em belonging to the loop} associated with $D$ and $D'$, respectively. Modulo 2, this is the same as $N_{\sf x}(D)-N_{\sf x}(D')=n_c(D)-n_c(D')$. This proves Lemma 1. \begin{figure}[tbp] \includegraphics[width=0.666\textwidth]{dimer_sectors} \caption{Expectation values for various local singlet operators for different lattice sizes $L\times L$. Shows are values for each of the projections of the wave function \eqref{state} onto the four topological sectors. For the smallest, $L=4$ lattice (circles), strong discrepancies between sectors are visible for each of the local bond operators (horizontal, vertical, and the two types of diagonal bonds.) These discrepancies decrease rapidly with system size (triangles and squares for $L=8,12$, respectively). Dotted, dashed, and solid lines indicate the averages for the different sectors for the different respective lattice sizes. Note that the link orientation of Fig. \ref{figure_1}c) does not preserve $\pi/2$ degree rotational symmetry. This is clearly shown by the difference between vertical and horizontal bonds. Remaining mirror symmetries relate the different diagonal links, however. This, too, is clearly seen at all system sizes. } \label{average_heisenberg} \end{figure} \begin{table}[t] \begin{tabular}{ | c | c || r | r | r | r |} \hline $L$ & $sec$& $\langle \vec{S}_{i} \vec{S}_{i+1}\rangle_{hor}$ & $\langle \vec{S}_{i} \vec{S}_{i+1}\rangle_{ver}$ & $\langle \vec{S}_{i} \vec{S}_{i+1}\rangle_{dia_{14}}$ & $\langle \vec{S}_{i} \vec{S}_{i+1}\rangle_{dia_{23}}$ \\ \hline \hline 4 & $(e,e)$ & -0.06718 $\pm$ 0.00019 & -0.20628 $\pm$ 0.00017 & -0.14045 $\pm$ 0.00015 & -0.14066 $\pm$ 0.00015 \\ \hline 4 & $(o,e)$ & -0.12020 $\pm$ 0.00011 & -0.17439 $\pm$ 0.00011 & -0.11156 $\pm$ 0.00009 & -0.11142 $\pm$ 0.00013 \\ \hline 4 & $(e,o)$ & -0.14849 $\pm$ 0.00015 & -0.10516 $\pm$ 0.00011 & -0.13851 $\pm$ 0.00011 & -0.13849 $\pm$ 0.00016 \\ \hline 4 & $(o,o)$ & -0.11855 $\pm$ 0.00013 & -0.14896 $\pm$ 0.00012 & -0.12456 $\pm$ 0.00012 & -0.12459 $\pm$ 0.00015 \\ \hline \hline 8 & $(e,e)$ & -0.05506 $\pm$ 0.00118 & -0.23239 $\pm$ 0.00103 & -0.08543 $\pm$ 0.00087 & -0.08517 $\pm$ 0.00066 \\ \hline 8 & $(o,e)$ & -0.04509 $\pm$ 0.00071 & -0.23987 $\pm$ 0.00082 & -0.08701 $\pm$ 0.00056 & -0.08645 $\pm$ 0.00062 \\ \hline 8 & $(e,o)$ & -0.06617 $\pm$ 0.00064 & -0.22224 $\pm$ 0.00072 & -0.08202 $\pm$ 0.00055 & -0.08219 $\pm$ 0.00057 \\ \hline 8 & $(o,o)$ & -0.05168 $\pm$ 0.00088 & -0.22849 $\pm$ 0.00119 & -0.08523 $\pm$ 0.00056 & -0.08471 $\pm$ 0.00062 \\ \hline \hline 12 & $(e,e)$ & -0.05482 $\pm$ 0.00541 & -0.23911 $\pm$ 0.00633 & -0.08569 $\pm$ 0.00481 & -0.08308 $\pm$ 0.00224 \\ \hline 12 & $(o,e)$ & -0.05151 $\pm$ 0.00259 & -0.23516 $\pm$ 0.00359 & -0.08559 $\pm$ 0.00234 & -0.08334 $\pm$ 0.00233 \\ \hline 12 & $(e,o)$ & -0.05547 $\pm$ 0.00245 & -0.23593 $\pm$ 0.00323 & -0.08387 $\pm$ 0.00234 & -0.08284 $\pm$ 0.00251 \\ \hline 12 & $(o,o)$ & -0.05405 $\pm$ 0.00401 & -0.23120 $\pm$ 0.00531 & -0.08221 $\pm$ 0.00352 & -0.08689 $\pm$ 0.00394 \\ \hline \end{tabular} \caption{Raw data of local operator expectation values in Fig. \ref{average_heisenberg}. It is seen that these local operators cannot distinguish different topological sectors at sufficient system sizes. } \label{} \end{table} {\em Lemma 2:} \Eq{qphase} holds if $D$ and $D'$ are related by a single ``bowtie'' resonance move, i.e., the shortest (length 4) self-intersecting resonance move taking two parallel dimers into a cross and vice versa (Fig. \ref{bowtie}). Proof: Following the logic of the proof of Lemma 1, it is now to be shown that if we traverse the bowtie in either direction, we encounter an even number of arrows that are oriented opposite to the direction of the traversal. One may easily check directly that the $\tilde \Theta$-orientation has this property, as shown in Fig. \ref{bowtie}). {\em Lemma 3:} Any two dimer coverings $D$ and $D'$ that are each without crosses can be related by a sequence of resonance moves along non-self-intersecting loops. Proof: Note that the loops forming the overlap graph between $D$ and $D'$ may still intersect, and each individual loop may be self-intersecting. However, unlike in the situation of lemma 2, at each cross exactly one dimer belongs to $D$ while the other belongs to $D'$. The situation is illustrated in Fig. \ref{selfcross}. If we follow the branch emerging from the South-West corner of the crossing, the next corner of the crossing we arrive at cannot be North-East. For otherwise, the crossing would be between two separate loops. The figure assumes (without loss of generality) that starting South-West, we first arrive again at the South-East corner. Then, a horizontal cut and re-gluing as shown in Fig. \ref{selfcross}a) will split the loop into two separate loops, which are, of course, not corresponding to legitimate resonance moves yet (a corner occurs at the center of a plaquette). However, the two separate loops in Fig. \ref{selfcross}b) correspond to two legitimate resonance moves, which, when carried out in the right order, reproduce the effect of a resonance move along the original single loop (see caption). The two new loops have in total one less self-intersection than the original loop. One may thus proceed inductively. Together, these three lemmas prove \Eq{qphase}, as one can of course connect any two dimer patterns $D$ and $D'$ first to ones without crosses via bowtie moves, and can then connect the resulting patterns via non-self-intersecting resonance moves because of Lemma 3. By Lemmas 1 and 2, \Eq{qphase} applies to each step of this procedure, and so applies to $D$ and $D'$ as well. Note that in the standard situation of planar graphs and Kasteleyn orientations, Lemma 1 is all one needs, and moreover, absent any momomers, $N_{\sf in}$ is automatically even. In the following, we will ignore the topological sector dependent phases mentioned above, and drop the primes in \Eq{RVBp}. We then have the possibility to compute the coefficients $a_I$ in polynomial time, and feed them into a sign-problem-free Monte-Carlo scheme via \Eq{corr2}. That the frustrated square lattice RVB wave function \Eq{state} indeed has this property was already pointed out in Ref. \cite{YangYao}, where a correlation length was stated. In the following, we will first present Monte-Carlo data on correlation functions for completeness, and then turn to the study of entanglement properties of the square-lattice RVB state \Eq{state}. \subsection{Correlations} We begin by investigating the connected dimer-dimer (i.e., singlet-singlet) correlation functions, both for the original RVB wave-function \eqref{state} as well as for its projections onto individual topological sectors. Such projections can be achieved by means of operations that flip the pre-Kasteleyn orientation $\tilde\Theta$ along non-trivial loops, which change the relative signs between topological sectors. Note that in the preceding section, we have defined topological sectors through the dynamics associated with (arbitrarily large) contractible resonance loop moves. Alternatively, one may define them in terms of two ``winding numbers'' \cite{ReadChakraborty89} that each count the parity of the number of dimers crossed by one of two fixed non-contractible loops (which together generate the fundamental group of the torus). This defines precisely four topological sectors, all of which are invariant under the resonance-move dynamics of the first definition. To the extent that these dynamics are ergodic in each sector, the two definitions are identical. It is not difficult to argue that this is the case, since the loops for the resonance moves considered need only be contractible and are otherwise arbitrary. We will now denote the winding numbers associated along two non-contractible loop, on along the $x$-direction and on along the $y$-direction, and going through the crosses of the frustrated square lattice, by $n_{x}$ and $n_{y}$, respectively. We denote their possible values by $e$ (even) and $o$ (odd), and thus have the notation $(n_{x},n_{y})=(e,e), (o,e), (e,o), (o,o)$ for the sectors, which we will also abbreviate through an index $i$ running from $1$ through $4$, writing $\ket{\Psi_i}$ for the projections of the RVB state onto the topological sectors. The winding numbers $n_{x}$, $n_{y}$ correspond to 't Hooft magnetic loop operators in Ising gauge theories, which are readily associated to quantum dimer models through well-studied mappings \cite{MoessnerSondhiFradkin, Fradkin_book}, though the connection with RVB wave functions is, of course, somewhat less immediate. Figure \ref{figure_3} shows that the dimer-dimer correlations are indistinguishable within error bars, both for the equal amplitude superposition over all topological sectors \Eq{state}, as well as for the projections onto any one of the four topological sectors. A linear fit shows that indeed correlations fall off exponentially, as expected for a gapped $\mathbb{Z}_{2}$ spin liquid. We also computed the correlations $\langle \vec{S}_{i}\cdot \vec{S}_{i+\kappa} \rangle$ and $\langle \vec{S}_{i}^{z} \vec{S}_{i+\kappa}^{z} \rangle$. Again, all correlations decay exponentially, topological sectors remain indistinguishable. Additionally, we checked that $\langle \vec{S}_{i} \vec{S}_{i+\kappa} \rangle = 3 \times \langle \vec{S}_{i}^{\alpha} \vec{S}_{i+\kappa}^{\alpha} \rangle$ with $\alpha = x,y,z$. This is an obvious consequence of $SU(2)$-invariance, which, however, is not manifest in the Pfaffian Ising-basis formulation, and thus serves as a consistency check. To complete the investigation of the behavior of local operators, we have measured the expectation value $\langle \vec{S}_{i} \vec{S}_{i+1} \rangle$ along the horizontal, vertical, and diagonal bonds. For each type, measured values were indistinguishable within error bars between different topological sectors for the larges systems sizes studied (Fig. \ref{average_heisenberg} and Table \ref{}). Indistinguishable topological sectors are of a piece with the unbroken symmetry of the Hamiltonian in the ground states, and the topological nature of the ground state degeneracy. \begin{figure}[tbp] \centering \mbox{\subfigure{\includegraphics[width=3.5in]{corr_INSET}}\quad \quad \subfigure{\includegraphics[width=3.5in]{corr_SS_SzSz_16} }} \caption{Correlation functions of various local operators. a) The connected singlet-singletcorrelation $\langle (\vec{S}_{i} \vec{S}_{i+1})(\vec{S}_{i+\kappa} \vec{S}_{i+1+\kappa})\rangle - \langle \vec{S}_{i} \vec{S}_{i+1}\rangle \langle \vec{S}_{i+\kappa} \vec{S}_{i+1+\kappa}\rangle$ for a system of size $(L_{x}, L_{y}) = (20,20)$ within a topological sector $|\Psi_{i}\rangle$, $i = 1,\ldots,4$. We observe that the correlations are the same within error bars. Thus, as for the local expectation values of Fig. \ref{average_heisenberg}, the four topological sectors to these local correlators. The insets demonstrates exponential decay on a logarithmic scale. b) The correlation functions $\langle \vec{S}_{i} \vec{S}_{i+\kappa}\rangle$ and $\langle \vec{S}^{z}_{i} \vec{S}^{z}_{i+\kappa}\rangle$ for each topological sector. Again, the correlations decay exponentially and they are indistingushible within errors. SU(2)-invariance, while not manifest in the Pfaffian-formulation, is restored in the correlators, as we numerically confirm $\langle \vec{S}_{i} \vec{S}_{i+\kappa} \rangle = 3 \times \langle \vec{S}_{i}^{\alpha} \vec{S}_{i+\kappa}^{\alpha} \rangle$ with $\alpha = x,y,z$. } \label{figure_3} \end{figure} \subsection{Entanglement} After verifying the exponential decay of the correlations, we now turn our attention to computing topological properties. It is now well established that much interesting information about the universal characteristics of any physical phase are encoded in entanglement properties. This is true for topological phases \cite{lw,KP}, holographic quantum matter \cite{hartnoll}, and many others. In the case of bipartite entanglement, where the lattice is divided into a region $A$ and its complement $B$, one may consider the Renyi entropy of order $n$ defined as $S_{n} = \text{ln Tr}(\rho_{A}^{n})/(1-n)$, where $\rho_{A} = \text{Tr}_{B}|\Psi \rangle \langle \Psi|$ is the reduced density matrix of region $A$. Ground states of gapped local Hamiltonians are known to exhibit an {\it area law} scaling in region size, which in two dimensions can generically be written as, $S_n(\rho_{A}) = \alpha_n L_{A} - \gamma + \cdots $ \cite{Eisert}. Here, the leading term is dependent on the ``area'' (or boundary) of region $A$. The second term, the topological entanglement entropy (TEE) $-\gamma$ \cite{lw,KP}, is characterized by the total quantum dimension $\mathcal{D}$, which is defined through the quantum dimensions of the individual quasiparticles $d_{i}$ of the underlying theory: $\mathcal{D} = \sqrt{\sum_{i} d_{i}^2}$ \cite{lw,KP,Dong}. Conventionally ordered phases have $\mathcal{D} = 1$, while topologically ordered phases have $\mathcal{D} > 1$ with the TEE given by $\gamma = \text{ln}(\mathcal{D})$. The TEE is a fingerprint of the underlying topological phase, but it does not in general uniquely identify the topological phase. For sufficiently large areas, $\gamma$ is a universal value independent of the shape of the regions involved {\em as long as} the area $A$ is {\em contractible}. However, in the case where the area $A$ has at least one non-contractible boundary the TEE $\gamma$ becomes state-dependent, i.e., is sensitive the specific linear combination of degenerate ground states it is evaluated for. As shown in Ref.~\cite{zhang}, further developing ideas of Ref. ~\cite{Dong}, if one expresses any ground state in the basis of the {\it minimum entropy states} (MES-states), $| \Psi_{\alpha} \rangle = \sum_{j} c_{j} | \Xi_{j} \rangle$, then the sub-leading constant to the area law from cut of the torus into two cylinders is \begin{eqnarray}\label{gamma_torus2} \gamma^{\prime}(\{p_{j}\}) = 2\gamma + \text{ln}\bigg(\sum_{j} \frac{p_{j}^{2}}{d_{j}^{2}} \bigg) \;. \end{eqnarray} for $S_{2}$, where $p_{j} = |c_{j}|^2$. We further discuss MES-states in the results to follow. \begin{figure}[tbp] \centering \mbox{\subfigure{\includegraphics[width=3.5in]{Together_S2_M12_N10_8_sectors_INSET_4}}\quad \quad \subfigure{\includegraphics[width=3.5in]{Together_S2_M12_N10_8_MES} }} \caption{Renyi entropies $S_2$ for cylindrical cuts of the torus and various system sizes, topological sectors, and linear combinations thereof. a) $S_{2}$ as a function of cylindrical height for all topological sectors and their equal amplitude superposition. All data agree well for system size 96, in agreement with the MES Ansatz \Eq{MES_Z2}, though small discrepancies can still be resolved. These are interpreted as finite size effects. Error bars are necessarily larger for larger system size, as the expectation value of the SWAP operator \cite{hastings,RatioT} becomes exponentially small. b) $S_2$ for the four minimum entropy states (MES) $|\Xi_{i}\rangle$ according to the Ansatz \Eq{MES_Z2}. Stronger discrepancies between the different MES-states are noted at smaller system size, but agreement within error bars is again attained for 120 sites. The inset in a) illustrates the procedure from which value for the topological entanglement entropy is extracted via the difference in $S_2$ for MES-states and the $\ket{\psi_i}$. See text for details. } \label{figure_5} \end{figure} In the following, we investigate entanglement properties of the frustrated spin-$\frac{1}{2}$ RVB wave functions using the variational Pfaffian MC scheme for lattices of up to $120$ sites. As mentioned above, the Pfaffian MC scheme allows one to project onto each topological sector, and every linear combination thereof. We will again use this feature in the following results. To obtain the second Renyi entropy $S_{2}$ for noncontractible regions, we employ the standard QMC replica-trick \cite{hastings,RatioT}. We have calculated the TEE for cylindrical regions $A$ that are wrapped around the torus like ``ribbons'', and whose boundaries thus correspond to generators of the fundamental group of the torus. Within our method, this approach has proven advantageous compared to schemes extracting the TEE from contractible regions, such as the Levin-Wen \cite{lw} and the Kitaev-Preskill \cite{KP} constructions. These constructions isolate the TEE from the entanglement $S_{2}$ by canceling the area dependent contributions when $S_{2}$ is computed for several suitably chosen areas. However, these procedures usually lead to values for the TEE $\gamma$ that are plaqued by rather large error bars. For better accuracy, we only consider regions with noncontractible boundaries. We examine frustrated RVB states for square lattices of dimensions $12\times 8$ and $12 \times 10$, respectively and calculate the Renyi entropy $S_{2}$ for cylindrical bipartitions. As the cylinder length increases, $S_{2}$ quickly saturates (Figure \ref{figure_5}a). This type of behavior is consistent with the system having a gap. For the smaller lattice size of 96, Figure \ref{figure_5}a) shows the $S_{2}$-values of different topological sectors to be in near agreement, despite some small discrepancies that are nonetheless resolved by the very small error bars . We interpret these discrepancies as finite size effects. While \Eq{gamma_torus2} predicts discrepancies in TEE for {\em some} linear combinations of the MES-states, the expected topological phase in question has $d_{i}=1$ for all $i$. Therefore, for the MES-states themselves, all TEE values are expected to be degenerate. The same will be true for any other ground state basis with coefficients in the MES-basis whose absolute values differ from one basis state to another only by permutation. The data for 120 lattice sites do not resolve any discrepancy in $S_{2}$ for different topological sectors within the (somewhat larger) error bars. We will argue any such discrepancies to be due to finite size effects and error bars, as we will explain in the following. Our discussion will be based on an informed guess for the MES-states. As mentioned above, at least for hard-core dimer versions of our wave functions to be discussed in the following Section, the topological sectors $(n_{x},n_{y})$ can be identified with 't Hooft magnetic loop eigenspaces of Ising gauge theory, or, for that matter, the toric code. For the latter, the linear relations between the ground states carrying well-defined 't Hooft magnetic loop quantum numbers and the MES-ground-states is exactly known \cite{zhang}. At least for the kagome quantum dimer model of \cite{misguich}, the same calculation goes through \cite{wildeboer15}, yielding the same relations. One expects these relations to be universal for the $\mathbb{Z}_{2}$ topological phase. Thus, as long as the 't Hooft magnetic loop quantum number carrying ground states can be unambiguously identified, the MES-states are known. For the RVB-states, this is the case, as long as we can lift these quantum numbers from the hard-core dimer picture associated to these wave functions. Based on the above assumptions, one arrives at the following Ansatz for MES-states associated with cuts along the $y$-direction: \begin{equation} | \Xi_{1,2}\rangle = \frac{1}{\sqrt{2}} (| \Psi_{ee} \rangle \pm | \Psi_{eo} \rangle) , \qquad | \Xi_{3,4}\rangle = \frac{1}{\sqrt{2}} (| \Psi_{oe} \rangle \pm | \Psi_{oo} \rangle)\;. \label{MES_Z2} \end{equation} Note that phase ambiguities regarding the $\ket{\Psi_{n_x,n_y}}$ are essentially eliminated by time-reversal symmetry, i.e., the requirement of working with ``real'' wave functions. Note also that the inverse relations \begin{equation}\label{singlesec} | \psi_{ee(eo)}\rangle = \frac{1}{\sqrt{2}} (| \Xi_{1} \rangle \pm | \Xi_{2} \rangle) , \qquad | \psi_{oe(oo)}\rangle = \frac{1}{\sqrt{2}} (| \Xi_{3} \rangle \pm | \Xi_{4} \rangle)\;. \end{equation} indeed give the same TEE $\gamma'_{\sf magnetic}=\gamma$ for $d_{i}\equiv 1$ in \Eq{gamma_torus2}. While the above relations are based on some assumptions, in particular, that the RVB-states do indeed belong to the $\mathbb{Z}_{2}$ topological phase, a strong consistency requirement is given by the fact that when $\gamma$ is now calculated based on the above identifications of MES-states, the value $\ln(2)\approx 0.69$ must be obtained. To this end, we compare Monte-Carlo results for $\gamma'$ for both the ``magnetic'' eigenstates $\ket{\Psi_{n_x,n_y}}$ (Figure \ref{figure_5}a) and the conjectured MES-states $\ket{\Xi_i}$ (Figure \ref{figure_5}b). We first observe a pronounced reduction in the $S_2$ discrepancies between different $\ket{\Xi_i}$ when comparing data for 96 and 120 lattice sites. This is again consistent with these discrepancies being due to finite system size. Indeed, from \Eq{gamma_torus2}, we expect $\gamma'_{\sf MES}=2\gamma$ for all MES-states. We now proceed by evaluating the {\em difference} in $S_2$ between the $\ket{\Xi_{i}}$ and the $\ket{\Psi_{n_x,n_y}}$ for the size 120 lattice, which consistency requires to be equal to $\gamma'_{\sf MES}-\gamma_{\sf magnetic}=2\gamma-\gamma=\gamma$. In the following, we extract this value by averaging this difference over cylinder lengths $x = 5, \ldots, 7$, and moreover, we may average over different topological sectors and MES-states, respectively. For example, working with $\ket{\Xi_{1}}$ and $\ket{\Xi_{2}}$ their linear combinations $\ket{\Psi_{ee}}$ and $\ket{\Psi_{eo}}$, respectively, one extracts six independent measurements of $\gamma$ with average $0.73$ and standard deviation $0.06$. The measurement of the logarithm of an exponentially small value is highly demanding. We can improve the accuracy by replacing $S_2$-values from the $\ket{\Psi_{eo}}$-sector with the values for $\ket{\Psi_{oe}}$, which are, at the time of this writing, better converged. The resulting average over six values and their standard deviation are then found to be $0.70\pm0.04$. There is thus encouraging agreement with the expected value of $\ln(2)$, indicating the consistency of our assumptions in identifying the MES-states. Another consistency check lies in the fact that the equal amplitude superposition over all $\ket{\Psi_i}$ has essentially the same $S_2$ value as the individual $\ket{\Psi_i}$ for the $96$ site lattice (Figure \ref{figure_5}a), as implied by Eqs. \eqref{singlesec} and \eqref{gamma_torus2}. Overall, these data well support the hypothesis that the wave function \eqref{state} describes a state in the $\mathbb{Z}_2$ topological phase. \section{A Quantum-Dimer Model\label{QDMSEC}} \subsection{General Hamiltonian} In Sec. \ref{RVBdef} we emphasized that the definition of the RVB-state \eqref{state} utilizes only local rules. Though we will not systematically elaborate here, it is likely that the \Eq{state} can be written as a tensor-network state, as have other short-range RVB wave functions \cite{schuch}. Related to that, it can be given a parent Hamiltonian comprised of local operators acting on sufficiently large cells, following the recipe of Ref.\cite{seidel09} or equivalently \cite{zhou} Ref.\cite{schuch}. Here we will chose the route of quantum dimer models \cite{Kivelson89} to motivate that the amplitudes in \Eq{state} may arise naturally from a local Hamiltonian at least in this context. To this end, we will now simplify the Hilbert space of the original spin-$1/2$ degrees of freedom of \eqref{state} according to the traditional quantum dimer model philosophy \cite{Kivelson89}. The quantum dimer model is defined on a Hilbert space with distinct, orthonormal states corresponding to each allowed dimer covering of the lattice, in our case, the frustrated square lattice. While representing singlets in an abstract sense, these dimers are more properly understood as hard-core bosons living on the medial lattice. In particular, different hard-core boson states have different inner products from the valence bond states associated with the same dimerizations. In the following, we will re-interpret the kets $\ket{D}$ as the hard-core dimer configuration associated with the dimerization $D$, and \underline{not} the valence bond configuration associated to $\ket{D}$ in preceding sections. We define a quantum dimer model (QDM) Hamiltonian via matrix elements between at most locally differing dimer states. There are two different types of matrix elements: the ``potential'' terms, in the following labeled $V,\ V^{\prime},\ V^{\prime\prime}$ and $V^{\prime\prime\prime}$, which are diagonal in the dimer basis and associate an interaction energy with various local arrangements of dimers, and ``kinetic'' terms, in the following referred to as $t,\ t^{\prime}, t^{\prime\prime}$, and $t^{\prime\prime\prime}$, which involve a local rearrangement of a small number of dimers. On the nonbipartite square lattice, we represent the Hamiltonian graphically as: \begin{eqnarray}\label{} {\cal H}&=&\sum_{\square}\Bigg[\!-\!t\left(\ket{\dimer{12-34}} \bra{\dimer{13-24}} \!-\!\frac{1}{2}\ket{\dimer{12-34}} \bra{\dimer{14-23_julia}} \!-\!\frac{1}{2}\ket{\dimer{13-24}} \bra{\dimer{14-23_julia}} \!+\!h.c. \right) \!+\!V\left(\frac{3}{2}\ket{\dimer{12-34}}\bra{\dimer{12-34}} + \frac{3}{2}\ket{\dimer{13-24}}\bra{\dimer{13-24}} + \ket{\dimer{14-23_julia}}\bra{\dimer{14-23_julia}} \right)\Bigg]\nonumber \\ \!&+&\!\sum_{\{\input{rect.latex}\}} \Bigg[\!-\!t^{\prime}\left(\ket{\dimer{13-45}}\bra{\dimer{14-35}} +h.c.\right)\!+\!V^{\prime}\left(\ket{\dimer{13-45}} \bra{\dimer{13-45}} +\ket{\dimer{14-35}}\bra{\dimer{14-35}}\right)\!-\! t^{\prime\prime}\left(\ket{\dimer{13-46}} \bra{\dimer{14-36}} \!+\!h.c.\right)\!+\! V^{\prime\prime}\left(\ket{\dimer{13-46}}\bra{\dimer{13-46}} \!+\! \ket{\dimer{14-36}}\bra{\dimer{14-36}}\right) \Bigg] \nonumber \\ \!&+&\!\sum_{\{\input{rectt.latex}\}} \Bigg[\!+\!t^{\prime\prime\prime}\left(\ket{\dimer{14-36_julia}}\bra{\dimer{14-36_julia3}} +h.c.\right)\!+\!V^{\prime\prime\prime}\left(\ket{\dimer{14-36_julia}} \bra{\dimer{14-36_julia}} +\ket{\dimer{14-36_julia3}}\bra{\dimer{14-36_julia3}}\right) \Bigg], \end{eqnarray} where black bonds represent dimers. In \eqref{}, the first sum runs over all plaquettes, and the second sum runs over all pairs of adjacent plaquettes (with both orientations) while the third sum runs over all ``3-by-3''-plaquettes. For the ``3-by-3''-plaquettes the braces symbolize that we take into account all four possible orientations that the dimer emanating from the center site can have while occupying a diagonal link. The circumstance that, somewhat atypically, kinetic terms appear with positive sign in some places accounts for the fact that the ground state to be stabilized, i.e., \Eq{state} re-interpreted as quantum dimer wave function, has negative amplitudes associated with crosses. The model \eqref{} contains kinetic and potential terms that involve dimers living on diagonal links and pairs of crossed dimers. It is thus similar to the one previously investigated by Yao and Kivelson \cite{yao_kivelson}, but expands the latter by including configurations and matrix elements involving crossed pairs of dimers. We emphasize that despite similarities in notation and terminology, QDM differ decisively from spin-1/2 systems. Whether the exploration of the phase diagram of QDMs yields actual insights into that of $SU(2)$-invariant spin-1/2 systems on the same lattice has been of great interest since the original QDM of Rokhsar and Kivelson \cite{Kivelson89}. In particular, RVB wave functions of the form \Eq{RVBp} have been studied extensively both for bipartite \cite{alet, tang} and non-bipartite lattices \cite{wildeboer12, schuch, YangYao, wildeboer15}, and have always been found to have qualitatively similar features as their QDM counterparts. The present case of interest, \Eq{state}, may be viewed as one step further removed from the original RVB-state when regarded as a quantum dimer wave function. Indeed, while in many lattice geometries with planar graphs, the valence bond state for this graph are now well-known to be linearly independent \cite{wildeboer11}, this is not so for non-planar case discussed here. In particular, crossed pairs of valence bonds and the two parallel configurations on the same square are linearly dependent: $| \dimer{14-23} \rangle \propto -( | \dimer{12-34} \rangle + | \dimer{13-24} \rangle )$. (It is currently an unknown problem whether the valence bond configurations discussed here exhibit further linear dependences other than the aforementioned for large lattices.) Linear independence has often been employed to strengthen the case of correspondence between QDMs and frustrated $SU(2)$-invariant quantum antiferromagnets, e.g., by means of systematic derivation of QDMs from antiferromagets via formal expansion in the ``overlap parameter'' \cite{schwandt}, which characterizes non-orthogonality between different valence bond configurations. Such expansions are well-defined only when linear independence is given. In the context of the present study, it is natural to extend questions of general correspondence between QDMs and anti-ferromagnets and/or their prototypical RVB-type trial states to situations where the linear independence is absent. \subsection{The RK point.} \label{rk_point_sec} The parameter space of the model \eqref{} has a ``generalized Rokhsar-Kivelson'' line where the exact ground states is know. In the following, we specialize to this line defined by the following choce of parameters: \begin{eqnarray} t=V, ~t'=V', t^{\prime\prime}=V^{\prime\prime} ~\textrm{and}~ t^{\prime\prime\prime}=V^{\prime\prime\prime} \;\;\mbox{(all couplings positive)}\;. \end{eqnarray} One can show that at this generalized RK line, \eqref{} can be expressed as a sum over positive semi-definite operators with ground state(s) formally analogous to \eqref{state}, which we repeat here as \begin{eqnarray}\label{eq:rk} \ket{\Psi}=\sum_{D}(-1)^{n_{c}} \ket{D} \;\;\mbox{(quantum dimers)}, \end{eqnarray} emphasizing that this is {\em not} the same as the RVB wave function \eqref{state}, but rather its quantum dimer version. As in the RVB case, for periodic boundary conditions four topological sectors may be introduced. As the dynamics of the model do not mix these sectors, the sum over $D$ may be restricted to any one sector, giving four ground states on the torus. It is indeed not difficult to see that each of the four types of local terms in \Eq{}, proportional to one of the fours independent couplings at the RK line, individually annihilates \Eq{eq:rk}. The quantum dimer RK state, and its projections onto topological sectors, thus simultaneously minimize each of these local terms. One expects \Eq{eq:rk} to describe a $\mathbb{Z}_2$ topological liquid just as the original RVB state \Eq{state} does, based on the results of the preceding section. Consider the dimer-dimer correlation fuction $\langle \Psi|n_i n_j|\Psi\rangle/\langle\Psi|\Psi\rangle$, where $n_i$ is the dimer occupation operator on a link labeled $i$. It is easy to see that this does map onto a classical dimer problem, and so can be addressed by Monte-Carlo evaluation. We will leave this to future studies. The point we wish to make here, however, is that this classical problem is {\em oblivious} to the phases in \Eq{eq:rk}. It is therefore not addressable through the Pfaffian methods that have been so successful in solving many classical dimer problems, and their quantum dimer analogues. Specifically, the Pfaffian formulation based on the pre-Kasteleyn orientation of the preceding section {\em does} always reproduce the phases of \Eq{eq:rk}, as we have seen, and therefore cannot produce the infinite temperature classical partition function that effectively describes the dimer-dimer correlator. We leave the evaluation of correlation functions of the QDM-state \eqref{eq:rk} as an interesting problem for the future. We point out that that the pre-Kasteleyn orientation discussed here has been previously utilized by Yao and Kivelson to gain analytic insight into correlations of a crossing-free variant of the QDM on the frustrated square lattice\cite{yao_kivelson}. Again, exact Pfaffian evaluation is not possible, but an exact reformulation as an {\em interacting} fermionic theory allows controlled mapping onto an effective massive Thierring model. We do not rule out that variants of local lattice Hamiltonians exists whose correlators reduce exactly to Pfaffians of pre-Kasteleyn matrices. We leave this as another interesting problem for the future. \section{Conclusion\label{conclusion}} In this work, we have formally defined the notion of a pre-Kasteleyn orientation. A special instance on the frustrated square lattice, which had already appeared in the study a ``crossing-free'' quantum dimer models \cite{yao_kivelson}, has been utilized for a Pfaffian re-formulation of a next-nearest neighbor spin-1/2 RVB wave function {\em with} crossings and a negative sign rule on the square lattice. This allowed us to evaluate the $S_{2}$-Renyi entropy on the torus using Pfaffian Monte-Carlo \cite{wildeboer12, wildeboer15}, confirming the wave functions ``minimum entropy state'' \cite{zhang} by consistent extraction of topological entanglement entropy of $\ln(2)$. This corroborates the fact that this wave function lies in the $\mathbb{Z}_2$ topological phase \cite{YangYao}. In addition, local expectation values, their dependence on topological sector, and correlation functions have been evaluated in some detail. As a by-product, various useful statements surrounding pre-Kasteleyn orientations were proven, which in particular address technicalities from the possibility of (self-)intersecting loops in the overlap graph between two different dimerizations. We also discussed a quantum dimer model related to the RVB wave function. We conjecture that pre-Kasteleyn orientations may prove to have broader applications facilitating combined analytic-numerical techniques for lattice wave functions defined on non-planar graphs, and are hopeful that our work will stimulate future investigation. \begin{acknowledgments} JW thanks NSF DMR 1306897 and NSF DMR 1056536 for partial support. AS gratefully acknowledges insightful discussions with K. Yang and Z. Nussinov. Our Monte Carlo codes are partially based upon the ALPS libraries \cite{alps1, alps2}. The simulations were run on the SHARCNET clusters. \end{acknowledgments}

\section{Introduction} Video streaming has become one of the most popular online activities in recent years. Today's video streaming market has become highly competitive and it has been crowded with various content providers. For example, in China, there are Youku, IQiyi, Sohu, LeTV and Tencent Video, etc. In North America, there are Netflix, Hulu and Amazon Video, etc. For these content providers, predicting the future video popularity is very important, for example, in planning advertisements. Although there are many types of contents such as movie, TV series, drama, cartoon, MTV and other user generated contents, we focus on predicting the popularity of TV series in this paper. One of the most used measurements of online video popularity is the view count. Suppose a TV series will be released on Youku\footnote{} in $7$ days, is it possible to predict the total view count of all episodes of this TV series? Generally, this problem is challenging for the following reasons. First, the prediction problem is both ``wide'' and ``deep''. Similar to the wide and deep learning for recommender systems \cite{widedeep}, the prediction problem not only depends on a wide range of features but also be highly non-linear and complex. Modern deep learning provides a very powerful framework for supervised learning. With more layers and units, deep neural network can represent complex and highly nonlinear models. Also, deep neural networks perform better in generalizing to unseen feature combinations. However, over-fitting becomes a serious issue when the data available is limited. Multi-task learning (MTL) \cite{MTL-original}\cite{MTL97} is one way to improve the generalization power of the original task, by jointly training multiple related tasks and sharing representations between them. Multi-task learning is widely used with success in machine learning, and it acts as a regularizer and it reduces the risk of over-fitting \cite{MTL-overfitting} and its ability to fit random noise. In addition to MTL, dropout \cite{dropout} and $\ell_2$ regularization are also useful to prevent over-fitting. Second, multiple competitors are involved in the prediction problem. Users often migrate among different content providers for many possible reasons \cite{migrate}. To predict a TV series's video views, one must consider the influence of other related TV series from various content providers. Some special structure should be involved in the prediction model to represent the relations of multiple competitors. Then relation networks (RN) introduced in \cite{RN} become useful, where originally a visual reasoning problem is considered. An RN is a simple and effective module for the case when input ``objects'' (usually hidden units of a neural network) have relations which each other. Relation network is also found to be useful in neural machine translation \cite{RN-nmt}, few-shot learning \cite{rn-fewshot} and complex relational reasoning \cite{RN-recurrent}. Relation networks are also related to relational learning in statistics \cite{statRelational}. In this paper, we introduce a general model for predicting the popularity of online videos. This paper is organized as follows. In Section \ref{sec_back}, we give background knowledge on multi-task learning and relation networks. In Section \ref{sec_prop}, we give the details of our proposed prediction model with the DNN module, the RN module and the MTL module. The experimental results are discussed in Section \ref{sec_experiment} and the related work is discussed in Section \ref{sec_related}. \section{Background} \label{sec_back} In this section, we briefly introduce multi-task learning and relation networks, which are used as separate modules in our prediction model. \subsection{Multi-task learning} Multi-task learning is typically done by hard or soft parameter sharing. In this paper, we use hard parameter sharing. It shares the hidden layers between all tasks, while keeps several task-specific output layers \cite{MTL}. The idea is shown in Figure \ref{fig-MTL}. \begin{figure}[h] \begin{center} \includegraphics[width=0.35\textwidth]{MTL.jpg} \end{center} \caption{Hard parameter sharing for multi-task learning in deep neural networks.}\label{fig-MTL} \end{figure} Suppose $\mathcal{L}$ is the loss for the main task and $\mathcal{L}'_i$s are the losses for some other auxiliary tasks, then they can be trained jointly with the total loss \begin{equation} \mathcal{L}_{total} = \mathcal{L} + \sum_{i}\lambda_i \mathcal{L}_i \end{equation} where $\lambda_i$ is the weight for the $i$-th auxiliary task. The value of $\lambda_i$ is used to control the importance of an auxiliary task. \subsection{Relation network} Given a set of ``objects'' $\{\bm{o_1},\bm{o_2},...,\bm{o_n}\}$, which are hidden units of a neural network. A relation network is defined as \begin{equation} \label{eqnRN1} f\left(\sum_{1 \leq i < j \leq n}g (\bm{o_i},\bm{o_j})\right) \end{equation} where $\bm{o_i}$ is the $i$-th object and $f$ and $g$ are differentiable functions. Here, the pair-wise relations are represented by common function $g$ and the combined relation of $\{\bm{o_1},\bm{o_2},...,\bm{o_n}\}$ is fed to the function $f$. Our prediction problem is slightly different. Suppose we want to predict one particular object, say $\bm{o}$. Instead, we consider the relations of $\bm{o}$ with the other objects $\bm{o_1},\bm{o_2},...\bm{o_n}$. Our relation network is defined by modifying (\ref{eqnRN1}) as \begin{equation} \label{eqnRN2} f\left( \sum_{i=1}^n g (\bm{o},\bm{o_i}) \right) \end{equation} and it is shown in Figure \ref{fig-RN}. \begin{figure}[h] \begin{center} \includegraphics[width=0.4\textwidth]{RN.jpg} \end{center} \caption{The relation network to represent relation of $\bm{o}$ with $\bm{o_1}$, $\bm{o_2}$, ..., $\bm{o_n}.$}\label{fig-RN} \end{figure} \section{Proposed Model}\label{sec_prop} The architecture of our proposed model is shown in figure \ref{fig-model}. Our model contains the deep neural network (DNN) module, the relation network (RN) module and the multi-task learning (MTL) module. We will describe the details in this section. \subsection{Problem definition} We define our problem mathematically. The input vector representing features is denoted as $\bm{x}$. For notation simplicity, we assume only one auxiliary task in our MTL module is jointly trained, and the corresponding outputs of the main task and the auxiliary task are denoted as $y$ and $y'$, respectively. In practice, more than one auxiliary tasks may be trained jointly depends on the problem given. Let $(\bm{x},y,y')$ denote the input vector, the main task output and the axillary task output. Suppose our training dataset is \begin{equation} \mathcal{D} = \{(\bm{x},y,y')\} \end{equation} Assume any sample $\bm{x} \in \mathcal{D}$ has relation with some other $n$ samples in $\mathcal{D}$, say $\bm{x_1},\bm{x_2},..,\bm{x_n} \in \mathcal{D}$ where $n$ is a fixed integer. Then the original training dataset $\mathcal{D}$ can be transformed to a new dataset \begin{equation} \mathcal{{D}}^* = \{(\bm{x^*}, y, y')\} \end{equation} where $\bm{x^*} = concat(\bm{x},\bm{x_1},\bm{x_2},..,\bm{x_n})$. We call these related samples $\bm{x},\bm{x_1},\bm{x_2},..,\bm{x_n}$ as ``objects'', which coincide with the terminology used in relation networks \cite{RN}. \begin{figure}[h] \begin{center} \includegraphics[width=0.45\textwidth]{graph.jpg} \end{center} \caption{Our proposed prediction model. It consists of three modules: the deep neural network (DNN) module, the relation network (RN) module and the multi-task learning (MTL) module.}\label{fig-model} \end{figure} \subsection{Our prediction model} Consider the main input vector $\bm{x}$ (main object) and the related input vectors $\bm{x_i}$ (related objects), for $i=1,2,...,n$, where $n$ is a fixed integer corresponding to $n$ related objects. The value of $n$ depends on the specific problem given. For instance, if we want to predict the popularity of a TV series, then $n$ related TV series should be considered, usually popular TV series with the closest releasing dates. The first part is the DNN module. We use deep multi-layer perceptrons (MLPs) to get their high-level representations \begin{equation} \bm{o}= e(\bm{x}),\quad \bm{o_i}=e(\bm{x_i}),\quad i=1,2,...,n. \end{equation} We use many layers of MLPs for the function $e$ to capture the complex combinations of features (more than $10$ layers). The next part is the RN module. Let the function $f$ and $g$ be two shallow MLPs (about $2$ or $3$ layers). The relation of $\bm{o}$ and $\bm{o_i}$ is represented by the function $g$, \begin{equation} \bm{r_i} = g(\bm{o},\bm{o_i}),\quad i=1,2,...,n. \end{equation} The combined relation is fed to $f$ as an input, \begin{equation} \bm{r} = f(\sum_{i=1}^n \bm{r_i}) \end{equation} where the summation is element-wise. The last part is the MTL module. The vector $\bm{r}$ from RN module is fed to the MTL module as a input. Note the shared dense layers of MTL in Figure \ref{fig-MTL} are skipped here, as there are already MLPs in the RN module. Therefore, the vector $\bm{r}$ is directly fed to two separate task-specific MLPs, say $h$ for the main task and $h'$ for the auxiliary task. Finally, we get the main output and the auxiliary output, respectively, \begin{equation} \hat{y} = h(\bm{r}),\quad \hat{y}' = h'(\bm{r}). \end{equation} In conclusion, for any input vector $(\bm{x^*},y,y')$ from the training dataset $\mathcal{D}^*$, we output $\hat{y}(\bm{x^*},\bm{\theta})$ and $\hat{y}'(\bm{x^*},\bm{\theta})$. The loss function is \begin{eqnarray} \mathcal{L}(\mathcal{D}^*,\bm{\theta}) &= &\sum_{(\bm{x^*},y,y') \in \mathcal{D}^*} |\hat{y}(\bm{x^*},\bm{\theta})-y|^2 \nonumber \\ & &+\lambda\sum_{(\bm{x^*},y,y') \in \mathcal{D}^*}|\hat{y}'(\bm{x^*},\bm{\theta})-y'|^2 \nonumber \\ & & +\gamma || \bm{\theta}||_2^2 \end{eqnarray} where $\lambda$ and $\gamma$ are hyper-parameters for MTL and $\ell_2$ regularization, respectively. \section{Experiments} \label{sec_experiment} In this section, we discuss the experimental results on predicting the total view counts of TV series on Youku. We use more than 200 dimensional features as our input, including the actor/actress name, genre and director name, etc. For the DNN module, we use $15$ layers of MLPs. For the RN module, we fix the number of related objects as $n=3$. We use $3$ layers of MLPs to represent a ``relation'' and the combined relation is fed to another $3$ layers of MLPs. For the MTL module, we jointly predict the total view count and the popularity index of Youku, which represents the TV series' popularity. The ReLU activation is used to provide the non-linearity. We also use batch normalization \cite{BN} to prevent gradients vanishing, and dropout \cite{dropout} and $\ell_2$ regularization to reduce over-fitting and increase the model's generalization power. Our model is end-to-end differentiable. The loss function is optimized by back propagation and accelerated with Momentum, using our self-developed JAVA framework of deep learning called DeepDriver \footnote{}. Note that the training dataset vary in time, hence, the date when the prediction is made is important. Typically, we predict $7$, $180$ and $360$ days before the TV series' releasing date. To evaluate the effectiveness of our prediction model, we compare the full model ``DNN+RN+MTL'' with the partial models ``DNN'' and ``DNN+MTL''. All TV series before Jan 1, 2017 are used as training data, and the other TV series from Jan 1, 2017 are used for evaluating. The results are shown in terms of coefficients of determination ($\mathcal{R}^2$) in Table \ref{tableResult}. For example, the $\mathcal{R}^2$ predicted $7$ days before the releasing date, the result is improved by $0.03$ with the MTL module and another $0.04$ with the additional RN module. \begin{table}[h] \begin{center} \begin{tabular}{|c|c|c|c|} \hline &DNN&DNN+MTL&DNN+RN+MTL\\ \hline 7 days& $0.80$& $0.83$ & $0.87$ \\ \hline 180 days& $0.75$& $0.80$ & $0.85$\\ \hline 360 days& $0.74$& $0.79$ & $0.85$\\ \hline \end{tabular} \caption{The results of predicting future view counts of TV series are shown in terms of coefficients of determination ($\mathcal{R}^2$). It is trained on TV series data before Jan 1, 2017 and evaluated on TV series data from Jan 1, 2017. The first, second and third row corresponds to predictions results made $7$, $180$ and $360$ days before the TV series' releasing date, respectively. }\label{tableResult} \end{center} \end{table} \section{Related Work} \label{sec_related} Much work has been done for predicting the popularity of web content, a survey is done in \cite{surveyPred}. In particular, there are a lot of approaches for predicting the popularity of online videos. In \cite{predSVR}, the popularity of Youtube and Facebook videos is predicted using support vector regression with Gaussian radial basis functions. In \cite{predSentiment}, video popularity is predicted with sentiment propagation via implicit network. A dual sentimental Hawkes process (DSHP) is proposed, and it is evaluated on four types of videos: movies, TV episodes, music videos, and online news. In \cite{predYoutube}, four methods are discussed for predicting the popularity of Youtube videos: univariate linear model, multivariate linear model, radial basis functions and a preliminary classification method. In \cite{predYouku}, both online video future popularity level and online video future view count are predicted for Youku videos. In \cite{predYouku2}, how the popularity of Youku videos evolved over time is analyzed, and multivariate regression models with different parameters are used. In \cite{predUGC}, predicting the popularity of user-generated contents (UGC) is discussed. In \cite{predSerial}, the popularity of online serials is predicted with autoregressive models, including naive auto-regressive (NAR) model based on the correlations of serial episodes, and transfer auto-regressive (TAR) model to capture the dynamic behaviors of audiences. In \cite{predFoursquare}, multivariate linear regression is used to predict the micro-review popularity using the location-based social networking platform Foursquare. The ``wide'' and ``deep'' prediction problem is similar to analyzing the high dimension low sample size (HDLSS) data. Feature selection is a powerful and widely used tool for HDLSS data. For instance, Lasso \cite{lasso} minimizes the objective function penalized by the $\ell_1$ norm and leads a sparse model. Unfortunately, Lasso considers only the linear dependency but ignores the nonlinearity. In \cite{dnp}, deep neural pursuit (DNP) is proposed with advantages in high nonlinearity, the robustness to high dimensionality, the capability of learning from a small number of samples. In \cite{widedeep}, a Wide\&Deep model is proposed for recommender system and it achieves both memorization and generalization. Wide linear models can effectively memorize sparse feature interactions using cross-product feature transformations, while deep neural networks can generalize to previously unseen feature interactions through low-dimensional embeddings. Our work is related to the above methods, but we extend the problem by considering the competition relation of objects, which is very common in practice. \section{Conclusion} In this paper, we propose a model for predicting the popularity of online videos via deep neural networks with MTL and RN modules. Our model can apply to situations where the problem is both ``wide'' and ``deep'', and with multiple competitions. The experimental results on predicting the total view counts of TV series suggest our model is effective. In the future work, we plan to improve our model with recurrent neural networks (RNN) related structures such as long short-term memory (LSTM) networks \cite{lstm}. Approaches in time series prediction may be also helpful to improve our model. Furthermore, we think even more features can be added to our current inputs, e.g., video frame information from the trailer and text information from video description, etc. Besides online videos, our work can be used to predict other types of web content. Our work can even apply to other similar applications, such as predicting stock prices and predicting the winner of a presidential election, etc. \newpage

\section{General Overview} The \aipcls{} is a \LaTeXe{} document class for conference proceedings of the American Institute of Physics and other documents with similar layout requirements. Your file will be used to reproduce your paper as is, the only modifications done by the publisher are adding appropriate page numbers and the copyright line. It is therefore essential that you embed all fonts when saving your file. This version of the guide explains all features of the class some of which are only applicable to certain proceeding layouts. The class provides essentially the same markup as implemented by \LaTeX's standard \texttt{article} class. In addition to this it implements the following: \begin{itemize} \item extended set of front matter commands, \item automatic placement of floats into column or page areas including turning of table floats by 90\textdegree{} if necessary, \item allows mixing column and page-wide floats without getting the numbering out of sync, \item footnotes will appear below bottom floats, \item extended set of citation commands if the \texttt{natbib} system is installed, \item support for table notes, \item support for textual page references like ``on the next page''. \end{itemize} Due to the extended functionality an article written for \LaTeX{}'s standard article class might need adjustments in the following places before it can be used with the \aipcls{} (a more detailed description is given in later sections): \begin{itemize} \item In the preamble, since the \aipcls{} requires a |\layoutstyle| declaration. \item In the front matter, since the \aipcls{} uses an extended set of title/author declarations. \item In the body of floats, since the \aipcls{} only allows a single |\caption| command and processes the body in horizontal mode. \end{itemize} \section{Checking your \LaTeX{} distribution} To ensure that your installation of \LaTeX{} contains everything necessary to successfully use the \aipcls{}, run the file \texttt{aipcheck.tex} through \LaTeX, e.g., \begin{verbatim} latex aipcheck \end{verbatim} It will try to determine if everything necessary is available and if not, will make recommendations what can be done about it. In certain cases you might be able to use the class if you follow the suggestions, in other cases the only solution is to upgrade your \LaTeX{} installation. Unfortunately it is impossible to check for all potential problems. If \texttt{aipcheck.tex} claims everything is fine, but you nevertheless have difficulties, consult the ``Frequently Asked Question'' (\texttt{FAQ.txt}) and the readme file in the distribution. \section{Class details} \subsection{Selecting the target layout} The class supports different layouts. These are selected by placing a |\layoutstyle| declaration in the preamble of the document. \BDefC{layoutstyle}[m]{layout name} This command is required. With version 1.3 of the \aipcls{} the following \Larg{layout name}s can be specified. \begin{description} \item[6x9] Layout for the AIP Conference Proceedings with 6 x 9 inches single column format (short name |6s|). \item[8x11single] Layout for the AIP Conference Proceedings with 8.5 x 11 inches single column format (short name |8s|). \item[8x11double] Layout for the AIP Conference Proceedings with 8.5 x 11 inches double column format (short name |8d|). \item[arlo] Layout for the ``Acoustics Research Letters Online'' --- ARLO. \end{description} For example, the current guide was produced using the declaration |\layoutstyle{|\texttt{\selectedlayoutstyle}|}|. \subsection{Supported options}\label{suppopt} As the class is based on the article class of standard \LaTeX{} all reasonable\footnote{Reasonable means not conflicting with fixed requirements for the AIP class, e.g., as this class requires 10pt body size option \texttt{11pt} and \texttt{12pt} are ignored and produce a warning.} options of this class are supported automatically. In addition there are a number of options unique to the \aipcls. \subsubsection{Paper selection} Two options control the placement of the text on the physical page. Choose the one that corresponds to your printer paper. \begin{description} \item[letterpaper] Directs the class to assume that the output is printed on US letter sized paper (default). \emph{Please note that the paper format is typically also specified in the program that turns the \LaTeX{} output into PostScript. For example, some \texttt{dvips} installations have A4 as their default paper (typically those in Europe). In that case you have to call the \texttt{dvips} program with the option \texttt{-t letter} to ensure that the resulting PostScript file has the correct margins!} \item[4apaper] Directs the class to assume that the output is printed on A4 sized paper. \end{description} \subsubsection{Font selection} Five options control the selection of fonts in the document; use at most one of them. \begin{description} \item[mathptmx] Directs the class to use PostScript Times and Symbol fonts (a few missing glyphs are taken from Computer Modern) for math by loading the \texttt{mathptmx} package. This option is the default. This option does not support the |\boldmath| command since there exists no PostScript Symbol font in bold. It is possible, however to use |\mathbf| which allows you to get at least a bold Latin Alphabet. \item[mathptm] Directs the class to use PostScript Times and Symbol fonts but used the older package \texttt{mathptm} which has upright greek lowercase letters. This option does not support the |\boldmath| command since there exists no PostScript Symbol font in bold. It is possible, however to use |\mathbf| which allows you to get at least a bold Latin Alphabet. \item[mathtime] Directs the class to use MathTime fonts for math by loading the \texttt{mathtime} package. These fonts are commercial so that this option will not work if you don't own them. If this option is chosen one can also use the options for this package as global options to the class. \item[mtpro] Directs the class to use MathTime Professional fonts for math by loading the \texttt{mtpro} package. These fonts are commercial (the successors to the MathTime fonts from the previous option) so that this option will not work if you don't own them. If this option is chosen one can also use the options for this package as global options to the class. \item[nomathfonts] Directs the class not to set up math fonts (which means using the installation default which is usually Computer Modern). This option is intended in case a special math font setup is loaded in the document preamble. \item[cmfonts] Directs the class to use standard Computer Modern fonts for math and text. This does not conform to the specification for this class and is intended for draft preparation in environments where the required fonts are unavailable. \end{description} \subsubsection{Textual references} The next options enable textual references; if this is desired select one of them: \begin{description} \item[varioref] Loads the \texttt{varioref} package (see \cite[p.68ff]{}) allowing to produce textual page references. See section on Cross-references~\vpageref{xref} for details. \item[nonvarioref] Disables the |\reftextvario| command so that the strings produced by \texttt{varioref} commands will not depend on the number of references seen so far. Implies the varioref option. \end{description} \subsubsection{Table note markers} Notes to tables can be influenced as follows: \begin{description} \item[tnotealph] Produce raised lower case alphabetic marks to indicate table notes. \item[tnotesymbol] Use footnote symbols to indicate table notes (default). \end{description} \subsubsection{Citation mode} The citation mode can be influenced with the following two options: \begin{description} \item[numcites] Citations are typeset using numbers. Depending on the proceeding style these might appear raised or in brackets, etc.~(default). \item[bibliocites] Citations are typeset using an author/year scheme. This requires the installation of the \texttt{natbib} system. \end{description} In some layout styles these options might be without effect. \subsubsection{Heading numbers} Heading numbers can be turned on or off with the following two options: \begin{description} \item[numberedheadings] Headings are numbered. \item[unnumberedheadings] Headings are unnumbered (default). \end{description} In some layout styles these options might be without effect. \subsubsection{Drafts} Finally there is one standard \texttt{article} class option which has its functionality extended: \begin{description} \item[draft] Allows |\tableofcontents| and similar commands to work without error message (during development of article). It marks overfull boxes and also provides page numbers in the printout. \textbf{Remove this option when producing the final paper.} \end{description} \subsection{Front matter} The class supports an extended set of front matter commands. These commands differ from those used by standard \LaTeX's \texttt{article} class. Thus, if an article already written is adapted to be used with the \aipcls{}, the front matter has to be modified somewhat. Some of the commands below are required only for certain proceedings. Declarations that are not required will be silently ignored. \BDefC{title}[om]{short title}{title text} In standard \LaTeX{} this command has no optional argument. In the \aipcls{} one can specify an abbreviated title text which is used, for example, in the running footer in draft mode. \BDefC{author}[mm]{author name}{author information} In standard \LaTeX{} this command had only one argument containing both author name and address information. In this class it has two arguments and the second argument contains data structured using key/value pairs separated by commas. For example, the authors of this paper have been specified as: \begin{verbatim} \author{F. Mittelbach}{ address={Zedernweg 62, Mainz}, ,email= {}} \author{D. P. Carlisle}{ address={Willow House, Souldern}, ,email={}} \end{verbatim} Supported keywords will be \texttt{address}, \texttt{email}, \texttt{altaddress}, \texttt{homepage}, and \texttt{thanks}. (With release 1.3 of \aipcls{} only \texttt{address}, \texttt{altaddress} and \texttt{email} should be used; support for the other keywords will be added later.) Depending on the layout of the target proceedings some of the keys may get ignored! \BDefC{classification}[m]{data} Some proceedings require classification data, e.g., PACS numbers. If not, this declaration is ignored. \BDefC{keywords}[m]{data} Some layouts require keyword data. If not, this declaration is ignored. \BDefC{copyrightholder}[m]{name} Some layouts require copyright information. Normally a default is provided by the class. With this declaration the copyright holder can be overwritten. \BDefC{copyrightyear}[m]{year} Some layouts require copyright data. With this declaration the copyright year can be specified. (If such data is required the current year is provided as default). \BDefE{abstract} In contrast to standard \LaTeX{} the abstract environment has to appear before the |\maketitle| command. \BDefC{maketitle} This command inserts the actual front matter data. It has to follow the above declarations. \subsubsection{Multiple authors} Multiple authors are entered by specifying one |\author| command per author. Care needs to be taken when specifying shared addresses: they have to be absolutely identical. Depending on the chosen layout the class will merge such addresses but will recognize them only as identical, if the input including spaces is the same! The |\and| command as defined in the \texttt{article} class to separate multiple authors is not supported. \subsubsection{Dates} \BDefC{received}[m]{date} \BDefC{revised}[m]{date} \BDefC{accepted}[m]{date} Some layouts require specification of date of arrival, revision, and/or acceptance. The above declarations provide a way to specify such dates if necessary. \BDefC{date}[m]{date} The article class provides the |\date| command which is not used by \aipcls. If supplied it will be ignored unless the \texttt{draft} option is specified in which case it will show up in a footer line together with the title and the page number to ease document development. \subsubsection{Other front matter commands} The |\tableofcontents|, |\listoffigures|, and |\listoftables| commands are provided but produce (beside output) an error message unless the \texttt{draft} option was selected. This is done since the \aipcls{} does not support page numbering and thus the above commands essentially produce incorrect data. \subsection{Headings} The \aipcls{} officially supports three heading levels, i.e., |\section|, |\subsection|, and |\subsubsection|. It also supports the commands |\paragraph| and |\subparagraph| although the latter heading levels are not part of the \aipcls{} specification and are therefore discouraged. In some layouts |\section| headings are changed to UPPERCASE. Special care is taken not to uppercase math material, but this support is only available if the package |textcase| is part of the \LaTeX{} distribution. \subsection{Cross-references}\label{xref} Cross-references to page numbers are not possible with the \aipcls{} as the page numbers are determined after production. For this reason the |\pageref| command of \LaTeX{} is disabled by default. Since headings in most layouts do not carry numbers they can't be referenced either. References to tables, figures, and equations are possible using the \LaTeX{} commands |\label| and |\ref|. However if the class option \texttt{varioref} or \texttt{nonvarioref} is used, references to page numbers are possible again as they will generate textual references of the form ``on the following page'' or ``on an earlier page'' etc. The produced strings are customizable as described in detail in the \texttt{varioref} package documentation or in \cite[p.68ff]{}. The class defaults are as follows and can be changed with |\renewcommand| in the document preamble. The \texttt{varioref} package normally distinguishes between reference to facing pages and references to pages that need turning over using different strings in these cases. However, since with \aipcls{} class page numbers are not determined at the time of production no assumption can be made that page $x$ and $x+1$ actually fall onto the same double spread. For this reason the defaults used here do not produce strings containing the word ``facing'' or ``opposite''. \begin{verbatim} \renewcommand\reftextfaceafter {on the next page} \renewcommand\reftextfacebefore {on the \reftextvario{previous} {preceding} page} \renewcommand\reftextafter {on the \reftextvario{next} {following} page} \renewcommand\reftextbefore {on the \reftextvario{previous page}{page before}} \renewcommand\reftextcurrent {on \reftextvario{this} {the current} page} \end{verbatim} Normally, text for references which are ``far away'' are produced using |\reftextfaraway| in \texttt{varioref}. However, to produce textual references without referring to actual page numbers even in this case, this command was hijacked in the \aipcls{} and redefined to determine whether or not this is a reference to some earlier or later page. So instead of changing this command the class provides the following two commands for customization: \begin{verbatim} \renewcommand\reftextearlier {\reftextvario{on an earlier page}{earlier on}} \renewcommand\reftextlater {\reftextvario{later on} {further down}} \end{verbatim} To illustrate the result of this package all references in this document are made using |\vref| or |\vpageref|, e.g., references to Figure~\vref{fig:b} and Figure~\vref{fig:a}. These commands work best if used only for important references. Be careful when using them several times close to each other as the automatically generated texts then may sound strange (as they do in the example in this paragraph). \BDefC{eqref}[m]{label} For reference to equation numbers |\eqref| can be used instead of the standard |\ref| command. The |\eqref| command will automatically add any frills required by the layout style, while |\ref| will only typeset the plain number. For example, in the \texttt{arlo} style it will print ``Eq.~(1)'' while |\ref| would result in ``1''. \subsection{Lists} The \aipcls{} supports all standard list environments like \texttt{itemize}, \texttt{enumerate}, etc. \subsection{Graphics support} Support for including and manipulating graphics is provided as the standard \LaTeX{} \texttt{graphicx} package is automatically loaded by the \aipcls. For detailed descriptions of the commands made available by this package see~\cite{} or the package documentation coming with the \LaTeX{} release. A sufficient introduction is also given by~\cite{} although there only the \texttt{graphics} package (a subset of the \texttt{graphicx} package) is described. A typical application is given in the following example where a picture is resized to span 70\% of one column: \begin{verbatim} \begin{figure}[!b] \resizebox{.7\columnwidth}{!} {\includegraphics{escher}} \source{Guy Shaw} \caption{An illustration taken from~\cite{}} \label{fig:a} \end{figure} \end{verbatim} resulting in figure \vref{fig:a}. \begin{figure}[!b] \resizebox{.7\columnwidth}{!} {\includegraphics[draft=false]{escher}} \source{Guy Shaw} \caption{An illustration taken from~\cite{}} \label{fig:a} \end{figure} \subsection{Floats} Floats are objects which do not have to stay in sync with the running text but are allowed to move from their original place to some other position where they fit better for page breaking reasons. Such objects they are typically numbered so that they can be referenced from within the running text. \LaTeX{} by default supports two float types: figures and tables. These float types are also supported by the \aipcls{} although their internal implementation is quite different resulting in a number of important differences in behavior:\footnote{There exist packages that extend the number of float types. (This information is given as a footnote to show that footnotes in this class come out below a bottom float.)} \begin{itemize} \item The position of the float caption is determined automatically, independently of the placement of the |\caption| command within the float body. \item Depending on its width the float automatically spans two columns. In case of a table the whole object (including its caption) might be rotated automatically if its exceeds |\textwidth|. \item The body of the float environments are processed in L-R mode and not in paragraph mode as in standard \LaTeX. This is necessary for measuring its width. Thus if paragraph mode is needed one has to put a \texttt{minipage} environment of the appropriate width (e.g., |\columnwidth|) into the body. \item Only one |\caption| command per float is allowed. \end{itemize} \subsubsection{Figures} \BDefE{figure}[o]{pos} Like with standard \LaTeX{} the optional \Larg{pos} argument can be used to specify into which float areas this float is allowed to migrate (default is |tbp|). The environment \texttt{figure*} is not supported as figures that need to span both columns are automatically recognized in two column mode. \BDefC{source}[m]{text} Command to specify the origin of the picture shown. The \Larg{text} will be printed in small italics below the illustration. A typical example of a figure float would be \begin{verbatim} \begin{figure} \resizebox{.8\textwidth}{!} {\includegraphics{outline}} \caption{PostScript example taken from~\cite{}} \label{fig:b} \source{F. Mittelbach} \end{figure} \end{verbatim} The result is shown in Figure~\vref{fig:b}. \begin{figure} \resizebox{.8\textwidth}{!}{\includegraphics[draft=false]{outline}} \caption{PostScript example taken from~\cite{}} \label{fig:b} \source{F. Mittelbach} \end{figure} \BDefC{spaceforfigure}[mm]{horizontal}{vertical} If the illustration is to be manually pasted into the final document one can leave the right amount of space by using this command as follows: \begin{verbatim} \begin{figure} \spaceforfigure{2in}{1cm} \caption{Caption for a figure to be pasted in later} \label{fig:3} \source{F. Mittelbach} \end{figure} \end{verbatim} All standard \TeX{} units can be used to specify the space needed. The above example make room for an illustration that is two inches wide and one centimeter high. The result is shown as Figure~\vref{fig:3}. \begin{figure} \spaceforfigure{2in}{1cm} \caption{Caption for a figure to be pasted in later} \label{fig:3} \source{F. Mittelbach} \end{figure} \subsubsection{Tables} \BDefE{table}[o]{pos} Like with standard \LaTeX{} the optional \Larg{pos} argument can be used to specify into which float areas this float is allowed to migrate (default is |tbp|). The environment \texttt{table*} is not supported as tables that need to span both columns are automatically recognized in two column mode. Typically the body of the environment would consist of a \texttt{tabular} environment responsible for producing the actual table including the table and stub headers. \BDefC{tablehead}[mmmm]{cols}{h-pos}{v-pos}{heading text} To ease the production of tables the command |\tablehead| is provided which is essentially and abbreviation for a |\multicolumn| command that additionally boldens its text argument. I.e., \Larg{cols} specifies the number of columns the \Larg{heading text} should span and \Larg{h-pos} defines the horizontal positioning of the text of the column(s), e.g., |l|, |r|, |c|, or |p{...}|. In contrast to a simple |\multicolumn| command the \Larg{heading text} can be split vertically by using |\\| to denote the line breaks. The \Larg{v-pos} argument should contain either |t|, |c|, or |b| denoting the vertical placement of the text in relation to other cells of that row. It is only relevant if the \Larg{heading text} consists of more than one line. See the example table \vpageref[below]{} that demonstrates the use of this command. \BDefC{source}[m]{text} Command to specify the origin of the data given in the table. The \Larg{text} will be printed in small italics below the table. \BDefC{tablenote}[m]{text} Command to produce a note to the table. It can only be used within a \texttt{table} environment and should be used only at the right end of a table cell. The command produces a raised footnote symbol at the place used which sticks into the right margin. As far as \LaTeX{} is concerned this symbol does not occupy any space. Thus is will not modify the alignment of table columns. The \Larg{text} will appear below the table. In the current release notes to |\caption| or |\source| are not possible. \BDefC{tablenote*}[m]{text} Like |\tablenote| but this time the raised footnote symbol will occupy space. This version is intended to be used in the middle of cells. An example showing the use of all commands described above is shown in Table~\vref{tab:a}. It was produced by the following input:\label{} \begin{verbatim} \begin{table} \begin{tabular}{lrrrr} \hline &\tablehead{1}{r}{b}{Single\\outlet} &\tablehead{1}{r}{b}{Small\tablenote {2-9 retail outlets}\\multiple} &\tablehead{1}{r}{b}{Large\\multiple} &\tablehead{1}{r}{b}{Total} \\ \hline 1982 & 98 & 129 & 620 & 847\\ 1987 & 138 & 176 & 1000 & 1314\\ 1991 & 173 & 248 & 1230 & 1651\\ 1998\tablenote{predicted} & 200 & 300 & 1500 & 2000\\ \hline \end{tabular} \source{Central Statistical Office, UK} \caption{Average turnover per shop: by type of retail organisation} \label{tab:a} \end{table} \end{verbatim} \begin{table} \begin{tabular}{lrrrr} \hline & \tablehead{1}{r}{b}{Single\\outlet} & \tablehead{1}{r}{b}{Small\tablenote{2-9 retail outlets}\\multiple} & \tablehead{1}{r}{b}{Large\\multiple} & \tablehead{1}{r}{b}{Total} \\ \hline 1982 & 98 & 129 & 620 & 847\\ 1987 & 138 & 176 & 1000 & 1314\\ 1991 & 173 & 248 & 1230 & 1651\\ 1998\tablenote{predicted} & 200 & 300 & 1500 & 2000\\ \hline \end{tabular} \source{Central Statistical Office, UK} \caption{Average turnover per shop: by type of retail organisation} \label{tab:a} \end{table} \BDefC{setlength}[mm]{\texttt{\upshape\string\hlinesep}}{value} Vertical spacing between horizontal lines produced from |\hline| inside a tabular environment is controlled by the length parameter |\hlinesep| in this class. The default value (1pt) gives one point extra space above such lines and three times as much (i.e. 3pt) extra space below. This is done to implement the layout requirements for tables in the AIP proceedings (which are not supposed to have vertical lines in the tables). If tables with vertical lines are necessary for some reason, then the value of this parameter should be set to \texttt{0pt} either globally for the whole document or locally within the \texttt{table} environment. Otherwise the vertical lines will have strange gaps whenever a |\hline| command is used to produce a horizontal line. \subsubsection{Counters} The |\alph| and |\fnsymbol| commands to represent counter values have extended ranges. For example |\alph| will now count up to 52 (zz) and the |\fnsymbol| command will produce the following symbols \makeatletter \@fnsymbol{1}, \@fnsymbol{2}, \@fnsymbol{3}, \@fnsymbol{4}, \@fnsymbol{5}, \@fnsymbol{6}, \@fnsymbol{7}, \@fnsymbol{8}, \@fnsymbol{9}, \@fnsymbol{10}, \@fnsymbol{11}, \@fnsymbol{12}, \@fnsymbol{13}, \@fnsymbol{14}, \@fnsymbol{15}, and \@fnsymbol{16}. \makeatother This will allow for up to 16 table notes per table. For documents that need a larger number of table notes select the option \texttt{tnotealph} to switch to lower case alphabetic letters to mark such notes. \subsubsection{Long tables} Tables which are longer than one page cannot be placed into a \texttt{table} environment as floats cannot have a size larger than a page. Such tables are supported by the standard \LaTeX{} package \texttt{longtable} written by David Carlisle. However this package only works in single column mode. With two-column layouts, such as the one for the AIP 8x11 double column proceedings, such tables can only be added at the end of the paper by preceding the |longtable| environments with a |\onecolumn| declaration. The package is supported by the class in the sense that captions within a \texttt{longtable} environment will be formatted using the appropriate style; however in contrast to the \texttt{table} environment it is the responsibility of the user to place the caption at the top of the table. The commands |\source| and |\tablenote| are not supported within this environment, but the |\tablehead| command can be used to produce column heads if desired. Refer to the \texttt{longtable} package documentation or to \cite[p.122ff]{} for a detailed description of the syntax of the \texttt{longtable} environment. A possible alternative is the package \texttt{supertabular} written by Johannes Braams; however in this case no attempt has been made to ensure that a table produced with \texttt{supertabular} conforms to the layout specification for the \aipcls{}. Be aware that this package defines its own |\tablehead| command (with a completely different function). Refer to the package documentation for the syntax description. A detailed comparison between \texttt{supertabular} and \texttt{longtable} can be found in Chapter~5 of \cite{}. \subsubsection{Building floats manually} The original \LaTeX{} environments \texttt{figure} and \texttt{table} as well as their star forms are still available under the names \texttt{ltxfigure} and \texttt{ltxtable}. They should not be used in normal circumstances but are provided in case the automatism of the \aipcls{} needs overwriting. Please note that if these environments are used the position of the |\caption| command determines the placement of the caption within the float body and that the special commands for figures and tables, e.g., |\tablenote|, etc.\ as provided by this class are not available within these environments. \begin{table}[!t] \makeatletter \if8\expandafter\@car\selectedlayoutstyle\@nil\relax \else \fontsize{7}{8}\selectfont \fi \makeatother \begin{tabular}{rrrp{.6\textwidth}} \hline \tablehead{1}{r}{b}{File} & \tablehead{1}{c}{b}{Date} & \tablehead{1}{c}{b}{Version} & \tablehead{1}{c}{b}{Description} \\ \hline aipproc.cls & 2000/08/31 & v1.2a & AIP Proceedings (FMi) \\ fixltx2e.sty & 1999/12/01 & v1.0b & fixes to LaTeX \\ calc.sty & 1998/07/07 & v4.1b & Infix arithmetic (KKT,FJ) \\ ifthen.sty & 1999/09/10 & v1.1b & Standard LaTeX ifthen package (DPC) \\ graphicx.sty & 1999/02/16 & v1.0f & Enhanced LaTeX Graphics (DPC,SPQR) \\ keyval.sty & 1999/03/16 & v1.13 & key=value parser (DPC) \\ graphics.sty & 1999/02/16 & v1.0l & Standard LaTeX Graphics (DPC,SPQR) \\ trig.sty & 1999/03/16 & v1.09 & sin cos tan (DPC) \\ graphics.cfg & \\ dvips.def & 1999/02/16 & v3.0i & Driver-dependant file (DPC,SPQR) \\ url.sty & 1999/03/28 & ver 1.5x & Verb mode for urls, etc. \\ article.cls & 2000/05/19 & v1.4b & Standard LaTeX document class \\ size10.clo & 2000/05/19 & v1.4b & Standard LaTeX file (size option) \\ aipxfm.sty & \\ mathptm.sty & 2000/01/12 &PSNFSS-v8.1 &Times + math package (SPQR) \\ times.sty & 2000/01/12 &PSNFSS-v8.1 &Times font as default roman(SPQR) \\ ot1ptm.fd & 2000/01/12 &PSNFSS-v8.1 & font definitions for OT1/ptm. \\ fontenc.sty & \\ t1enc.def & 2000/08/30 & v1.91 &Standard LaTeX file \\ t1ptm.fd & 2000/01/12 &PSNFSS-v8.1 & font definitions for T1/ptm. \\ textcomp.sty & 2000/08/30 &v1.91 &Standard LaTeX package \\ ts1enc.def & 1998/06/12 & v3.0d & (jk/car/fm) Standard LaTeX file \\ varioref.sty & 1999/12/02 &v1.2c &package for extended references (FMi) \\ aip-8s.clo & \\ ttct0001.sty & \\ shortvrb.sty & 2000/07/04 &v2.0m & Standard LaTeX documentation package (FMi) \\ hyperref.sty & 2000/05/08 &v6.70f & Hypertext links for LaTeX \\ pd1enc.def & 2000/05/08 &v6.70f & Hyperref: PDFDocEncoding definition (HO) \\ hyperref.cfg & \\ hdvips.def & 2000/05/08 &v6.70f & Hyperref driver for dvips \\ pdfmark.def & 2000/05/08 &v6.70f & Hyperref definitions for pdfmark specials \\ ts1cmr.fd & 1999/05/25 &v2.5h & Standard LaTeX font definitions \\ nameref.sty & 2000/05/08 &v2.18 & Cross-referencing by name of section \\ t1pcr.fd & 2000/01/12 &PSNFSS-v8.1 & font definitions for T1/pcr. \\ ot1ptmcm.fd & 2000/01/03 &Fontinst v1.801 & font definitions for OT1/ptmcm. \\ omlptmcm.fd & 2000/01/03 &Fontinst v1.801 & font definitions for OML/ptmcm. \\ omspzccm.fd & 2000/01/03 &Fontinst v1.801 & font definitions for OMS/pzccm. \\ omxpsycm.fd & 2000/01/03 & Fontinst v1.801 &font definitions for OMX/psycm. \\ ts1ptm.fd & 2000/01/12 & PSNFSS-v8.1 &font definitions for TS1/ptm. \\ escher.eps & && Graphic file (type eps) \\ outline.eps & & & Graphic file (type eps) \\ \hline \end{tabular} \caption{Files used by the \aipcls{}} \label{tab:b} \source{Output of \texttt{\string\listfiles} when processing \texttt{aipguide.tex}} \end{table} \subsection{Urls} \BDefC{url}[m]{data} For documenting URLs and related data the |\url| command is provided. It allows breaking the URL in certain places and typesets it in an adequate font and format. Instead of using curly brackets the argument can be delimited by two identical characters not used in the argument. \subsection{Bibliography} Referring to other articles, books, etc.\ can be done using the |\cite| command of standard \LaTeX{}. The list of references itself can either be produced using standard \LaTeX{} methods or using \textsc{Bib}\TeX. If installed, the \aipcls{} class includes the \texttt{natbib} system which offers an extended set of citation commands. These commands have been originally developed to support author/year citation styles but are also useful with numerical citation styles. The \texttt{natbib} system has two basic citation commands, |\citet| and |\citep| for \emph{textual} and \emph{parenthetical} citations, respectively. There also exist the starred versions |\citet*| and |\citep*| that print the full author list, and not just the abbreviated one. All of these may take one or two optional arguments to add some text before and after the citation. Table~\vref{} shows some examples. \begin{table} \begin{tabular}{@{}l@{\quad$\Rightarrow$\quad}l} \hline \multicolumn{2}{@{}l}{\bfseries Author/year style} \\ \hline |\citet{jon90}| & Jones et al. (1990)\\ |\citet[chap.~2]{jon90}| & Jones et al. (1990, chap.~2)\\[0.5ex] |\citep{jon90}| & (Jones et al., 1990)\\ |\citep[chap.~2]{jon90}| & (Jones et al., 1990, chap.~2)\\ |\citep[see][]{jon90}| & (see Jones et al., 1990)\\ |\citep[see][chap.~2]{jon90}| & (see Jones et al., 1990, chap.~2)\\[0.5ex] |\citet*{jon90}| & Jones, Baker, and Williams (1990)\\ |\citep*{jon90}| & (Jones, Baker, and Williams, 1990) \\ \hline \multicolumn{2}{@{}l}{\bfseries Numerical style} \\ \hline |\citet{jon90}| & Jones et al. [21]\\ |\citet[chap.~2]{jon90}| & Jones et al. [21, chap.~2]\\[0.5ex] |\citep{jon90}| & [21]\\ |\citep[chap.~2]{jon90}| & [21, chap.~2]\\ |\citep[see][]{jon90}| & [see 21]\\ |\citep[see][chap.~2]{jon90}| & [see 21, chap.~2]\\[0.5ex] |\citep{jon90a,jon90b}| & [21, 32]\\ \hline \end{tabular} \caption{Example of \texttt{natbib} commands and their results} \label{} \end{table} There are many more commands and variants, see \cite{} or \cite{} for further details. \subsubsection{Bibliography produced manually} \BDefE{thebibliography}[m]{widest-label} Environment to hold the list of references. \BDefC{bibitem}[m]{label} Command to start a bibliographical entry having the label \Larg{label} for use in |\cite| commands. Refer to the publishers manual, e.g., \cite{}, for information on how to lay out individual entries. For example: \begin{verbatim} \bibitem{Brown2000} M.~P. Brown and K. Austin, \emph{The New Physique}, Publisher Name, Publisher City, 2000, pp. 212--213. \end{verbatim} If commands from \texttt{natbib} (e.g., from table~\ref{}) should be usable, then additional information has to be passed to the |\bibitem| via an optional argument. \BDefC{bibitem}[om]{display-info}{label} The optional argument \Larg{display-info} should then, and only then, contain the author(s) name(s) followed by the year in parentheses without any spaces, for example: \begin{verbatim} \bibitem[Brown and Austin(2000)] {Brown2000} ... \end{verbatim} The essential feature is that the label (the part in brackets) consists of the author names, as they should appear in the citation, with the year in parentheses following. There must be no space before the opening parenthesis! This will be automatically produced if \BibTeX{} is used. \subsubsection{Bibliography produced using \textsc{Bib}\TeX} The \aipcls{} is accompanied by \BibTeX{} style files which can be used to produce compliant reference lists from \BibTeX{} database files. To use \BibTeX{} one first has to run the source file through \LaTeX{} then run \BibTeX{} and then rerun \LaTeX{} twice to get all references resolved. \BibTeX{} is described in more detail in appendix B of \cite{} and in chapter~13 of \cite{}. \BDefC{bibliographystyle}[m]{style-name} This declaration specifies to \BibTeX{} that the style \Larg{style-name} should be used. It can be placed anywhere within the document but is usually positioned directly in front of the command described below. For a discussion which of the supplied \BibTeX{} styles should be used for which proceedings see the section ``Special requirements\ldots'' below. \BDefC{bibliography}[m]{bib-list} This command denotes the position where the reference list produced by \BibTeX{} will be included in the document. The \Larg{bib-list} is a comma separated list of \BibTeX{} database files. \section{General requirements and restrictions} This class was designed to work with \LaTeXe{} release 1999/06/01 or a later version. Earlier releases may work but have not been tested. With the exception of the packages \texttt{natbib} and \texttt{url} it only requires files which are part of a standard \LaTeX{} distribution, i.e., it should work if your installation contains the following components: \texttt{base}, \texttt{tools}, \texttt{graphics}, and \texttt{psnfss}, see \vref{tab:b} for files used to produce this document. The most recent \LaTeX{} distribution as well as \texttt{natbib} and \texttt{url} can be obtained from CTAN sites (Comprehensive \TeX{} Archive Network). Refer to \url{} for more information on CTAN and \TeX{} in general. A ready to run \TeX{} system for various platforms which has everything required is available on CD-ROM, look into \url{}. This \TeX{} implementation is also made available as an add-on to several books on \LaTeX, e.g., \cite{}. For loading individual packages from a CTAN archive refer to \url{} and search for the package name. Please omit extensions such as \texttt{.sty} when searching, e.g., search for \texttt{natbib} rather than \texttt{natbib.sty}, as such packages are often distributed in source form only, e.g., as a \texttt{.dtx} file. It is also possible to download a complete \TeX/\LaTeX{} installation from CTAN, e.g., Miktex + Winedit + Ghostview. Finally, it is also possible to download a CD-ROM image of the \TeX-live CD from CTAN (roughly 300MB): search for \texttt{texlive} (and make sure you select a suitable mirror near you). \section{Special requirements for individual layouts} \subsection{AIP proceeding layout 6x9} \begin{itemize} \raggedright \item The entire paper will be reduced 15\% in the printing process. Please make sure all figures as well as the text within the figures are large enough in the manuscript to be readable in the finished book. \item The use of the |\source| command is discouraged. \item Compliant \BibTeX{} styles are \texttt{aipproc} (for use with \texttt{natbib}) and \texttt{aipprocl} (if \texttt{natbib} is missing at the site). \item The options \texttt{bibliocites} and \texttt{numberedheadings} have no effect. \end{itemize} \subsection{AIP proceeding layout 8x11 single/double} \begin{itemize} \raggedright \item The use of the |\source| command is discouraged. \item Compliant \BibTeX{} styles are \texttt{aipproc} (for use with \texttt{natbib}) and \texttt{aipprocl} (if \texttt{natbib} is missing at the site). \item The options \texttt{bibliocites} and \texttt{numberedheadings} have no effect. \end{itemize} \subsection{ARLO} Note: the ARLO layout is no longer supported. \begin{itemize} \raggedright \item A copyright year (|\copyrightyear|) needs to be provided. \item Pacs numbers should be provided (|\classification|). \item The \texttt{arlo} layout offers one additional environment to specify multimedia files: \begin{verbatim} \begin{multimedia} \multimediauid{523} \multimediatype{gif} \multimediasize{1.2Mb} \multimediaurl{ } \multimediacaption{Fancy video} \label{fv} \end{multimedia} \end{verbatim} References to a multimedia file can be made using |\label| and |\ref|. Instead of the latter command |\multimediaref| can be used to automatically get the appropriate frills, e.g., `Mm.~2' instead of just `2' as produced by |\ref|. \item Select the \texttt{draft} option for the initial submission and the copy-editing stage. Replace it by the \texttt{final} option when producing the final paper, so that page numbers and other items are stripped away. \item To conform to the layout specification for citations the \texttt{natbib} system has to be installed. \item For ARLO two compliant \BibTeX{} styles are available: \texttt{arlonum} should be used together with the class option \texttt{numcites}, while \texttt{arlobib} should be used together with the option \texttt{bibliocites}. \item The options \texttt{bibliocites} and \texttt{numberedheadings} can be used to switch to author/year citation scheme and numbered headings, respectively. \end{itemize} \section{Introduction} Amongst the major shortcomings of the Standard Model are the neutrino mass and the dark matter stability problem. Underpinning the origin of neutrino mass and elucidating the nature of dark matter would constitute a gigantic step forward in particle physics. Here we focus on the possibility that the \textbf{neutrino nature and dark matter stability problem may be closely interconnected}. Concerning neutrinos a major unknown is whether they are their own anti - particles, an issue which has remained an open challenge ever since Ettore Majorana had his pioneering idea on the quantum mechanics of spin \cite{}. On the other hand, since many years, physicists have pondered about what is the dark matter made of, and what makes it stable, a property usually assumed in an ad-hoc fashion. Indeed, although the existence of non-baryonic dark matter is well established by using cosmological and astrophysical probes, its nature has otherwise remained elusive. Recently it has been argued in \cite{} that the Dirac or Majorana nature of neutrinos is intimately connected to the breaking pattern of $U(1)_L$ Lepton number symmetry, which is accidentally conserved in the Standard Model, to its residual conserved $Z_n$ subgroups. Interestingly enough, neutrinos can only be Majorana for specific $Z_n$ groups and even within these specific $Z_n$ groups, only for special choices of the charges, while they will be Dirac in any other case. Here we focus on the $ U(1)_L \rightarrow {Z}_4$ scenario, which we call \textbf{Quarticity}. We show that \textbf{a WIMP dark matter candidate can naturally emerge, stabilized by Quarticity, the same symmetry associated to the Diracness of neutrinos}. Moreover, the smallness of neutrino masses can be understood through the Dirac analogue of type-I seesaw. We also show that many different symmetry groups can be used to give the flavour structure of neutrino mixings and therefore \textbf{a rich variety of models arise}, showing different features and predictions but all maintaining the principal result, \textbf{the connection between Dark Matter and the Dirac nature of neutrinos}. \section{The model} The general structure of the model is given in \cite{}. Concerning the particle content of the model, we need to introduce some new fields in addition to those in the Standard Model. In the fermionic sector we introduce \textbf{3 generations of right handed neutrinos}, $\nu_{i,R}$, to obtain neutrino masses, as well as \textbf{3 more generations of heavy neutral leptons} with the two chiralities, wich we will call $N_{i, L}$ and $N_{i,R}$, in order to have a seesaw mechanism. We also have to introduce a new scalar, $\chi$, which has a \textbf{non - zero vev}. In addition there are two other scalars carrying non-trivial $Z_4$ charges, a real scalar $\eta$, and a complex scalar $\zeta$, which is the \textbf{DM candidate}. All of them are gauge singlets. The set of transformation rules under $Z_4$ for the fields in the model is: \pagebreak \begin{table}[ht] \centering \begin{tabular}{c c || c c} \hline \hline \textbf{Fields} \hspace{5mm} & ${Z}_4 $ \hspace{5mm} & \textbf{Fields} \hspace{5mm} & $ Z_4 $ \hspace{5mm} \\ \hline \hline \textbf{Fermions} \hspace{5mm} & \hspace{5mm} & \hspace{5mm} & \hspace{5mm} \\ \hline $\bar{L}_{i,L}$ \hspace{5mm} & $\mathbf{z}^3$ \hspace{5mm} & $\nu_{i,R}$ \hspace{5mm} & $\mathbf{z}$ \hspace{5mm} \\ $ l_{i,R} $ \hspace{5mm} & $\mathbf{z}$ \hspace{5mm} & $\bar{N}_{i,L}$ \hspace{5mm} & $\mathbf{z}^3$ \hspace{5mm} \\ $N_{i,R}$ \hspace{5mm} & $\mathbf{z}$ \hspace{5mm} & \hspace{5mm} & \hspace{5mm} \\ \hline \textbf{Scalars (vev $\neq $0)} \hspace{5mm} & \hspace{5mm} & \hspace{5mm} & \hspace{5mm} \\ $\Phi$ \hspace{5mm} & $\mathbf{1}$ \hspace{5mm} & $\chi$ \hspace{5mm} & $\mathbf{1}$ \hspace{5mm} \\ \hline \textbf{Scalars (vev=0)} \hspace{5mm} & \hspace{5mm} & \hspace{5mm} & \hspace{5mm} \\ $\zeta$ \hspace{5mm} & $\mathbf{z}$ \hspace{5mm} & $\eta$ \hspace{5mm} & $\mathbf{z}^2$ \hspace{5mm} \\ \hline \hline \end{tabular} \caption{Charge assignments for leptons, quarks, scalars ($\Phi_i^u$, $\Phi_i^d$ and $\chi_i$) as well as ``dark matter sector'' ($\zeta$ and $\eta$). Here $\mathbf{z}$ is the fourth root of unity, i.e. $\mathbf{z}^4 = 1$. } \label{tab1} \end{table} The only requirement for the flavour structure of the model is that \textbf{the flavour symmetry group $G$ forbids the tree level neutrino mass}, this is, $\bar{L} \phi^c \nu_R$. Then \textbf{neutrino masses are generated via a type-I seesaw}. \subsection{Neutrino Diracness} \textbf{The symmetry rules under ${Z}_4$ ensures that neutrinos are Dirac particles by forbidding all the possible Majorana terms at all orders. } Note that the scalars which have a non-zero vev, $\Phi$ and $\chi$, transform trivially under ${Z}_4$, thus implying that \textbf{the Quarticity symmetry ${Z}_4$ is not spontaneously broken}. Moreover \begin{equation} (\Phi^\dagger \Phi)^n \chi^m \sim 1 \end{equation} \noindent This implies that if all the Majorana mass terms are forbidden at tree level, then they are forbidden at all orders. This is indeed the case, because \textbf{all neutral fermionic field transform as $z$ under ${Z}_4$}. Therefore, all the Majorana masses at tree level are forbidden. For example \begin{equation} \bar{\nu}_R^c \nu_R \sim z^2 \end{equation} \noindent The higher order mass terms -via non-zero vevs of the scalars $\Phi$ and $\chi$- also transform non-trivially under the quarticity symmetry \begin{equation} \bar{\nu}_R^c \nu_R (\Phi^\dagger \Phi)^n \chi^m \sim z^2 \end{equation} Therefore, these terms are forbidden and \textbf{neutrinos remain Dirac particles at all orders, while the symmetry ${Z}_4$ is not broken after spontaneous symmetry breaking}. \subsection{A stable Dark Matter candidate} To ensure that neutrinos remain Dirac particles, the ${Z}_4$ quarticity symmetry should remain exact, so that no scalar carrying a ${Z}_4$ charge should acquire any vev. Thus \textbf{both $\zeta$ and $\eta$ which carry ${Z}_4$ charge can potentially be stable as a result of the unbroken quarticity symmetry}. However, owing to the ${Z}_4$ charge assignments, the $\eta$ field has cubic couplings to both scalars and fermions. These couplings lead to the decay of $\eta$ and thus, despite carrying a ${Z}_4$ charge, \boldmath $\eta$ \unboldmath \textbf{is not stable}. On the other hand, in the lagrangian there is no term of the form $\zeta \rho_i \rho_j$, where $\rho$ is a generic scalar, nor of the form $\zeta \psi_i \psi_j$, where $\psi$ is a generic fermion. This implies that \boldmath $\zeta$ \unboldmath \textbf{is an stable particle, thus it is a viable dark matter candidate}. \vspace{3mm} \noindent The detection could be direct, via \textbf{nuclear recoil}. The WIMP particle will also contribute to the \textbf{invisible decay of the Higgs}. Both features are shown in the following diagrams \ref{} \begin{figure}[!h] \centering \includegraphics[height=4cm]{dirdet.pdf} \hspace{10mm} \includegraphics[height=4cm]{invis_dec.pdf} \caption{Left panel: Leading order Feynman diagram of the nuclear recoil of the dark matter candidate via a exchange of the Higgs boson. Right panel: Tree level Feynman diagram of the invisible decay of the Higgs boson to a pair of DM particles.} \label{} \end{figure} \noindent \textbf{The constraints from LHC for invisible decay and the for direct searches coming from LUX experiment rules out part of the parameter space regarding the higgs-dark matter quartic coupling $\lambda_{h\zeta}$ and mass of dark matter $m_\zeta$}, which can be seen in the figure \ref{}. \vspace{3mm} \begin{figure}[!h] \centering \includegraphics[height=4cm]{Searches.pdf} \caption{Parameter space excluded by LHC invisible decay data and LUX direct searches in the $\lambda_{h\zeta}$ vs $m_\zeta$ plane.} \label{} \end{figure} \section {Flavour predictions} The model includes, apart from the Standard Model gauge group, a global symmetry $Z_4 \otimes G$, where $Z_4$ is the Quarticity symmetry already discussed while $G$ is the flavour symmetry. It's role is to forbid the tree level coupling between left and right handed neutrinos and this can be trivially done using $Z_2$, as done in \cite{}. However, one can use non-abelian discrete symmetries to extract very interesting flavour predictions \subsection{$\Delta_{27}$ as flavour symmetry} This model was discussed in \cite{}. \textbf {When we take \boldmath $\Delta(27)$ \unboldmath as the flavour symmetry}, $G$, an interesting structure appears in the neutral fields mass matrix, which can then be diagonalized and the masses and mixing parameters can be extracted. The parameters in the model have enough freedom to \textbf{fit all the mixing angles and neutrino masses constraints}. The charged assignments are shown in Table \ref{tab2}. \begin{table}[h] \centering \begin{tabular}{c c c || c c c} \hline \hline Fields \hspace{1cm} & $\Delta(27)$ \hspace{1cm} & $Z_4$ \hspace{1cm} & Fields \hspace{1cm} & $\Delta(27)$ \hspace{1cm} & $Z_4$ \\ \hline \hline $\bar{L}_e$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{z}^3$ \hspace{1cm} & $\nu_{e,R}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{z}$ \\ $\bar{L}_\mu$ \hspace{1cm} & $\mathbf{1''}$ \hspace{1cm} & $\mathbf{z}^3$ \hspace{1cm} & $\nu_{\mu, R}$ \hspace{1cm} & $\mathbf{1'}$ \hspace{1cm} & $\mathbf{z}$ \\ $\bar{L}_\tau$ \hspace{1cm} & $\mathbf{1'}$ \hspace{1cm} & $\mathbf{z}^3$ \hspace{1cm} & $ \nu_{\tau,R}$ \hspace{1cm} & $\mathbf{1''}$ \hspace{1cm} & $\mathbf{z}$ \\ $ l_{i,R} $ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{z}$ \hspace{1cm} & $\bar{N}_{i,L}$ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{z}^3$ \\ $N_{i,R}$ \hspace{1cm} & $\mathbf{3'}$ \hspace{1cm} & $\mathbf{z}$ \hspace{1cm} & \hspace{1cm} & \hspace{1cm} & \\ \hline $\Phi_i$ \hspace{1cm} & $\mathbf{3'}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\chi_i$ \hspace{1cm} & $\mathbf{3'}$ \hspace{1cm} & $\mathbf{1}$ \\ $\zeta$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{z}$ \hspace{1cm} & $\eta$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{z}^2$ \\ \hline \end{tabular} \caption{ The $\Delta (27)$ and $Z_4$ charge assignments for leptons, the Higgs scalars ($\Phi_i$, $\chi_i$) and the dark matter sector scalars ($\zeta$ and $\eta$). Here $\mathbf{z}$ is the fourth root of unity, i.e. $\mathbf{z}^4 = 1$.} \label{tab2} \end{table} The interesting scenario arises when we consider the following alignment for the vevs of the scalars \begin{eqnarray} \langle \Phi_i \rangle = v \hspace{1mm}(1,1,1)\\ \langle \chi_i \rangle = u \hspace{1mm}(1, 1+\epsilon, 1+\alpha) \end{eqnarray} \noindent Where $\epsilon << 1$. In this limit, one can find a \textbf{correlation between the CP violation parameter, \boldmath $\delta_{CP}$ \unboldmath, and the deviation from the alignment limit which is parametrized by \boldmath $\epsilon$ \unboldmath}. This is shown in the figure \ref{delta}. \begin{figure}[!h] \centering \includegraphics[height=5cm]{jcp_inv_e3_12.pdf} \caption{Leptonic CP violation phase $\delta_{CP}$ versus $\epsilon$, the deviation from the reference alignment. We have taken $\alpha = 1.2$. See text. } \label{delta} \end{figure} \subsection{$A_4$ as flavour symmetry} This realization of the quarticity model was done in \cite{}. Taking \boldmath $A_4$\unboldmath \textbf{ as the flavour symmetry} $G$ gives very different results. On the one hand, the model gives some \textbf{strong predictions} \begin{eqnarray} \theta^\nu_{23} & = & 45^\circ, \quad \quad \delta^\nu \, = \, \pm \, 90^\circ \nonumber \\ \theta^\nu_{12} & = & \rm{arbitrary} \quad \quad \theta^\nu_{13} \, = \, \rm{arbitrary} \label{unu} \end{eqnarray} \noindent While $\theta_{12}^\nu$ and $\theta_{13}^\nu $ are strongly correlated. However, \textbf{this correlation does not allow the fitting of the experimental constraints} for these two angles. The model can be \textbf{minimally extended} to include another set of Higgs doublet, which only couples to the up quarks and the neutrinos, while the other Higgs doublet only couples to the d-type quarks and the charged leptons. This way, \textbf{not only the mixing parameters can be fitted but also a series of interesting predictions arise}. The charged assignments are \pagebreak \begin{table}[!h] \centering \begin{tabular}{c c c c || c c c c} \hline \hline Fields \hspace{1cm} & $SU(2)_L$ \hspace{1cm} & $A_4$ \hspace{1cm} & $Z_4$ \hspace{1cm} & Fields \hspace{1cm} & $SU(2)_L$ \hspace{1cm} & $A_4$ \hspace{1cm} & $Z_4$ \\ \hline \hline $\bar{L}_i$ \hspace{1cm} & $\mathbf{2}$ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{z}^3$ \hspace{1cm} & $\nu_{e,R}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{z}$ \\ $\bar{N}_{i,L}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{z}^3$ \hspace{1cm} & $\nu_{\mu, R}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{1'}$ \hspace{1cm} & $\mathbf{z}$ \\ $N_{i,R}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{z}$ \hspace{1cm} & $ \nu_{\tau,R}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{1''}$ \hspace{1cm} & $\mathbf{z}$ \\ $ l_{i,R} $ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{z}$ \hspace{1cm} & $ d_{i,R} $ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{z}$ \\ $ \bar{Q}_{i,L} $ \hspace{1cm} & $\mathbf{2}$ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{z}^3$ \hspace{1cm} & $ u_{i,R} $ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{z}$ \\ \hline $\Phi_1^u$ \hspace{1cm} & $\mathbf{2}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\chi_i$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{1}$ \\ $\Phi_2^u$ \hspace{1cm} & $\mathbf{2}$ \hspace{1cm} & $\mathbf{1'}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\eta$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{z}^2$ \\ $\Phi_3^u$ \hspace{1cm} & $\mathbf{2}$ \hspace{1cm} & $\mathbf{1''}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\zeta$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & $\mathbf{z}$ \\ $\Phi_i^d$ \hspace{1cm} & $\mathbf{2}$ \hspace{1cm} & $\mathbf{3}$ \hspace{1cm} & $\mathbf{1}$ \hspace{1cm} & \hspace{1cm} & \hspace{1cm} & \hspace{1cm} & \\ \hline \end{tabular} \caption{Charge assignments for leptons, quarks, scalars ($\Phi_i^u$, $\Phi_i^d$ and $\chi_i$) as well as ``dark matter sector'' ($\zeta$ and $\eta$). Here $\mathbf{z}$ is the fourth root of unity, i.e. $\mathbf{z}^4 = 1$. } \label{tab3} \end{table} A strong correlation between $\delta_{CP}$ and $\delta_{23}$ appears as shown in figure \ref{} \begin{figure}[!h] \centering \includegraphics[scale=0.38]{deltath23Regionmodgreen.pdf} \includegraphics[scale=0.36]{JCPsinth23mod.pdf} \caption{CP violation and $\theta_{23}$ predictions within the model. Left panel: $\delta_{CP}$ vs $\theta_{23}$. The green regions are the $1\sigma$ (dark) and $3\sigma$ (light) regions for $\theta_{23}$ from current oscillation fit. Right panel: Same correlation, now showing $J_{CP}$ vs $\sin^2 \theta_{23}$ and zooming in the region allowed by the model, fully consistent in the $2\sigma$ experimental range.} \label{} \end{figure} \noindent Also, \textbf{only normal hierarchy of neutrinos is realized}. Moreover, an interesting \textbf{relation between the masses of the quarks and the charged leptons} is also present in the model. \begin{equation} \label{} \frac{m_b}{{\sqrt{m_d m_s}}} \approx \frac{m_\tau}{{\sqrt{m_e m_\mu}}} \end{equation} \noindent In this case, the flavour symmetry predicts \textbf{(i)} a generalized \textbf{bottom-tau mass relation} involving all families, \textbf{(ii)} small neutrino masses are induced via a \textbf{type-I seesaw}, \textbf{(iii)} \textbf{CP must be significantly violated} in neutrino oscillations, \textbf{(iv)} the atmospheric angle \boldmath $\theta_{23}$ \unboldmath \textbf{lies in the second octant} while showing a \textbf{strong correlation with \boldmath $\delta_{CP}$ \unboldmath} and \textbf{(v)} only the \textbf{normal neutrino mass ordering is realized.} A more detailed phenomenological analysis of the oscillation predictions of the model has been recently done in \cite{}. \section{Conclusions} We have shown \textbf{a very rich family of predictive models} which can accommodate many different setups while conserving the central features of the model, this is \begin{itemize} \item The quarticity symmetry ${Z}_4$ ensures that neutrinos are Dirac particles and the stability of the DM candidate. \item Naturally small neutrino masses are generated via a type-I seesaw. \end{itemize} \noindent Then, the freedom to select the flavour symmetry $G$ leads to a \textbf{whole variety of models with rich predictions}, not only in the neutrino sector but also in the quark and charged lepton sector. \begin{theacknowledgments} This article is based on the poster presentation done in the Pontercorvo school 2017 in Prague. The original works were done in collaboration with Jos\'e W. F. Valle, Rahul Srivastava and Ernst Ma. Work supported by the Spanish grants FPA2017-85216-P and SEV-2014-0398 (MINECO), PROMETEOII/2014/084 (Generalitat Valenciana) and Severo Ochoa IFIC FPI. \end{theacknowledgments} \bibliographystyle{aipproc} \IfFileExists{\jobname.bbl}{} {\typeout{} \typeout{******************************************} \typeout{** Please run "bibtex \jobname" to obtain} \typeout{** the bibliography and then re-run LaTeX} \typeout{** twice to fix the references!} \typeout{******************************************} \typeout{} } \section{Introduction} \ifthenelse{\equal\selectedlayoutstyle{6x9}}{\par\bfseries Note: The entire paper will be reduced 15\% in the printing process. Please make sure all figures as well as the text within the figures are large enough in the manuscript to be readable in the finished book.\par\bfseries Note: The entire paper will be reduced 15\% in the printing process. Please make sure all figures as well as the text within the figures are large enough in the manuscript to be readable in the finished book.\par\bfseries Note: The entire paper will be reduced 15\% in the printing process. Please make sure all figures as well as the text within the figures are large enough in the manuscript to be readable in the finished book.\normalfont}{} Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. Some url test \url{}. \section{Important words of the universe are shown in this section} \subsection{A B-head} Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile\footnote{Here we test footnotes.} regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere nostros et\footnote{Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos.} breviter Troiae supremum audire laborem, quamquam animus meminisse\footnote{Here we test footnotes.} horret, luctuque refugit, incipiam. Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere\footnote{Here we test footnotes.} nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. \subsubsection{A C-head} Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus\doingARLO{ \mmref{fv}} cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. \doingARLO{ \begin{multimedia} \mmuid{523} \mmtype{gif} \mmsize{1.2Mb} \mmurl{} \mmcaption{Fancy video} \label{fv} \end{multimedia}} Et iam nox umida caelo praecipitat $J_{ion}$ suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere nostros et breviter Troiae $J_{ion}$ supremum audire laborem: \begin{equation} J_{ion}=A\frac{exp\left[-\frac{E_a}{kT}\right]}{kT}j \label{ionflux} \end{equation} lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. $A$ talia fando, $E_a$ iam nox umida, $k$ caelo praecipitat, suadentque cadentia sidera somnos. See \eqref{ionflux}. \paragraph{And a D-head} Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui \cite{Brown2000,}. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet \cite{Mittelbach/} a lacrimis? Et iam nox umida caelo praecipitat, suadentque \cite{Wang} cadentia sidera somnos. \doingARLO{ \begin{multimedia} \mmuid{524} \mmtype{wav} \mmsize{300k} \mmurl{} \mmcaption{Fancy Wave} \end{multimedia} } \begin{figure} \reflectbox{\includegraphics[height=.3\textheight]{golfer}} \caption{Quis talia fando Myrmidonum} \end{figure} \begin{figure}[b] \resizebox{12pc}{!}{\includegraphics{outline}} \caption{Quis talia fando Myrmidonum} \end{figure} \begin{figure} \caption{Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui.} \includegraphics[height=.5\textheight]{golfer} \end{figure} Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? \begin{table}[!t] \begin{tabular}{rrrp{.7\textwidth}} \hline \tablehead{1}{r}{b}{File} & \tablehead{1}{c}{b}{Date} & \tablehead{1}{c}{b}{Version} & \tablehead{1}{c}{b}{Description} \\ \hline aipproc.cls & 1998/08/21 & v0.9a & AIP Proceedings (FMi) \\ calc.sty & 1997/11/11 & v4.0e & Infix arithmetic (KKT,FJ) \\ graphicx.sty & 1997/06/09 & v1.0d & Enhanced LaTeX Graphics (DPC,SPQR) \\ keyval.sty & 1997/11/10 & v1.10 & key=value parser (DPC) \\ graphics.sty & 1997/09/09 & v1.0f & Standard LaTeX Graphics (DPC,SPQR) \\ trig.sty & 1994/10/16 & v1.08 & sin cos tan (DPC) \\ dvips.def & 1997/09/09 & v3.0e & Driver-dependant file (DPC,SPQR) \\ article.cls & 1998/05/05 & v1.3y & Standard LaTeX document class \\ size10.clo & 1998/05/05 & v1.3y & Standard LaTeX file (size option) \\ mathptm.sty & 1998/01/07 & PSNFSS v.7 & Times + math package from fontinst : S Rahtz \\ fontenc.sty & 1998/06/12 & v1.9p & Standard LaTeX package \\ t1enc.def & 1998/06/12 & v1.9p & Standard LaTeX file \\ times.sty & 1998/01/07 & PSNFSS v.7 & Times font as default roman : S Rahtz \\ textcomp.sty & 1998/06/12 & v1.9p & Standard LaTeX package \\ ts1enc.def & 1998/06/12 & v3.0d & (jk/car/fm) Standard LaTeX file \\ varioref.sty & 1997/12/06 & v1.1d & package for extended references (FMi) \\ fix2col.sty & 1998/08/17 & v0.03 & Output Routine fixes for two column mode (DPC,FMi) \\ shortvrb.sty & 1998/05/19 & v2.0b & Standard LaTeX documentation package (FMi) \\ ts1cmr.fd & 1998/03/27 & v2.5g & Standard LaTeX font definitions \\ t1ptm.fd & 1997/09/30 & Fontinst v1.6 & font definitions for T1/ptm. \\ t1pcr.fd & 1997/09/30 & Fontinst v1.6 & font definitions for T1/pcr. \\ ot1ptmcm.fd & 1997/09/30 & Fontinst v1.6 & font definitions for OT1/ptmcm. \\ omlptmcm.fd & 1997/09/30 & Fontinst v1.6 & font definitions for OML/ptmcm. \\ omspzccm.fd & 1997/09/30 & Fontinst v1.6 & font definitions for OMS/pzccm. \\ omxpsycm.fd & 1997/09/30 & Fontinst v1.6 & font definitions for OMX/psycm. \\ ot1ptm.fd & 1997/09/30 & Fontinst v1.6 & font definitions for OT1/ptm. \\ ts1ptm.fd & 1997/09/30 & Fontinst v1.6 & font definitions for TS1/ptm. \\ \hline \end{tabular} \caption{Sample of a wide table that is automatically being turned sideways} \label{tab:b} \end{table} \begin{figure} \resizebox{6pc}{!}{\includegraphics{escher}} \resizebox{6pc}{!}{\reflectbox{\includegraphics{escher}}} \caption{The caption} \end{figure} Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia\footnote{A few more footnotes} fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia\footnote{Here we test footnotes.} sidera somnos. Sed si tantus amor casus cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. In the following we test itemize environments up to the forth level. \begin{itemize} \item An item with more than a line of text. Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai. \item Another item with sub entries \begin{itemize} \item A sub entry. \item Second sub entry. \begin{itemize} \item A sub sub entry. \begin{itemize} \item A sub sub sub entry. \item Second sub sub sub entry. \end{itemize} \item Second sub sub entry. \end{itemize} \end{itemize} \item A final item. \end{itemize} \begin{table} \begin{tabular}{lrrrr} \hline & \tablehead{1}{r}{b}{Single\\outlet} & \tablehead{1}{r}{b}{Small\tablenote{2-9 retail outlets}\\multiple} & \tablehead{1}{r}{b}{Large\\multiple} & \tablehead{1}{r}{b}{Total} \\ \hline 1982 & 98 & 129 & 620 & 847\\ 1987 & 138 & 176 & 1000 & 1314\\ 1991 & 173 & 248 & 1230 & 1651\\ 1998\tablenote{predicted} & 200 & 300 & 1500 & 2000\\ \hline \end{tabular} \caption{Average turnover per shop: by type of retail organisation} \label{tab:a} \end{table} Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a \cite{} lacrimis? In the following we test enumrerate environments up to the second level. In addition we look how ridiculous large labels look. \begin{enumerate} \item An item \cite{} \item Another item with sub entries \begin{enumerate} \item A sub entry \cite{Wang} \item Second sub entry \end{enumerate} \item The final item with normal label. \end{enumerate} Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? \begin{description} \item[Infandum] regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai. \item[Sed] si tantus amor casus cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. \item[Lamentabile] regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? \end{description} Sed si tantus amor casus cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus \cite{} cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a \cite{} lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere nostros et breviter Troiae supremum audire laborem, quamquam animus meminisse horret, luctuque refugit, incipiam. \section{Infandum, regina, iubes renovare dolorem} Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Et iam nox umida caelo praecipitat, suadentque cadentia sidera somnos. Sed si tantus amor casus cognoscere nostros et breviter Troiae supremum audire \cite{} laborem, quamquam animus meminisse horret, luctuque refugitum, refugit, incipitat, suadenovare dolorem, Troianas ut opes Ulixi temperet breviter Troiaeque ipse nostros et a lacrimis? Infandum, regina, iubes renovare dolor casus cognosceret a lacrimis? Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint \cite{} Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? \begin{theacknowledgments} Infandum, regina, iubes renovare dolorem, Troianas ut opes et lamentabile regnum cruerint Danai; quaeque ipse miserrima vidi, et quorum pars magna fui. Quis talia fando Myrmidonum Dolopumve aut duri miles Ulixi temperet a lacrimis? \end{theacknowledgments} \doingARLO[\bibliographystyle{aipproc}] {\ifthenelse{\equal{\AIPcitestyleselect}{num}} {\bibliographystyle{arlonum}} {\bibliographystyle{arlobib}} }

\section{Organized Behavior Characteristics} \label{} In Twitter approximately $6,000$ tweets are posted per second \footnote{}. With such huge numbers of tweets and users, getting messages seen is not a trivial task. In order to increase the likelihood of the message's observation, some strategies are required. To gain insight about how organized behavior is manifested in Twitter, specifically in the political domain, we manually inspected tweets with chosen top trend hashtags. These hashtags are selected as candidate organized hashtags, if they seemed suspicious after reading some of their tweets. To get larger numbers of such tweets, Twitter Search API is used to fetch tweet sets for each candidate hashtag. To gain an overall impression, the following values are computed for each tweet set: the percentage of distinct words ($DW(\%)$), the average tweet count per user ($TPU(\mu)$), the percentage of retweets($RT(\%)$), and the variance and standard deviation of hashtags ($HT(\sigma^2)$ and $HT(\sigma)$). This enabled a comparison among tweet sets, which are shown in Table~\ref{}. A low value of $DW(\%)$ indicates a lack of diversity in vocabulary, while a high value of $TPU(\mu)$ shows that the users who posted in the collection tend to repeatedly post the same hashtag. Also, high values of $HT(\sigma^2)$ \& $HT(\sigma)$ indicate the use of multiple hashtags in the collection. This can be observed in the so-called viral activity, where multiple hashtags are used to reach as many users as possible. Similarly, high $RT(\%)$ values indicate less original content. In addition to the statistical computations, for each user in the analyzed tweet set, other tweet sets with similar hashtag are analyzed to see the user's participation with other hashtag also. In this step, two hashtags are thought as similar if manual inspection of their tweets \& hashtag names reveals that they have a common point based on content and targeted audience. The amount of users, who exist in multiple tweet set, also considered during manual inspection of organized tweet sets(see~Table~\ref{}) . \input{collection-manual-inspection} The manual inspection of these tweet sets (especially the ones about political propaganda) revealed that the organized behavior\ characteristics can be summarized as: sharing a common goal, temporal synchronization among users, and the dissemination of messages. Similar characteristics are reported in studies about digital and social media activism\cite{CanadianAdvocacy2.0,TheseDaysWillNeverBeForgotten,DigitalActivismNonViolentConflict}. Basically, to make an impact on Twitter, the message has to get seen and spread. The more it is seen the more it can spread. The more it spreads the more it persists and is likelier to be seen. In order to detect organized behavior, we benefit from tweet sets with two criterias: hashtag use and bot account activity. Hashtags are used in organized behaviors~\cite{NazisVsISISTwitter,ISISTwitterCensus}, because they facilitate a shared context by grouping disjoint tweets. During the US election 2016, the \#FeelTheBern hashtag is used by supporters of the democratic candidate Bernie Sanders~\cite{feelTheBern}. Twitter's top trend functionality also makes hashtag use preferable for organized behavior\ \cite{NazisVsISISTwitter,ISISTwitterCensus}. Furthermore, the prevailing presence of bot existence in a tweet set suggests the presence of organized behavior. During 2016 US elections, approximately 400,000 bots generated around 3.8 million tweets\cite{PresedentialElectionOnline}. Similarly, in the first presidential debate, $1/3$ of 1.8 million pro-Trump posts and $1/4$ of 600,000 pro-Hillary posts were generated by bots ~\cite{BotsAndAutomationOverTwitter}. The degree of these bot tweets and the concentration of their messages on same hashtags~\cite{botNews5} suggest that these bots post to serve a goal. Likewise, use of pro-Trump bots for spreading fake news~\cite{botNews4} also supports this idea. These bot activities include all three characteristics of organized behaviors, which were discovered upon our manual inspection, such as sharing a common goal, the temporal synchronization of tweets, and the dissemination of messages. \section{BACKGROUND} \label{} This section provides information about the tools and techniques utilized during the prototype development of the model. \subsection{Twitter} \label{} There are many Twitter related terms\cite{twitterGlossary}, the most relevant of which are described here. Twitter is a social networking application that supports posting short messages called \textit{tweet}s. The main connection among users is the \textit{follow} relations. The act of propagating a tweet is called \textit{retweet}ing. The more a tweet is retweeted the more visible it becomes. However, since the volume of tweets is very high, the chances of an individual tweet being seen is very low. A reference to another user is made using a \textit{mention}\footnote{For more information about Twitter entities see: \url{}} , which is a user name preceded by the \@ character (i.e. @theRealDonaldTrump). A \textit{reply} to a tweet, is a tweet whose first token is a mention corresponding to the replier. Twitter sends notifications to users when they are mentioned, bringing attention to who did the mentioning and why. A \textit{hashtag}, a string preceeded by \#, is used to tag a tweet. When used in multiple tweets by multiple people, it serves to create a relationship among tweets. This is widely performed during events as well as with people and topics. The Twitter REST API\footnote{Twitter REST API: \url{} ().} provides a great variety of methods with some access limits to prevent excessive queries \footnote{For rate limits see:\url{}.} In this study, the User Timeline API is used for fetching the tweets of a user in a given time interval. The Twitter Search API supports queries subject to various criteria, such as a hashtag, or a geographic location. Queries return results from a sampling of tweets which are posted in recent 7 days. Results of queries, are therefore, subject to the limitations of the sampling algorithms of the Twitter API, which claims to be optimized for relevance. \subsection{Supervised Classification Methods} \label{} In machine learning, classification refers to the task of automatically identifying the category of an entity. There are three main approaches to learning: supervised, semi-supervised, and unsupervised\cite{}. This work relies on supervised learning methods where a model is trained with datasets consisting of labeled data samples (observations) that identify the class of the data. The models generated with these approaches are used to predict the classes in unlabeled datasets. This work examines three supervised learning methods to detect organized behavior: random forest, support vector machine (SVM), and logistic regression. The random forest approach creates a forest of random decorrelated decision trees from a trained feature set \cite{PracticalDataScienceWithR}. The resulting forest is used to predict a classification based on the most predicted class by its decision trees. The support vector machine method aims to identify a hyperplane that best divides a dataset into two classes. It finds hyperplanes in high dimensional spaces with the greatest possible margin between it and the data points, which increases the confidence\cite{}. SVMs work better on smaller and cleaner datasets. Logistic regression is a statistical approach that is applied to machine learning\cite{PracticalDataScienceWithR}. It is well suited for binary classification problems. In supervised learning methods, there is the risk of generating overfitted models when training data sets are small or there are a large number of features. An overfitted model results in good accuracy with training data, but low accuracy with test data. The use of resampling techniques\cite{BootstrappingAdvisingResearchMethods} makes it possible to train non-biased classifier models by increasing the variance in the training data set. In statistics, resampling methods may be used to generate random subsets to test the success of classification models. The ten fold cross validation divides the training data set into ten subsamples. During the training phase, it generates models from nine subsamples of the training data, which are tested with the tenth subsample. This process is repeated ten times for each subsample in training data and test data. The results of these tests are averaged to determine whether the generated model is satisfactory \cite{CrossValidationkohavi1995study}. Principal component analysis (PCA) is a statistical method for explaining data with large number of variable using a smaller number of variables called the principal components. PCA aims to find the minimum number of uncorrelated features with the highest variances to reduce dimension in the data set~\cite{}. The Apache Spark\cite{sparkMlLibguide} machine learning library provides a wide range of support including all the classifications methods described here. In the prototype implementation, classification experiments have relied on Spark while PCA and feature selection processes are done by Weka\cite{wekaDefinitions}, which is a data mining software. \section{Discussion} \label{} The results presented in Section~\ref{} are promising for detecting various kinds of organized behavior. There are numerous directions to pursue in terms of extending the model, implementation techniques, and data collection. The three classification experiments (organized vs organic; political vs non-political; and pro-Trump vs pro-Hillary vs None) are somewhat different, focusing on the presence of any, topic-based, and group-based organization respectively. All three models are trained with the same features and all yield promising results. Based on these results we speculate that the proposed features may fingerprint tweet collections, at least in terms of their nature. It could be that our data set is too small even though it contains millions of tweets\footnote{For details see: \url{}. On the other hand, user features may indeed be most significant ones by shining light on the the characteristic of those who are in collusion. A classifier model generated solely based on user features may be worth developing, since it is fairly easy and cost effective, which would be beneficial for real-time classification. Collection labeling is performed by manual inspection, which is enormously time consuming and error prone. While the visualization tools that we created to summarize collections help, the task remains difficult. We spent endless hours inspecting thousands of tweets and in deliberation. Better approaches for labeling training data sets should be explored. The approach to constructing and processing collections must be systematically investigated to study the impact of various choices. While the use of hashtags is simple and powerful, they are not the only mechanism for coordination and organization on Twitter. During our experiments, by using data sets, which ignore \textit{tracedHT}\ related features~(Section~\ref{}), we aimed to test performance of our approach in case of tweets are collected based on other mechanisms such as community detection\cite{ozer2016community,bakillah2015geo} algorithms or topic detection\cite{} methods. The aim of the present work was to develop a basic model for detecting patterns of organized behavior on Twitter. Limited resources also played a role in keeping it simple, however, it is interesting to learn how simple approaches perform. We examined many features that are not in the current model that are related to tweet content, users relations, and information flow patterns. We observed the presence of unusually similar content posted by the same persons or those connected to them during the same interval, presumably serving a shared agenda. Similarity computation among all tweets in a set is very costly (complexity of $\mathcal{O}(n^2)$)-- a task that exhausted our resources. Efficient approaches to compare large sets of posts is an interesting research direction. Tightly connected communities can be very effective in propagating information. In popular context, such as political campaigns, follow relations exceeding 50K are not uncommon. When the followers of followers of thousands of people are considered, the computation of closeness centrality becomes challenge. Limiting the number of hops to 2 nodes, the users stored in a Neo4j\footnote{Neo4j } graph database became overwhelming. Alternative approaches to address closeness centrality is an interesting direction. Another observation is information flow patterns, for example recurring message paths, such as $A$ tweets $t_x$, which is retweeted by $B$, which is retweeted by $C$. Here the users $A,B,C$ remain the same, wheras the message $t_i$ may vary. Such pattern suggests presence of coordination, quite possibly automated. Although, the current model does not include many features we examined, they remain of great interest. The features we did include are those we considered to be significant and whose computation was in the realm of our resources. Unfortunately, tweet sentiment detection, which is a highly explored area, has limited success due to the nature of tweets being short, messy, unconventional, and satirical. The performance of Stanford NLP tool\cite{} used in this work is insufficient as of yet. A great deal of effort went into developing the feature extraction system to produce the training data set. Many performance issues were experienced related to processing big data. Big Data systems used for real-time detection require high memory and high processing power (GPU) to run complex algorithms on very large data. To prevent excessive database I/O, in memory graph databases needs to be used. Big data frameworks and tools are appropriate for social media data, which is vast and created in a continuous flow. Our prototype implementation uses MongoDB. As the table sizes and I/O requests increased the performance of queries dramatically decreases (in spite of index use). To prevent bottlenecks in one table, temporary tables were created with indexes to reduce the wait-time. Unfortunately, I/O bottlenecks persisted. For real time decision and detection systems there is a need for tools like Apache Spark\cite{}, which supports terabyte-scale in-memory data processing. Our repository is more than $480$ gigabytes with over $200$ million tweets, summary tables, indexes. A Spark cluster can process all that data in memory. Also, the Spark GraphX API would increase the feature extraction performance, especially for graph based features. Our observations have strengthened our belief in how relevant it is to detect misinformation and manipulation. With manipulators on social media getting more sophisticated, we expect that the detection of their activities will get more difficult. To summarize, the results of this work are encouraging with many future directions to pursue. The main future directions are with respect to validation, improved features, performance, and domain specific detections. \section{Experiments and Results} \label{} To assess the performance of the proposed approach, the classifier models are tested on the training data set using 10-fold cross validation. The results were promising for the organized behavior\ classification. This motivated us to proceed with the generation of two more classifiers that use the same features for two different classification problems: \textit{political} vs. \textit{non-political} and \textit{pro-Trump} vs \textit{pro-Hillary}. For each classification task, the training data set is labeled differently. \subsection{Data: Tweet Sets} \label{} Training data sets are generated by using extracted features from expanded tweet sets(\textit{ET}) In the training data set a row is labeled with the label of the collection it is generated from. Table~\ref{} show the sizes of \textit{ST}\ (\#\textit{ST}), ET (\#ET), the number of tweets used for temporal features extraction (\#TFT) and number of users(\#Users) for some collections \footnote{Sizes of the all collections in data set: \url{}. \input{training-data-size} \subsubsection{Organized Tweet Sets} \label{} To create a training data set, first \textit{ST}\ consisting of tweets with a hashtag must be gathered. A hashtag is labeled as organized if the presence of organized behavior\ is observed within its tweets based on our manual inspection (i.e. \#LadyGagaNewVideo, \#podestaemails, \#podestaemails13, \#podestaemails20, \#BoycottHamilton). Suitable hashtags reported the news or literature are also labeled organized (i.e. \#tcot \& \#pjnet based on a study\cite{NazisVsISISTwitter}). Hashtags used by large amount of bots during the 2016 US election are also labeled as organized\footnote{In the first presidential debate, 32.7 percent of pro-Trump hashtags and 22.3 percent of pro-Hillary hashtags were posted by bots\cite{BotsAndAutomationOverTwitter}}. These hashtags are reported in studies regarding the prevalence of bot accounts during the 2016 US election~\cite{BotsAndAutomationOverTwitter,PresedentialElectionOnline,BotsAndAutomationOverTwitter2}. For externally discovered hashtags, if their bot based content generation is greater than 20\%, a collection is created for them (i.e. \#AmericaFirst, \#MakeAmericaGreatAgain, \#TrumpPence16, \#NeverHillary, \#VoteTrump, \#CrookedHillary, \#LoveTrumpsHate, \#NeverTrump, \#StrongerTogether, \#ImWithHer, \#UniteBlue.) When seeking for candidate hashtags, we create a backlog. The ones that are not yet manually inspected are compared with collections known to be organized. If there are significant number of overlapping users, the hashtag is considered as organized. This is due to the observation that colluding users are active over time and use several persisting hashtags to manipulate others. Table~\ref{} shows the users who participated in multiple pro-Trump hashtags in different times. Furthermore, the percentage of users who registered after July 2015\footnote{The official nomination of Trump for presidency.} are in increasing trend~(Figure~\ref{}). A suspected hashtag is labeled as organized if at least 20 percent of its users participated in bot based collections. Some of the tweet sets labeled in this manner are based on hashtags of \#benghazi, \#obamacarefail, \#imvotingbecause, \#draintheswamp, \#trumpwon, \#clintonemails, \#auditthevote, and \#hillaryemails. In the tweet sets collected at different timess based on the bots cited in \cite{BotsAndAutomationOverTwitter} and \cite{PresedentialElectionOnline} numerous common hashtags were observed (i.e. \#election2016,\#tcot, \#imwithher, \#nevertrump, \#neverhillary, \#trumppence2016, \#p2). This suggests that the use of some hashtags persist throughout a campaign, such as the observed in the US election 2016 campaign. This is not surprising, since hashtags serve as group specific identifiers that serve to unite messages and people. \input{user-analysis-table} \begin{figure} \begin{center} \includegraphics[height= 0.25\paperheight]{sameUserCreationDate} \end{center} \caption{Percentage of same users in Table~\ref{} (users in \#benghazi \& \#crookedHillary) based on Twitter registration date} \vskip\baselineskip \label{} \end{figure} \subsubsection{Organic Tweet Sets } Hashtags are labeled as organic when they are deemed to be spontaneously posted. In case of trending topics, such hashtags emerge due to events like holidays, natural disaster, and news about popular people. For example, the hashtags related the United Airlines, such as \#unitedairlines, \#boycottUnitedAirlines, \#newunitedairlinesmottos, instantly became popular after a video of passenger, who was forcibly removed from a plane due to over booking, went viral. Similarly, during the 2017 Oscars ceremony, the \#oscarsfail hashtag became top trending after the best picture award was accidentally given to wrong movie. Hashtags related to special days like \#Thanksgiving, \#LaborDay, \#NationalSiblingsDay, and \#WomansDay are also tagged as \textit{organic}, since these hashtags are used by a large number of diverse people. Likewise, topic centric hashtags are widely used by diverse sets of users, such as \#news, \#deeplearning, and \#bigdata are labeled as organic. However, recall that during the labelling of organized hashtags, we labelled hashtags as organized in case of considerable amount of their contents are created by automated bots. For the hashtags of \#Thanksgiving, \#LaborDay, \#NationalSiblingsDay, and \#WomansDay, it is also possible that they may contain bot activity, especially when they are trending topics. In order to understand the level of bot existence in these organic hashtags, we examined characteristics of their tweets by calculating statistical values of $DW(\%)$, $TPU(\mu)$, $RT(\%)$, $HT(\sigma^2)$, and $HT(\sigma)$ (the metrics used during the manual inspection of tweets in Section \ref{}). As a result of these calculations, we conclude that these hashtags are used spontaneously. For example, given the values for \#womansDay shown in Table~\ref{}, higher value of $DW(\%)$ indicates less similar messages, and lower value of $TPU(\mu)$ shows less multiple posts by the same users. Likewise, lower values of $HT(\sigma^2)$ \& $HT(\sigma)$ indicate less hashtag use in \#womansDay tweets. Furthermore, the overlapping users in organic hashtags inspected with the idea that if a user is present in numerous organic hashtags, it may be an incidence of bots hijacking trending hashtags. Table~\ref{} shows overlapping users in multiple hashtags regarding special days, however their percentage is very low compared to those in organized hashtags(see~Table~\ref{}). \input{organic-sample-users} \subsubsection{Political vs. Non-Political} The labeling of political and non-political hashtags are done by manual inspecting \textit{ST}. The data set labeled as political is created from 2016 US election data. The determination of a label was fairly straightforward in comparison to organized behavior. \subsubsection{Pro-Trump, Pro-Hillary, Neither} Specifically in political campaigns, there are strong groups who are pro and against certain candidates. Meticulous attention is given to reaching people and delivering propaganda that supports their candidate and damages other candidates. In the 2016 US election the political campaigns were fierce and quite different than earlier campaigns. Heavy use of social media was deployed from within the country as well as from other countries. Just what happened during the campaign and elections has been a topic of heated debate that is unfolding at the time of writing this paper. The hashtags are mainly selected and labeled based on given \textit{pro-Trump} and \textit{pro-Hillary} hashtags in the work on bots and automation in Twitter\cite{BotsAndAutomationOverTwitter}, The remaining hashtags are manually labeled as usual (with use of Table~\ref{} and Table~\ref{}), except for hashtags that do not belong to either class that are labeled as \textit{None}. \subsection{Training Data Sets} \label{} The labels of training data sets for each category are as follows\footnote{Training data sets can be found in \url{}}: \begin{description} \item [organic vs organized]: $851$ records of which $625$ are labeled as organized and $226$ organic. \item [political vs non-political]: $879$ records of which $231$ are labeled as non-political and $648$ political. \item [pro-Trump vs pro-Hillary]: $854$ records of which $311$ are labeled as pro-Trump, $171$ pro-Hillary, and $371$ None. \end{description} For each category, four different training data sets are generated based on the features of all collections: (1) all extracted features, (2) principal components of all extracted features obtained from PCA(see~Section~\ref{}), (3) all extracted features except user features based on \textit{tracedHT}\ (see~ Equations~\ref{}), and (4) principal components of the features in (3). Data sets (3) and (4) ignore \textit{tracedHT}\ related user features for the purpose of assessing how the model behaves when the features of the hashtag that were used to fetch the initial data are eliminated. Since hashtags are not the only mechanism for coordination on Twitter, one would expect the method to remain successful. By removing the \textit{tracedHT}\ related features in Equations~\ref{}, we can see how reliant the detection is to hashtags. In the second and fourth data sets, PCA is used in order to examine impact of dimension reduction on the performance of the classification, with consideration of the need for real-time processing. Table~\ref{} provides information about the size of some collections in the collection repository. Even though the size of collections are highly limited by Twitter Search API, (\#ET) sizes are in the millions. \subsection{Results of three Classifers} \label{} The model is evaluated for each classifier and data set, using 10-fold cross validation. The organic vs organized and political vs non-political are binary classification models, whereas the pro-Trump vs pro-Hillary vs None classification is a multi-classification problem. Results show that the random forest algorithm results in high scores with full features, while logistic regression and SVM algorithms give better results when PCA is applied. Furthermore, the performance with Data~Set~3~and~4 shows that the presence or absence of \textit{tracedHT}\ is not essential to capture organized behavior. Tables~\ref{}, \ref{}, and \ref{} show the results for each category. SVM results are not provided for pro-Trump vs pro-Hillary classification due to lack of support for multi-classification with SVM in Spark MlLib\cite{sparkMlLibguide}. \input{evaluation-organic-organized} \input{political-nonpolitical-table.tex} \subsection{Top five features of classifications} \label{} Feature selection is important for improving accuracy and decreasing training time. The following sections enumerate the top five features\footnote{Features are selected using the \emph{ClassifierSubsetEval} attribute evaluator with the Random Forest classifier and the Best First search method in the Weka\cite{wekaDefinitions} by applying 10 fold cross validation} for each classification. References to related functions and feature equations are provided next to the features. $C$ refers to a collection and $U = users(C)$. All features assume $u \in U$, additional constraints are expressed when relevant. Percentages, frequencies, maximum, minimum, $\sigma^2$, and $\sigma$ values are computed over $t \in C$ and $u \in tweets(u,C)$. For details regarding the functions and features see Section~\ref{}. The most important five features during classifications for each data set are : \begin{description} \item{\textbf{Organic vs. Organized}} \begin{description} \item{All features:} \begin{enumerate} \item $\sigma^2$ of $\funcname{media-use}(u,C)$ [Eq.~\ref{}] \item $\%$ of users such that $\funcname{follower-degree}(u) > 1$ [Eq.~\ref{}] \item $\%$ of users such that $1 \leq favorite\#(u) \leq 100$ \item Minimum of $tweet\#(u)$ \item Maximum $\funcname{tweet-hashtag\#}_{\mu/d}(u,D,C)$ [Eq.~\ref{} \end{enumerate} \item{All features excluding the hashtags:} \begin{enumerate} \item $\#$ of users such that $0.6 \leq \funcname{media-use}(u,C) \leq 0.9$ [Eq.~\ref{}] \item $\#$ of users such that $\funcname{hashtag-use}(u,C)=7$ [Eq.~\ref{}] \item $\#$ of users such that $\funcname{mention-use}(u,C)=10$ [Eq.~\ref{}] \item Minimum of $\funcname{follower-degree}(u)$ [Eq.~\ref{}] \item $\#$ of users such that $1 \leq favorite\#(u) \leq 100$ \end{enumerate} \end{description} \item{\textbf{Political vs. Non-Political classification results}} \begin{description} \item{All features:} \begin{enumerate} \item $\sigma^2$ of $\funcname{media-use}(u,C)$ [Eq.~\ref{}] \item Maximum of $\funcname{mention-use}(u,C)$ of all users [Eq.~\ref{}] \item $\mu$ of $\funcname{follower-degree}(u)$ [Eq.~\ref{}] \item $\mu$ of $10.001 \leq favorite\#(u) \leq 20.000$ \item Maximum of $tweet\#(u)$ \end{enumerate} \item{All features excluding the hashtags:} \begin{enumerate} \item $\sigma^2$ of $\funcname{media-use}(u,C)$ [Eq.~\ref{}] \item $\mu$ of $\funcname{follower-degree}(u)$ [Eq.~\ref{}] \item $\%$ of users such that $10.001 \leq favorite\#(u) \leq 20.000$ \item $\%$ of users such that $10.001 \leq tweet\#(u) \leq 20.000$ \item Maximum of $tweet\#(u)$ \end{enumerate} \end{description} \item{\textbf{pro-Hillary vs. pro-Trump vs. None}} \begin{description} \item{All features:} \begin{enumerate} \item $\%$ of users such that $0.0001 \leq \funcname{media-use}(u,C) \leq 0.5$ [Eq.~\ref{}] \item $\%$ of users such that $\funcname{url-use}(u,C)=1$ [Eq.~\ref{}] \item $\%$ of users such that $\funcname{mention-use}(u,C)=7$ [Eq.~\ref{}) \item$\sigma^2$ of $\funcname{url-use}(u,C)$ [Eq.~\ref{}] \item $\%$ of users such that $\frac{\funcname{mention-use}(u,C)}{\funcname{tweet}\#_{\mu/d}(u)}=1$ [Eq.~\ref{} \& Section~\ref{}] \end{enumerate} \item{All features excluding the hashtags:} \begin{enumerate} \item $\%$ of users such that $\funcname{mention-use}(u,C)=0$ [Eq.~\ref{}] \item $\%$ of users such that $0.0001 \leq \funcname{media-use}(u,C)= \leq 0.5$ [Eq.~\ref{}] \item $\%$ of users such that $0.0001 \leq \funcname{mention-use}(u,C) \leq 0.5$ [Eq.~\ref{}] \item $\%$ of users such that $\funcname{url-use}(u,C)=1$ [Eq.~\ref{}] \item $\%$ of users such that $1 \leq favorite\#(u) \leq 100$ \end{enumerate} \end{description} \end{description} It is surprising that none of the temporal features are in top five features, suggesting that organized behavior\ detection could be achieved by extracting user features only. One of the most common advise for increasing engagement is the timing of posts to increase likelihood of visibility. The top user-based are also reported in other studies\cite{DetectingSpamURLCao,OrganicorOrganizedUrlCao}, which is encouraging. Compared to temporal features, extracting user features is easier and more efficient. A classifier based only on user features might be useful. The use of media stands out as an important factor, appearing in the top 5 of every classification. This is not surprising, since it is generally known that tweets with media are viewed the most. Bots and trolls use regularly use them. The consistent use of media entities can be observed from the variance. Furthermore, consistent use of media results in high media ratio for the user (Equation~\ref{}), which may signal bots or users in some organized behavior. A follower degree of $1$ means that the user does not follow anyone. To be effective, organized users are expected to follow users. Therefore, a follower degree value $1$ would likely identify an organic collection. On the contrary, a follower degree of $0$ implies not being popular -- having $0$ followers. Newly deployed bots often have $0$ followers~\cite{botNews4}. Furthermore, a user's tweet\# and favorite\# are helpful in classification. When these values are in the thousands they are more likely to be associated to an automated user, whereas low values indicate regular users. Very high URL, hashtag, and mention ratios imply the occurrence of multiple such entities in most of their tweets (Equation~\ref{}). Dedicating the small space provided by a tweet to deliver several entities suggests an intent to become more visible. Mentions trigger notification to the mentioned and it may also noticed by the follower of the mentioned person. URLs extend the information that can be delivered by luring a user to an external site. Finally, a high mention rate points to interaction with other users. The daily hashtag rate (Eq.~\ref{}) shows the daily use of the \textit{tracedHT}\ during the days that hashtag is used in Twitter. A higher frequency of tweets than typical on a given day compounded by the occurence of the \textit{tracedHT}\ suggests an organized behavior. \section{Feature Selection} \label{} Based on the characteristics of organized behavior and the manual examination of large numbers of tweets (Section~\ref{}), two kinds of main features are identified for detecting the presence of organized behavior: \textit{user} and \textit{temporal} features. The user features capture information about the characteristics of users. The temporal features focus on the timing of tweets (independent of who posted them) to detect the presence of a synchronization. Recall that, the posting time and synchronization are useful for increasing the visibility of a message. In other words, user features focus on identifying the participants and the temporal features focus on detecting the presence of organized behavior. In order to clearly describe the features consider the sets $ST$ (seed tweets, Section~\ref{}) to be the initial set of tweets fetched with a query for a given hashtag (\textit{tracedHT}). Let $ET$ (Expended Tweet Set, Section~\ref{}), be the set of tweets obtained by expanding $ST$ with additional tweets of users who have posted a tweet $t \in ST$ (see~Figure~\ref{}). Table~\ref{} defines a set of functions used to formulate the features. Essentially these functions correspond to data fetched using the Twitter API, thus, refer to Twitter specific types like Tweet, Hashtag, and Mention. Typed sets are denoted with a type name subscripted with $_s$. such as \emph{Hashtag}$_s$ that represents a set of hashtags. \begin{table}[H] \centering \caption{Descriptions of user and tweet functions, where $T$ is set of tweets, $t$ is a tweet, $h$ is a hashtag, $u$ is a user, $D$ is a set of Days, $d$ is a day, $\Delta$ is a duration} \label{} \begin{tabular}{|l|l|l|} \hline \thead{\textbf{function}}&\thead{\textbf{Type}}&\thead{\textbf{Description}}\\ \hline \multicolumn{3}{|c|}{\textbf{User functions}}\\ \hline $\funcname{reg-date}(u)$ & $\mathit{Date}$ & date $u$ registered\\ \hline $\funcname{following\#}(u)$ & $\mathit{Integer}$ & number of users that $u$ follows\\ \hline $\funcname{follower\#}(u)$ & $\mathit{Integer}_s$ & number of users who follow $u$\\ \hline $\funcname{favorite\#}(u)$ & $\mathit{Integer}$ & number of tweets favorited by $u$\\ \hline $\funcname{tweets}(u,T)$ & $\mathit{Tweet}_s$ & set of tweets posted by $u$ in $T$\\ \hline $\funcname{tweet\#}(u)$ & $\mathit{Integer}$ & number of tweets posted by $u$\\ \hline $\funcname{tweetsD}(u,T,D)$ & $\mathit{Tweet{_s}_s}$ & set of tweet sets \\ & & (posted by $u$ on $d \in D$)\\ \hline $\funcname{users}(T)$ & $\mathit{User}_s$ & set of users in $T$\\ \hline $\funcname{hashtag?}(u,T,h)$ & $(0|1)$ & $1$ if user has posted a tweet with $h$ in $T$,\\ & & $0$ otherwise\\ \hline \multicolumn{3}{|c|}{\textbf{Tweet functions}}\\ \hline $\funcname{hashtag\#}(t)$ & $\mathit{Integer}$ & number of hashtags that occur in $t$ \\ \hline $\funcname{hashtag?}(t,h)$ & $(0|1)$ & $1$ if $h \in \funcname{hashtags}(t)$, $0$ otherwise\\ \hline $\funcname{mention\#}(t)$ & $\mathit{Integer}$ & number of mentions that occur in $t$ \\ \hline $\funcname{hashtagD}(T,h)$ & $\mathit{Date}_s$ & set of days when tweets with\\ & & $h$ occur in $T$ \\ \hline $\funcname{media\#}(t)$ & $\mathit{Integer}$ & number of media that occur in $t$\\ \hline $\funcname{retweeted?}$(t) & $( 0 | 1 )$ & $1$ if $t$ is retweeted, $0$ otherwise\\ \hline $\funcname{sentiment}(t)$ & \multicolumn{2}{l|}{(\texttt{\small VeryNegative|Negative|Neutral|}} \\ & \multicolumn{2}{l|}{\texttt{\small Positive|VeryPositive})} \\ \hline $\funcname{mentions}(t)$ & $\mathit{Mention}_s$ & the set of mentions that occur in $t$\\ \hline $\funcname{url\#}(t)$ & $\mathit{Integer}$ & number of links (URL) that occur in $t$\\ \hline $\funcname{temporalTweets}(T,\Delta)$ & $\mathit{Tweet}_s$ & set of temporal tweets in $T$ based on $\Delta$\\ \hline \end{tabular} \end{table} \subsection{User Features} \label{} As the creators of the tweets that may be part of some collusion, the characteristics of the contributors are expressed with a set of user features. The following features are computed for each user, which corresponds to lines $13-16$ of Algorithm~\ref{}. \noindent\textbf{Tweet Count}: The total number of tweets posted by a user: $tweet\#(u)$. Used to indicate how active a user is. \noindent\textbf{Favorited tweet count}: The number of tweets that are marked as \textit{favorite}\footnote{Twitter has renamed Favorite to Like during the final stages of the preparation of this article.}: $favorite\#(u)$. Marking tweets as a favorite is an approach used to gain attention from others to increase follower count (users might follow those who favorite their tweets). \noindent\textbf{Average tweets/day}: The average number of tweets per day, based on the number of days since the user registered. Used to understand the user's daily tweet frequency, since higher values might indicate the behavior of an automated account. \begin{equation} \label{} \funcname{tweet}\#_{\mu/d}(u) = \frac{tweet\#(u)} {\funcname{today}()-\funcname{regDate}(u)} \end{equation} \noindent where $\mathit{today}()$ returns current date and subtraction of days returns the number of days. \noindent\textbf{Follower degree}: The degree of a user's followers: \begin{equation} \label{} \funcname{follower-degree}(u) = \frac{\funcname{follower\#}(u)} {\funcname{follower\#}(u) + \funcname{following\#(u)}} \end{equation} Follower degrees that approach $1$ indicate a high degree of followers that is typical with popular persons, while degrees approach to $0$ indicate the opposite. Newly created bots tend to follow numerous users and have very few followers. \noindent\textbf{Entity use}: Entities are used to relate a tweet to other tweets, users, external resources, or media. They are used to gain attention, therefore a higher rate would be expected in propaganda. The use of the entity types: \footnote{For more information about Twitter entities see: \url{}} \textsc{URL}, \textit{Mention}, \textit{Hashtag}, and \textit{Media} are computed for each $u \in \funcname{users}(ET)$: \begin{subequations} \label{} \begin{equation} \label{} \funcname{hastag-use}(u,ET) = \frac{\displaystyle\sum_{t \in tweets(u,ET)} \mathit{hashtag\#}(t)}{\mid tweets(u,ET) \mid} \end{equation} \begin{equation} \label{} \funcname{url-use}(u,ET) = \frac{\displaystyle\sum_{t \in tweets(u,ET)} \mathit{url\#}(t)}{\mid tweets(u,ET) \mid} \end{equation} \begin{equation} \label{} \funcname{mention-use}(u,ET) = \frac{\displaystyle\sum_{t \in tweets(u,ET)} \mathit{mention\#}(t)}{\mid tweets(u,ET) \mid} \end{equation} \begin{equation} \label{} \funcname{media-use}(u,ET) = \frac{\displaystyle\sum_{t \in tweets(u,ET)} \mathit{media\#}(t)}{\mid tweets(u,ET) \mid} \end{equation} \end{subequations} \noindent\textbf{Traced hashtag use}: The number user's tweets that contain the \textit{tracedHT}\ within ET. Used to understand how focused the user is to the \textit{tracedHT}. In organized behavior, it more likely that users concentrated on a hashtag (Section~\ref{}) \begin{equation} \label{} \funcname{user-hashtag-use}(u,ET) = \displaystyle\sum_{t \in tweets(u,ET)} \mathit{hashtag?}(t,\textit{tracedHT}) \end{equation} \noindent\textbf{Average daily tweets of \textit{tracedHT}}: Focuses on the daily use of the \textit{tracedHT}, considering that a participant in propaganda (or similar activity) would have a persistent use. \begin{equation} \label{} \funcname{tweet-hashtag\#}_{\mu/d}(u,D,T) = \frac{\displaystyle\sum_{t \in \funcname{tweetsD}(u,T,D)} hashtag?(t,\textit{tracedHT})} {\mid D \mid} \end{equation} \noindent where $T=ET$, $u \in \funcname{users(T)}$, $D = \funcname{hashtagD}(T,\textit{tracedHT})$ \noindent\textbf{Average Tweets/Day vs. Average daily tweets of \textit{tracedHT}}: This value is used to understand how the user's behavior is in the days when she used \textit{tracedHT} compared to her general daily behavior. If a user posts much more tweets than she normally does for a hashtag, this can be indicator of an organized behavior. \begin{equation} \label{} \funcname{user-daily-tweet-comparison}(u,D,ET) = \frac{\funcname{tweet-hashtag\#}_{\mu/d}(u,D,ET)} {\funcname{tweet}\#_{\mu/d}(u)} \end{equation} \noindent\textbf{User creation date:} $userReg(u)$, which can be useful in understanding of collective behavior of users(see~Figure~\ref{}). \subsection{Temporal Features} \label{} The temporal aspects of tweets focus on the characteristics of tweets independent of who posted them with the aim of detecting the presence of a coordinated effort. As was explained in Section\ref{} the dissemination of a message calls for making it visible, which is achieved by synchronous targeted posting to increase the odds of delivery. Therefore, temporal features of tweets posted within a certain time period are computed and tracked. Figure~\ref{} shows the algorithm for computing these features. Unlike with the user features\ which consider all of the tweets of users to characterize them, in the case of the temporal features only tweets that contain the \textit{tracedHT}\ are used. The initial set of tweets that contain \textit{tracedHT}\ ($ST$) are extended with tweets posted by $u \in \funcname{users}(ST)$ resulting in $ET$. The relationship and frequency of users to tweets are expected to be different in coordinated posting behavior. This section describes the temporal features of tweets posted during an interval $I$, where: \begin{equation} \label{} T= \bigcup\limits_{\substack{t \in T \\ \funcname{hashtag?}(t,\textit{tracedHT}) = 1}} \funcname{temporalTweets}(ET,I): \end{equation} \noindent\textbf{Entity use}: These are just like the computations made for a user, but in this case the tweet set includes the tweets that were posted during the time interval by many users: \begin{subequations} \label{} \begin{equation} \funcname{hashtag-use}(T) = \frac{\displaystyle\sum_{t \in T} \funcname{hashtag\#}(t) \;} { \mid T \mid} \end{equation} \begin{equation} \funcname{url-use}(T) = \frac{\displaystyle\sum_{t \in T} url\#(t) \; } { \mid T \mid} \end{equation} \begin{equation} \funcname{mention-use}(T) = \frac{\displaystyle\sum_{t \in T} mention\#(t) \; } { \mid T \mid} \funcname{media-use}(T) = \frac{\displaystyle\sum_{t \in T} media\#(t) \; } { \mid T \mid} \end{equation} \end{subequations} \noindent\textbf{Temporal Tweet Per User (TPU)}: This feature is used to check tweet per user in temporal basis. It is extracted because higher values of TPU in $T$ can be an indicator of organized behavior. \begin{equation} \label{} \funcname{temporal-tpu\#}(T)= \frac{\displaystyle\sum_{t \in T} 1} { \mid \funcname{users}(T) \mid} \end{equation} \noindent\textbf{Retweeted tweet features}: Features based on retweets can be sign of bot account existence and organized behavior, because in case of retweets there is no need to generate a content, which can be challenging for the automated accounts. The following are computed for the retweeted tweets : \begin{description} \item[Retweet frequency:]{The number of retweets ((i.e. if $100$ users have retweeted $t_x \in T$, this counts as $100$ retweets): \begin{equation} \funcname{retweet\#}(T) = \displaystyle\sum_{t \in T} retweeted?(t) \end{equation} \begin{equation} \label{} \funcname{retweet\%}(T) = \frac{\funcname{retweet\#}(T) } { \mid T \mid} \end{equation} } \item[Unique retweeted frequency:]{The percentage of distinct retweets among retweeted tweets. It shows the diversity in the retweets. If most of the retweets are retweets of the same tweet, it can be a sign of a collective behavior. \begin{equation} \label{} \funcname{original-retweeted-tweet\%}(T) = \frac{\Bigg\lvert \bigcup\limits_{\substack{t \in T\\retweeted?=1}} \{ t \} \Bigg\lvert} {\displaystyle\sum_{t \in T} retweeted?(t) } \end{equation} } \item[Retweeting user frequency:]{Show how many of the users in the temporal tweet set are participated in retweets about \textit{tracedHT}. If in a temporal tweet set, most of the users are retweeted a tweet about \textit{tracedHT}, this can be a sign of collective behavior. \begin{equation} \funcname{retweeted-users\#}(T) = \displaystyle\sum_{u \in users(T)}\ceil{\frac{\displaystyle\sum_{t \in tweets(u,T)} retweeted?(t)}{\mid tweets\#(u,T) \mid}} \end{equation} \begin{equation} \label{} \funcname{retweeted-users\%}(T) = \frac{\funcname{retweeted-users\#}(T)} { \mid \funcname{users}(T) \mid} \end{equation} } \end{description} \noindent\textbf{Features for tweets that are not retweeted:} These features focus on tweets and users who are not subject to retweeting. They are extracted because higher values in these features may suggest organic behavior\ due to difficulty of creating an original tweet. \begin{description} \item[Unretweeted tweet frequency:] { \begin{equation} \label{} \funcname{unretweeted\%}(T) = 1 - \funcname{retweet\%}(T) \end{equation} } \item[Users with no retweets frequency:]{ \begin{equation} \label{} \funcname{unretweeted-users\%}(T) = 1 - \funcname{retweeted-users\%}(T) \end{equation} } \item[Unretweeted tweet count:] { \begin{equation} \label{} \funcname{unretweeted\#}(T) = \mid T \mid - \funcname{retweet\#}(T) \end{equation} } \item[Count of users with no retweets frequency:]{ \begin{equation} \label{} \funcname{unretweeted-users\#}(T) = \mid \funcname{users}(T) \mid - \funcname{retweeted-users\%}(T) \end{equation} } \item[Ratio of unretweets and users with no retweets:]{ \begin{equation} \label{} \funcname{unretweeted-tweet_user_ratio}(T) = \frac{unretweeted\#(T)}{unretweeted-users\#(T)} \end{equation} } \end{description} \noindent\textbf{Mention Features:} Since there are usually groups who are pro and against the candidates in political domain, it is likely that users from different sides can have a dispute in Twitter. In these disputes, users use mentions to target other users. Furthermore, in order to propagate messages, bot users may use mentions to get attention of other users. Therefore, mention related features could be beneficial for organized behavior\ detection. \begin{description} \item[Mention Ratio:]{Frequency of mentions to distinct mentions: \begin{equation} \label{} \funcname{mention-ratio}(T) = \frac{\mid \bigcup\limits_{t \in T} mentions(t) \mid} {\displaystyle\sum_{t \in T} mention\#(t)} \end{equation} } \item[Ratio of mentions in retweets:]{Frequency of mentions that occur in retweeted tweets. \begin{equation} \label{} \funcname{mention-RT}(T) = \frac{\mid \bigcup\limits_{\substack{t \in T \\ \funcname{retweeted?}(t) = 1}} mentions(t) \mid} {\displaystyle\sum_{t \in T} mention\#(t) \; \funcname{retweeted?}(t)} \end{equation}} \item[Ratio of mentions in unretweeted tweets:]{Frequency of mentions in tweets that are not retweeted. \begin{equation} \label{} \funcname{mention-notRT}(T) = \frac{\mid \bigcup\limits_{\substack{t \in T \\ \funcname{retweeted?}(t) = 0}} mentions(t) \mid} {\displaystyle\sum_{t \in T} mention\#(t) - \displaystyle\sum_{t \in T} mention\#(t) \funcname{retweeted?}(t)} \end{equation} } \end{description} \noindent\textbf{Sentiment Features} Within the tweets that contains a mention, the percentage of tweets with the sentiments \textit{VeryNegative}, \textit{Negative}, \textit{Neutral}, \textit{Positive}, \textit{VeryPositive} are computed. The sentiment analysis of tweets are determined using the Stanford Core NLP tool~\cite{StanfordCoreNLP}. In our study, tweet sentiments are only calculated for reply tweets because of the likelihood of disputes in political domain. In these disputes, large amount of negative tweets directed to a user may be sign of organized behavior. \begin{equation} \label{} \funcname{mentioned-sentiment\%}(s,T) = \frac{\mid \{ t | t \in T, \funcname{sentiment}(t) = s \}\mid }{|T|} \end{equation} \subsection{Collection Representation} \label{} To gain information about a collection its users and tweets (within a time interval) are characterized by numerous features extracted from data fetched from Twitter. This feature extraction phase yields a large set of features representing each user and temporal analysis. Any assessment about the collection requires the representation of the collection as a whole, which will be used in the training data set. Collections are uniformly represented by distributing the extracted values for each user feature into buckets defined for each feature. Each feature value is placed in the bucket whose criteria it matches. As an example, consider Figure~\ref{}, which shows the hashtags posted by the users of the \#draintheswamp collection based on Equation~\ref{}. These values are summarized as follows: hashtagPostCount\_1~:~12.2\%, hashtagPostCount\_2~:~6.9\%,..., hashtagPostCount\_100-...~:~3.44\%. Note that, the buckets are defined based on the observations from inspecting many collections. In addition to the feature buckets, the mean, variance, standard deviation, minimum and maximum values are computed for each feature. All these values associated with a collection are stored in the Feature Repository~(Algorithm~\ref{}). On the other hand, in the summarization of temporal features, the values are not separated to the bucket values. For each temporal interval feature, statistical values(mean, variance, standard deviation, min \& max values) are calculated. Figure~\ref{} and Figure~\ref{} illustrate the extracted temporal features(Equation~\ref{} - Equation~\ref{} ) of \#draintheswamp collection. In the training data set, the resulting collection representation is used. \begin{figure} \begin{center} \includegraphics[width=0.6\paperwidth,height=0.25\paperheight]{draintheswamp_userPostCounts} \end{center} \caption{Percentage of users based tweet counts with \#draintheswamp hashtag (Equation~\ref{})} \vskip\baselineskip \label{} \end{figure} \begin{figure} \begin{center} \includegraphics[width=0.6\paperwidth,height=0.25\paperheight]{draintheswamp_hourlyEntityRatios} \end{center} \caption{Hourly entity ratios of \#draintheswamp hashtag(Equation~\ref{})} \vskip\baselineskip \label{} \end{figure} \begin{figure} \begin{center} \includegraphics[width=0.65\paperwidth,height=0.25\paperheight]{draintheswamp_totalNonRetweetDistinctUser} \end{center} \caption{Hourly Distribution of non-retweet count and distinct user count of \#draintheswamp hashtag(Equation~\ref{} )} \vskip\baselineskip \label{} \end{figure} \section{Conclusions} \label{} This work proposes a supervised learning based model for automatically classifying tweet sets that exhibits organized behavior patterns. Towards this end models are trained with user and temporal features extracted from over $200$ million tweets that were mostly gathered during the 2016 US presidential election. Three types of classifications were performed among the categories: [organic,organized], [political, non-political], and [pro-Trump,pro-Hillary,None]. In each case, the random forest algorithm consistently resulted in high accuracy and f-measure scores with an average of $0.95$. The results of classifying tweet sets suggest that neither content nor user relation features are required to successfully classify them. Furthermore, that user features are the most significant regarding our classification tasks. Further investigated with larger training data sets should be perform to further validate. Features like the closeness centrality and tweet similarity are very costly, therefore it is encouraging that the detection of organized behavior can be achieved without them. Organized vs organic behavior classification is well suited for politics and campaigns, as the organizing efforts carried out in Twitter are well captured in the model and features. The results are encouraging and we intend to pursue this work and its domain specific applications. \section{Introduction} \label{} Twitter is one of the most popular microblogging platforms with nearly 328 million users~\footnote{(April 2017)}. In recent years, Twitter has become an alternative platform to main stream media for getting news. Politics is one of the most prominent uses of social media platforms due to their facilitation of reaching the masses. Numerous guides and tips are disseminated for political campaigns for the effective use of Twitter\footnote{.} Most politicians have a verified Twitter accounts, which they use to directly communicate with the public. Twitter provides a lucrative platform for diffusing (mis)information at a very low cost \cite{crossChannelMediaCosts}, which can be easily utilized for targeting and manipulating users. The 2016 US presidential election, where Hillary Clinton and Donald Trump were presidential candidates, demonstrated the effectiveness of using Twitter very well\cite{FM7090,}. During this campaign approximately 400,000 social bots (automated agents) generated around 3.8 million tweets corresponding to 19 percent of all campaign related posts \cite{PresedentialElectionOnline}. Furthermore, recently Fake News has gained much attention due to its prevalence and persuasive capacity\cite{allcott2017social}. The impact of believing such news can be dramatic. For example, during US election 2016, a news, which states Hillary Clinton used a pizza restaurant, as a front for a pedophile sex ring became a viral, and resulted in an arrest of a person who entered the restaurant with an assault rifle~\cite{pizzaGate2,pizzaGate1}. During the 2016 US election, there were numerous fake news such as ``Donald Trump Protester Speaks Out: I Was Paid \$3,500 To Protest Trump's Rally."~\cite{fakeNewsProtestors}. In order to propagate messages in accordance with an agenda, there is an increasing incidence of employing social media users \cite{BotsStrongerInBrexit}. An investigation of fake news sites during the 2016 US election revealed that Veles (a Macedonian town with a population of 45,000) was the source of many pro-Trump fake news \cite{fakeNewsMacedonia2,fakeNewsMacedonia1,fakeNewsMacedonia3}. In order to rapidly produce fake news, hundreds of teenagers were employed in the city. Similarly, massive presence of bot accounts in pro-Trump and pro-Hillary messages~\cite{BotsAndAutomationOverTwitter} suggests existence of an organized effort to disseminate a message. Propaganda calls for coordination and organization. When carried out effectively it can be quite deceiving and manipulate opinions and decision. It can sway election results and impact the future of a country or the world. Within collaborative and organized use, terrorist groups and extremists may harm security and peace. For example, ISIS\footnote{ISIS is an extremist terrorist organization} and White Supremacy Extremists (FBI declares as ``Domestic Extremist Ideologies''\footnote{FBI known violent extreme groups: \url{}) strategically use social media to recruit new members and to disseminate their propaganda \cite{NazisVsISISTwitter,ISISTwitterCensus}. The automatic detection of propaganda on social media may help to protect social media users and services from malicious activity. This work proposes an approach for the automatic detection of organized behavior, which is likely to be observed in online propaganda. The term organized behavior\ is used to refer to a collaborative and coordinated posting behavior involving multiple users, who serve an agenda. On the other hand, tweets that are posted spontaneously without any predetermined agenda are referred to organic behavior\footnote{The organic term is borrowed from \cite{OrganicorOrganizedUrlCao}}\enlargethispage{-\baselineskip}. For the automatic detection of organized behavior, we propose a supervised learning method that utilizes features of collective behavior in hashtag based tweet sets, which are collected by querying hashtags of interest. Numerous user and temporal features are extracted from the collected tweets to feed machine learning algorithms to train a model for detecting organized behavior. After the analysis of over 200 million tweets which are mainly posted during the 2016 US presidential election The model is trained using a training data set with 851 records. The 851 records in the training data set reflects the analysis of over 200 million tweets which are posted mainly during the 2016 US presidential election. All of the features in the training data set are extracted by analyzing user characteristics \& temporal tweeting patterns, which are independent of the content (except for sentiment based feature) and graph related features. In other words, they are based on the characteristics of a poster and the presence of temporal coordination among the tweets. This is akin to sensing that \textit{something is up} without knowing \textit{what is up}. To discover the latter further analysis would be required on content. The content and graph related features are not used, because extraction of these features are not cheap and easy compared to user \& temporal features. Thus, this might affect the performance of near real time automatic detection. Given the massive quantity of content, such detection is deemed to be of significance, since malicious content can quickly become trending and spread to millions of users. Finally, the extracted features are used to train models to predict tweet set behavior in each of the three categories [organized behavior\ vs. organic behavior], [pro-Trump vs. pro-Hillary vs. none] and [political vs. non-political]. The trained models gave promising results for each category, where the Random Forest algorithm consistently resulted in high scores with an average accuracy and f-measure greater than $0.95$. Although the training data size, 851 records, is somewhat small for a supervised learning approach, the resulting classification models are able to distinguish tweet sets among the categories and are encouraging for use in larger systems. This work proposes a proof of concept method for the near real time detection of propaganda and organized behavior. The main contributions of this work are as follows: a classification model to detect organized behavior in tweet sets based on only user and temporal posting features, a prototype implementation for feature extraction \& classification, qualitative and quantitative description of features of tweet sets involving the controversial 2016 US election. The remainder of the paper is structured as follows: Section~\ref{} describes background information required to follow this work; Section~\ref{} presents related work; Section~\ref{} describes the overall approach; Section~\ref{} presents the proposed model; Section~\ref{} provides definitions of the extracted features; Section~\ref{} presents the experiments and results; Section~\ref{} discusses the results and future work; Section~\ref{} concluding remarks. \section{Acknowledgment} \label{} This work is partially supported by the Turkish State Planning Organization (DPT) under the TAM Project, number 2007K120610. We thank the members of the Complex Systems Lab (SosLab) at the Computer Engineering Department of Bogazici Universitysupport for their support and constructive comments. \bibliographystyle{achemso} \section{An Organized Behavior Detection Model} \label{} The automated detection of organized behavior on Twitter calls for methods that scale to the data corresponding to massive numbers of posts and users. Machine learning approaches are promising in many classification problems related to complex large datasets of social big data\cite{BELLOORGAZ201645}. This work proposes a simple model based on supervised learning methods. Figure~\ref{} shows the overall approach that consists of two main phases: feature extraction and model generation. The feature extraction phase handles collecting the tweets of interest and extracting their features. A set of tweets of interest is referred to as a \emph{collection}. Tweets of interest are chosen to be those that contain a specific hashtag, since hashtags are widely used to increase engagement. Feature extraction is performed on each collection, resulting in rows of features that are used to train the classifiers. \begin{figure} \begin{center} \includegraphics[height= 0.25\paperheight, width=1\columnwidth]{organizedBehaviorDiagrams} \end{center} \caption{The overview of generating a classification model for tweet sets.} \vskip\baselineskip \label{} \end{figure} \subsection{Feature Extraction Phase} \label{} \input{feature-extraction-algorithm} The feature extraction phase performs two main tasks: preparation of a collection and extracting its features (Algorithm~\ref{}). A collection is created by fetching tweets with occurrences of a \textit{tracedHT}\ (hashtag of interest). Additional tweets of the users who posted these tweets are fetched to expand the collection in order to get more information about them. The resulting collection is referred to with a hashtag (i.e. \#lockHerUp collection). A \textit{tracedHT}\ may be associated with multiple collections if they are created at different times that it is used (i.e. \#guncontrol). Collection expansion consists of fetching additional tweets that were posted within a given time before and after the time of a post captured in \textit{seed tweets}\ (line~11~in~Algorithm~\ref{}). In our experiments we chose this duration to be one week. The intent is to capture whether there is a significant difference in the behavior of a user before and after the time of their post in \textit{seed tweets}. The collection of \textit{seed tweets}\ expanded with the user tweets is referred to as the \textit{expanded tweet set} (\textit{ET}). An unqualified reference to a collection should be interpreted as the \textit{expanded tweet set}. A reference to the \emph{users in a collection} should be interpreted as the users who contributed to the posts in that collection. Expanding the collection does have a considerable impact on the size of the collection. Table~\ref{} shows size comparisons of a few sample collections. All the user and temporal features, as defined in Section~\ref{}, are extracted for each collection and stored in the \emph{Features Repository}. User features are computed based on individual characteristics as well as those that compare a user with others in the collection. The temporal features of tweets are computed based on time intervals corresponding to the time period of the collection (Section~\ref{}). In our experiments this size was chosen as one hour. Inspecting the tweets according the time periods attempts to detect the presence of coordinated posts, since organized activities typically strategically schedule posting times. Finally, in order to have an overall view of collections, the mean, variance, standard deviation, minimum, and maximum values for all the features are computed and stored in the \textit{Features Repository} as a row of the training data set. \subsection{Model Generation Phase} Supervised learning methods are used to train models using the features extracted for the collections. Three classifiers are generated using the supervised learning methods of random forest, SVM, and logistic regression. Furthermore, three kinds of classifications are generated: organized behavior\ vs. organic behavior\ , \textit{pro-Trump} vs. \textit{pro-Hillary}, and \textit{political} vs. \textit{non-political}. The models were chosen to explore different contexts of organized behavior. More specifically, the presence of any kind of organization, a specific kind of group based organization, and topic based organization. Section~\ref{} provides details about the specific collections, the features generated from them, the performance of the generated classifiers, and a comparison of the chosen supervised learning methods. \section{Related Work} \label{} The work in the area of detecting online behavior can be categorized into three main types: behavior detection, spam detection, content and group detection. Cao \cite{OrganicorOrganizedUrlCao} examines the URL sharing behavior in Twitter and distinguishes the sharing behaviors as organic and organized. For the detection of organized urls, a graph is generated with nodes representing users and edges representing the use of the same URLs in posts. A community detection algorithm using the Louvain method\cite{comlargenet} is applied to the graph to extract user groups. Then, URL and posting time based features of those users are extracted and supervised classifications methods, such as Random Forest, NBTree, SMO and LogitBoost, are applied on manually labeled training data with 406 organic and 406 organized records. After the training phase with 10-fold cross validation, it is shown that applied approach gives a performance around 0.8, in all methods, while Random Forest works best with high F measure (0.836) and ROC area (0.921). Similarly, Ratkiewicz \cite{DetectingAndTrackingPoliticalAbuse} studied astroturf political campaigns, which are run by politically-motivated individuals and organizations that use multiple centrally-controlled accounts. During the study, topological, content-based and crowdsourced features of information diffusion networks on Twitter are extracted. These features are mainly extracted from composed directed graph whose nodes represent individual user accounts and edges represent the retweet, mention, reply events between users. Supervised Learning Methods are used for understanding early stages of viral diffusion. It is presented that their approach gives accuracy results better than 96\% in the detection of astroturf content in a dataset of 2010 US midterm elections. In another study about the classification of group behaviors \cite{ClassificationofGroupBehaviors}, it is aimed to detect criminal and anti-social activities in social media. To detect these behaviors, the theories of group behaviors and interactions were developed. Also, graph matching algorithms are applied to explore consistent social interactions, which suggest that complex collaborative behaviors can be modeled and detected using a concept of group behavior grammars. Also, in order to disclose spam URLs, Cao \cite{DetectingSpamURLCao} analyzed behavioral features in three categories : click-based, posting based, clicking statistics. With these features, a training data set is created with 1,049 spam and benign urls by checking the labels of urls in the tweets from a category website. Another training data set is created by manual labelling of 219 benign and 79 spam urls. For the both behavioral features in two training data sets, random forest algorithm is used as supervised learning method by using 10-fold cross validation. The algorithm is trained by using all feature sets together and separately such as click-based, posting based, clicking statistics. With this approach, 86\% of accuracy is found by using all features for the training. In another research, political bot accounts, who take place in Brexit Referendum and play strategic role in referendum conversations, are analyzed~\cite{BotsStrongerInBrexit}. It is found that these bots use excessively the family of hashtags associated with the argument for leaving the European Union, and utilize different levels of automation. It is stated that these bots, which compose one percent of sampled accounts, generate almost a third of all the messages for the Brexit referendum contents in Twitter. When Cao's study about organized urls \cite{OrganicorOrganizedUrlCao} is compared with our study, it will be seen that the main focus of this study \cite{OrganicorOrganizedUrlCao} is to detect organized URL behaviors, while our study aims to detect organized behaviors. Also, the extracted and derived features are different from each other. However, in both study, same classification method approaches are used and random forest algorithm gave the most promising results. Furthermore, when Ratkiewicz's study about astroturf political campaigns \cite{DetectingAndTrackingPoliticalAbuse} is compared with our study, same supervised classification methods are used in both studies, while feature extraction phases are different from each other. In their study, features are extracted from a constructed graph, while in our study they are extracted from users and temporal tweets. Also in the study about group classification \cite{ClassificationofGroupBehaviors}, group behaviors are identified as a result of graph matching algorithms. In our study no graph related algorithm is used nor a graph related feature is extracted due to performance concerns for real time detection. In the work about \cite{DetectingSpamURLCao} identifying spam urls, similar works have been done considering feature extraction and classification methods. However, our work focuses on detection of organized behavior patterns and features of user \& temporal tweets. There are also studies about content and group detection which are in parallel with our study \cite{AutoDetectionElectionRelatedTweets,DetectingJihadistMessages,AutomaticClassificationMuseumTweets}. In all studies either underlying content or group is detected. However, in all of them the detection is done by using language or topic related features. On the other hand, in our study this detection has been made with features which do not contain network and content related features except the sentiment features.

\section{Introduction}\label{} We are in the new avenue of the modern cosmology immediately after the observational evidence~\cite{Riess1998,Perlmutter1999,Bernardis2000,Perlmutter2003} of late-time acceleration of the Universe. This invites a serious theoretical challenge to Einstein's general theory of relativity, which was proven unquestionable for the last several decades. To address this situation in the framework of GR many authors~\cite{Caldwell2002,Nojiri2003,Odinstov2003,Padmanabhan2002,Kamenshchik2001,Bento2002} introduced presence of exotic entity, widely known as dark energy. On the other hand, a class of authors proposed different modified theories time and again to address the present accelerating phase of the universe. Essentially by replacing the standard Einstein-Hilbert action by arbitrary function of Ricci scalar $R$, i.e., $f(R)$ it was successfully possible to explain the late time cosmic acceleration, unified inflation with dark energy and also galactic dynamics of massive test particles without introducing any exotic dark energy~\cite{Carroll2004,Nojiri2007,Nojiri2008,Cognola2008,Cognola2011,Nojiri2011}. For the detailed work on $f(R)$ gravity one may consult the following literature~\cite{Lobo2008,Sotiriou2010,Capozziello2010}. Recently, Harko et al.~\cite{harko2011} have presented more generalized form of the $f(R)$ gravity theory by choosing the Lagrangian density as the arbitrary function $f (R, \mathcal{T} )$ where $R$ is the Ricci scalar and $\mathcal{T}$ is the trace of the energy-momentum tensor. This is known as $f (R, \mathcal{T} )$ theory of gravity and depends on a source term, which represents the variation of the energy-momentum tensor with respect to the metric, due to coupling of the matter and geometry. Here, for the presence of an extra force perpendicular to the four velocity, the test particles do not follow the geodesic path. However, lateron Chakraborty~\cite{SC2013} showed that due to the linear form of the function $f(R,\mathcal{T})$ given as $f(R,\mathcal{T})=R+h(T)$, the whole system behaves like a non-interacting 2-fluid system and test particles follow the geodesic path. The interaction between the matter and curvature terms can be interpreted as the origin of the 2nd type of fluid. In his detailed work Chakraborty~\cite{SC2013} also explained how rapidly the 2nd type of fluid becomes exotic though the first type is still a normal fluid. In this line of thinking many literature can be referred~\cite{Harko-2008,Bisabr-2012,Jamil12,Alv13,shabani13,shabani14,Zare,shabani16a,shabani16b} where $f(R,\mathcal{T})$ theory of gravity have been used successfully. Sharif et al.~\cite{sharif2014} indicated the factors which are effecting the stability of the isotropic and spherically symmetric stellar system in $f(R,\mathcal{T})$ gravity. To find the collapse equation for the Newtonian and post-Newtonian eras by employing a perturbation scheme on the dynamical equations Noureen et al.~\cite{noureen2015} imposed condition on adiabatic index. Further, by employing the perturbation method on the modified field equations Noureen and Zubair~\cite{zubair2015a,noureen2015b,noureen2015c} in a series of papers studied dynamical analysis of a spherically symmetric collapsing star under different physical situations. Ahmed et al.~\cite{Ahmed2015} studied the effect of $f(R,\mathcal{T})$ gravity on gravitational lensing and compared the yielded result with the standard result of GR. Using Runge-Kutta 4th-order method to solve the TOV equation in $f(R,\mathcal{T})$ gravity Moraes et al.~\cite{Moraes2015} studied hydrostatic equilibrium configurations for the neutron stars and strange stars. Later on Das et al.~\cite{Amit2016} used the results to study spherically symmetric and isotropic compact stellar system in $f(R,\mathcal{T})$ gravity by adopting the Lie algebra with conformal Killing vectors. Under the $f(R,\mathcal{T})$ gravity Das et al.~\cite{Amit2017} also presented a model for Gravastar to avoid singularity and thus provided an alternative to black hole. In this article, keeping the above works in mind, we have presented an exact solution of the TOV equation for the spherically symmetric and isotropic strange star candidates in the framework of $f(R,\mathcal{T})$ theory of gravity. The outline of the present study is organized as follows: In section~\ref{sec1} we present basic mathematical formulation of $f(R,\mathcal{T})$ theory. The Einstein field equations are shown in section~\ref{sec2}. Solutions of the field equations for $f\left(R,\mathcal{T}\right)$ gravity are obtained in section~\ref{sec3}. In section~\ref{sec4} we discuss some physical properties of our stellar system, such as the energy conditions~\ref{subsec4.1}, mass-radius relation~\ref{subsec4.2}, compactification factor and redshift~\ref{subsec4.3} along with the stability conditions~\ref{subsec4.4}, viz., modified TOV equation~\ref{subsubsec4.4.1}, causality condition~\ref{subsubsec4.4.2} and adiabatic index~\ref{subsubsec4.4.3}. Finally, in section~\ref{sec5} we make some concluding remarks. \section{Basic mathematical formulation of $f(R,\mathcal{T})$ Theory}\label{sec1} Following Harko et al.~\cite{harko2011}, we propose the action due to $f(R,\mathcal{T})$ gravity which reads \begin{equation}\label{1.1} S=\frac{1}{16\pi}\int d^{4}xf(R,\mathcal{T})\sqrt{-g}+\int d^{4}x\mathcal{L}_m\sqrt{-g}, \end{equation} where $f(R,\mathcal{T})$, $g$ and $\mathcal{L}_m$ are the general function of the Ricci scalar $R$ and the trace of the energy-momentum tensor $\mathcal{T}$, the determinant of the metric $g_{\mu\nu}$ and the matter Lagrangian density, respectively. Further we consider the geometrical units $G=c=1$. Now variation of the action (\ref{1.1}) with respect to $g_{\mu\nu}$ yields the field equations of the $f(R,\mathcal{T})$ gravity model as follows: \begin{eqnarray}\label{1.2} &\qquad\hspace{-3cm} f_R (R,\mathcal{T}) R_{\mu\nu} - \frac{1}{2} f(R,\mathcal{T}) g_{\mu\nu} + (g_{\mu\nu}\DAlambert - \nabla_{\mu} \nabla_{\nu}) f_R (R,\mathcal{T})\nonumber \\ &\qquad\hspace{5cm} = 8\pi T_{\mu\nu} - f_\mathcal{T}(R,\mathcal{T}) T_{\mu\nu} - f_\mathcal{T}(R,\mathcal{T})\Theta_{\mu\nu}, \end{eqnarray} where {{$f_R (R,\mathcal{T})= \partial f(R,\mathcal{T})/\partial R$, $f_\mathcal{T}(R,\mathcal{T})=\partial f(R,\mathcal{T})/\partial \mathcal{T}$. Again, $\DAlambert \equiv\partial_{\mu}(\sqrt{-g} g^{\mu\nu} \partial_{\nu})/\sqrt{-g}$}} is the D'Alambert operator, $R_{\mu\nu}$ is the Ricci tensor, $\nabla_\mu$ represents the covariant derivative associated with the Levi-Civita connection of $g_{\mu\nu}$, $\Theta_{\mu\nu}= g^{\alpha\beta}\delta T_{\alpha\beta}/\delta g^{\mu\nu}$ and the stress-energy tensor defined as $T_{\mu\nu}=g_{\mu\nu}\mathcal{L}_m-2\partial\mathcal{L}_m/\partial g^{\mu\nu}$~\cite{Landau2002}. The covariant divergence of eq. (\ref{1.2}) gives~\cite{barrientos2014} \begin{eqnarray}\label{1.3} \hspace{-0.5cm}\nabla^{\mu}T_{\mu\nu}&=&\frac{f_\mathcal{T}(R,\mathcal{T})}{8\pi -f_\mathcal{T}(R,\mathcal{T})}[(T_{\mu\nu}+\Theta_{\mu\nu})\nabla^{\mu}\ln f_\mathcal{T}(R,\mathcal{T}) +\nabla^{\mu}\Theta_{\mu\nu}-\frac{1}{2}g_{\mu\nu}\nabla^{\mu}\mathcal{T}]. \end{eqnarray} The above eq. (\ref{1.3}) features that the energy-momentum tensor is not conserved from the point of view of the General Relativity (GR). Basically in the covariant derivative of the stress-energy tensor as $f_\mathcal{T}(R,\mathcal{T}) \neq 0$ (for our model $f_\mathcal{T}(R,\mathcal{T})=2\chi$) and hence $\nabla^{\mu}T_{\mu\nu}\neq 0$. As a result, the system will not be conserved as usual in Einsteinian gravity. This aspect can be shown in details with another treatment as follows: The equation of motion for a test particle~\cite{barrientos2014} in our case can be provided as \begin{eqnarray}\label{2} \frac{d^2x^{\mu}}{ds^2}+\Gamma^{\mu}_{\lambda\nu}{u^{\lambda}}{u^{\nu}}=f^{\mu}, \end{eqnarray} where $f^{\mu}$ is the effective force and is given by \begin{eqnarray}\label{3} f^{\mu}=\frac{\left(8\pi{\nabla}_{\nu}p-\frac{1}{2}f_\mathcal{T}(R,\mathcal{T}){\nabla}_{\nu}\mathcal{T} \right)}{\left(\rho+p\right)\left[8\pi+f_\mathcal{T}(R,\mathcal{T})\right]} h^{\mu\nu}, \end{eqnarray} where $h_{\mu\nu}=g_{\mu\nu}-{u_{\mu}}{u_{\nu}}$ is a projection operator. Introducing the trace $\mathcal{T}=\rho-3p$ and for the vanishing pressure $p$ we find \begin{eqnarray}\label{4} f^{\mu}=-\frac{f_\mathcal{T}(R,\mathcal{T}){\nabla}_{\nu}\rho}{2 \rho \left[8\pi+f_\mathcal{T}(R,\mathcal{T})\right]} h^{\mu\nu}. \end{eqnarray} Clearly unlike the usual Einsteinian gravity the particles do not follow a geodesic path. Hence, the energy-momentum tensor is not conserved in $f(R,\mathcal{T})$ gravity from the point of view of GR. Now, we are considering the energy-momentum tensor of a perfect fluid given as \begin{equation}\label{1.4} T_{\mu\nu}=(\rho+p)u_\mu u_\nu-{p}g_{\mu\nu}, \end{equation} where ${v_{\mu}}$ and ${u_{\nu}}$ are the radial-four vectors and four velocity vectors, respectively. Here $\rho$ and $p$ represent the matter density and pressure of the fluid, respectively. In the present study we choose $\mathcal{L}_m=-p$ and also we have $\Theta_{\mu\nu}=-2T_{\mu\nu}-p g_{\mu\nu}$. The logic behind the choice of $\mathcal{L}_m=-p$ is as follows: here we have followed the calculations as presented by Harko et al.~\cite{harko2011} where they assumed that ${\mathcal{L}}_m = -p$. Again, Faraoni~\cite{Faraoni2009} showed that both the Lagrangian ${\mathcal{L}}_m=p$ and ${\mathcal{L}}_m= -\rho$ are equivalent when the fluid couples minimally to the gravity. Also, Harko in his study~\cite{Harko2010} showed that the Lagrangian expressed in terms of $\rho$ and $p$ only are completely equivalent, as $\rho$ and $p$ are freely interchangeable thermodynamic quantities. Thus the assumption ${\mathcal{L}}_m = -p$ used in the present work is quite arbitrary and does not change the subsequent results whatever the form one adopts. We assume the simplified and linear functional form of $f(R,\mathcal{T})$ as suggested by Harko et al.~\cite{harko2011} which is given as $f(R,\mathcal{T})=R+2\chi\mathcal{T}$, where $\chi$ is a constant. Many authors successfully used this functional form in their $f(R,\mathcal{T})$ models~\cite{singh2015,moraes2014b,moraes2015a,moraes2015b,moraes2017,singh2014,baffou2015,shabani13,shabani14,sharif2014b,reddy2013b,kumar2015,shamir2015,Fayaz2016}. Now substituting the specific form of $f(R,\mathcal{T})$ in eq.~(\ref{1.2}) we have the modified form of the Einstein field equation as follows \begin{equation}\label{1.5} G_{\mu\nu}=8\pi T_{\mu\nu}+\chi \mathcal{T}g_{\mu\nu}+2\chi(T_{\mu\nu}+p g_{\mu\nu})=8\pi T^{eff}_{\mu\nu}, \end{equation} where $G_{\mu\nu}$ is the usual Einstein tensor and $T^{eff}_{\mu\nu}=T_{\mu\nu}+\frac{\chi}{8\pi} \mathcal{T}g_{\mu\nu}+\frac{\chi}{4\pi}(T_{\mu\nu}+p g_{\mu\nu})$. Putting $\chi=0$ in eq.~(\ref{1.5}) the standard general relativistic results can easily be retrieved. Substituting $f(R,\mathcal{T})=R+2\chi\mathcal{T}$ in eq.~(\ref{1.3}) yields \begin{equation}\label{1.6} \left(4\pi+\chi\right)\nabla^{\mu}T_{\mu\nu}=-\frac{1}{2}\chi\left[g_{\mu\nu}\nabla^{\mu}\mathcal{T}+2\nabla^{\mu}(p g_{\mu\nu})\right]. \end{equation} We may also represent eq.~(\ref{1.6}) as \begin{equation}\label{1.7} \nabla^{\mu}T^{eff}_{\mu\nu}=0. \end{equation} Again substituting $\chi=0$ in eqs.~(\ref{1.6}) and (\ref{1.7}) the standard form of the conservation of the energy momentum tensor for GR can be achieved. \section{The Einstein field equations for $f\left(R,\mathcal{T}\right)$ gravity}\label{sec2} We assume the interior spacetime of the spherically symmetric static stellar configuration which can be described by the metric given as \begin{equation}\label{2.1} ds^2=e^{\nu(r)}dt^2-e^{\lambda(r)}dr^2-r^2(d\theta^2+\sin^2\theta d\phi^2), \end{equation} where $\nu$ and $\lambda$ are the metric potentials and functions of the radial coordinate only. Now using eq.~(\ref{1.5}) along with eqs. (\ref{1.4}) and (\ref{2.1}) we have the explicit form of the Einstein field equations in $f\left(R,\mathcal{T}\right)$ gravity as~\cite{Moraes2015,Amit2016,Amit2017} {{\begin{eqnarray}\label{2.2} &\qquad\hspace{-1cm} {{\rm e}^{-\lambda}} \left( {\frac {\lambda^{{\prime}}}{r}}-\frac{1}{{r}^{2}} \right) +\frac{1}{{r}^{2}}= \left( 8\pi +3\chi \right) \rho-\chi p=8\pi{{\rho}^{eff}},\\ \label{2.3} &\qquad\hspace{-1cm} {{\rm e}^{-\lambda}} \left( {\frac {\nu^{{\prime}}}{r}}+\frac{1}{{r}^{2}} \right) -\frac{1}{{r}^{2}}= \left( 8\pi +3\chi \right) p-\chi\rho=8\pi{{p}^{eff}}, \end{eqnarray}}} where $`\prime$' denotes the differentiation with respect to the radial coordinate. Also, the effective density and pressure ${\rho}^{eff}$ and ${p}^{eff}$ respectively, are given as \begin{eqnarray}\label{2.2a} \rho^{eff}=\rho+\frac{\chi}{8\,\pi}\left( 3\rho-p \right), \\ \label{2.2b} p^{eff}=p-\frac{\chi}{8\pi}\left( \rho-3p \right). \end{eqnarray} Considering all the three flavors of quarks are confined in a bag and are non-interacting in nature, we have assumed in the present study that the quark matter distribution is governed by the simplest form of the EOS given by \begin{equation}\label{2.4} {p}+{B}={\sum_f}{p^f}, \end{equation} where $B$ is the total external bag pressure which is a constant quantity under certain numerical range and $p^f$ is the individual pressure due to the quark flavors of up, down and strange quarks. The de-confined quarks inside the bag has the total energy density $\rho$ as follows \begin{equation} \rho={\sum_f}{{\rho}^f}+B, \label{2.5} \end{equation} where ${{\rho}^f}=3{p^f}$ is the energy density of individual quark flavors. Using eqs.~(\ref{2.4}) and (\ref{2.5}) we have the EOS of quark matter as follows \begin{equation} {p}=\frac{1}{3}({{\rho}}-4B),\label{2.6} \end{equation} which is the well known MIT bag EOS. It is worthwhile to mention here that without introducing the critical aspects of quantum mechanical particle physics the MIT bag EOS is used successfully in many literature~\cite{1,2,3,4,5,6,7,8} to study the strange star candidates in the framework of GR. Here the factor $1/3$ in eq. (\ref{2.6}) is arising due to the fact that the $u$, $s$ and $d$ quarks are considered to be massless. Farhi and Jaffe in their study~\cite{Jaffe1984} showed that for the massless and non-interacting quarks the strange quark matter (SQM) might be the true ground state of the strongly interacting matter distribution when $B$ is between $57~MeV/{fm}^3$ and $94~MeV/{fm}^3$. However, in the present study we have chosen value of the bag constant as $B=83~MeV/{{fm}^3}$~\cite{Rahaman2014}. Let us define mass of the spherically symmetric stellar system as \begin{equation}\label{2.7} m \left( r \right) =4\pi \int_{0}^{r}\!{{\rho}^{eff}} \left( r \right) {r}^ {2}{dr}. \end{equation} In order to obtain a non-singular monotonic decreasing matter density inside the stellar system, following Mak and Harko~\cite{MH2002}, we consider the simplified form of $\rho$ given as \begin{equation}\label{2.8} \rho(r)=\rho_c\left[1-\left(1-\frac{\rho_0}{\rho_c}\right)\frac{r^{2}}{R^{2}}\right], \end{equation} where $\rho_c$ and $\rho_0$ are the constants and they represent maximum and minimum values of $\rho$, respectively. Now from the modified form of the Einstein field equation (\ref{1.5}) one can note that in the exterior spacetime $T_{\mu\nu}$ is zero as there is no matter in the vacuum spacetime. Hence we find from eq.~(\ref{1.5}) that $G_{\mu\nu}{\mid}_{ext}=0$ which means that there will be no change of GR solution for the exterior spacetime metric even in the $f(R,\mathcal{T})$ theory of gravity. Hence we adopt the exterior Schwarzschild metric given by \begin{eqnarray}\label{7} {ds}^2=\left(1-\frac{2M}{r}\right)dt^2- {\left(1-\frac{2M}{r}\right)}^{-1} dr^2-r^2(d\theta^2+\sin^2\theta d\phi^2). \end{eqnarray} Using eqs. (\ref{2.2a}) and (\ref{2.6})-(\ref{7}) we have the total mass of the stellar system as \begin{eqnarray}\label{2.10} M=\frac{2R}{45\pi}}{\left(72B{\pi}^{2}{R}^{2}+56B\pi {R}^{2}\chi+9B{R}^{2}{\chi}^{2}+12{\pi }^{2}{R}^{2}\rho_{{c}}+4\pi {R}^{2}\chi\rho_{{c}} \right). \end{eqnarray} Substituting eq.~(\ref{2.7}) into eq.~(\ref{2.2}) we have eventually \begin{eqnarray}\label{2.11} {{\rm e}^{-\lambda \left(r \right) }}=1-{\frac {2m(r)}{r}}. \end{eqnarray} Based on the above mathematical formalism, in the following Section 4, we shall solve the field equations to find out different physical parameters of the system. \section{Solution of Einstein field equations for $f\left(R,\mathcal{T}\right)$ gravity}\label{sec3} Now solving the Einstein field equations~(\ref{2.2}) and (\ref{2.3}) and using the eqs. (\ref{2.2a}), (\ref{2.2b}), (\ref{2.6}) and (\ref{2.10}) we obtain different physical parameters of the system given as \begin{eqnarray}\label{3.1} & \qquad\hspace{-1cm} \lambda \left( r \right) =-\ln \left[ {\frac { \left( 16\lambda_{{1 }}{r}^{2}+\pi \right) {R}^{5}-16\,\lambda_{{1}}{r}^{4}{R}^{3}-5\,M \pi {R}^{2}{r}^{2}+3\,M\pi {r}^{4}}{{R}^{5}\pi }} \right],\\\label{3.2} & \qquad\hspace{-1cm} \nu \left( r \right) =\frac {1}{ \left( 12\pi +4\chi \right) \nu_ {{1}}} \Bigg[-8\nu_{{1}} \left( \pi +\frac{\chi}{8} \right) \ln \Big\lbrace \left( 16\lambda_{{1}}{r}^{2}+\pi \right) {R}^{5}-16 \lambda_{{1}}{r}^{4}{R}^{3}-5M\pi {R}^{2}{r}^{2}+3M\pi {r}^{4} \Big\rbrace\nonumber \\ & \qquad\hspace{-1cm} +12\nu_{{2}}{\rm arctanh} \left({\frac {\lambda_{{1}}{R}^{5}-2\,\lambda_{{1}}{r}^{2}{R}^{3}-{\frac {5M\pi {R}^{2}}{16}}+\frac{3}{8}M\pi {r}^{2}}{\nu_{{1}}} }\right)-12\nu_{{2}}\,{\rm arctanh}\left\lbrace{\frac {{R}^{2} \left( -16{R}^{3}\lambda_{{1}}+M\pi \right) }{16\nu_{{1}}}}\right\rbrace \nonumber \\ & \qquad\hspace{-1cm}+8\nu_{{1}} \Big\lbrace \left(\pi +\frac{\chi}{8}\right) \ln \lbrace {R}^{5}\pi \left( 1-\frac{2M}{R}\right)\rbrace+\frac{3}{2}\left( \pi +\frac{\chi}{3}\right) \ln \left( 1-{\frac {2M}{R}} \right) \Big\rbrace \Bigg] ,\\\label{3.3} & \qquad \rho^{eff}\left( r \right)={\frac {-48\,\lambda_{{1}}{R}^{5}+80\,\lambda_ {{1}}{r}^{2}{R}^{3}+15\,M\pi \,{R}^{2}-15M\pi {r}^{2}}{8{\pi }^{2}{ R}^{5}}},\\\label{3.4} & \qquad p^{eff} \left( r \right) =-{\frac {10 \left( {R}^{3}\lambda_{{1}}-3/16M\pi \right) \left( R-r \right) \left( R+r \right) }{{R}^{5} \left( 3 \pi +\chi \right) \pi }}, \end{eqnarray} where $\lambda_{{1}}= \left( \pi +\frac{\chi}{4}\right) \left( \pi +\frac{\chi}{2}\right)B$, $\nu_{{1}}={R}^{2}\sqrt { {\lambda_{{1}}}^{2}{R}^{6}+\frac{1}{4}\lambda_{{1}}\pi {R}^{4}-\frac{5}{8} M\lambda_{{1}}\pi {R}^{3}-{\frac {3\,{\pi }^{2}RM}{64}}+{\frac {25{\pi }^{2}{M}^{2}}{256}}}$ and $\nu_{{2}}=\frac{4}{3}\left[ B{\pi }^{2}{R}^{3}+ \left( \frac{5}{8}B{R}^{3}\chi-{\frac {5M}{32}} \right) \pi +\frac{1}{16}B{R}^{3}{\chi}^{2} \right] \left(\pi +\frac{\chi}{4}\right) {R}^{2}$. \begin{figure}[!htp]\centering \includegraphics[scale=.3]{lambda.pdf} \includegraphics[scale=.3]{nu.pdf} \caption{Variation of (i) ${e}^{\lambda(r)}$ (left panel) and (ii) ${e}^{\nu(r)}$ (right panel) as a function of the fractional radial coordinate $r/R$ for the strange star candidate $LMC~X-4$ are shown. Here and in what follows $B=83~ MeV/{{fm}^3}$.} \label{Fig1} \end{figure} \begin{figure}[!htp]\centering \includegraphics[scale=.3]{density.pdf} \includegraphics[scale=.3]{pressure.pdf} \caption{Variation of (i) the density $\mathbf{\rho^{eff}}$ (left panel) and (ii) pressure $\mathbf{p^{eff}}$ (right panel) as a function of the fractional radial coordinate $r/R$ for the strange star candidate $LMC~X-4$ are shown.} \label{Fig2} \end{figure} The variation of the physical parameters, viz., $e^{\lambda}$, $e^{\nu}$, $\mathbf{{\rho}^{eff}}$ and $\mathbf{p^{eff}}$ are featured in figures~\ref{Fig1} and \ref{Fig2}. We find from eq. (\ref{1.6}) the modified form of equation of conservation due to $f(R,\mathcal{T})$ gravity as follows \begin{eqnarray}\label{3.5} -p^{{\prime}}-\frac{1}{2}\nu^{{\prime}} \left( \rho+p \right) +{\frac {\chi}{8\,\pi +2\,\chi}}\left( \rho^{{\prime}}-p^{{\prime}} \right)=0. \end{eqnarray} Using eqs. (\ref{2.3}), (\ref{2.7}), (\ref{2.11}) and (\ref{3.5}) we have hydrostatic equilibrium equation for $f\left(R,\mathcal{T}\right)$ gravity which can be given by \begin{eqnarray}\label{3.6} p^{{\prime}}=-{\frac{(\rho+p) \left[ 4\pi pr+{\frac {m}{{r}^{2}}}-\frac{1}{2}\chi\, \left( \rho-3p \right) r \right]}{\left( 1-{\frac {2m}{r}} \right) \left[ 1+{\frac {\chi}{8\pi +2\chi}}\left( 1-{\frac {\rm d\rho}{{\rm d} p}}\right) \right]}}, \end{eqnarray} where we have considered that the matter-energy density $\rho$ depends on the pressure $p$, i.e., $\rho=\rho\left(p\right)$. By substituting $\chi=0$ in eq.~(\ref{3.6}) the standard form of the Tolman-Oppenheimer-Volkoff (TOV) equation can be achieved as applicable in GR. Now using MIT bag EOS~(\ref{2.6}) and eq.~(\ref{2.8}) we have presented exact solution for the hydrostatic equilibrium equation~(\ref{3.6}) for $f\left(R,\mathcal{T}\right)$ gravity theory. Here, as mentioned earlier, the value of the bag constant is assumed as $B=83~MeV/{{fm}^3}$~\cite{Rahaman2014}. We have predicted radius of the strange star candidate $LMC~X-4$ for the parametric values of $\chi$ in Table~\ref{Table 1} and values of the observed mass as shown in Table~\ref{Table 2}. {\footnotesize{\begin{table}[htp] \centering \caption{Numerical values of the physical parameters for the strange star candidate $LMC~X-4$ having mass~$1.29 \pm 0.05~{M_{\odot}}$~\cite{dey2013} due to the different values of $\chi$.} \label{Table 1} \resizebox{\columnwidth}{!}{ \begin{tabular}{ *6{>{\centering\arraybackslash}m{1in}} @{}m{0pt}@{}} \hline Values of $\chi$ & $\chi=-1.6$ & $\chi=-0.8$ & $\chi=0$ & $\chi=0.8$ & $\chi=1.6$ &\\ [2ex] \hline Predicted Radius \textit{(Km)} & 11.205 & 10.285 & 9.511 & 8.841 & 8.249 &\\[1ex] ${\rho}^{eff}_c$ ($gm/{{cm}^3}$) & $5.090\times {{10}^{14}}$ & $6.795\times {{10}^{14}}$ & $8.892\times {{10}^{14}}$ & $1.148\times {{10}^{15}}$ & $1.468\times {{10}^{15}}$ &\\ [1ex] ${\rho}^{eff}_0$ ($gm/{{cm}^3}$) & $3.855\times {{10}^{14}}$ & $4.843\times {{10}^{14}}$ & $5.927\times {{10}^{14}}$ & $7.107\times {{10}^{14}}$ & $8.383\times {{10}^{14}}$ &\\ [1ex] ${p}^{eff}_c$~($dyne/{{cm}^2}$) & $4.455\times {{10}^{34}}$ & $6.391\times {{10}^{34}}$ & $8.883\times {{10}^{34}}$ & $1.207\times {{10}^{35}}$ & $1.613\times {{10}^{35}}$ &\\ [1ex] ${2M}/{R}$ & 0.34 & 0.37 & 0.40 & 0.43 & 0.46 &\\[1ex] $Z_s$ & 0.23 & 0.26 & 0.29 & 0.33 & 0.36 &\\[1ex] \hline \end{tabular} } \end{table} }} \begin{table}[htp] \centering \caption{Numerical values of physical parameters for the different strange star candidates for $\chi=-0.8$ } \label{Table 2} \resizebox{\columnwidth}{!}{ \begin{tabular}{ *7{>{\centering\arraybackslash}m{1in}} @{}m{0pt}@{}} \hline Strange Stars & Observed Mass ($M_{\odot}$) & Predicted Radius ($Km$) & ${{\rho}^{eff}_c}$ $(gm/{cm}^3)$ & ${{p}^{eff}_c}$ $(dyne/{cm}^2)$ & Surface Redshift & $\frac{2M}{R}$ \\ \hline $Vela~X-1$ & $1.77 \pm 0.08$~\cite{dey2013} & $11.180 \pm 0.122$ & $7.756\times {{10}^{14}}$ & $9.534\times {{10}^{34}}$ & 0.370 & 0.467 \\ $4U~1820-30$ & $1.58 \pm 0.06$~\cite{guver2010b} & $10.866 \pm 0.107$ & $7.339\times {{10}^{14}}$ & $8.172\times {{10}^{34}}$ & 0.323 & 0.429 \\ $Cen~X-3$ & $1.49 \pm 0.08$~\cite{dey2013} & $10.700 \pm 0.153$ & $7.159\times {{10}^{14}}$ & $7.580\times {{10}^{34}}$ & 0.303 & 0.411 \\ $LMC~X - 4$ & $1.29 \pm 0.05$~\cite{dey2013} & $10.285 \pm 0.112$ & $6.796\times {{10}^{14}}$ & $6.393\times {{10}^{34}}$ & 0.260 & 0.370 \\ $SMC~X - 1$ & $1.04 \pm 0.09$~\cite{dey2013} & $9.667 \pm 0.247$ & $6.387\times {{10}^{14}}$ & $5.054\times {{10}^{34}}$ & 0.210 & 0.317 \\ \hline \end{tabular} } \end{table} \section{Physical properties of the stellar model under $f\left(R,\mathcal{T}\right)$ gravity}\label{sec4} In this section we discuss physical validity of the achieved solutions under $f\left(R,\mathcal{T}\right)$ gravity. By studying the energy conditions, modified form of the TOV equation, the adiabatic index etc., we try to explore different physical properties of the stellar model. \subsection{Energy conditions}\label{subsec4.1} To be consistent with the energy conditions, viz., Null Energy Condition (NEC), Weak Energy Condition (WEC), Strong Energy Condition (SEC) and Dominant Energy Condition (DEC) in $f\left(R,\mathcal{T}\right)$ gravity theory, following inequalities should be hold simultaneously for the system given as~\cite{SC2013} {{\begin{eqnarray}\label{4.1.1} &\qquad\hspace{-1.5cm} ~NEC:~~{{\rho}^{eff}}+{{p}^{eff}}\geq 0,\\ \label{4.1.2} &\qquad\hspace{-1.5cm} ~WEC:~~{{\rho}^{eff}}+{{p}^{eff}}\geq 0,~{{\rho}^{eff}}\geq 0, \\ \label{4.1.3} &\qquad\hspace{-1.5cm} ~SEC:~~{{\rho}^{eff}}+{{p}^{eff}}\geq 0,~{{\rho}^{eff}}+3{{p}^{eff}}\geq 0, \\ \label{4.1.4} &\qquad\hspace{-1cm} ~DEC:~~{{\rho}^{eff}}\geq 0,~{{\rho}^{eff}}-{{p}^{eff}}\geq 0. \end{eqnarray}}} \begin{figure}[!htpb] \centering \includegraphics[width=5cm]{E_16} \includegraphics[width=5cm]{E_08} \includegraphics[width=5cm]{E0} \includegraphics[width=5cm]{E08} \includegraphics[width=5cm]{E16} \caption{Variation of the energy conditions with the radial coordinate $r/R$ for the strange star candidate $LMC~X-4$ due to different chosen values of $\chi$. } \label{Fig3} \end{figure} Figure~\ref{Fig3} shows that our model is consistent with all the energy conditions for all the chosen parametric values of $\chi$ and hence confirms that the physical validity of the solution. \subsection{Mass-Radius relation}\label{subsec4.2} Using eqs.~(\ref{2.2a}),~(\ref{2.6})~and~(\ref{2.7}) we obtain mass function for our system as \begin{eqnarray}\label{4.2.1} m \left( r \right) = \left({1+{\frac { \chi }{3\pi }}}\right){{\tilde{m}}}+\frac{2}{9}\chi\,B{r}^{3}, \end{eqnarray} where $\tilde{m}=4\pi \int_{0}^{r}\!{{\rho}} \left( r \right) {r}^{2}{dr}$ representing the usual mass term due to the quark matter, whereas the extra term ${\frac {\chi}{3\pi }}\left( \tilde{m}+\frac{2}{3}\pi B{r}^{3} \right)$ is due to the effect of the modification of GR. Hence the total mass of the system as given in eq.~(\ref{2.10}) becomes \begin{eqnarray}\label{4.2.1b} M=\frac{2R}{45\pi}}{\left(72B{\pi}^{2}{R}^{2} +56B\pi {R}^{2}\chi+9B{R}^{2}{\chi}^{2}+12{\pi }^{2}{R}^{2}\rho_{{c}}+4\pi {R}^{2}\chi\rho_{{c}} \right). \end{eqnarray} In figure~\ref{Fig4}, for the chosen values of $\chi$, viz. $-1.6,~-0.8,~0,~0.8$ and $1.6$, we have presented the total mass~$M$ (normalized in solar mass) versus the radius~$R$ relation for the strange star candidates. This figure~\ref{Fig4} features that the maximum value of mass~$\left(M_{max}\right)$ gradually decreases for the increasing values of $\chi>0$. On the other hand, $M_{max}$ increases with the decreasing values of $\chi<0$. \begin{figure}[!htpb]\centering \includegraphics[width=8cm]{MR_curve.pdf} \caption{Mass~$(M/{M_{\odot}})$ vs Radius~($R$~in km) curve of the strange star candidates for different values of $\chi$. Solid circles are representing maximum mass for the strange star candidates.} \label{Fig4} \end{figure} \begin{figure}[!htpb] \centering \includegraphics[width=6cm]{max_M.pdf} \includegraphics[width=6cm]{max_R.pdf} \caption{Variation of the (i) mass~$M/{M_{\odot}}$ (left panel) and (ii) radius~$R$~in km (right panel) of the strange star candidates as a function of the central density $({\rho^{eff}_c})$ are shown. Solid circles are representing maximum mass for the strange star candidates.} \label{Fig5} \end{figure} In figure~\ref{Fig5} we have shown variation of mass~($M/{M_{\odot}}$ in the left panel) and radius~($R$ in the right panel) for the strange star candidates with respect to the central density~$\left({\rho}^{eff}_c\right)$. We find that as the value of $\chi$ increases gradually from the negative to the positive values, the maximum mass points are achieved for the higher values of ${\rho}^{eff}_c$. Also, we find that due to decreasing values of $\chi$ from $\chi=0$, the value of the maximum mass gradually increases and with the increasing values of $\chi$ from $\chi=0$, the value of the maximum mass gradually decreases. For example, when $\chi=-1.6$ and $1.6$ the maximum masses~$3.52~{M_{\odot}}$ and ~$2.17~{M_{\odot}}$ are achieved for ${\rho}^{eff}_c=4.774\times{{10}^{14}}$~$gm/{{cm}^3}$ and ${\rho}^{eff}_c=1.230\times{{10}^{15}}$~$gm/{{cm}^3}$, respectively. When $\chi=0$ the maximum mass~$2.70~{M_{\odot}}$ is attained for ${\rho}^{eff}_c=2.508\times{{10}^{15}}$~$gm/{{cm}^3}$. Figure~\ref{Fig5}~(right panel) also features that the total radius $R$ gradually decreases with the increase of $\chi$ from the negative to the positive values. When $\chi=-1.6$ and $1.6$, the radius corresponding to the maximum mass points are given as $12.630~km$ and $7.955~km$, respectively. On the other hand, for $\chi=0$ we find the radius corresponding to the maximum mass point is $9.797~km$. So, as $\chi$ gradually increases from $\chi=0$ the strange star candidates become less massive and smaller but they become more compact due to increase of the density, however when $\chi$ decreases from $\chi=0$ the strange star candidates become more massive, larger in radius but less dense stellar system. Thus, we are able to justify the main motivation for working with the alternative gravity theories in the context of stellar hydrostatic equilibrium that our model allows more massive objects (in comparison to GR) to exist. \subsection{Compactification factor and redshift}\label{subsec4.3} The compactification factor for our model in $f\left(R,\mathcal{T}\right)$ gravity is given as \begin{eqnarray}\label{4.3.1} u\left(r\right)=\frac{m\left(r\right)}{r}={\frac {{r}^{2} \left( -16\lambda_{{1}}{R}^{5}+16\lambda_{{1} }{r}^{2}{R}^{3}+5M\pi {R}^{2}-3M\pi {r}^{2} \right) }{2\pi{R}^{5}}}. \end{eqnarray} \begin{figure}[!htpb]\centering \includegraphics[scale=.3]{compact.pdf} \includegraphics[scale=.3]{redshift.pdf} \caption{Variation of the (i) compactification factor (left panel) and (ii) redshift (right panel) as a function of the radial coordinate $r/R$ for the strange star candidate $LMC~X-4$.} \label{Fig6} \end{figure} The surface redshift is defined as \begin{eqnarray} & \qquad {Z_s}={{\rm e}^{-\frac{\nu \left( R \right)}{2} }}-1={\frac {1}{\sqrt {1-{\frac {288B{\pi }^{2}{R}^{2}+224B\pi {R}^{2}\chi+36B{R}^{2}{\chi}^{2}+48{\pi }^{2}{R}^{2}\rho_{{c}}+16\pi {R}^{2}\chi \rho_{{c}}}{45\pi }}}}}-1. \end{eqnarray} It is known that one can quantify the gravitational redshift of an absorption line that only allows the measurement of the compactness, but not mass and radius distinctly unless other advanced methods are used. However, in the present study we find from figure~\ref{Fig6} that as the value of $\chi$ decreases from $\chi=0$, the compactification factor as well as the redshift of the stellar system decrease consequently. On the other hand, when $\chi$ increases from $\chi=0$ we find that both the factors increase accordingly. The profiles of both the compacification factor and redshift function are featured in figure~\ref{Fig6}. \subsection{Stability of the system}\label{subsec4.4} In the following subsections we have discussed stability of the stellar system under $f\left(R,\mathcal{T}\right)$ gravity theory by studying (i)~Modified TOV equation, (ii)~Causality condition and (iii)~Adiabatic index. \subsubsection{Modified TOV equation}\label{subsubsec4.4.1} We have shown modified form of the energy conservation equation for $f\left(R,\mathcal{T}\right)$ gravity in eq.~(\ref{1.6}) and later in a more concise form in eq.~(\ref{1.7}). This clearly invites modification of the standard form of the Tolman-Oppenheimer-Volkoff (TOV) equation~\cite{Varela2010,Ponce1993} for GR. In the present case modified form of the TOV equation is shown in eq.~(\ref{3.5}), given as \begin{equation*} -p^{{\prime}}-\frac{1}{2}\nu^{{\prime}} \left( \rho+p \right) +{\frac {\chi}{8\pi +2\chi}}\left( \rho^{{\prime}}-p^{{\prime}} \right)=0, \end{equation*} where the first term represents the hydrodynamic force $(F_h)$, second term is gravitational force~$(F_g)$ and the last term represents a force~$(F_m)$ which arises due to modification of the gravitational Lagrangian of the Einstein-Hilbert action. Above eq.~(\ref{3.5}) predicts that for the system to be in equilibrium the sum of different forces for our system must be zero, i.e., ${F_h}+{F_g}+{F_m}=0$. For our system the forces are as follows: \begin{eqnarray}\label{4.4.1.1} &\qquad\hspace{-1cm} {F_{{g}}=-\frac{1}{2}{{\nu}^{\prime}}\left(\rho+p\right)=\big[10 \big\lbrace \left[ B{R}^{3}{\pi }^{2}+ \left( \frac{3}{4}B{R}^{3}\chi-\frac{3}{16}M \right) \pi +\frac{1}{8}B{R}^{3}{\chi}^{2} \right] {r}^{2}-{\frac {7f_{{1}}}{10}} \big\rbrace} \nonumber\\ &\qquad {\big\lbrace \left( \frac{\chi}{8}+\pi \right) \left( -\frac{16}{3} {R}^{3}\lambda_{{1}}+M\pi \right) {r}^{2}+\frac{16}{3}f_{{2}} \big\rbrace r \big]{\Big/}\big[3 \left( \pi +\frac{\chi}{3}\right) ^{2}\pi {R}^{5} \big\lbrace \left( -{R}^{3}\lambda_{{1}}+\frac{3}{16}M\pi \right) {r}^{4}}\nonumber\\ & \qquad {+\left[ B{R}^{3}{\pi }^{2}+ \left( \frac{3}{4}B{R}^{3}\chi-{\frac {5M}{16}} \right) \pi +\frac{1}{8}B{R}^{3}{\chi}^{2} \right]{R}^{2}{r}^{2}+\frac{1}{16}\,{R}^{5}\pi \big\rbrace\big]},\\\label{4.4.1.2} & \qquad\hspace{-5.5cm} F_{{h}}=-\frac{dp}{dr}=-{\frac {5r \left( 16B{R}^{3}{\pi }^{2}+12B\pi {R}^{ 3}\chi+2B{R}^{3}{\chi}^{2}-3M\pi \right) }{4\pi{R}^{5} \left( 3 \pi +\chi \right) }},\\\label{4.4.1.3} &\qquad\hspace{-4.8cm} F_{{m}}={\frac {\chi}{8\pi +2\chi}}\left(\rho^{{\prime}}-p^{{\prime}} \right)={\frac {5\chi r \left( 16B{R}^{3}{\pi }^{2}+12B\pi {R}^{3}\chi+2B{R}^{3}{\chi}^{2}-3M\pi \right) }{ 4\left( 4\pi + \chi \right) \pi {R}^{5} \left( 3\pi +\chi \right) }}, \end{eqnarray} where $f_{{1}}= \left( \frac{\chi}{2}+\pi \right) \left( \pi +\frac{3\chi}{14} \right) B{R}^{5}-{\frac {15}{56}}M\pi {R}^{2}$ and $f_{{2}}= \lambda_{{1}}\left( \frac{\chi}{8}+\pi \right) {R}^{5}-{\frac {15}{64}}M \left( \pi +\frac{\chi}{6} \right) \pi {R}^{2}$. \begin{figure}[!htpb] \centering \includegraphics[width=5cm]{f_16} \includegraphics[width=5cm]{f_08} \includegraphics[width=5cm]{f0} \includegraphics[width=5cm]{f08} \includegraphics[width=5cm]{f16} \caption{Variation of forces with the radial coordinate $r/R$ for the strange star candidate $LMC~X-4$ due to different chosen values of $\chi$.} \label{Fig7} \end{figure} Figure~\ref{Fig7} features that the equilibrium of forces are achieved and this supports stability of the system. For $\chi>0$ the combined effect of $F_g$ and $F_m$ is counterbalanced by $F_h$ and for $\chi<0$ the effect of $F_g$ is counterbalanced by the combined effect of $F_h$ and $F_m$. For $\chi>0$ the force $F_m$ is attractive in nature and acts along the inward direction, however for $\chi<0$ we find $F_m$ is repulsive in nature and acts along the outward direction. For $\chi=0$ the effect of $F_g$ is counterbalanced by $F_h$ and the standard result of GR is retrieved. \subsubsection{Causality condition}\label{subsubsec4.4.2} According to the causality condition for a physically stable stellar model, the square of the sound speed~$\left({v^2_s}=\frac{d{p^{eff}}}{d{{\rho}^{eff}}} \right)$ should lie within the limit $\left[0,1\right]$, i.e., $0\leq {v^2_s} \leq 1$. In the present study square of the sound speed is given by \begin{eqnarray}\label{4.4.2.1} {v^2_s}={\frac {\pi }{3\pi +\chi}}. \end{eqnarray} \begin{figure}[!htpb] \centering \includegraphics[width=6cm]{sound_vel} \caption{Variation of ${v^2_s}$ with the radial coordinate $r/R$ for the strange star candidate $LMC~X-4$.} \label{Fig8} \end{figure} Figure~\ref{Fig8} features that for $\chi >0$ the value of ${v^2_s}$ decreases than the reference value of ${v^2_s}$ in GR which is $\frac{1}{3}$ and for $\chi<0$ the value of ${v^2_s}$ increases than the value of ${v^2_s}$ as in $\chi=0$. For all the chosen values of $\chi$, we find that ${v^2_s}$ is well within the limit $\left[0,1\right]$. Hence, our model is consistent with the causality condition and is stable. \subsubsection{Adiabatic Index}\label{subsubsec4.4.3} The adiabatic index characterizes stiffness of the EOS for a given density and it can be used to study the stability of both relativistic and non-relativistic compact stars. After the pioneering work by Chandrasekhar~\cite{Chandrasekhar1964}, later on many scientists~\cite{Hillebrandt1976,Horvat2010,Doneva2012,Silva2015} have studied dynamical stability of the stellar system against an infinitesimal radial adiabatic perturbation. Heintzmann and Hillebrandt~\cite{Heintzmann1975} predicted that the adiabatic index should exceed $\frac{4}{3}$ inside a dynamically stable stellar system. For our model adiabatic index $\Gamma$ reads as \begin{eqnarray} & \qquad\hspace{-7cm} \Gamma=\frac{{p^{eff}}+{{\rho}^{eff}}}{{p^{eff}}}\,\frac{d{p^{eff}}}{d{{\rho}^{eff}}}=\frac{{p^{eff}}+{{\rho}^{eff}}}{{p^{eff}}}\,v^2_s\nonumber \\ & \qquad =\frac{14 \left[ \left( \pi +\frac{3\chi}{14}\right) \left( \pi +\frac{\chi}{2}\right) B{ R}^{5}-{\frac {10}{7}}\lambda_{{1}}{r}^{2}{R}^{3}-{\frac {15}{56}}M\pi {R}^{2}+{\frac {15}{56}}M\pi {r}^{2} \right] \left( \pi +\frac{\chi}{4}\right)}{15\left( \lambda_{{1}}{R}^{3}-\frac{3}{16}\,M\pi \right) \left( R-r \right) \left( R+r \right) \left( \pi +\frac{\chi}{3}\right)}. \label{4.4.3.1} \end{eqnarray} \begin{figure}[!htp] \centering \includegraphics[width=6cm]{adia_index} \caption{Variation of adiabatic indeces $\Gamma$ with the radial coordinate $r/R$ for the strange star candidate $LMC~X-4$.} \label{Fig9} \end{figure} The variation of $\Gamma$ is shown in figure~\ref{Fig9} and it reveals that presnt model is completely stable against an infinitesimal radial adiabatic perturbation for all the selected values of $\chi$ . \section{Discussion and conclusion}\label{sec5} In the present article following Harko et al.~\cite{harko2011} we choose a Lagrangian density as a simplified linear function of $f\left(R,\mathcal{T}\right)$ given as $f\left(R,\mathcal{T}\right)=R+2\chi T$, where $R$ is the Ricci scalar and $\mathcal{T}$ is the trace of the energy momentum tensor $T_{\mu\nu}$. Equation~(\ref{1.6}) clearly suggests that the system is not conserved in the point of view of GR as $\nabla^{\mu}T_{\mu\nu} \neq 0$. Again, eq.~(\ref{1.5}) indicates that the matter part is made of not only one type of fluid, which in our case is supposed to be the quark matter, but also a second type of fluid due to the effect of the coupling of matter and geometry. If we consider the energy-momentum tensor due to the effective fluid then the stellar system is conserved as shown in eq.~(\ref{1.7}) and field equations become equivalent to the standard form of the Einstein gravity. For the detailed study, the work by Chakraborty~\cite{SC2013} (and the references therein) can be consulted where it has been explicitly discussed about the origin and nature of the second type of fluid due to $f\left(R,\mathcal{T}\right)$ gravity. However, in the present article our aim has been to study a model of strange stars under the framework of $f\left(R,\mathcal{T}\right)$ gravity and examine the stability as well as the physical properties of the ultra-dense compact stellar system. Considering the simplified MIT bag EOS for quark matter, we have solved Einstein field equations using the observed values of the mass of some of the strange star candidates. To feature different properties and results of the solutions for the chosen parametric values of $\chi$, we considered $LMC~X-4$ as a representative of the strange star candidates. We choose the values of $\chi$ as $\chi=-1.6,~-0.8,~0,~0.8$~and$~1.6$. Variation of the metric potentials with respect to the radial coordinate are shown in figure~\ref{Fig1} which suggests that our system is free from any sort of physical or geometrical singularity. In figure~\ref{Fig2} variation of ${\rho}^{eff}$ and $p^{eff}$ with respect to the radial coordinate are presented. It shows that both the physical parameters are maximum at the center and decrease monotonically throughout the stellar system to become minimum at the surface. Our model is also consistent with all the energy conditions for all the chosen values of $\chi$ and they are shown in figure~\ref{Fig3}. Further, in figure~\ref{Fig4} we have presented the normalized total mass~$M$ versus the radius~$R$ relationship due to chosen values of $\chi$ for a specific value of bag constant given as $B=83~MeV/{fm}^{3}$~\cite{Rahaman2014}. We find that the maximum mass points (as indicated by the solid circles) and their radii decrease gradually with the increasing value of $\chi>0$ and they increase with the decreasing value of $\chi<0$. For example, when $\chi=-1.6$ the maximum mass and the corresponding radius increase to $30.49\%$ and $28.92\%$ than the respective values at $\chi=0$ and become $3.51~{M_{\odot}}$ and $12.63~km$, respectively, whereas the maximum mass and the corresponding radius decrease to $19.63\%$ and $18.8\%$ than the respective values as found in GR and become $2.17~{M_{\odot}}$ and $7.955~km$, respectively for $\chi=1.6$. In figure~\ref{Fig5} we have also shown the variation of the normalized total mass~$M$ and $R$ with respect to the central density of the effective fluid distribution (${{\rho}^{eff}_c}$). We find that with $\chi=-1.6$ the maximum mass point $3.52~{M_{\odot}}$ is obtained for ${{\rho}^{eff}_c}=2.076~{{\rho}_{nuclear}}$, which is $40.57\%$ lower than the value as found in GR. For $\chi=1.6$ the maximum mass point, $2.17~{M_{\odot}}$ is achieved for ${{\rho}^{eff}_c}=5.349~{{\rho}_{nuclear}}$, which is $53.15\%$ higher than the value found in GR. Therefore, essentially it can be predicted that as the values of $\chi$ increase from $\chi=0$, the stellar configuration of a strange star get shrinked and the density inside the system increases gradually to produce an ultra-dense compact stellar system. On the other hand, as $\chi$ decreases from $\chi=0$, the stellar system becomes more massive and larger in size so that the density decreases gradually to produce a less dense compact stellar system. We have studied the present system for the specific form of the density profile (eq. \ref{2.8}) which is advantageous to use, as it is providing singularity free solution at the center. Moreover, applying this known form of the density profile and MIT bag EOS we are able to provide exact solutions of the modified TOV equation (basic eq. (2.9) and hence modified eq. (4.6)) for the $f(R,\mathcal{T})$ theory of gravity, without employing any complex method of the numerical analysis. It is to note that due to this assumption of the density profile we can present the typical mass~$M$ (normalized in solar mass) versus radius~$R$ relationship for the strange star candidates in figure~\ref{Fig4} which may encourage further study in this line for all other suitable density profiles. For the sake of interest of a comparative study, in Table~\ref{Table 1} we have predicted values of different physical parameters for the strange star candidate $LMC~X-4$ due to parametric values of $\chi$. The results from Table~\ref{Table 1} strongly suggest the same conclusions which we already predicted from figures~\ref{Fig4} and~\ref{Fig5}. In figure~\ref{Fig6} we have shown the variation of compactification factor and redshift with the radial coordinate. For the chosen values of $\chi$ we find from Table~\ref{Table 1} that the redshift of $LMC~X-4$ lies within the range $0.23-0.36$. In Table~\ref{Table 2} we have presented values of different physical parameters for the different strange star candidates due to a specific values of $\chi$, viz. $\chi=-0.8$. In the present study high values of the surface and central densities, and the high redshift value confirm that the chosen compact stellar candidates are actually strange quark stars~\cite{Ruderman1972,Glendenning1997,Herjog2011}. Also, from both the Tables.~\ref{Table 1} and \ref{Table 2} we find that for all values of $\chi$ the values of $2M/R$ are well within the Buchdahl limit~\cite{Buchdahl1959}, i.e., $2M/R>8/9$. To discuss the stability of the system in details we have studied the modified TOV equation for $f\left(R,\mathcal{T}\right)$ gravity as given in eq.~(\ref{3.5}). Figure~\ref{Fig7} features that for both $\chi<0$ and $\chi>0$, the sum of all the forces are zero and the system attains equilibrium. It also shows an interesting fact that an extra force $F_m$ arises due to $f\left(R,\mathcal{T}\right)$ gravity. This factor $F_m$ is repulsive in nature and acts along the outward direction for $\chi<0$, however the force is attractive in nature which acts along the inward direction when $\chi>0$. The square of the sound speed ($v_s^2$) due to different values of $\chi$ is constant throughout the stellar configuration and well within the limit $\left[0,1\right]$, which confirms stability of the system. The variation of $v_s^2$ with respect to the radial coordinate is shown in figure~\ref{Fig8}. Also, from figure~\ref{Fig9} we find that the adiabatic indices ($\Gamma$) for the chosen values of $\chi$ are greater than $4/3$ and thus provide stability of the system against an infinitesimal radial adiabatic perturbation. In a nutshell, in the present study by providing an exact solution of the Einstein field equations and examining different physical tests under $f\left(R,\mathcal{T}\right)$ theory of gravity we have presented a stable stellar model suitable for the strange star candidates. \section*{Acknowledgments} SR and FR are thankful to the Inter-University Centre for Astronomy and Astrophysics (IUCAA), Pune, India for providing Visiting Associateship under which a part of this work was carried out. SR is also thankful to the authority of The Institute of Mathematical Sciences, Chennai, India for providing all types of working facility and hospitality under the Associateship scheme. FR is also thankful to DST-SERB (EMR/2016/000193), Govt. of India for providing financial support. A part of this work was completed while D.D. was visiting IUCAA and the author gratefully acknowledges the warm hospitality and facilities at the library there. We all are thankful to the anonymous referee for several pertinent suggestions which have enabled us to modify the manuscript substantially.

\section{Introduction} The theoretical interest in the $3.9$ GeV/c$^2$ energy region of the positive-parity charmonium spectra began in 2003, after the discovery of the $X(3872)$ by the Belle Collaboration~\cite{}. This state, with $J^{PC}=1^{++}$, decays through the $J/\psi\rho$ and $J/\psi\omega$ channels which are, respectively, forbidden and OZI-suppressed for a $c\bar c$ configuration. Since then, many other structures have been observed in this energy region, like the $Y(3940)$ one, discovered by Belle~\cite{}, the $X(3940)$~\cite{} or the $X(3930)$, a $J^{PC}=2^{++}$ peak measured in 2006 by the same Collaboration in the mass spectrum of the $D\bar D$ mesons produced by $\gamma \gamma$ fusion, rapidly assigned to the $\chi_{c2}(2P)$ state at the PDG. Additionally, a recent controversy has appeared concerning the $X(3915)$ resonance~\cite{}, traditionally assigned to a $J^{PC}=0^{++}$ state. Recent works favor a $J^{PC}=2^{++}$ nature for the latter, which as a side effect would imply a large non-$q\bar q$ component for the $X(3915)$ state~\cite{}. Furthermore, a new charmonium-like state decaying to $D\bar D$, called $X(3860)$, has been recently reported by the Belle Collaboration~\cite{} (stimulated by the early work of Ref.~\cite{}) with a mass of $3862_{-32\,-13}^{+26\,+40}$ MeV/$c^2$ and a width of $201_{-67\,-82}^{+154\,+88}$ MeV. In this case, the $J^{PC}=0^{++}$ option is preferred over the $2^{++}$ one, but its quantum numbers are not determined yet. To study this abundance of states, all belonging to the same energy region, it is necessary to include all the nearby thresholds. Based on previous calculations of the $J^{PC}=1^{++}$ sector by Refs.~\cite{}, which used a coupled-channels approach including the bare $c\bar c\,(2^3P_1)$ and $D\bar D^\ast+h.c.$ components, in this work we will perform a calculation for the $0^{++}$ and $2^{++}$ channels~\cite{} including the $c\bar c\,(2^3P_{\{0,2\}})$ bare charmonium states, which are coupled to the closest thresholds. Recently, this formalism has also been successfully applied to the $c\bar s$~\cite{} and $b\bar s$ sectors~\cite{}. \vspace*{-.25cm} \section{Coupled-channels model} \label{} \vspace*{-.25cm} In order to describe mixed molecular-$c\bar c$ states we need two ingredients. On the one hand, a model for the $q-q$ interactions to describe the naive charmonium spectra and the hadron-hadron interactions. On the other hand, a formalism to describe the coupling between Fock spaces must be developed, through a mechanism that couples the two and four quark systems. Regarding the first ingredient, the constituent quark model (CQM) employed in this work has been extensively detailed in Ref.~\cite{}, thus only its most relevant aspects will be briefly described. Further details of the model and explicit expressions for these interactions are given in the latter reference. This CQM has been successful in describing the meson~\cite{} and baryon phenomenology~\cite{}. The present CQM is based on the assumption that quarks acquire a dynamical mass as a consequence of the spontaneous chiral symmetry breaking. In order to compensate the mass term in a chiral-invariant QCD Lagrangian we must include chiral fields. Then, the simplest Lagrangian that fulfills the previous properties can be written as \begin{equation} \label{lagrangian} {\mathcal L} =\overline{\psi }(i\, {\slashed{\partial}} -M(p^{2})U^{\gamma_{5}})\,\psi \end{equation} where $U^{\gamma_5}=e^{i\frac{\lambda _{a}}{f_{\pi }}\phi ^{a}\gamma _{5}}$ is the Goldstone boson fields matrix and $M(p^2)$ the dynamical constituent mass. The Goldstone boson field matrix $U^{\gamma_{5}}$ can be expanded in terms of boson fields, from where one-boson exchange interaction between quarks naturally emerges. Above the chiral symmetry breaking scale the interaction through Goldstone bosons does not act. However, quarks still interact through gluon exchange diagrams. Besides, the model is completed with the confinement, a nonperturbative QCD effect. This interaction can be modeled with a screened potential~\cite{PhysRevD.40.1653}, which takes into account the saturation of the potential at some interquark distance due to the spontaneous creation of light-quark pairs. About the second ingredient, the aforementioned CQM is insufficient to describe the interplay between charmonium and meson-meson channels, hence, we need to develop a complete coupled-channels formalism. For that purpose we will follow the formalism of Ref.~\cite{}. That way, the influence of the $c \bar c$ states on the dynamics of the two meson states is studied. This allows to generate new states through hadron-hadron interaction, described by the coupling with heavy quarkonium states and the underlying $q-q$ interaction. The global hadronic state, combination of heavy quarkonium and meson-meson channels, is \begin{equation} \label{} | \Psi \rangle = \sum_\alpha c_\alpha | \psi_\alpha \rangle + \sum_\beta \chi_\beta(P) |\phi_{M_1} \phi_{M_2} \beta \rangle \end{equation} being $|\psi_\alpha\rangle$ the $c\bar c$ eigenstates of the two body Hamiltonian and $\phi_{M_i}$ the $c\bar n$ (or $\bar c n$)~\footnote{In this work we will denote $n$ as the light quarks $\{u,d,s\}$} eigenstates describing the $M_i$ mesons. Moreover, $|\phi_{M_1} \phi_{M_2} \beta \rangle$ is the two meson state with $\beta$ quantum numbers coupled to total $J^{PC}$ quantum numbers and $\chi_\beta(P)$ is the relative wave function between the two mesons ${M_1}{M_2}$ in the molecule. The description of hadron-hadron interaction from the fundamental $q-q$ interactions is obtained from Resonating Group Method (RGM). The wave functions of the mesons (either $c\bar c$ or $c\bar n$) are calculated by solving the Sch\"odinger equation for the quark-antiquark problem using the Gaussian Expansion Method (GEM). We employ Gaussian trial functions whose ranges are in geometric progression~\cite{}, which helps to optimize the ranges using few free parameters. This choice generates a dense distribution of Gaussians at small range, which is adequate for short-range potentials. The coupling between the $c\bar c$ and the $c\bar n-n\bar c$ channels requires the creation of a light quark pair $q\bar q$. We will use the quark pair creation $^3P_0$ model~\cite{}, which describes the strong decay processes from simple principles, giving equivalent results compared to microscopic calculations~\cite{PhysRevD.54.6811}. Accordingly, the transition potential $h_{\beta \alpha}(P)$ describing the $c\bar c\to M_1M_2$ process within the $^3 P_0$ model can be defined as $\langle \phi_{M_1} \phi_{M_2} \beta | \mathcal{T}| \psi_\alpha \rangle = P \, h_{\beta \alpha}(P) \,\delta^{(3)}(\vec P_{\mbox{cm}})$, being $P$ the relative momentum of the two meson state and $\mathcal{T}$ the transition operator, non-relativistic reduction of the $^3P_0$ pair creation Hamiltonian. Now, the coupled-channels system for the bare charmonium states and the meson-meson channels can be expressed as an Schr\"odinger-type equation~\cite{} where, apart from the potential coming from $q-q$ interactions described by the CQM, we have an effective potential coding the coupling with $c\bar c$ states, \begin{equation}\label{} V^{\rm eff}_{\beta'\beta}(P',P;E)=\sum_{\alpha}\frac{h_{\beta'\alpha}(P') h_{\alpha\beta}(P)}{E-M_{\alpha}}. \end {equation} In order to describe states above thresholds, a more suitable formalism is to describe resonances as poles in the complex $T(\vec p,\vec p',E)$ matrix, solution of the Lippmann-Schwinger equation. Employing the total potential $V_T(P',P;E)=V(P',P)+V^{\rm eff}(P',P;E)$, where $V(P',P)$ is the potential from CQM and $V^{\rm eff}_{\beta'\beta}(P',P;E)$ the effective potential of Eq.~\ref{}, the $T$-matrix can be factorized as \begin{equation} \label{} T^{\beta'\beta}(P',P;E)=T^{\beta'\beta}_V(P',P;E) +\sum_{\alpha,\alpha'}\phi^{\beta'\alpha'}(P';E) \Delta_{\alpha'\alpha}^{-1}(E)\phi^{\alpha\beta}(P;E) \end{equation} where $T^{\beta'\beta}_V(P',P;E)$ the $T$ matrix of the CQM potential excluding the coupling to the $c\bar c$ pairs and $\Delta_{\alpha'\alpha}(E)=\left((E-M_{\alpha})\delta^{\alpha'\alpha}+\mathcal{G}^{\alpha'\alpha}(E)\right)$ is the propagator of the mixed state, where $M_\alpha$ is the bare mass of the heavy quarkonium and $\mathcal{G}^{\alpha'\alpha}(E)$ the complete mass-shift. Then, resonances emerge as zeros of the latter propagator ($\left|\Delta_{\alpha'\alpha}(\bar{E})\right|=0$). Additionally, the $\phi^{\beta\alpha}(P;E)$ functions are the $^3P_0$ vertex dressed by the RGM meson-meson interaction. \vspace*{-.25cm} \section{Results} \label{} \vspace*{-.25cm} For the present calculation, the bare $c\bar c\,(2^3P_{\{0,2\}})$ meson states will be coupled to $D\bar D$ (3734 MeV/c$^2$), $\omega J/\psi$ (3880 MeV/c$^2$), $D_s\bar D_s$ (3937 MeV/c$^2$) and $D^\ast\bar D^\ast$ (4017 MeV/c$^2$) channels, where the threshold energies for each channel are indicated in parenthesis. For $J^{PC}=2^{++}$ we add the $D\bar D^\ast+h.c.$ (3877 MeV/c$^2$) channel, which is closed for $0^{++}$ sector. The $D^\ast \bar D^\ast$ channel is included because it is the only threshold which couples to the $c\bar c\,(2^3P_0)$ in $S$ wave and, hence, its effect is expected to be essential to describe the $0^{++}$ sector. Furthermore, Heavy Quark Spin Symmetry tells us that the $D^\ast\bar D^\ast$ in the $2^{++}$ sector has the same dynamics as the $D\bar D^\ast+h.c.$ channel in the $1^{++}$ sector, which is fundamental in describing the $X(3872)$. Table~\ref{ A} shows the masses and widths of the states found in the coupled-channels calculation using the original parameters of Ref.~\cite{} (denoted as {\it model A}). In addition, to explore the sensitivity of the results, a second calculation (named {\it model B}) is shown in the same table, where the bare mass of the $c\bar c\,(2^3P_{\{0,2\}})$ pairs has been slightly changed ($0.25\%$) and the value of the $^3P_0$ strength parameter $\gamma$ has been taken from Ref.~\cite{}. Such changes in the initial parameters are within the uncertainties of the model. \begin{table*}[!t] \begin{tabular}{cccccccccc} Model &$J^{PC}$ & Mass & Width & ${\cal P} [c\bar{c}] $ & ${\cal P}[D\bar D]$ & ${\cal P}[D \bar D^{\ast}]$ & ${\cal P}[\omega J/\psi]$ & ${\cal P}[D_s\bar D_s]$ & ${\cal P}[D^{\ast} \bar D^{\ast}]$\\ \hline \multirow{3}{*}{A}&$0^{++}$ & $3890.3$ & $6.7$ & $44.1\%$ & $21.6\%$ & $-$ & $28.4\%$ & $2.6\%$ & $3.3\%$ \\ &$0^{++}$ & $3927.4$ & $229.8$ & $19.2\%$ & $66.3\%$ & $-$ & $5.3\%$ & $3.7\%$ & $5.5\%$ \\ &$2^{++}$ & $3925.6$ & $19.0$ & $42.2\%$ & $11.3\%$ & $37.0\%$ & $4.0\%$ & $0.4\%$ & $5.1\%$ \\ \hline \multirow{3}{*}{B}&$0^{++}$ & $3889.0$ & $11.8$ & $43.5\%$ & $27.3\%$ & $-$ & $20.4\%$ & $3.8\%$ & $4.9\%$ \\ &$0^{++}$ & $3947.5$ & $201.6$ & $19.4\%$ & $66.0\%$ & $-$ & $3.7\%$ & $8.0\%$ & $2.9\%$ \\ &$2^{++}$ & $3915.1$ & $19.8$ & $37.8\%$ & $14.1\%$ & $36.4\%$ & $5.12\%$ & $0.4\%$ & $6.1\%$ \\ \end{tabular} \caption{\label{ A} Mass and decay width, in MeV, and probabilities of the different channels, for models A and B.} \end{table*} The same number of states is predicted by both models: two resonances with $J^{PC}\!\!=\!0^{++}$, a broad and a narrow one, and one with $J^{PC}\!\!=\!2^{++}$. The mass and width of the $J^{PC}\!\!=\!2^{++}$ state is compatible with those measured for the $X(3930)$ for model A, but its mass moves closer to the mass region of the $X(3915)$ for model B, so its assignment remains unclear. In the $J^{PC}\!\!=\!0^{++}$ region both models show a similar phenomenology. On the one hand, the mass of the narrow state is more compatible with the new $X(3860)$ resonance than to the $X(3915)$. However, its predicted width is smaller than the experimental one, which is most likely connected to the position of the node in the $c\bar c\,(2^3P_0)$ bare wave function, which makes the width sensitive to small variations in the wave function structure or, equivalently, to the mass of the $X(3860)$. The second $J^{PC}\!\!=\!0^{++}$ state is more intriguing. Its mass suggests a preliminary assignment to the $Y(3940)$ resonance, but, as already mentioned, its width is far from the experimental value. Alternatively, one can claim the second $0^{++}$ state is actually the $X(3860)$, as it is a broad state whose width agrees with the experimental one and its mass is close to the experiments, whose value reaches up to $3900$ MeV. Within this picture, the first state seems too narrow, which can complicate its measurement in the experiments of Ref.~\cite{}. What appears to be clear is that none of the $J^{PC}\!\!=\!0^{++}$ states seems to favor a $X(3915)$ assignment, neither by mass nor by width. Nevertheless, the predicted $J^P=2^+$ state is compatible with both $X(3915)$ and $X(3930)$, within the uncertainties of the CQM and the value of the $^3P_0$ quark-pair creation parameter. So, in order to further analyze the properties of the $X(3915)$ resonance, we can explore its decay properties. Under a $J^{PC}\!\!=\!0^{++}$ hypothesis for the $X(3915)$ state we are unable to describe the experimental branching fractions. In contrast, the product of the two-photon decay width and the branching fraction to $\omega J/\psi$ and $D\bar D$ channels can be calculated assuming the $X(3915)$ and $X(3930)$ are two faces of the same $J^{PC}\!\!=\!2^{++}$ resonance. Such results are quoted in Table~\ref{tab:BR J2}, where the decay to the $D\bar D^*$ channel is also shown. The results for both models A and B are very similar, not far from the experimental values. Consequently, our results suggest that both the $X(3915)$ and $X(3930)$ resonances fit within a $J^{PC}\!\!=\!2^{++}$ assignment, pointing to the conclusion that they are two faces of the same state, which agrees with Ref.~\cite{}. \begin{table*}[!t] \begin{tabular}{ccccc} & Belle & BaBar & model A & model B \\ \hline $\Gamma_{\gamma \gamma} \times {\cal B}(2^{++}\to \omega J/\psi )$ & $18(5)(2)$ & $10.5(1.9)(0.6)$ & $20.9$ & $24.9$ \\ $\Gamma_{\gamma \gamma} \times {\cal B}(2^{++}\to D\bar D )$ & $180(50)(30) $ & $249(50)(40)$ & $75.4$ & $81.4$ \\ $\Gamma_{\gamma \gamma} \times {\cal B}(2^{++}\to D\bar{D^\ast})$ & - & - & $196.0$ & $151.9$ \\ \end{tabular} \caption{\label{tab:BR J2} Product of the two-photon decay width and the branching fraction (in eV) for the $J^{PC}=2^{++}$ sector for each model, compared with Belle and BaBar Collaboration experimental results~\cite{}} \end{table*} \vspace*{-.35cm} \section*{Acknowledgments} \vspace*{-.35cm} This work has been partially funded by EU's Horizon 2020 programme under the Marie Sk\l{}odowska-Curie grant agreement no. 665919; by Spanish MINECO under Contract nos. FPA2014-55613-P and SEV-2016-0588; and by Junta de Castilla y Le\'on and ERDF under Contract no. SA041U16. \vspace*{-.35cm} \bibliographystyle{JHEP}

\section{Introduction} The Tully-Fisher relation (TFR) \citep{tully1977new} is the fundamental scaling relation between stellar mass \citep{mcgaugh2000baryonic} (originally luminosity) and rotation velocity. It has been shown to hold consistently in the nearby universe for regular disk galaxies [e.g. \citet{bell2001stellar, kassin2006stellar}]. The traditional use of the TFR has been as a distance indicator, via the establishment of the tightest possible scaling relation. This usage requires the exclusion of galaxies outside limited morphology ranges, or that show signs of interaction [e.g. \citet{pierce1992luminosity,bureau1996new,giovanelli1996band,haynes1999band,tully2000distances}]. Scatter below the TFR has been linked to low stellar mass and perturbation using a variety of methods, including optical rotation curves \citep{barton2001tully,kannappan2002physical} and integral field spectroscopy (IFS)\citep{cortese2014sami}. In previous work, we found an inverse relationship between stellar mass and kinematic asymmetry for galaxies in the Sydney--AAO Multi-object Integral field spectrograph (SAMI) Galaxy Survey sample \citep{bloom2016sami}. This result was in agreement with other work, using a variety of metrics for disturbance [e.g. \citet{van1998neutral,cannon2004complex,lelli2014dynamics}], demonstrating that low stellar mass galaxies have complex kinematics, deviating from ordered rotation. Rotation curve measurements of kinematic asymmetry \citep{barton2001tully,garrido2005ghasp} found dwarf galaxies to be both disordered and to have low {measured} rotational velocity. It is also well known that stellar mass is linked to morphological type, with low mass galaxies tending to have irregular morphologies \citep{roberts1994physical,mahajan2015galaxy}. In the past, spectroscopic measurements were mostly taken using a single fibre or slit \citep{york2000sloan,percival20012Df,driver2009gama}. {There have been extensive TFR studies with single slit and fibre measurements [\citet{tully1977new}, \citet{courteau1997optical}, \citet{bohm2004tully}, \citet{mocz2012tully} and others]}. Despite the success of fibre and slit-based measurements in determining the TFR, they are vulnerable to potential errors introduced by slit placement {\citep{spekkens2005cusp,oh2011dark,simons2015transition}} and aperture effects. It is also difficult to investigate spatial variation across an extended source using a single slit. Two dimensional spatially resolved kinematics provide a means to circumvent this problem, allowing for greater robustness in kinematic measurements. SAMI is a multiplexed integral field spectrograph, enabling the production of sample sizes of the order of thousands of galaxies on a much shorter timescale than would be possible with a single integral field spectrograph \citep{croom2012sydney}. Two dimensional spatially resolved spectroscopy has been used to explore the TFR in a variety of ways. The K-band Multi-Object Spectrograph (KMOS) Redshift One Spectroscopic Survey (KROSS) use spatially resolved H$\alpha$ emission to study evolution of the $M_{*}/M_{\odot}$ TFR to $z\sim1$ \citep{tiley2016kmos}. At low redshift, \citet{davis2011atlas3d} use the ATLAS$^{3D}$ sample to show the carbon monoxide (CO) TFR for early-type galaxies. {Recently, the CALIFA Survey have produced a TFR at low redshift using rotation curve fitting to stellar velocity fields \citep{bekeraite2016space}.} The HI TFR has also been thoroughly studied using kinematic maps [e.g. \citet{begum2008baryonic,stark2009first,trachternach2009baryonic,oh2011dark}]. In this work, we show the stellar mass TFR for galaxies in the SAMI Galaxy Survey Sample, and highlight trends with asymmetry and stellar mass. Section~\ref{} details the sample selection, instrumentation, data reduction, and methods. Section~\ref{} presents the TFR for our sample and gives the main relationships between stellar mass, kinematic asymmetry and scatter off the TFR and Section~\ref{} explores the physical and observational causes for scatter below the TFR. We conclude in Section~\ref{}. We assume a standard cosmology, with $\Omega_m = 0.3,\ \Omega_\lambda = 0.7$ and H$_0 = 70 \mathrm{km/s/Mpc}$. \section{Sample and Methods} \label{} \subsection{The SAMI instrument and SAMI Galaxy Survey sample} The SAMI instrument uses imaging fibre bundles, or hexabundles \citep{bland2011hexabundles,bryant2014sami}, to take simultaneous integral field spectra for multiple objects. 61 optical fibres comprise each SAMI hexabundle, with each core subtending $\sim1.6$ arcseconds on sky, yielding a total bundle diameter of $15$ arcseconds. {13 hexabundles can be used simultaneously on sky - 12 on galaxies and 1 on a standard star.} The instrument is installed at the prime focus of the Anglo--Australian Telescope (AAT), with fibre cables feeding the double-beamed AAOmega spectrograph \citep{sharp2006performance,croom2012sydney}. The 3400 galaxies in the SAMI Galaxy Survey sample were selected from those in the GAMA survey \citep{driver2009gama}, with the addition of 8 galaxy clusters {(Owers et al., in prep)}. The galaxies selected from GAMA sample a broad range in stellar mass and environment density. The final SAMI Galaxy Survey sample consists of four stellar mass, volume-limited sub-samples, supplemented by additional low mass and filler samples as detailed in \citet{bryant2015sami}. The SAMI Galaxy Survey data reduction pipeline produces datacubes for each galaxy with adequate signal to noise. For a full description of the SAMI Galaxy Survey data reduction pipeline, we refer the reader to \citet{sharp2015sami}. {For the early SAMI data release, see \citet{allen2015sami}, and for the upcoming full first data release, see Green et al. (in prep.).} This work builds on that in \citet{bloom2016sami}, and uses similar data products from the SAMI Galaxy Survey including spectral fits using {\small LZIFU}{\sc}. {\small LZIFU}{\sc} is a spectral fitting pipeline written in Interactive Data Language ({\small IDL}{\sc}) that performs flexible emission line fitting of IFS data cubes. The {\small LZIFU}{\sc} pipeline produces 2D emission line strength and kinematic maps for user-assigned lines. For a more detailed explanation of the {\small LZIFU}{\sc} pipeline, see \citet{ho2016sami}. At the start of this work, 827 galaxies had been observed and processed through the {\small LZIFU}{\sc} pipeline. In previous work, we used a H$\alpha$ signal to noise (S/N) cut of 10 to exclude noisy spaxels from the velocity fields output by {\small LZIFU}{\sc}. Here, we relax the S/N cut to 6, in order to increase the size of the sample. The median velocity error did not increase, {and the decrease in mean spaxel S/N, from $56.4\pm0.1$ to $48.6\pm0.1$ was deemed to be acceptably small}. It was thus found that this increased the number of galaxies in the sample without compromising data quality. 813 galaxies met the S/N cut requirements and yielded results from kinemetry, a 230$\%$ increase on the sample size in \citet{bloom2016sami}. {As in \citet{bloom2016sami} the H$\alpha$ S/N cut resulted in a sample with a broad range of morphological types, but a bias towards late-type galaxies. } \subsection{TFR and asymmetry measurements} \subsubsection{Stellar mass} \label{} The stellar mass measurements are from the GAMA Survey catalogue StellarMasses \citep{taylor2011galaxy}, and are based on $ugriz$ SEDs constructed from matched aperture photometry. Typical errors {on individual stellar mass values} are 0.1dex. {This is an internal error to our data set and does not account for systematic offsets in stellar mass measures between our data and other papers that adopt different methods of stellar mass estimation, potentially using different model stellar populations.} \subsubsection{Inclination} \label{} In order to correct for velocity projection, we calculate galaxy inclination such that: \begin{equation} cos^{2}(i)=\frac{(1-\eta)^{2}-(1-\eta_{max})^2}{1-(1-\eta_{max})^2} , \end{equation} where $i$ is the angle of inclination, $\eta$ is the ellipticity and $\eta_{max}=0.8$ is taken as the ellipticity exhibited by an edge-on disk. Ellipticities are calculated using 400 arcsec r-band cutout image from the Sloan Digital Sky Survey (SDSS). The input image is processed with SExtractor \citep{bertin1996sextractor} to mask out neighbouring objects. PSFEx \citep{bertin2011automated} is used to build a model point spread function (PSF) at the location of the galaxy centre. The ellipticity is then found as a light-weighted mean for the whole galaxy, using the Multi-Gaussian Expansion (MGE) technique \citep{emsellem1994multi} and code from \citet{cappellari2002efficient}. For more detail, we refer to D'Eugenio et al. (in prep.). \begin{figure} \centering \includegraphics[width=9cm]{hist_class_corr_cutoff.pdf} \caption{Histogram of $\overline{v_{asym}}$ for the full sample in this work (grey), galaxies visually classified as normal (black, dashed) and perturbed (red) in previous work. {The asymmetry cutoff from \citet{bloom2016sami} is shown in blue, for reference.} Visually classified asymmetric and normal galaxies are seen to have distinct distributions of $\overline{v_{asym}}$.} \label{} \end{figure} \begin{figure} \centering \includegraphics[width=9cm]{inclination_k1_mass.pdf} \includegraphics[width=9cm]{inclination_k35_mass.pdf} \caption{Inclination correction ($\sin(i)$) versus $\overline{k1}$ and $\overline{k3+k5}$, with points coloured by $\log(M_*/M_{\odot})$. As expected, $\sin(i)$ is correlated with $\overline{k1}$ and $\overline{k3+k5}$. {Note that $\overline{k1}$ represents bulk rotation, whereas $\overline{k3+k5}$ carries kinematic asymmetry.}} \label{} \end{figure} \subsubsection{Kinemetry} \label{} Kinemetry is an extension of photometry to the higher order moments of the line of sight velocity distribution (LOSVD)\footnote{The kinemetry code is written in IDL, and can be found at \url{} \citep{krajnovic2006kinemetry}.}. It was developed as a means to quantify asymmetries in stellar velocity (and velocity dispersion) maps. Such anomalies may be caused by internal disturbances or by external factors, namely interactions \citep{krajnovic2006kinemetry}. In \citet{bloom2016sami}, we defined a quantitative asymmetry measure, $\overline{v_{asym}}$, derived from kinemetry \citep{krajnovic2006kinemetry} and justified by a thorough by-eye classification based on SDSS imaging. Fig.~\ref{} shows $\overline{v_{asym}}$ for the full sample in this work (grey), galaxies visually classified as normal (black, dashed) and perturbed (red) in \citet{bloom2016sami}. Visually asymmetric and normal galaxies are seen to have distinct distributions of $\overline{v_{asym}}$. The kinemetry algorithm fits a series of regularly spaced, concentric ellipses, from which radial moment values are taken. The ellipses are fit with parameters defined by the galaxy centre, kinematic position angle (PA) and photometric ellipticity. We note a change from \citet{bloom2016sami}, in which we used the photometric PA. The ellipticity is taken from MGE fitting, as described in Section~\ref{}. The kinematic PA is calculated by the FIT\_KINEMATIC\_PA routine from \citet{krajnovic2006kinemetry}. {We use the photometric centre to define the centre of the fitted ellipses. We fit a 2-dimensional Gaussian to the SAMI Galaxy Survey $r$-band continuum flux maps and took the centroid of the location of the 25 brightest spaxels in a 6x6 pixel area around the centre of the fitted Gaussian. We did not use the the H$\alpha$ emission maps because they contain clumps of star formation and other features, which make determining the centre from these maps potentially unreliable.} The combination of the fitted ellipses is used to construct individual moment maps. As defined in \citet{bloom2016sami}, following \citet{shapiro2008kinemetry}, the quantitative measure of kinematic asymmetry is found as the median asymmetry over the radii of the fitted ellipses from kinemetry: \begin{equation} \overline{v_{asym}}=\overline{\left({\frac{k3+k5}{2k1}}\right)} \end{equation} where $k1, k3, k5$ are the coefficients of the Fourier decomposition of the first, third and fifth order moments of the line of sight velocity distribution, respectively. Regular rotation is carried in the $k1$ term, and asymmetry in the higher order terms. {That is, a perfect disk with entirely regular rotation would have no power in the higher order terms, but would be completely represented by $k1$. We do not include $k0$, $k2$ or $k4$ as these terms are used for studying the velocity dispersion map. For a full explanation of our use of kinemetry, see \citet{bloom2016sami}.} Fig.~\ref{} shows inclination correction against $\overline{k{1}}$ and $\overline{k{3}+k{5}}$ with points coloured by stellar mass. As expected, $\overline{k1}$ and inclination correction are correlated, with a Spearman rank correlation test giving $\rho=0.2$ and a probability {of no correlation} $p=9.2\times10^{-8}$. Similarly, $\overline{k3+k5}$ and inclination correction have correlation coefficients $\rho=0.2, p=3.4\times10^{-8}$. If perturbations are randomly oriented with respect to the disk, for example due to turbulence, we would not expect a correlation with inclination. The observed correlation implies that perturbations tend to be in the plane of the disk. Given that both $\overline{k1}$ and $\overline{k3+k5}$ are correlated with $\sin(i)$, we do not apply an inclination correction to $\overline{v_{asym}}$. Due to the re-sampling process used to make the datacubes, the noise in neighbouring spaxels is correlated. This effect is negligible for spaxels spaced more than $\sim$2.5 spaxels apart. A full explanation of how covariance is handled in SAMI Galaxy Survey data can be found in \citet{sharp2015sami}. In \citet{bloom2016sami} we forced the separation between kinemetry ellipses to be $>2.5$ spaxels. This avoided error underestimation, but reduced the number of ellipses such that it was not possible to extract rotation curves with sufficient detail for our purposes here. Accordingly, we allow the spacing between ellipses to be as low as 1 spaxel and account for covariance by bootstrapping errors. In most cases, perturbation in the velocity field, rather than error on $k1$ terms, dominates the deviation of the fits from perfect, regular rotation. \subsubsection{$V_{rot}$} \label{} The $V_{rot}$ measurements are obtained from the radial $k1$ values fit by the kinemetry package \citep{krajnovic2006kinemetry}, with the $k1$ term carrying the bulk rotation of the disk. Using $k1$ also has the advantage of removing most disturbance, as it traces only the first order moment of the velocity map. We applied the Bayesian model inference routine outlined in \citet{Callingham2015} to fit the rotation curves of the various galaxies in this study as arctan curves, using the model from \citet{courteau1997optical}: \begin{equation} V=V_{{circ}}\times\bigg(\frac{2}{\pi}\bigg)\times \mathrm{arctan}\bigg(\frac{x}{r_t}\bigg) + c \label{} \end{equation} where $V_{circ}$ is the asymptotic velocity, $x$ the distance from centre along the major axis, $r_t$ the transition radius between the rising and flat part of the rotation curve and $c$ is an offset term occurring when the fits do not pass through $(0,0)$. Note that $c<5\mathrm{km/s}$ in all fits, and is typically $\sim1\mathrm{km/s}$. We take the rotation velocity, $V_{rot}$ as $V$ at $2.2r_{e}$, where $r_e$ is the effective radius from MGE fitting, because this is the location of peak rotational velocity of a pure exponential disk \citep{courteau1997optical}. {Choosing $2.2r_{e}$ as our cutoff value also ensured that we did not trace the fitted curve as far beyond the spatial coverage of the data as would be the case using the asymptotic velocity. In some cases, we do not reach out to $2.2r_{e}$, but we have found that these galaxies do not have a different distribution of $V_{rot}$ values from the rest of the sample and their $k1$ rotation curves are still well fit by Equation~\ref{}.} The posterior probability density functions of the model parameters of Equation~\ref{} were sampled using the Markov chain Monte Carlo (MCMC) algorithm $\tt{emcee}$ \citep{ForemanMackey2013}. As an affine-invariant ensemble sampler, $\tt{emcee}$ is suited to fitting the rotation curves because it is not significantly impacted by covariance between the parameters \citep{Goodman2010}. A Gaussian likelihood function was maximised using $\tt{emcee}$, with physically motivated uniform priors (i.e. no negative distances, $r_t>0$), by applying the simplex algorithm to direct the walkers \citep{Nelder1965}. The data were assumed to be Gaussian and independent in this process. The uncertainties derived from the method represent the difference between the mean of the posterior distribution and the 16$^{\mathrm{th}}$ and 84$^{\mathrm{th}}$ percentiles. \begin{figure*} \centering \includegraphics[width=15cm]{tfr_vc_re_asyms_2d_mge_3_log.pdf} \caption{The TFR, showing log($V_{rot}/km s^{-1}$) at 2.2$r_e$ against log($M_{*}/M_{\odot}$), with points coloured by $\log(\overline{v_{asym}})$. $V_{rot}$ is calculated from the radial $k1$ moment values produced by kinemetry. The TFR extends across the stellar mass range of the sample, although scatter increases at low stellar mass. Scatter is also correlated with asymmetry. The fit to low-asymmetry galaxies in Equation~\ref{} is shown in black (dashed), with 1$\sigma$ errors on the fit indicated by the orange region.} \label{} \end{figure*} \section{The SAMI Galaxy Survey TFR} \label{} \subsection{2D Spatially resolved TFR} Using the rotation curves generated by kinemetry (i.e. the $k1$ radial values), we produce a `best case' TFR in Fig.~\ref{}, with stellar masses from the GAMA Survey StellarMasses catalogue \citep{taylor2011galaxy} and points coloured by kinematic asymmetry. Rotation velocities were fit as in Section~\ref{}, with 729 of 813 galaxies fit successfully and passing quality control. Galaxies were rejected by the fitting routine if there were fewer than 5 points in the $k1$ radial array with errors $<\mathrm{12km/s}$. We choose $\mathrm{12km/s}$ because beyond this value the scatter in the fitted velocities increases substantially. The median number of input points was 13. After error cuts, the median number of input $k1$ values was 12. Fits failed quality control if there was no turnover point in the data, {and the data did not extend out to $2.2r_e$} (leading to unphysical fitted asymptotic velocities $>1000\mathrm{km/s}$) or if $\chi^2>200$. We discuss quality control and fitting further in Section~\ref{}. Henceforth, galaxies with fits failing quality control or with too few included spaxels are considered `not fit'. The median $\overline{v_{asym}}$ of the not fit and the fit sample are not statistically offset at $0.050\pm0.01$ and $0.049\pm0.002$, respectively. The highest value in the not fit sample is $\overline{v_{asym}}=0.87$, whereas in the fit sample it is $\overline{v_{asym}}=0.78$. The other influence on whether galaxies could be fit for circular velocity is H$\alpha$ S/N. {However, the median H$\alpha$ S/N for those that could have their velocity fit, and those that could not be fit are within a standard error of each other: $26.0\pm1.2$ and $25.8\pm2.6$, respectively. Similarly, the median number of spaxels is also within uncertainties: $590.0\pm6.2$ and $617.2\pm12.9$.} Table~\ref{} details the fractions of galaxies successfully fit and gives the numbers of galaxies that were excluded from the sample for various reasons. Poor spatial coverage and low H$\alpha$ S/N are shown to cause galaxies to be rejected by our fitting routine for circular velocity. However, given that the number of rejected galaxies is $\sim10\%$, and that the median $\overline{v_{asym}}$, number of spaxels and median S/N remain unchanged, we conclude that the influence of these factors is small and does not cause significant bias in our sample, despite a weak tendency to reject the most asymmetric galaxies. The TFR in Fig.~\ref{} extends through the low stellar mass region of the sample. We stress that, unlike TFRs produced as distance indicators, we include all galaxies that satisfy our fitting and quality control procedures. To define a fiducial TFR , we exclude asymmetric galaxies (being conservative, we set the cutoff to $\overline{v_{asym}}<0.04$) from the sample used to fit the TFR line because, in general, their gas velocities are not expected to follow the same relation as normal galaxies and they are seen to be outliers. We note that, within the range up to $\overline{v_{asym}}<0.06$, there is no significant change in slope of or scatter around the TFR. Using only this low-asymmetry sample of galaxies, we fit the TFR for our sample: \begin{equation} \mathrm{log}(V_{rot}/kms^{-1})=0.31 \pm {0.0092} \times \mathrm{log}(M_{*}/M_{\odot})- 0.93 \pm 0.10 \label{} \end{equation} with an RMS of 0.15 dex in $\log(V_{rot}/kms^{-1})$ for the whole sample. {The RMS for galaxies with $\overline{v_{asym}}<0.04$ is 0.091dex}. Due to the covariance between SAMI Galaxy Survey spaxels, we anticipate that the errors on individual $V_{rot}$ values will be underestimated by the fitting code. Accordingly, in order to obtain errors on fitted TFR parameters, we bootstrap the calculation. The bootstrapping involved 1000 iterations of adding Gaussian noise, defined by scatter from the best fit, to the sample and recalculating the linear fit parameters, with the final error taken from the scatter in the fitted parameters over all the iterations. Henceforth we define scatter as the difference between the expected velocity from the power law fit and the observed velocity, for a given stellar mass. As has been seen in previous work [e.g. \citet{barton2001tully,kannappan2002physical}], scatter below the TFR increases at low stellar masses ($\log(M_*/M_{\odot})<9.5$). We find, using a Spearman rank correlation test of scatter to stellar mass, $\rho=-0.21, p=2\times10^{-9}$. \begin{figure} \centering \includegraphics[width=9cm]{tfr_scatter_log.pdf} \caption{The TFR as in Fig.~\ref{}, with median rotation velocity in stellar mass bins (red line) and RMS in stellar mass bins for all galaxies (dashed, black). The TFR line from Equation~\ref{} is shown in dotted black.} \label{} \end{figure} Scatter below the TFR is correlated with kinematic asymmetry, with $\rho=0.44, p=4\times10^{-38}$, as can be seen in Fig.~\ref{} (and explicitly in Fig.~\ref{}). Fig.~\ref{} more explicitly shows the scale of the scatter in our TFR. Of the 153 galaxies scattered below the root mean square (RMS) line, calculated for the whole asymmetry range, in Fig.~\ref{} (lower dashed line), $49\%$ have $\overline{v_{asym}}>0.065$. This is a much higher proportion of asymmetric galaxies than is found for galaxies above the RMS line ($11\%$). The median $\log(V_{rot}/kms^{-1})$ in bins of stellar mass is also shown in red in Fig.~\ref{}. It is consistent with the fitted line over the full stellar mass range of the sample. There is more negative scatter in the low than higher mass cases: $25\pm3\%$ of galaxies with $\log(M_*/M_{\odot})<9.5$ lie below the RMS line, compared to $10\%\pm1$ of galaxies with $\log(M_*/M_{\odot})>9.5$. This result is in agreement with \citet{krajnovic2011atlas3d}, \citet{oh2016sami} and others. \begin{figure*} \includegraphics[height=9cm,width=\textwidth]{padiff_asym_residuals.pdf} \caption{From left: $\overline{v_{asym}}$ against PA offset, $\overline{v_{asym}}$ against scatter off the 2D spatially resolved TFR, and PA offset against scatter off the spatially resolved TFR. The correlation between PA offset and $\overline{v_{asym}}$, and the inverse relationships between $\overline{v_{asym}}$ and scatter off the TFR and PA offset and scatter off the TFR can be seen. However, the strongest trend is clearly the inverse correlation between scatter off the TFR and $\overline{v_{asym}}$. Red dashed lines are included at the median of $\overline{v_{asym}}$ in the leftmost plot and TFR residual in the rightmost plot as visual aids to show the correlations.} \label{} \end{figure*} An offset between the photometric and kinematic PA is a known indicator of interaction. As in \citet{fogarty2015sami}, this offset is calculated as: \begin{equation} \sin\Psi=|\sin(\rm{PA_{phot}-PA_{kin}})| \end{equation} where $\Psi$ is the offset, and $\rm{PA_{phot}}$ and $\rm{PA_{kin}}$ are the photometric and kinematic PAs, respectively. The offset lies between $0^{\circ}-90^{\circ}$, and is insensitive to $180^\circ$ differences in the PAs. The kinematic PA found as in Section~\ref{} and photometric PA from single MGE fits to the SDSS $r$-band images in the GAMA Survey DR2 catalogue SersicCat \citep{kelvin2012galaxy}. All galaxies scattered below the RMS line have an offset between the kinematic and photometric PA $>15^{\circ}$. Fig.~\ref{} shows the TFR as in Fig.~\ref{}, but with points coloured by inclination correction, ellipticity and offset between the kinematic and photometric PAs. There is no relationship between inclination correction and scatter off the TFR either upwards or downwards ($\rho=-0.022, p=0.56$), or between ellipticity and scatter off the TFR ($\rho=-0.021, p=0.55$). There is, however, a correlation between PA offset and scatter, with $\rho=0.18, p=6.3\times10^{-7}$. We discuss this relationship further in Section~\ref{}. {We note that the population of galaxies with ellipticity$<0.1$ have slightly higher dispersion around the TFR line than the rest of the sample, with an RMS of 0.21 in $\log(V_{rot}/kms^{-1})$, compared to 0.15 for the whole sample. However, these galaxies do not exclusively lie off the TFR, and represent only $4\%$ of the sample, so do not contribute significantly to statistical scatter. Given that we aim to be as complete as possible in our analysis, we do not perform an ellipticity cut.} \begin{figure*} \centering \includegraphics[width=18cm]{tfr_vc_re_asyms_biases.pdf} \caption{The TFR as in Fig.~\ref{}, with points coloured (left to right) by ellipticity, inclination correction and offset between the kinematic and photometric PA. There is no trend in the first two cases, but a there is one in the third. This indicates that scatter off the TFR is not associated with geometric correction, but it is linked to a physical process: the offset between PAs. } \label{} \end{figure*} Fig.~\ref{} shows examples of fits to low mass galaxies, to demonstrate that galaxies are fit well even at the low mass end of the TFR. {In some cases, there is patchy data with S/N$>6$ around the edges of the velocity field. This does not affects the fits, as kinemetric ellipses are constrained by PA and ellipticity, and are only included if $<75\%$ of points in the ellipse contain valid data. There may, however, be some uncertainty introduced by unreliable $r_e$ measurements, due to the difficulties inherent in performing photometry on small, low surface brightness galaxies. We discuss this more fully in Section~\ref{}. } \begin{figure*} \centering \includegraphics[width=7cm]{tfr_totmass_lowm_examples.pdf} \includegraphics[width=7cm]{24433_mosaic_nosigma.pdf} \includegraphics[width=7cm]{321022_mosaic_nosigma.pdf} \includegraphics[width=7cm]{105573_mosaic_nosigma.pdf} \includegraphics[width=7cm]{71099_mosaic_nosigma.pdf} \includegraphics[width=7cm]{15220_mosaic_nosigma.pdf} \includegraphics[width=7cm]{54382_mosaic_nosigma.pdf} \includegraphics[width=7cm]{273951_mosaic_nosigma.pdf} \caption{The TFR as in Fig.~\ref{}, with examples of low mass galaxies to illustrate that we are able to fit low stellar mass galaxies with a variety of velocities. The top left panel shows the TFR with the examples in colours. For each galaxy, the SDSS image is shown with the 15 arcsec SAMI field of view as a white circle (left), and the $H\alpha$ velocity field with S/N$>6$ (right). The $k1$ values from kinemetry are plotted, as well as the fitted rotation curve. The red horizontal line gives $V_{rot}$ at 2.2$r_e$, uncorrected for inclination. Corrected rotation velocities and inclination corrections are given within the figures. } \label{} \end{figure*} Fig~\ref{} shows our TFR in comparison with those from \citet{green2013dynamo} and \citet{simons2015transition} and our linear fit from Fig.~\ref{} in comparison to that from \citet{reyes2011calibrated}. Our slope value is in agreement with the results from \citet{reyes2011calibrated}, who take a sample of 189 galaxies from the SDSS sample with $z<0.1$ over mass range $9.0<\log(M_*/M_{\odot})<11.0$ to find a slope of $0.278\pm0.13$ (see Fig.~\ref{}). At high stellar mass, our TFR is in agreement with the results of \citet{green2013dynamo} who, similarly to this work, use 2D spatially resolved H$\alpha$ kinematics to fit rotation velocities at 2.2$r_e$ for low redshift galaxies. {We also show the TFR line from \citet{bekeraite2016space}, who use IFS to calculate rotation velocity from stellar kinematic maps. Comparison to this work is directly relevant, due to the closeness of their redshift ($0.005 < z < 0.03$) and stellar mass ($9.0<\log(M_*/M_\odot)<12.0$) ranges to the SAMI Galaxy Survey. Their work shows SDSS Petrosian absolute magnitude ($M_r$) against circular velocity. In order to directly compare our measurements, we derive a conversion factor by taking the mean ratio of $M_r$ to stellar mass for our galaxies with $\overline{v_{asym}}<0.04$ and $\log(M_*/M_\odot)<9.0$. We use these cuts to exclude high asymmetry galaxies (as for producing Equation~\ref{}) and so as not to extend below the CALIFA mass range. The resulting converted CALIFA stellar mass TFR is given by: \begin{equation} \mathrm{log}(V_{rot}/kms^{-1})=0.26 \pm 0.017 \times \mathrm{log}(M_{*}/M_{\odot})- 0.50 \pm 0.13 \end{equation} This line is shown in Fig.~\ref{}. Despite the differences in method, the \citet{bekeraite2016space} fit agrees with the fit to our work within the errors. {We note that that our conversion factor was derived by taking a simple mean, which ignores differences in mass/light ratio for galaxies of different ages, which, if not ignored, may introduce a change in the derived slope.} {When comparing between different TFR studies, it is also important to note that many studies use different definitions of velocity. For example, \citet{bekeraite2016space} define rotation velocity at $r_{opt}$, the radius containing $83\%$ of the light. Different definitions of rotation velocity can lead to different TFR slopes if measured at a radius where the rotation curve has not flattened \citep{bradford2016slippery}. We note that at $r_{opt}$, most galaxies' rotation curves will have flattened, and depending on structural morphology it is a similar radius to $2.2r_e$, which was selected for the same reason. } \citet{simons2015transition} found a transition mass in the TFR at $\log(M_{*}/M_{\odot})\sim9.5$, below which galaxies were almost all scattered lower than the relation, in contrast to our results. Fig.~\ref{} shows histograms of residuals around the fitted TFRs for each data set, in bins of stellar mass. In the $\mathrm{log}(M_{*}/M_{\odot})<9.0$ and $9.0<\mathrm{log}(M_{*}/M_{\odot})<10.0$ cases, there is an excess of negative residuals in the \citet{simons2015transition} distribution, in comparison to the distributions from the SAMI Galaxy Survey data, with two-sample Kolmogorov-Smirnov tests giving $p$-values of 0.018 and 0.023, respectively. In both data sets, there is a progressive increase in negative scatter as stellar mass decreases. The discrepancy may have been caused by the difficulty in accurate slit placement for low mass, asymmetric galaxies. \citet{simons2015transition} themselves discuss the difficulty of accurate slit placement, and given the higher proportion of asymmetric galaxies at low stellar mass seen in Fig.~\ref{}, a number of the galaxies in \citet{simons2015transition} may have inaccurate values for $V_{rot}$. Finally, the \citet{simons2015transition} sample is at significantly higher redshift than ours ($\overline{z}=0.26$, rather than $\overline{z}=0.03$), leading to lower spatial resolution. {There may be redshift evolution between the \citet{simons2015transition} sample and our sample. However, the TFR slope at the high mass end is similar for both samples and therefore does not show the evolution suggested in \citet{cresci2009sins} and \citet{ziegler2001evolution}. Although change in TFR slope has been found in several studies of TFR evolution, there has been no work to date investigating scatter at the lower mass end as an effect of evolution of galaxy kinematics. The differences between slit and IFU data discussed above may mask any signatures of evolution of the TFR between \citet{simons2015transition} and our sample at low stellar mass. Comparable data types are needed to conduct a valid study of evolution, such as between the KROSS and SAMI samples in Tiley et al. (in prep.).} In order to compare the low mass end of our TFR with other samples, we convert our stellar masses to baryonic masses. We approximate the total baryonic mass by adding the fractional gas mass to the stellar mass, but as we do not have direct gas masses for our sample, gas fractions are estimated from equation 5 in \citet{cortese2011effect}: \begin{equation} \log\bigg(\frac{M(HI)}{M_{*}/M_{\odot}}\bigg)=-0.33(NUV-r)-0.40log(\mu_{*})+3.37 \end{equation} where $NUV-r$ is the {colour in magnitudes} and $\mu_{*}$ is the stellar mass surface density {in $\mathrm{M_*/pc^2}$}. The relationships between stellar mass, scatter and asymmetry observed in Fig.~\ref{} remain in the baryonic mass TFR. Fig.~\ref{} shows our TFR with baryonic mass, instead of stellar mass. We compare with the HI rotation velocity and baryonic mass TFRs from the low-redshift, gas dominated ($M_{gas}>M_{star}$) sample in \citet{stark2009first}, {and the low mass compilation in \citet{papastergis2015big}, as discussed in \citet{sales2017low}}. These samples have selected against disturbed galaxies, so it is not surprising that our TFR shows more downwards scatter than theirs. However, it is notable that the main distributions are within a similar $V_{rot}$ range. \begin{figure} \centering \includegraphics[width=9cm]{comparison_tfr_sm.pdf} \caption{The 2D TFR from this work shown with the stellar mass/H$\alpha$ TFR from \citet{green2013dynamo} and \citet{simons2015transition} and the fitted TFR lines from \citet{reyes2011calibrated} and \citet{bekeraite2016space}. The linear fit to our data from Fig.~\ref{} is also shown. Our points and fit are consistent with the comparison samples at high mass. Below $\log(M_{*}/M_{\odot})\sim9.5$, however, there is disagreement between our results and those in \citet{simons2015transition}. Unlike \citet{simons2015transition}, we do not see a decay of the TFR below this cutoff, although there is an increase in negative scatter.} \label{} \end{figure} \begin{figure} \centering \includegraphics[width=9cm]{comparison_tfr_bar_2.pdf} \caption{ The 2D TFR with baryonic masses, shown with comparison samples from \citet{stark2009first} and the compilation in \citet{papastergis2015big}. These comparison samples follow a similar distribution to our work, taking into account that they selected against disturbed galaxies, and so show less scatter than our complete sample.} \label{} \end{figure} \begin{figure*} \centering \includegraphics[width=\textwidth]{simons_sami_allhists.pdf} \caption{Histograms of residuals around the TFR lines (shown in Fig.~\ref{}) for the \citet{simons2015transition} data (red, dashed) and that in this work (black, solid), with the zero line in each case shown in dashed blue. In all cases, there is an excess of negative residuals in the \citet{simons2015transition} distribution, in comparison to this work. } \label{} \end{figure*} \subsection{`Simulated slit' TFRs} \label{} In order to examine the effects of slit measurement compared with 2D kinematics, we placed a `simulated slit' on the observed SAMI Galaxy Survey H$\alpha$ velocity fields. Using either photometric or kinematic position angles we selected spaxels from SAMI Galaxy Survey H$\alpha$ velocity maps along a `slit' of width 1 arcsec and the given PA. This slit width allows our results to be compared to other low-redshift TFR studies, including \citet{raychaudhury1997tests}, \citet{kannappan2002physical}, \citet{reyes2011calibrated} and others, that use slit widths 2 arcsec and smaller, and are comparable to our work on a physical scale. We note that slit widths of 2 arcsec or 3 arcsec do not alter the global scatter of the resulting TFR. We also select a narrow slit width in order to isolate the effects of changing the PA, and minimise spaxels selected in common when the PAs are offset. Velocity values for spaxels falling within the slit were plotted against radial distance from the fitted central spaxel, and arctan rotation curves were fit to the slit data, as above. We note that slit placement for our sample is not vulnerable to placement error, as is discussed in \citet{simons2015transition}, who consider acceptable slit placement to be within 40$^\circ$ of the photometric PA. Of course, some galaxies may have poorly defined PAs, such as in the case of strong bars (in the photometric case) or asymmetric galaxies (in the photometric and kinematic cases). However, once the PA is calculated, there is no error in placement. Of the total parent sample of 813 galaxies, the highest fraction were fit when the input velocities were from 2D spatially resolved kinematics. Table~\ref{} gives fraction of galaxies fit for each method, and the number of galaxies not fit because they had $<5$ points with errors $<12$km/s, no turnover point or $\chi^2>200$, respectively. We applied the same cuts to the 2D and simulated slit samples. We choose a limit of $\chi^2=200$ because the distribution of $\chi^2$ {was strongly peaked at low values, and} for all three samples has a significant break between $200$ and $550$, making this a natural cutoff. {The reduced $\chi^2$ distribution also has a clear break, and separates out the same galaxies.} We note that no galaxies in the 2D spatially resolved sample were cut because they had $\chi^2>200$. In fact, 4 galaxies had $\chi^2>200$, but also had no turnover, which was the cut performed first. $82\%$ of galaxies in the 2D spatially resolved sample had $\chi^2<14$, and $90\%$ had $\chi^2<50$. Galaxies with higher $\chi^2$ were asymmetric galaxies with rotation curves deviating from the arctan model. However, the generally low $\chi^2$ values for fits to kinematically normal galaxies provide justification for the use of the arctan curve from \citet{courteau1997optical}, as it is a good match for the rotation curves derived from the whole kinematic map. Typical degrees of freedom for each rotation curve are $\sim10$ for rotation curves from kinemetry and $\sim20$ for those from the simulated slit. \begin{center} \begin{table*} \begin{tabular}{ b{2.5cm} l b{2cm} l b{2cm} l b{2.5cm} l b{1.5cm} l b{2.5cm} l b{2.5cm} } \hline & Fraction fit & Too few points & No turnover & $\chi^2>200$ & \vline ~RMS (dex) & Median abs. residual from TFR (dex) \\ \hline 2D Spat. resolved & $90\%$ & 59 & 21 & 0 &\vline ~0.15 & 0.058 \\% $\pm5\times10^{-4}$\\% & 31$\%$ \\ Kin. PA slit & $74\%$ & 96 & 66 & 51 &\vline~ 0.17 & 0.063 \\%$\pm6\times10^{-4}$\\ Phot. PA slit & $68\%$ & 102 & 42 & 101&\vline~ 0.25 & 0.071 \\%$\pm7\times10^{-4}$ \\% & 42$\%$ \\ \hline \end{tabular} \caption[Table caption text]{This table gives the fraction of galaxies fit for each method, and the number of galaxies not fit because they had $<5$ points with errors $<12$km/s, no turnover point or $\chi^2>200$, respectively. We also show global scatter measurements for the TFRs produced using our three methods (see Fig.~\ref{}). Scatter (both RMS and median absolute residual from TFR as in Equation~\ref{}, in log space) is increased by using a slit, instead of the 2D spatially resolved map. Scatter is further increased when the slit is positioned at the photometric, rather than the kinematic PA.} \label{} \end{table*} \end{center} \begin{center} \begin{table*} \begin{tabular}{ b{2.5cm} l b{2.5cm} l b{2.5cm} l b{2cm}} & \multicolumn{2}{c}{$\Delta V_{rot}$/Stellar Mass correlation} & \multicolumn{2}{c}{$\Delta V_{rot}$/$\overline{v_{asym}}$ correlation}& \multicolumn{2}{c} {TFR fit parameters} \\ \hline & $\rho_{SM}$ & p-value$_{SM}$ &\vline ~ $\rho_{asym}$& p-value$_{asym}$&\vline ~ Slope & $y$-int. \\ \hline 2D Spat. resolved & $ -0.28$ & $2\times10^{-6}$&\vline ~ $ -0.30$ & $1.4\times10^{-8}$& \vline ~0.31 & -0.93\\ Kin. PA slit & $-0.25$ & $4\times10^{-5}$ &\vline ~ $-0.32$ & $1.2\times10^{-9}$ &\vline ~0.30 & -0.84 \\ Phot. PA slit & $-0.36$ & $2\times10^{-3}$ &\vline ~ $-0.25$ & $2.5\times10^{-6}$ & \vline ~0.29 & -0.83 \\ \hline \end{tabular} \caption[Table caption text]{This table gives Spearman rank correlation parameters (subscript $SM$) between scatter off the TFR and stellar mass for the three TFRs in Fig.~\ref{}. In all three cases, scatter below the TFR is inversely correlated with stellar mass. We also give correlation parameters (subscript $asym$) for $\overline{v_{asym}}$ and scatter off the TFR for the three samples in Fig.~\ref{}. The relationship holds in all three cases. Note that the correlations are negative because the main scatter is below the TFR. {Finally, we give TFR fit parameters for all three samples, as shown in Fig.~\ref{}.}} \label{} \end{table*} \end{center} The reason for the larger number of fitted galaxies from 2D kinematics is that the $k1$ values used in the 2D spatially resolved case represent bulk rotation, without asymmetry. Further, using the 2D spatially resolved data allows for more spaxels to be included, reducing the proportional influence of individual, noisy data points or fluctuation in the velocity field. We note that there is no systematic difference in the kinematic asymmetry of the three samples, with the median asymmetries for the 2D spatially resolved maps, kinematic and photometric slits being $0.040\pm0.002$ $0.038\pm0.002$ and $0.039\pm0.002$, respectively. Typical uncertainties are also equivalent. Fig.~\ref{} shows example plots for a normal galaxy, with low kinematic asymmetry and aligned photometric and kinematic PAs. Fig.~\ref{fig:pa}, by contrast, shows a kinematically normal galaxy for which there is significant misalignment between the photometric and kinematic PA. This misalignment causes a $\sim40$km/s difference in calculated rotation velocity when using slits placed at the different PAs and illustrates how using the photometric PA to measure rotation velocity can increase scatter off the TFR. Fig.~\ref{} shows an asymmetric, low mass galaxy, for which the PAs are also misaligned. In this case, asymmetry makes determination of an accurate PA in either case difficult. However, the comparative smoothness of the rotation curve generated from the 2D spatially resolved map, compared to that from the slits, demonstrates the value of the map method in producing a better rotation curve fit. We note that the slit rotation curves are not always centred on 0. This is because the slit is centred on the photometric centre of the galaxy, which can be offset from the kinematic centre of the map. {Because the ellipses used to generate the rotation curves from kinemetry use the photometric, rather than the kinematic, centre, the curves do not always go through $(0,0)$, as there are often slight offsets between the centres of the flux and H$\alpha$ velocity maps. Given that the offset is, in all cases, less than the error on the fitted velocity, we do not consider this a significant cause of scatter. {We also note that the offsets are not proportional to $\overline{v_{asym}}$, making it unlikely that small differences between the photometric and kinematic centres contribute significantly to asymmetry.}} \begin{figure*} \centering \includegraphics{301346_tfr_mosaic_paper5.pdf} \caption{We show fit results for GAMA310346, as an example of a kinematically normal galaxy, in which the photometric and kinematic PAs are aligned. The top row shows: the H$\alpha$ velocity map from {\small LZIFU}{\sc} (with the kinematic and photometric PAs overplotted) cut by S/N, the simulated slit spaxels at the kinematic and photometric PAs and the SDSS image. The SAMI instrument field of view is shown as a white circle. We then show the fitted rotation curve (green) and position on the TFR (red) for the simulated slits and 2D spatially resolved maps, respectively. On the plots of the rotation curves, the red dashed line shows the velocity (without inclination correction) at 2.2$r_e$. {The 2D rotation curve is a symmetric one-sided rotation curve, which has been plotted on the same axis scale as the others for ease of comparison.}} \label{} \end{figure*} \begin{figure*} \centering \includegraphics{623144_tfr_mosaic_paper5.pdf} \caption{A galaxy with offset photometric and kinematic PAs. The circular velocity using the 2D spatially resolved map is 113.0 km/s, similar to the value calculated when using the slit at the kinematic PA (115.2km/s). However, when the slit is placed at the photometric PA, the circular velocity is found to be 77.0 km/s. Note that this galaxy is not centred in the SAMI instrument field of view.} \label{fig:pa} \end{figure*} \begin{figure*} \centering \includegraphics{290396_tfr_mosaic_paper5.pdf} \caption{A highly asymmetric ($\overline{v_{asym}}=0.17$) galaxy at low stellar mass. The photometric and kinematic PAs are significantly offset (although, given the level of disturbance, calculating an accurate PA is difficult). The rotation curve from the 2D spatially resolved map is much smoother than those from the simulated slits. This galaxy is also not centred in the SAMI instrument field of view.} \label{} \end{figure*} Fig.~\ref{} and Fig.~\ref{} show the TFR produced using a slit placed along the photometric and kinematic PA, respectively. \begin{figure} \centering \includegraphics[width=9cm]{tfr_vc_re_asyms_kinpa_2_log.png} \caption{The TFR with log($M_{*}/M_{\odot}$) against log($V_{rot}/km s^{-1}$) at 2.2$r_e$, calculated using rotation values from a slit placed along the kinematic PA. Points are coloured by $\log(\overline{v_{asym}})$. The only change between this plot and Fig.~\ref{} is the angle at which the slit is placed. \label{} \end{figure} \begin{figure} \centering \includegraphics[width=9cm]{tfr_vc_re_asyms_photpa_2_log.png} \caption{The TFR with log($M_{*}/M_{\odot}$) against log($V_{rot}/km s^{-1}$) at 2.2$r_e$, calculated using rotation values from a slit placed along the photometric PA. Points are coloured by $\log(\overline{v_{asym}})$. There is more scatter than in Fig.~\ref{}, particularly at low stellar mass. The main trends of Fig.~\ref{}, however, are still visible: the TFR extending to low stellar mass and a correlation between scatter and $\overline{v_{asym}}$.} \label{} \end{figure} \section{Causes of scatter} \label{} Scatter off the TFR can be caused either by observational effects or by physical properties of galaxies. Understanding the scatter is thus crucial, both to more accurately plot the TFR and to understand why galaxies scatter. Processes known to correlate with scatter, such as asymmetry \citep{kannappan2002physical}, are also important for understanding galaxy evolution. Fig.~\ref{} shows the TFRs produced by the three methods (2D kinematics, a slit at the kinematic PA and one at the photometric PA), including only the 456 galaxies common to all three. Table~\ref{} gives the RMS and median absolute residuals off the TFR line for the three TFRs in Fig.~\ref{}. There are a variety of causes of scatter, which affect the TFRs in different ways: \begin{itemize} \item{Photometric uncertainties} \item{Kinematic asymmetry} \item{Scatter caused by using a slit} \item{Underestimation of $V_{rot}$ caused by not measuring velocity along the kinematic PA} \end{itemize} We discuss these effects in turn in the following sections. \begin{figure*} \centering \includegraphics[width=\textwidth]{tfr_common_all_log.pdf} \caption{TFRs produced using three methods, showing only galaxies common to all three. Points are coloured by kinematic asymmetry, and the black line shown gives the fitted TFR from Fig.~\ref{}. The decrease in scatter from left to right is demonstrative of the influence of PA and the kinemetry method of moment extraction. {The red dashed lines show the TFR fit line for the galaxies common to all three using the kinematic and photometric PA, respectively. Slopes and intercepts are given in Table~\ref{}.}}. \label{} \end{figure*} \subsection{Photometric uncertainties} \label{} Small, low surface brightness galaxies may have less reliable photometry than large, bright galaxies, as has been noticed in \citet{hill2011galaxy} and \citet{simons2015transition}. The typical error value for MGE fits is 0.05 on $\log(r_e)$, but it is anticipated that there may be small increases in error on PA and ellipticity at low stellar masses. In order to quantify the photometric uncertainty, we calculated the difference in PA, $r_e$ and ellipticity from MGE photometry and the GAMA Sersic catalogue \citep{kelvin2012galaxy}. The RMS of the differences in PA, $r_e$ {(normalised to the MGE $r_e$)} and ellipticity are $14.6^{\circ}$, {0.24} and 0.12, respectively. The fractions of galaxies with scatter $>3\sigma$ are $3\%, 2\%$ and $1\%$, respectively. {We note that we expect the slightly larger scatter and RMS values for difference in the PA because of the large uncertainties associated with calculating a PA for very circular galaxies. Nevertheless, {the fact that} the scatter and RMS remain small in all cases indicates that the Sersic and MGE values are largely equivalent. } It is possible that a small amount of scatter at low stellar mass is enhanced by errors in photometry. {However, causes of scatter are discussed below and, in particular, errors in stellar mass are small even at low stellar mass, averaging $0.1$dex, and are thus insufficient to explain the large scatter that we observe.} \subsection{Kinematic asymmetry} In all cases in Fig.~\ref{}, there remains the correlation between kinematic asymmetry and scatter below the TFR, as in Fig.~\ref{} and Fig.~\ref{}. Table~\ref{} gives correlation parameters for this relationship in all three cases. There is also an inverse relationship between stellar mass and scatter off the TFR. Using the linear fit TFR from Fig.~\ref{}, we calculate the negative scatter correlation with stellar mass and present the results for the common galaxies in Table~\ref{}. Given the known inverse relationship between stellar mass and kinematic asymmetry \citep{van1998neutral,cannon2004complex,lelli2014dynamics,bloom2016sami} and the relationships shown here between scatter off the TFR and kinematic asymmetry, this is not surprising. However, that the degree to which scatter depends on stellar mass changes between the presented TFRs (see Table~\ref{}) is indicative that it is more than a purely physical relationship, and could be at least partly a function of measurement. \citet{oh2016sami} demonstrated an inverse relationship between asymmetry and rotation velocity from stellar kinematics. {The $S_{0.5}$ relation gives $\log(M_*/M_\odot)$ against $\sqrt{0.5V_{rot}^2+\sigma^2}$, where $\sigma$ is the mean velocity dispersion.} {It has been observed that asymmetric galaxies scatter less from the $S_{0.5}$ relation than they do from the TFR \citep{cortese2014sami,simons2015transition}. This indicates that asymmetry is related to the extent to which systems are dispersion, rather than rotation, supported. We shall investigate this further in future work.} \subsection{Scatter caused by using a slit} The slightly decreased scatter in the TFR from 2D kinematics compared with the kinematic slit TFR (see Table~\ref{}), is due to the derivation of the $V_{rot}$ values in the 2D kinematics TFR from the $k1$ moment map, that contains bulk rotation without asymmetry. $k1$ shows underlying disk rotation, calculated over all spaxels in the input velocity map, specifically excluding asymmetry and perturbation. By contrast, the velocity values used to produce the $V_{rot}$ values in the kinematic slit TFR are taken directly from the total velocity map, that includes asymmetries and fluctuations between spaxels. This can be seen in the difference between rotation curves from kinemetry and the kinematic PA slit in Figs.~\ref{fig:pa} and \ref{}, in which the rotation curve from the kinematic slit is significantly less smooth than that from 2D kinematics. Fig.~\ref{} shows $\log(V_{rot}/kms^{-1})$ from the kinematic PA slit against $\log(V_{rot}/kms^{-1})$ from the 2D kinematic maps, with points coloured by asymmetry. A Spearman rank correlation test of asymmetry and absolute difference between $\log(V_{rot}/kms^{-1})$ from the kinematic PA slit against $\log(V_{rot}/kms^{-1})$ from the full kinematic maps gives $\rho=0.2, p=3\times10^{-5}$. The relationship between the $V_{rot}$ values is tight at high velocity, but below $\log(V_{rot, 2D}/kms^{-1})\sim1.7$ the relation becomes much less tight. This indicates that the effects described above disproportionately affect asymmetric (and, by extension, low mass) galaxies. The inherent lack of smoothing entailed by using a slit also contributes some of the increased scatter in the photometric slit TFR (see Table~\ref{}), as the rotation curves are produced in the same way as the kinematic slit TFR. \begin{figure} \centering \includegraphics[width=9cm]{vsvsvs_2d_log.pdf} \caption{$\log(V_{rot}/kms^{-1})$ from the kinematic PA slit against $\log(V_{rot}/kms^{-1})$ from the full kinematic maps, with points coloured by asymmetry. Difference between the measured $V_{rot}$ values is correlated with asymmetry, with an otherwise tight relation showing increased scatter below $\log(V_{rot, 2D}/kms^{-1})\sim1.7$. } \label{} \end{figure} \subsection{Offset between photometric and kinematic PA} \label{} Other than input PA, there is no difference in velocity fitting or input data between the photometric and kinematic PA slit TFRs. From the increased scatter in the photometric slit, compared to the kinematic slit, TFR we conclude that scatter in the photometric slit TFR can be at least partially attributed to the fact that $23\%$ of galaxies common to all three TFR samples have misalignment $>30^{\circ}$. This misalignment leads to systematic underestimation of the rotation velocity when calculated along the photometric PA. In extreme cases, the photometric PA may lie along the zero line of the velocity map, entirely `missing' the galaxy's rotation (see Figs.~\ref{}, \ref{fig:pa}, \ref{}). Fig.~\ref{} shows the underestimation of $V_{rot}$ for some galaxies produced by using the photometric, rather than kinematic PA. There are multiple causes of kinematic and photometric misalignment. Table~\ref{} gives the results of a by-eye classification of misaligned galaxies, done by examining rotation maps and SDSS images of misaligned galaxies. Of the misaligned galaxies, those with bars and those which are face-on can be considered to be misaligned due to observational, rather than physical effects. That is, the bar leads to the measured photometry being significantly different from the kinematics. If they were excluded from the photometric PA TFR, scatter would decrease. \begin{figure} \centering \includegraphics[width=9cm]{vsvsvs_phot_log.png} \caption{$V_{rot}$ using the photometric against kinematic slit, with points coloured by asymmetry. Using the photometric slit produces a systematic underestimation of $V_{rot}$, as compared with the kinematic slit.} \label{} \end{figure} \begin{center} \begin{table} \begin{tabular}{ b{1.5cm} l b{1.5cm} l b{1cm l }} \hline Classification & Fraction of misaligned galaxies & Number \\ \hline Face-on & 9$\%$ & 10 \\ Barred & 29$\%$ & 30\\ Asymmetric & 24$\%$ & 25\\ Normal & 37$\%$ & 39\\ \hline \end{tabular} \caption[Table caption text]{This table gives results of a by-eye classification of misaligned galaxies in order to identify causes of misalignment. Face-on galaxies may have uncertain photometric PAs, whereas bars can lead to spurious photometry. Asymmetric galaxies may have variable or distorted PAs and kinematics. Lastly, `normal' galaxies are both photometrically and kinematically normal, but are nevertheless misaligned.} \label{} \end{table} \end{center} \begin{figure} \includegraphics[width=9cm]{vel_offset_scatter_3_log.png} \caption{Difference between $V_{rot, kin.PA}$ and $V_{rot, phot.PA}$, {normalised by $V_{rot, kin.PA}$}, against offset in kinematic and photometric PA, with points coloured by asymmetry. There is a correlation between difference in calculated $V_{rot}$ and PA offset, and between asymmetry and PA offset. } \label{} \end{figure} Fig.~\ref{} demonstrates this relationship (with a correlation $\rho=0.66, p=1\times10^{-15}$) between difference in $V_{rot, kin.PA}$ and $V_{rot, phot.PA}$ and PA offset (with $V_{rot, kin.PA}$ and $V_{rot, phot.PA}$ being the rotation velocities calculated using the kinematic and photometric PAs, respectively). We define velocity offset as: \begin{equation} \mathrm{V_{offset}}={\mathrm{V_{rot, kin.PA}}-\mathrm{V_{rot, phot.PA}}} \end{equation} There is a further correlation between PA offset and asymmetry, with $\rho=0.17, p=1\times10^{-4}$. As shown in Fig.~\ref{}, there is also a relationship in the 2D spatially resolved TFR between scatter off the TFR and PA offset. Fig.~\ref{} shows the three key relationships in this work: $\overline{v_{asym}}$ against PA offset, $\overline{v_{asym}}$ against residuals above and below the TFR, and PA offset against scatter off the 2D spatially resolved TFR for all galaxies in Fig.~\ref{}, with Table~\ref{} giving partial Spearman rank correlation test results. Note that galaxies scattered above the TFR with large PA offsets are asymmetric and barred galaxies. Although the inverse relationship between $\overline{v_{asym}}$ and scatter off the TFR is the strongest, the correlation between $\overline{v_{asym}}$ and PA offset and the inverse relationship between scatter off the TFR and PA offset are also evident. The relative strengths of these relationships point to kinematic asymmetry as the strongest driver of scatter off the TFR. However, the influence of PA choice on scatter is significant. Our photometric PA slit method is similar to placing a long slit at the photometric PA to calculate rotation velocity, as is seen in literature [e.g. \citet{neistein1999tully,pizagno2007tully,reyes2011calibrated,simons2015transition}]. Some scatter in the TFRs in these works can be attributed to their use of the photometric PA. Deeper photometry, or targeting more nearby galaxies, may be able to reduce scatter in these cases, as the derived PA would be less influenced by bars, and galaxies identified as belonging to some categories in Table~\ref{} could be excluded. \begin{center} \begin{table} \begin{tabular}{ b{4cm} l b{1.8cm} l b{1.8cm} l } \hline & $\rho$ & p-value \\ \hline $\overline{v_{asym}}$/ PA offset (! TFR scat.)& $ 0.19$ & $1.2\times10^{-7}$\\ $\overline{v_{asym}}$/ TFR scat. (! PA offset)& $-0.27$ & $1.9\times10^{-14}$\\ PA offset/TFR scat. (! $\overline{v_{asym}}$) & $-0.12$ & $2.1\times10^{-2}$ \\ \hline \end{tabular} \caption[Table caption text]{This table gives partial correlation parameters between $\overline{v_{asym}}$, PA offset and scatter off the TFR, with the variable held constant indicated by `!'. Tests were performed using all points in Fig.~\ref{}. Whilst the relationships persist in all three cases, the links between PA offset and $\overline{v_{asym}}$ and scatter off the TFR and $\overline{v_{asym}}$ are stronger than that between PA offset and scatter off the TFR. } \label{} \end{table} \end{center} \section{Conclusions} \label{} Without selecting based on asymmetry or Hubble type, we show the TFR for 729 kinematically and morphologically diverse galaxies in the SAMI Galaxy Survey sample and find: \begin{equation} \mathrm{log}(V_{rot}/kms^{-1})=0.31 \pm 0.092 \times \mathrm{log}(M_{*}/M_{\odot})- 0.93 \pm 0.10 \end{equation} Scatter increases at low stellar mass, but the TFR, particularly for low asymmetry galaxies, is still present. Scatter below the TFR is well correlated with asymmetry across the stellar mass range of the sample. Asymmetry is inversely correlated with stellar mass, confirming the result of \citet{bloom2016sami}. We find that slit-based, rather than 2D spatially resolved, measurements of rotation velocity increase scatter below the TFR. Further, if the slit is placed at the photometric PA, rather than the kinemetric, scatter increases again. Scatter is thus highly dependent on PA selection. This result may be relevant to other TFR works, particularly those aiming for the tightest possible TFR for use as a distance indicator. \section*{Acknowledgements} The SAMI Galaxy Survey is based on observations made at the Anglo-Australian Telescope. The Sydney-AAO Multi-object Integral field spectrograph (SAMI) was developed jointly by the University of Sydney and the Australian Astronomical Observatory. The SAMI input catalogue is based on data taken from the Sloan Digital Sky Survey, the GAMA Survey and the VST ATLAS Survey. The SAMI Galaxy Survey is funded by the Australian Research Council Centre of Excellence for All-sky Astrophysics (CAASTRO), through project number CE110001020, and other participating institutions. The SAMI Galaxy Survey website is . GAMA is a joint European-Australasian project based around a spectroscopic campaign using the Anglo-Australian Telescope. The GAMA input catalogue is based on data taken from the Sloan Digital Sky Survey and the UKIRT Infrared Deep Sky Survey. Complementary imaging of the GAMA regions is being obtained by a number of independent survey programs including GALEX MIS, VST KiDS, VISTA VIKING, WISE, Herschel- ATLAS, GMRT ASKAP providing UV to radio coverage. GAMA is funded by the STFC (UK), the ARC (Australia), the AAO, and the participating institutions. The GAMA website is . Support for AMM is provided by NASA through Hubble Fellowship grant $\#$HST-HF2-51377 awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS5-26555. SMC acknowledges the support of an Australian Research Council Future Fellowship (FT100100457). M.S.O. acknowledges the funding support from the Australian Research Council through a Future Fellowship (FT140100255). {We thank the referee for their helpful and insightful comments, which clarified and improved the paper.}

\section{Introduction} \label{intro} Quantum decays are a common phenomenon in particle, nuclear, and atomic physics \cite{ghirardi,fprev,pdg}. A typical starting point for the discussion of the decay of the amplitude for the survival probability of a certain unstable state $S$ \begin{equation} a_{S}(t)=\int_{m_{th}}^{+\infty}\mathrm{dm}d_{S}(m)e^{-imt}\text{ ,}\label{as \end{equation} where $d_{S}(m)$ is the so-called energy (or mass) distribution ($\mathrm{dm d_{S}(m)$ is the probability that the unstable state has an energy (or mass) between $m$ and $m+dm$ and $m_{th}$ the lowest energy of the system). Under general assumptions, one can show that the survival probability $p_{S (t)=\left\vert a_{S}(t)\right\vert ^{2}$ is not exponential both at short times (where $p_{S}^{\prime}(t\rightarrow0)=0$) and at long times (where a power low is realized). For these deviations to occur, it is enough that a low-energy threshold is present and that $d_{S}(m)$ is not of the Breit-Wigner type \cite{vw}, see also Refs. \cite{ghirardi,fprev,khalfin,facchiprl,winter,volya,urbanolast} and refs. therein. The short-time behavior leads to the so-called Quantum\ Zeno Effect (QZE): multiple collapse measurements freeze the time evolution, thus preventing the decay to take place \cite{dega,misra,koshinorev,gppra}. Note, very often $p_{S}(t)$ is expressed as $1-t^{2}/\tau_{Z}^{2}+...$($\tau_{Z}$ is the Zeno time), but the weaker requirement of a zero derivative of $p_{S}(t)$ at $t=0$ is sufficient. Experimentally, deviations from the exponential law have been measured at short times in Ref. \cite{raizen} (for the corresponding QZE see Ref. \cite{raizen2}) and at long times in\ Ref. \cite{rothe} (for an indirect proof through data on beryllium decays, see Ref. \cite{kelkar}). Lee Hamiltonians \cite{lee} (LH) represent a useful theoretical framework for the study of decays, e.g. Refs. \cite{fprev,facchiprl,vanhove,duecan,giacosapra} and refs. therein. This approach resembles very closely Quantum\ Field Theory (QFT). Similar Hamiltonians have been used in various areas of physics, which go from atomic physics and quantum optic \cite{vw,ford,jc} to QCD \cite{liu}. The issue of non-exponential decay in a pure QFT framework is still debated: while in\ Ref. \cite{maiani} a negative result was found (also in the case of super-renormalizable Lagrangian), in\ Ref. \cite{duecan,zenoqft} a different result is obtained: it is argued that also in QFT Eq. (1) holds and short- and long-time deviations take place. While the final goal is the derivation of Eq. (1), and hence of non-exponential decay, in a genuine QFT relativistic environment, in this work we take a more humble intent. We aim to recall in a detailed (and also didactical) way how Eq. (1) emerges when using Lee Hamiltonians. In particular, we shall also show that $d_{S}(m)$ is the mass distribution of the decaying particle. Moreover, we establish a link between a discrete and continuous base of final states and between the basis of the unperturbed and full Hamiltonians. This study is intended to be useful for further analyses on non-exponential decays. The article is organized as follows: in\ Sec. 2 we present the Lee Hamiltonian, both in the discrete and in the continuous cases. Then, in Sec. 3 we study the time evolution of an unstable state: the amplitude of the survival probability is expressed first as the Fourier transformation of the propagator and then of the energy distribution. Finally, in Sec. 4 we present conclusions and outlook. \section{The Lee Hamiltonian's approach} We present here the Lee Hamiltonian (LH), first using an infinite discrete set of decay products and then performing the limit to the continuous case. \subsection{Discrete LH} Let us consider the quantum state $\left\vert S\right\rangle $ as the unstable state that we aim to investigate. In particular, we study its time evolution after its preparation at $t=0.$ The state $\left\vert S\right\rangle $ interacts with an `infinity' of other states, denoted as: \begin{equation} \left\vert k_{n}\right\rangle \text{ with }k_{n}=\frac{2n\pi}{L}\text{ and }n=0,\pm1,\pm2,...\text{ , \end{equation} where the quantity $L$ (with the dimension of energy$^{-1}$) can be thought as the length of the linear box in which we place our system. The physical results should not depend on $L,$ if it is large enough. The quantities $k_{n}$ `look like' momenta, see below. Finally, the basis of the Hilbert space of our quantum problem reads \begin{equation} \text{Basis of the Hilbert space }\mathcal{H}\text{: }\left\{ \left\vert S\right\rangle ,\left\vert k_{0}\right\rangle ,\left\vert k_{1}\right\rangle ,\left\vert k_{-1}\right\rangle ,...\right\} \equiv\left\{ \left\vert S\right\rangle ,\left\vert k_{n}\right\rangle \right\} \end{equation} with the usual orthonormal and completeness relations: \begin{equation} \left\langle S|S\right\rangle =1\text{ , }\left\langle S|k_{n}\right\rangle =0\text{ },\text{ }\left\langle k_{n}|k_{m}\right\rangle =\delta_{nm}\text{ ; }\left\vert S\right\rangle \left\langle S\right\vert +\sum_{n}\left\vert k_{n}\right\rangle \left\langle k_{n}\right\vert =1_{\mathcal{H}}\text{ . \end{equation} The Lee Hamiltonian of the system consists of two pieces: \begin{equation} H=H_{0}+H_{1 \end{equation} where $H_{0}$ describes the free (non-interacting) part while $H_{1}$ mixes $\left\vert S\right\rangle $ with all $\left\vert k_{n}\right\rangle $ \begin{equation} H_{0}=M_{0}\left\vert S\right\rangle \left\langle S\right\vert +\sum _{n=0,\pm1,...}\omega(k_{n})\left\vert k_{n}\right\rangle \left\langle k_{n}\right\vert \text{ ; }H_{1}=\sum_{n=0,\pm1,...}\frac{gf(k_{n})}{\sqrt{L }\left( \left\vert S\right\rangle \left\langle k_{n}\right\vert +\left\vert k_{n}\right\rangle \left\langle S\right\vert \right) \text{ . \end{equation} Following comments are in order: \begin{itemize} \item The quantities $M_{0},$ $\omega(k_{n}),$ $gf(k_{n})$ are real. \item The Hamiltonian $H$ is Hermitian. \item Dimensions: $M_{0}$ and $\omega(k_{n})$ have dimensions [energy], while $g$ has dimensions [energy$^{+1/2}$] \item The energy $M_{0}$ is the bare energy of the level $\left\vert S\right\rangle $. In particle physics, it is the bare mass at rest. \item The energy $\omega(k_{n})$ is the bare energy of the state $\left\vert k_{n}\right\rangle ,$ see below. \item The coupling constant $g$ measures the strength of the interaction; the form factor $f(k_{n})$ modulates the interaction. In practice, each mixing $\left\vert S\right\rangle \longleftrightarrow\left\vert k_{n}\right\rangle $ has its own coupling constant $gf(k_{n}).$ \item The factor$\sqrt{L}$ is introduced for future convenience: it is necessary for a smooth continuous limit $L\rightarrow\infty$. \item For notational simplicity, $\sum_{n=0,\pm1,...}$can be also expressed simply as $\sum_{n}$. \end{itemize} Further discussion concerns the interpretation and the energy $\omega(k_{n})$. \textbf{Interpretation:} The state $\left\vert S\right\rangle $ represents an unstable particle $S$ in its rest frame and the state $\left\vert k_{n}\right\rangle $ represents a possible final state of the decay of $S.$ In the simplest case of a two-body decay, the state $\left\vert k_{n \right\rangle $ represents \textbf{two} particles emitted by $S$ flying back-to-back \begin{equation} S\rightarrow\varphi_{1}+\varphi_{2}\text{ . \end{equation} In the case of one spacial dimension, $k_{n}$ can be interpreted as the momentum of $\varphi_{1}$, while $-k_{n}$ is the momentum of $\varphi_{2}$. Schematically: $\left\vert k_{n}\right\rangle \equiv\left\vert \varphi _{1}(k_{n}),\varphi_{2}(-k_{n})\right\rangle $. In this way, the total momentum of $\left\vert k_{n}\right\rangle $ is still zero, as it must. (The 3D extension is straightforward). As examples of such a process, we may think of: (i) The neutral pion $\pi^{0}$ decays into two photons: $\pi ^{0}\rightarrow\gamma\gamma.$ Then, $\pi^{0}$ in its rest frame corresponds to $\left\vert S\right\rangle ,$ while $\gamma\gamma$ corresponds to $\left\vert k_{n}\right\rangle $ (one photon has momentum $k_{n},$ the other $-k_{n}$). (Note, a very large number of two-body decays is listed in the PDG \cite{pdg}). (ii) An excited atom $A^{\ast}$ decays into the-ground state atom $A$ emitting a photon $\gamma$: $A^{\ast}\rightarrow A\gamma.$ In this case, $A^{\ast}$ is the sate $\left\vert S\right\rangle ,$ while $\left\vert k_{n}\right\rangle $ represents the joint system of the ground-state atom $A$ and the photon. \textbf{Function }$\omega(k_{n})$: as mentioned above, the function $\omega(k_{n})$ represents the energy of the state $\left\vert k_{n \right\rangle $. In the case of a two-body decay its form is given by \begin{equation} \omega(k_{n})=\sqrt{k_{n}^{2}+m_{1}^{2}}+\sqrt{k_{n}^{2}+m_{2}^{2}}\text{ , \end{equation} where $m_{1}$ is the mass of $\varphi_{1}$ and $m_{2}$ of $\varphi_{2}$. Clearly, $\omega(k_{n})\geq m_{1}+m_{2}=m_{th}$, where $m_{th}$ represents the lowest energy of the $\left\vert k\right\rangle $ states. In the two-photon decay such as the process (i) described above, one has $m_{1}=m_{2}=0,$ hence $\omega(k_{n})=2\left\vert k_{n}\right\vert \geq0=m_{th}$. In an atomic decay of the type $A^{\ast}\rightarrow A+\gamma$, one has $m_{1}=0,$ and $m_{2}=M_{A},$ hence \begin{equation} \omega(k_{n})\simeq\left\vert k_{n}\right\vert +M_{A}. \end{equation} In this case, one could also subtract a constant term, $H_{0}\rightarrow H_{0}-M_{A}1_{\mathcal{H}},$ out of which $\omega(k_{n})\simeq\left\vert k_{n}\right\vert $. \subsection{Continuous LH} The limit $L\rightarrow\infty$ implies that the variable $k_{n}$ becomes continuos: \begin{equation} k_{n}=\frac{2\pi n}{L}\rightarrow k\subset(-\infty,+\infty). \end{equation} As usual, when $L$ is sent to infinity sums turn into integrals \begin{equation} \sum_{n}=\frac{L}{2\pi}\sum_{n}\frac{2\pi}{L}\rightarrow\frac{L}{2\pi \int_{-\infty}^{+\infty}\mathrm{dk}=L\int_{-\infty}^{+\infty}\frac {\mathrm{dk}}{2\pi}\text{ , \end{equation} where $\delta k=2\pi/L$ has been introduced in order to generate the differential $dk$. Next, we turn to the kets $\left\vert k\right\rangle $ in the continuous limit, for which we expect that $\left\langle k_{1 |k_{2}\right\rangle =\delta(k_{1}-k_{2})$. To this end, let us write down the following $L$-dependent discrete representation of the Dirac-delta function \begin{equation} \delta_{L}(k_{n})=\int_{-L/2}^{L/2}\frac{\mathrm{dx}}{2\pi}e^{ik_{n x}=\left\{ \begin{array} [c]{c 0\text{ for }n\neq0\\ \frac{L}{2\pi}\text{ for }n=0 \end{array} \right. \text{ . \end{equation} In the limit $L\rightarrow\infty$ one obtains (for an arbitrary function $u(k)$): \begin{equation} u(0)=\sum_{n}\delta k\delta_{L}(k_{n})u(k_{n})\rightarrow\int_{-\infty }^{+\infty}\mathrm{dk}\delta(k)u(k)=u(0) \end{equation} showing that $\delta(k)=\lim_{L\rightarrow\infty}\delta_{L}(k_{n})$ holds. Finally, the quite subtle link between $\left\vert k_{n}\right\rangle $ and $\left\vert k\right\rangle $ is given by \begin{equation} \left\vert k_{n}\right\rangle \overset{L\rightarrow\infty}{=}\sqrt{\frac{2\pi }{L}}\left\vert k\right\rangle \text{ . \end{equation} Namely: \begin{equation} \left\langle k_{1}|k_{2}\right\rangle =\lim_{L\rightarrow\infty}\frac{L}{2\pi }\left\langle k_{n_{1}}|k_{n_{2}}\right\rangle =\lim_{L\rightarrow\infty }\left\{ \begin{array} [c]{c 0\text{ for }n_{1}\neq n_{2}\\ \frac{L}{2\pi}=\delta_{L}(0)\text{ for }n_{1}=n_{2 \end{array} \right. =\delta(k_{1}-k_{2})\text{ , \end{equation} as desired. (Note, in 3D we have $\sum_{\mathbf{k}}\rightarrow V\int \frac{d^{3}k}{(2\pi)^{3}}$ ,where $V=L^{3},$ and $\left\vert \mathbf{k =2\pi\mathbf{n}/L\right\rangle \rightarrow(2\pi)^{3/2}/\sqrt{V}\left\vert \mathbf{k}\right\rangle $). It is also quite peculiar that the dimension of the ket changes when considering the limit $L\rightarrow\infty$ \begin{equation} \dim[\left\vert k_{n}\right\rangle ]=[\text{Energy}^{0}]\text{ (dimensionless) , }\dim[\left\vert k\right\rangle ]=[\text{Energy}^{-1/2}]\text{ . \end{equation} Then, the continuos Hilbert space is given by $\mathcal{H}=\left\{ \left\vert S\right\rangle ,\left\vert k\right\rangle \right\} $ wit \begin{equation} \left\langle S|S\right\rangle =1\text{ , }\left\langle S|k\right\rangle =0\text{ },\text{ }\left\langle k_{1}|k_{2}\right\rangle =\delta(k_{1 -k_{2})\text{ . \end{equation} We also check the completeness relation: \begin{equation} 1_{\mathcal{H}}=\left\vert S\right\rangle \left\langle S\right\vert +\sum _{n}\left\vert k_{n}\right\rangle \left\langle k_{n}\right\vert =\left\vert S\right\rangle \left\langle S\right\vert +\sum_{n}\delta k\left( \sqrt {\frac{L}{2\pi}}\left\vert k_{n}\right\rangle \left\langle k_{n}\right\vert \sqrt{\frac{L}{2\pi}}\right) \overset{L\rightarrow\infty}{\rightarrow }\left\vert S\right\rangle \left\langle S\right\vert +\int_{-\infty}^{+\infty }dk\left\vert k\right\rangle \left\langle k\right\vert =1_{\mathcal{H} \end{equation} Finally, we are ready to present the Lee Hamiltonian $H=H_{0}+H_{1}$ in the continuous limit \begin{equation} H_{0}=M\left\vert S\right\rangle \left\langle S\right\vert +\int_{-\infty }^{+\infty}\mathrm{dk}\omega(k)\left\vert k\right\rangle \left\langle k\right\vert \text{ , }H_{1}=\int_{-\infty}^{+\infty}\mathrm{dk}\frac {gf(k)}{\sqrt{2\pi}}\left( \left\vert S\right\rangle \left\langle k\right\vert +\left\vert k\right\rangle \left\langle S\right\vert \right) \text{ .}\nonumber \end{equation} One can verify that the dimensions is preserved. For instance: \begin{equation} \dim\left[ dk\omega(k)\left\vert k\right\rangle \left\langle k\right\vert \right] =\dim[dk]\dim[\omega(k)]\dim^{2}[\left\vert k\right\rangle ]=[\text{Energy}][\text{Energy}][\text{Energy}^{-1}]=[\text{Energy}]\text{ . \end{equation} \section{Determination of the survival probability} \subsection{Time evolution operator} The Schr\"{o}dinger equation (in natural units \begin{equation} i\frac{\partial\left\vert \psi(t)\right\rangle }{\partial t}=H\left\vert \psi(t)\right\rangle \end{equation} can be univocally solved for a certain given initial state \begin{equation} \left\vert \psi(0)\right\rangle =c_{S}\left\vert S\right\rangle +\sum_{n c_{n}\left\vert k_{n}\right\rangle \overset{L\rightarrow\infty}{\equiv c_{S}\left\vert S\right\rangle +\int_{-\infty}^{+\infty}\mathrm{dk c(k)\left\vert k\right\rangle \end{equation} with $c(k)\overset{L\rightarrow\infty}{\equiv}\sqrt{\frac{L}{2\pi}}c_{n}$ . The normalization $\left\langle \psi(t)|\psi(t)\right\rangle =1$ implie \begin{equation} 1=\left\vert c_{S}\right\vert ^{2}+\sum_{n}\left\vert c_{n}\right\vert ^{2}\overset{L\rightarrow\infty}{\equiv}\left\vert c_{S}\right\vert ^{2 +\int_{-\infty}^{+\infty}\mathrm{dk}\left\vert c(k)\right\vert ^{2}\text{ . \end{equation} In particular, one is typically interested to the case $c_{S}=1$ (but not only). A formal solution to the time evolution is obtained by introducing the time-evolution operator \begin{equation} U(t)=e^{-iHt}\rightarrow\left\vert \psi(t)\right\rangle =U(t)\left\vert \psi(0)\right\rangle \text{ . \end{equation} The time-evolution operator $U(t)$ can be expressed in terms of a Fourier transform (for $t>0$) \begin{equation} U(t)=\frac{i}{2\pi}\int_{-\infty}^{+\infty}\mathrm{dE}\frac{1 {E-H+i\varepsilon}e^{-iEt}\text{ }=\frac{i}{2\pi}\int_{-\infty}^{+\infty }\mathrm{dE}G(E)e^{-iEt}\text{ with }G(E)=\frac{1}{E-H+i\varepsilon}\text{ ,} \label{uoft \end{equation} where $\varepsilon$ is an infinitesimal quantity and $G(E)$ is the `propagator operator', which can be expanded as \begin{equation} G(E)=\frac{1}{E-H+i\varepsilon}=\sum_{n=0}^{\infty}\left( \frac{1 {E-H_{0}+i\varepsilon}H_{1}\right) ^{n}\frac{1}{E-H_{0}+i\varepsilon}\text{ , \end{equation} where we have used that $(AB)^{-1}=B^{-1}A^{-1}$ ($A,B$ arbitrary operators on the Hilbert space $\mathcal{H}$). \subsection{Propagator,\ Feynman rules, and survival probability} We are interested in the evaluation of the survival (or non-decay) probability amplitude $a_{S}(t)=\left\langle S\right\vert U(t)\left\vert S\right\rangle $, out of which the survival probability of the state $S$ reads $p_{S (t)=\left\vert a_{S}(t)\right\vert ^{2}.$ In the \textit{trivial limit,} in which $H=H_{0}$ ( $g\rightarrow0$), one has \begin{equation} a_{S}(t)=\left\langle S\right\vert U(t)\left\vert S\right\rangle =\left\langle S\right\vert e^{-iH_{0}t}\left\vert S\right\rangle =e^{-iM_{0}t}\rightarrow p_{S}(t)=1\text{ . \end{equation} Alternatively, one may use Eq. (\ref{uoft}): \begin{equation} a_{S}(t)=\left\langle S\right\vert U(t)\left\vert S\right\rangle =\frac {i}{2\pi}\int_{-\infty}^{+\infty}\mathrm{dE}\frac{1}{E-M_{0}+i\varepsilon }e^{-iEt}\text{ }=e^{-iM_{0}t}\text{ , \end{equation} where we have closed downwards and picked up the pole for $E=M_{0 -i\varepsilon$ (one is obliged to close downwards to guarantee convergence). In passing by, we note that the object \begin{equation} G_{S}^{\text{free}}(E)=G_{S}^{(0)}(E)=\left\langle S\right\vert \frac {1}{E-H_{0}+i\varepsilon}\left\vert S\right\rangle =\frac{1}{E-M_{0 +i\varepsilon \end{equation} is the free propagator of the state $S.$ In the \textit{interacting case }the evaluation of\textit{ }$a(t)$\textit{ }proceeds as follow: \begin{equation} a_{S}(t)=\frac{i}{2\pi}\int_{-\infty}^{+\infty}\mathrm{dE}G_{S}(E)e^{-iEt \text{ , where }G_{S}(E)=\left\langle S\right\vert G(E)\left\vert S\right\rangle =\left\langle S\right\vert \frac{1}{E-H+i\varepsilon}\left\vert S\right\rangle \label{as1 \end{equation} is the full propagator of $S.$ It is now necessary to evaluate $G_{S}(E)$ explicitly though a lengthy but straightforward calculation \cite{duecan}: \begin{equation} G_{S}(E)=\left\langle S\right\vert G(E)\left\vert S\right\rangle =\sum _{n=0}^{\infty}\left\langle S\right\vert \left( \frac{1}{E-H_{0 +i\varepsilon}H_{1}\right) ^{n}\frac{1}{E-H_{0}+i\varepsilon}\left\vert S\right\rangle =\sum_{n=0}^{\infty}G_{S}^{(n)}(E) \end{equation} wit \begin{equation} G_{S}^{(n)}(E)=\left\langle S\right\vert \left( \frac{1}{E-H_{0 +i\varepsilon}H_{1}\right) ^{n}\left\vert S\right\rangle \frac{1 {E-M_{0}+i\varepsilon}\text{ . \end{equation} Let us evaluate the first three terms \begin{equation} n=0\rightarrow G_{S}^{(0)}(E)=\left\langle S\right\vert 1\left\vert S\right\rangle \frac{1}{E-M_{0}+i\varepsilon}=\frac{1}{E-M_{0}+i\varepsilon }\text{ , \end{equation \begin{equation} n=1\rightarrow\left\langle S\right\vert \frac{1}{E-H_{0}+i\varepsilon H_{1}\left\vert S\right\rangle \frac{1}{E-M_{0}+i\varepsilon}=0\text{ , \end{equation} \begin{align} n & =2\rightarrow G_{S}^{(1)}(E)=\left\langle S\right\vert \left( \frac {1}{E-H_{0}+i\varepsilon}H_{1}\right) ^{2}\left\vert S\right\rangle \frac {1}{E-M_{0}+i\varepsilon}\\ & =\frac{1}{E-M_{0}+i\varepsilon}\left\langle S\right\vert H_{1}\frac {1}{E-H_{0}+i\varepsilon}H_{1}\left\vert S\right\rangle \frac{1 {E-M_{0}+i\varepsilon}=-\frac{\Pi(E)}{\left( E-M_{0}+i\varepsilon\right) ^{3}}\text{ . \end{align} The recursive quantity $\Pi(E)$ reads \begin{equation} \Pi(E)=-\left\langle S\right\vert H_{1}\frac{1}{E-H_{0}+i\varepsilon H_{1}\left\vert S\right\rangle \text{ . \end{equation} We introduce $1_{\mathcal{H}}=\left\vert S\right\rangle \left\langle S\right\vert +\int_{-\infty}^{+\infty}dk\left\vert k\right\rangle \left\langle k\right\vert $ two times, obtaining: \begin{align} \Pi(E) & =-\left\langle S\right\vert H_{1}1_{\mathcal{H}}\frac{1 {E-H_{0}+i\varepsilon}1_{\mathcal{H}}H_{1}\left\vert S\right\rangle =-\int_{-\infty}^{+\infty}\mathrm{dk}\int_{-\infty}^{+\infty}\mathrm{dq \left\langle S\right\vert H_{1}\left\vert k\right\rangle \left\langle k\right\vert \frac{1}{E-H_{0}+i\varepsilon}\left\vert q\right\rangle \left\langle q\right\vert H_{1}\left\vert S\right\rangle \nonumber\\ & =-\int_{-\infty}^{+\infty}\mathrm{dk}\int_{-\infty}^{+\infty \mathrm{dq}\frac{gf(k)}{\sqrt{2\pi}}\frac{\delta(k-q)}{E-\omega (k)+i\varepsilon}\frac{gf(q)}{\sqrt{2\pi}}=-\int_{-\infty}^{+\infty \frac{\mathrm{dk}}{2\pi}\frac{g^{2}f(k)^{2}}{E-\omega(k)+i\varepsilon}\text{ ,}\nonumber \end{align} where $\left\langle S\right\vert H_{1}\left\vert k\right\rangle =gf(k)/\sqrt {2\pi}$ was used. Going further, for $n=0,1,2,...$ we get $G_{S ^{(2n+1)}(E)=0$ and \begin{equation} G_{S}^{(2n)}(E)=\frac{\left[ -\Pi(E)\right] ^{n}}{\left( E-M_{0 +i\varepsilon\right) ^{n+1}}\text{ . \end{equation} Finally \begin{align} G_{S}(E) & =\sum_{n=0}^{\infty}G_{S}^{(2n)}(E)=\sum_{n=0}^{\infty \frac{\left[ -\Pi(E)\right] ^{n}}{\left( E-M_{0}+i\varepsilon\right) ^{2n+1}}=\frac{1}{\left( E-M_{0}+i\varepsilon\right) }\sum_{n=0}^{\infty }\frac{\left[ -\Pi(E)\right] ^{n}}{\left( E-M_{0}+i\varepsilon\right) ^{n}}\\ & =\frac{1}{\left( E-M_{0}+i\varepsilon\right) }\frac{1}{1+\frac{\Pi (E)}{E-M_{0}+i\varepsilon}}=\frac{1}{E-M_{0}+\Pi(E)+i\varepsilon}\text{ . \end{align} At this point, we can identify `Feynman rules' reminiscent of QFT \begin{align} \text{bare }S\text{ propagator } & \rightarrow\frac{1}{E-M_{0}+i\varepsilon }\\ \text{bare }k\text{ propagator (}k\text{ fixed)} & \rightarrow\frac {1}{E-\omega(k)+i\varepsilon}\\ kS\text{ vertex} & \rightarrow gf(k)\\ \text{internal }k\text{ line(}k\text{ not fixed)} & \rightarrow-\Pi (E)=\int_{-\infty}^{+\infty}\frac{dk}{2\pi}\frac{g^{2}f(k)^{2}}{E-\omega (k)+i\varepsilon \end{align} Note, the latter can be understood by applying $gf(k)$ at each vertex and the $k$-propagator in the middle, and by an overall integration $\int_{-\infty }^{+\infty}\frac{dk}{2\pi}$ due to the fact that $k$ is not fixed. The full propagator of $S$ determined above \begin{equation} \text{full }S\text{ propagator }\rightarrow\frac{1}{E-M_{0}+\Pi (E)+i\varepsilon}\text{ , \end{equation} can be also obtained in a very elegant way by using the Bethe-Salpeter equation obtained by using the Feynman rules listed above: \begin{equation} G_{S}(E)=\frac{1}{E-M_{0}+i\varepsilon}-\frac{1}{E-M_{0}+i\varepsilon \Pi(E)G_{S}(E)\text{ . \end{equation} Finally, the survival amplitude (\ref{as1}) can be expressed as \begin{equation} a_{S}(t)=\frac{i}{2\pi}\int_{-\infty}^{+\infty}\mathrm{dE}G_{S}(E)e^{-iEt =\frac{i}{2\pi}\int_{-\infty}^{+\infty}\mathrm{dE}\frac{1}{E-M_{0 +\Pi(E)+i\varepsilon}e^{-iEt}\text{ .} \label{as2 \end{equation} \subsection{Spectral function and survival probability} Let us denote the basis of eigenstates of the Hamiltonian $H$ as $\left\vert m\right\rangle $ wit \begin{equation} H\left\vert m\right\rangle =m\left\vert m\right\rangle \text{ for }m\geq m_{th}\text{ (}m_{th}\text{ is the low-energy threshold) . \end{equation} The existence of a minimal energy $m_{th}$ is a general physical and mathematical property. The states $\left\vert m\right\rangle $ form an orthonormal basis of the Hilbert space $\mathcal{H=}\{\left\vert m\right\rangle $ with $m\geq m_{th}\}$, whose elements fulfill standard relations: \begin{equation} 1_{\mathcal{H}}=\int_{m_{th}}^{+\infty}\mathrm{dm}\left\vert m\right\rangle \left\langle m\right\vert \text{ ; }\left\langle m_{1}|m_{2}\right\rangle =\delta(m_{1}-m_{2})\text{ . \end{equation} The link between the `old' basis $\{\left\vert S\right\rangle ,\left\vert k\right\rangle \}$ (eigenstates of $H_{0}$) and the `new' basis $\{\left\vert m\right\rangle \}$ (eigenstates of $H$) is not trivial. The state $\left\vert S\right\rangle $ can be expressed in terms of the basis $\{\left\vert m\right\rangle $ $\}$ as \begin{equation} \left\vert S\right\rangle =\int_{m_{th}}^{\infty}\mathrm{dm}\alpha _{S}(m)\left\vert m\right\rangle \text{ with }\alpha_{S}(m)=\left\langle S|m\right\rangle \text{ . \end{equation} The quantity \begin{equation} d_{S}(m)=\left\vert \alpha_{S}(m)\right\vert ^{2}=\left\vert \left\langle S|m\right\rangle \right\vert ^{2 \end{equation} is called the \textbf{spectral function (or energy/mass distribution) }of the state $S$. The normalization of the state $\left\vert S\right\rangle $ implies the normalization of the mass distribution $d_{S}(m)$: \begin{equation} 1=\left\langle S|S\right\rangle =\int_{m_{th}}^{\infty}d_{S}(m)\mathrm{dm \text{ .} \label{norm \end{equation} The simple intuitive interpretation is that $d_{S}(m)\mathrm{dm}$ represents the probability that the state $S$ has a energy (or mass) between $m$ and $m+dm.$ As a consequence, the time-evolution can be easily evaluated by inserting $1=\int_{m_{th}}^{+\infty}\mathrm{dm}\left\vert m\right\rangle \left\langle m\right\vert $ two times \begin{equation} a_{S}(t)=\left\langle S\right\vert U(t)\left\vert S\right\rangle =\left\langle S\right\vert e^{-iHt}\left\vert S\right\rangle =\int_{m_{th}}^{\infty }\mathrm{dm}d_{S}(m)e^{-imt}\text{ . \end{equation} This is all formally correct, but it does not help us further as long as we do not have a way to calculate $d_{S}(m)$. This is possible by using the propagator of $S$ studied in Sec. 3.2. In fact, the propagator can be re-expressed as (again, inserting $1=\int_{m_{th}}^{+\infty}dm\left\vert m\right\rangle \left\langle m\right\vert $ two times) \begin{equation} G_{S}(E)=\frac{1}{E-M_{0}+\Pi(E)+i\varepsilon}=\left\langle S\right\vert 1\frac{1}{E-H+i\varepsilon}1\left\vert S\right\rangle =\int_{m_{th}}^{+\infty }\mathrm{dm}\frac{d_{S}(m)}{E-m+i\varepsilon}\text{ .} \label{kl \end{equation} Its physical meaning can be understood by noticing that the dressed propagator $G_{S}(E)$ has been rewritten as the `sum' of free propagators, whose weight function is $d_{S}(m).$ As a next step, we need to invert Eq. (\ref{kl}). Let us first consider the case $g=0.$ In this limit, it is evident from Eq. (\ref{kl}) that: \begin{equation} d_{S}(E)=\delta(E-M_{0})\text{ . \end{equation} This is expected because in this case the state $\left\vert S\right\rangle $ is an eigenstate of the Hamiltonian, hence the mass distribution is a delta-function peaked at $M_{0}$. When the interaction is switched on, we evaluate the imaginary part of Eq. (\ref{kl}): \begin{equation} \operatorname{Im}G_{S}(E)=\int_{m_{th}}^{+\infty}\mathrm{dm}\frac{-\varepsilon d_{S}(m)}{(E-m)^{2}+\varepsilon^{2}}=-\int_{m_{th}}^{+\infty}\mathrm{dm d_{S}(m)\pi\delta(E-m)=-\pi d_{S}(E)\text{ }. \end{equation} Hence $d_{S}(E)$ is calculated as \begin{equation} d_{S}(E)=-\frac{\operatorname{Im}G_{S}(E)}{\pi}=\frac{1}{\pi}\frac {\operatorname{Im}\Pi(E)}{(E-M_{0}+\operatorname{Re}\Pi(E))^{2}+\left( \operatorname{Im}\Pi(E)\right) ^{2}}\text{ .} \label{ds \end{equation} The normalization of $d_{S}(E),$ Eq. (\ref{norm}), can be also proven by using Eq. (\ref{ds}), see details in\ Ref. \cite{gpprd}. In the end, once the spectral function $d_{S}(m)$ is known, the survival amplitude can be re-expressed as its Fourier transform by using Eqs. (\ref{as2}) and (\ref{kl}) \begin{align} a_{S}(t) & =\frac{i}{2\pi}\int_{-\infty}^{+\infty}\mathrm{dE}G_{S (E)e^{-iEt}=\frac{i}{2\pi}\int_{-\infty}^{+\infty}\mathrm{d}E\int_{m_{th }^{+\infty}\mathrm{dm}\frac{d_{S}(m)}{E-m+i\varepsilon}e^{-imt}\nonumber\\ & =\int_{m_{th}}^{+\infty}\mathrm{dm}d_{S}(m)e^{-imt}=\int_{m_{th}}^{+\infty }\mathrm{dm}d_{S}(m)e^{-imt}\text{ . \end{align} The latter expression coincides with Eq. (1), whose detailed determination was our goal. From here on, all the usual strategy can be applied \cite{ghirardi,fprev,koshinorev,duecan}. In particular, the (unphysical) Breit-Wigner limit is obtained for $\omega(k)=k$ (unlimited from below) and $f(k)=1,$ out of which $d_{S}(m)=\frac{\Gamma}{2\pi}\left( \left( m-M_{0}\right) ^{2}+\Gamma^{2}/4\right) ^{-1}$ with $\Gamma=g^{2}.$ In this case, $a_{S}(t)=e^{-i(M_{0}-i\Gamma/2)t}$ and $p_{S}(t)=e^{-\Gamma t}$ (see details in\ Refs. \cite{giacosapra,actalast}). \section{Conclusions} We have proven that Eq. (1) holds in the QFT-like approach of effective Lee Hamiltonians by showing all the main steps leading to it. However, a Lee Hamiltonian is not fully equivalent to QFT, since some features are still missing. In fact, the Lee approach does not contain transitions from the vacuum state to some particles (in genuine QFT, terms of the type $\left\vert 0\right\rangle \left\langle S\varphi_{1}\varphi_{2}\right\vert $+hc are also part of the interacting Hamiltonian and affect the results for finite time intervals). Moreover, quadratic expressions are not present in the propagator(s) of the Lee Hamiltonian but naturally appear in QFT. Hence, the main question for the future reads: is Eq. (1) as it stands valid also in\ QFT? If, as argued in\ Refs. \cite{duecan,zenoqft}, this is true, non-exponential decay is realized in QFT both at short and long times. Moreover, in the interesting case of a super-renormalizable Lagrangian, the short-time behavior is independent on the cutoff (this is so because the energy distribution $d_{S}(m)$ scales as $m^{-3}$ for large $m$). Note, this is different from the result of Ref. \cite{maiani} obtained by using perturbation theory at second order (that is, without resumming the propagator). A further interesting topic for the future is the study of the decay of a particle with a nonzero momentum \cite{khalfin2,stefanovich,shirokov,urbanowski,giunti,rel}. Contrary to naive expectations, the usual relativistic time dilatation formula does not hold (even in the exponential limit, a different analytical result is obtained, see details in Ref. \cite{rel}). The full understanding of the study of decay in QFT can also help to shed light on decays of moving particles. \bigskip \textbf{Acknowledgments: }F. G. thanks S. Mr\'{o}wczy\'{n}ski and G. Pagliara for very useful discussions. F. G. acknowledges support from the Polish National Science Centre (NCN) through the OPUS project no. 2015/17/B/ST2/01625.

\section{Introduction} Human visual system relies on a variety of depth cues to gain 3D perception. The most important ones are binocular, defocus, and motion cues. Binocular cues such as stereopsis, eye convergence, and disparity yield depth from binocular vision through exploitation of parallax. Defocus cue allows depth perception even with a single eye by correlating variation of defocus blurs with the motion of the ciliary muscles surrounding the lens. Motion parallax also provides useful input to assess depth, but arrives over time and depends on texture gradients. Computer vision algorithms such as stereo matching \cite{scharstein02, brown03} and depth-from-focus/defocus \cite{nayar92, nayar94, malik07, favaro05, favaro07} seek to directly employ binocular and defocus cues which are available without scene statistics. Recent studies have shown that the two types of cues complement each other to provide 3D perception \cite{held12}. In this paper, we seek to develop learning based approaches to emulate this process. To exploit binocular cues, traditional stereo matching algorithms rely on feature matching and optimization to maintain the Markov Random Field property: the disparity field should be smooth everywhere with abrupt changes at the occlusion boundaries. Existing solutions such as graph-cut, belief propagation~\cite{kolmogorov02, sun03}, although effective, tend to be slow. In contrast, depth-from-focus (DfF) exploits differentiations of sharpness at each pixel across a focal stack and assigns the layer with highest sharpness as its depth. Compared with stereo, DfF generally presents a low fidelity estimation due to depth layer discretization. Earlier DfF techniques use a focal sweep camera to produce a coarse focal stack due to mechanical limitations whereas more recent ones attempt to use a light field to synthetically produce a denser focal stack. Our solution benefits from recent advance on computational photography and we present an efficient and reliable learning based technique to conduct depth inference from a focal stack pair, emulating the process of how human eyes work. We call our technique binocular DfF or B-DfF. Our approach leverages deep learning techniques that can effectively extract features learned from large amount of imagery data. Such a deep representation has shown great promise in stereo matching \cite{zbontar15, zagoruyko15, luo16}. Little work, however, has been proposed on using deep learning for DfF or more importantly, integrating stereo and DfF. This is mainly due to the lack of fully annotated DfF datasets. \begin{figure}[t] \begin{center} \includegraphics[width=0.99\linewidth]{DFF_teaser.pdf} \end{center} \vspace{-8pt} \caption{\emph{BDfFNet} integrates \emph{FocusNet}, \emph{EDoFNet} and \emph{StereoNet} to predict high quality depth map from binocular focal stacks.} \label{} \end{figure} \begin{figure*}[t] \begin{center} \includegraphics[width=0.9\linewidth]{dataset_sample.pdf} \end{center} \vspace{-8pt} \caption{A binocular focal stack pair consists of two horizontally rectified focal stacks. The upper and lower triangles show corresponding slices focusing at respective depths. Bottom shows the ground truth color and depth images. We add Poisson noise to training data, a critical step for handling real scenes.} \label{} \end{figure*} We first construct a comprehensive focal stack dataset. Our dataset is based on the highly diversified dataset from \cite{mayer16}, which contains both stereo color images and ground truth disparity maps. Then we adopt the algorithm from \emph{Virtual DSLR} \cite{yang16} to generate the refocused images. \cite{yang16} uses color and depth image pair as input for light field synthesis and rendering, but without the need to actually create the light field. The quality of the rendered focal stacks are comparable to those captured by expensive DSLR camera. Next, we propose three individual networks: (1) \emph{FocusNet}, a multi-scale network to extract depth from a single focal stack (2) \emph{EDoFNet}, a deep network consisting of small convolution kernels to obtain the extended depth of field (EDoF) image from the focal stack and (3) \emph{StereoNet} to obtain depth directly from a stereo pair. The EDoF image from \emph{EDoFNet} serves to both guide the refinement of the depth from \emph{FocusNet} and provide inputs for \emph{StereoNet}. We also show how to integrate them into a unified solution \emph{BDfFNet} to obtain high quality depth maps. Fig. \ref{} illustrates the pipeline. We evaluate our approach on both synthetic and real data. To physically implement B-DfF, we construct a light field stereo pair by using two Lytro Illum cameras. Light field rendering is then applied to produce the two focal stacks as input to our framework. Comprehensive experiments show that our technique outperforms the state-of-the-art techniques in both accuracy and speed. More importantly, we believe our solution provides important insights on developing future sensors and companion 3D reconstruction solutions analogous to human eyes. \section{Related Work} Our work is closely related to depth from focus/defocus and stereo. The strength and weakness of the two approaches have been extensively discussed in \cite{schechner00, vaish06}. \noindent\textbf{Depth from Focus/Defocus} Blur carries information about the object's distance. Depth from Focus/Defocus (DfF/DfD) recovers scene depth from a collection of images captured under varying focus settings. In general, DfF \cite{nayar92, nayar94, malik07} determines the depth by analyzing the most in-focus slice in the focal stack, while DfD \cite{favaro05, favaro07} infers depth based on the amount of the spatially varying blur at each pixel. To avoid ambiguity in textureless region, Moreno-Noguer \emph{et al.} \cite{Moreno-Noguer07} used active illumination to project a sparse set of dots onto the scene. The defocus of the dots offers depth cue, which could be further used for realistic refocusing. \cite{hasinoff09} combined focal stack with varying aperture to recover scene geometry. Moeller \emph{et al.} \cite{moeller15} applied an efficient nonconvex minimization technique to solve DfD in a variational framework. Suwajanakorn \emph{et al.} \cite{suwajanakorn15} proposed the DfF with mobile phone under uncalibrated setting. They first aligned the focal stack, then jointly optimized the camera parameters and depth map, and further refined the depth map using anisotropic regularization. A drastic difference of these methods to our approach is that they rely on hand-crafted features to estimate focusness or blur kernel, whereas in this paper we leverage neural network to learn more discriminative features from focal stack and directly predict depth at lower computational cost. \noindent \textbf{Learning based Stereo} Depth from stereo has been studied extensively by the computer vision community for decades. We refer the readers to the comprehensive survey for more details \cite{scharstein02, brown03}. Here we only discuss recent methods based on Convolutional Neural Network (CNN). Deep learning benefits stereo matching at various stages. A number of approaches exploit CNN to improve the matching cost. The seminal work by {\v Z}bontar and LeCun~\cite{zbontar15} computed a similarity score from patches using CNN, then applied the traditional cost aggregation and optimization to solve the energy function. Han \emph{et al.}\cite{han2015matchnet} jointly learned feature representations and feature comparison functions in a unified network, which improved on previous results with less storage requirement. Luo \emph{et al.} \cite{luo16} speeded up the matching process by using a product layer, and treated the disparity estimation as a multi-class classification problem. ~\cite{chen2015deep, zagoruyko15, liu2016euclidean, park2016look} conducted similar work but with different network architecture. Alternatively, CNN can also help predict the confidence of disparity map to remove outliers. Seki and Pollefeys \cite{seki2016patch} leveraged CNN for stereo confidence measure, and incorporated predicted confidence into Semi-Global Matching by adjusting its parameters. In order to automatically generate the dataset for learning based confidence measure, Mostegel \emph{et al.} \cite{mostegel16} checked the consistency of multiple depth maps of the same scene obtained with the same stereo approach, and collected labeled confidence map as the training data. End-to-end network architectures have also been explored. Mayer \emph{et al.}~\cite{mayer16} adopted and extended the architecture of the FlowNet~\cite{dosovitskiy2015flownet}, which consists of a contractive part and an expanding part to learn depth at multiple scales. They also created three synthetic datasets to facilitate the training process. Kn{\"o}belreiter \emph{et al.}~\cite{knobelreiter2016end} learned unary and pairwise cost of stereo using CNNs, then posed the optimization as a conditional random field (CRF) problem. The hybrid CNN-CRF model was trained in image's full resolution in an end-to-end fashion. Combining DfF/DfD and stereo matching has also been studied, although not within the learning framework. Early work~\cite{klarquist95, subbarao97} attempted to utilize the depth map from the focus/defocus to reduce the search space for stereo and solve the correspondence problem more efficiently. \cite{Rajagopalan04} simultaneously recovered depth and restored the original focused image from a defocused stereo pair. Recently, Tao \emph{et al.}~\cite{tao13} analyzed the epipolar image (EPI) from light field camera to infer depth. They found that the horizontal variances after vertical integration of the EPI encodes defocus cue, while vertical variances encodes disparity cue. The two cues were then jointly optimized in an MRF framework. To obtain high resolution depth in a semi-calibrated manner, Wang \emph{et al.}~\cite{wang16} proposed a hybrid camera system that consists of two calibrated auxiliary cameras and an uncalibrated main camera. They first transfered the depth from auxiliary cameras to the viewpoint of the main camera by rectifying three images simultaneously, and further improved the depth map along occlusion boundaries using defocus cue. Aforementioned approaches leave the combination and optimization of focus and disparity cue to postprocessing. In contrast, we resort to extra layers of network to infer the optimized depth with low computational cost and efficiency. \section{Dual Focal Stack Dataset} With fast advances of the data driven methods, numerous datasets have been created for various applications. However, by far, there are limited resources on focal stacks. To this end, we generate our dual focal stack dataset based on FlyingThings3D from \cite{mayer16}. FlyingThings3D is an entirely synthetic dataset, consisting of everyday objects flying along randomized 3D paths. Their 3D models and textures are separated into disjointed training and testing parts. In total, the dataset contains about 25,000 stereo images with ground truth disparity. To make the data tractable, we select stereo frames whose largest disparity is less than 100 pixels, then we normalize the disparity to $0\sim1$. Takeda \emph{et al.} \cite{takeda2013fusing} demonstrate that in stereo setup, the disparity $d$ and the diameter of the circle of confusion $c$ have a linear relationship: \begin{equation} \frac{d}{c} = \frac{l}{D} \end{equation} where $l$ is the baseline length and $D$ is the aperture size. Based on above observation, we adopt the \emph{Virtual DSLR} approach from \cite{yang16} to generate synthetic focal stacks. \emph{Virtual DSLR} requires color and disparity image pair as inputs, and outputs refocused images with quality comparable to those captured from regular, expensive DSLR. The advantage of their algorithm is that it resembles light field synthesis and refocusing but does not require actual creation of the light field, hence reducing both memory and computational load. In addition, their method takes special care of occlusion boundaries to avoid color bleeding and discontinuity commonly observed in brute-force blur-based defocus synthesis. To better explain their approach, we list the formulation as below: \begin{equation} C_p = \frac{|s-s_p|}{s_p}D = sD|\frac{1}{z_p} - \frac{1}{z_s}|, \end{equation} To simulate a scene point $p$ with depth $z_p$ projected to a circular region on sensor, we assume the focal length $f$, an aperture size $D$, sensor to lens distance $s$,and the circular region diameter $C_p$. Here $z_s = (1/f - 1/s)^{-1}$ and $s_p = (1/f - 1/z_p)^{-1}$ according to the thin lens law. The diameter of the circular region $C_p$ measures the size of blur kernel and it is linear to the absolute difference of the inverse of the distances $z_p$ and $z_s$. For the scope of this paper, we use only circular apertures, although more complex ones can easily be synthesized. To emulate the pupil of the eye in varying lighting conditions, we randomly select the size of the blur kernel for each stereo pair, but limit the largest diameter of the blur kernel to 31 pixels. We also evenly separate the scene into 16 depth layers and render a refocused image for each layer. After generating the focal stacks, we add poisson noise to the images to simulate the real image captured by a camera. This turns out to be critical in real scene experiments, as described in section \ref{}. Finally, we split the generated dual focal stacks into 750 training data and 70 testing data. Figure \ref{} shows two slices from the dual focal and their corresponding color and depth image. \section{B-DfF Network Architecture} \label{} Convolutional neural networks are very efficient at learning non-linear mapping between the input and the output. Therefore, we aim to take an end-to-end approach to predict a depth map. \cite{simonyan14} shows that a deep network with small kernels is very effective in image recognition tasks. Although a small kernel has limited spatial support, a deep network by stacking multiple layers of such kernels could substantially enlarge the receptive field while reducing the number of parameters to avoid overfitting. Therefore, a general principle in designing our network is to use deep architecture with small convolutional kernels. As already mentioned, the input to our neural network are two rectified focal stacks. To extract depth from defocus and disparity, our solution is composed of three individual networks. We start in section \ref{} by describing the \emph{FocusNet}, a multi-scale network that estimates depth from a single focal stack. Then in section \ref{} we further enhance the result by the extended depth of field images from \emph{EDoFNet}. Finally we combine \emph{StereoNet} and \emph{FocusNet} in \ref{} to infer high quality depth from binocular focal stacks. \subsection{FocusNet for DfF/DfD} \label{} \begin{figure}[t] \begin{center} \includegraphics[width=0.99\linewidth]{network_multiscaleNets.pdf} \end{center} \vspace{-8pt} \caption{\emph{FocusNet} is a multi-scale network for conducting depth-from-focus.} \label{} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=1\linewidth]{n3.pdf} \end{center} \vspace{-8pt} \caption{Left: \emph{EDoFNet} consists of 20 layers of convolutional layers to form an extended depth-of-field (EDoF) image from focal stack. Right: \emph{FocusNet-v2} combines \emph{FocusNet} and \emph{EDoFNet} by using the EDoF image to refine the depth estimation.} \label{} \end{figure} \begin{figure*}[t] \begin{center} \includegraphics[width=0.85\linewidth]{hourglass_network_small.pdf} \end{center} \vspace{-8pt} \caption{(a) \emph{StereoNet} follows the Hourglass network architecture which consists of the max pooling layer (yellow), the deconvolution layer (green) and the residual module (blue). (b) shows the detailed residual module.} \label{} \end{figure*} Motivated by successes from multi-scale networks, we propose \emph{FocusNet}, a multiscale network to extract depth from a single focal stack. Specifically, \emph{FocusNet} consists of four branches of various scales. Except the first branch, other branches subsample the image by using different strides in the convolutional layer, enabling aggregation of information over large areas. Therefore, both the high-level information from the coarse feature maps and the fine details could be preserved. At the end of the branch, a deconvolutional layer is introduced to upsample the image to its original resolution. Compared with the traditional bicubic upsampling, deconvolution layer automatically learns upsampling kernels that are better suited for the application. Finally, we stack the multi-scale features maps together, resulting in a concatenated per-pixel feature vector. The feature vectors are further fused by layers of convolutional networks to predict the final depth value. An illustration of the network architecture is shown in Fig. \ref{}. We use $3\times3$ kernels for most layers except those convolutional layers used for downsampling and upsampling, where a larger kernel is used to cover more pixels. The spatial padding is also applied for each convolution layer to preserve the resolution. Following \cite{simonyan14}, the number of feature maps increases as the image resolution decreases. Between the convolutional layers we insert PReLU layer \cite{he15} to increase the network's nonlinearity. For the input of the network we simply stack the focal stack images together along the channel's dimension. \subsection{Guided Depth Refinement by EDoF Image} \label{} There exist many approaches \cite{ferstl13, hui16} to refine/upsample depth image with the guidance of an intensity image. The observation is that homogeneous texture regions often correspond to homogeneous surface parts, while depth edges often occur at high intensity variations. With this in mind, we set out to first extract the EDoF image from the focal stack, then guide the refinement of the depth image. Several methods \cite{kuthirummal11, suwajanakorn15} have been proposed to extract the EDoF image from the focal stack. However, the post processing is suboptimal in terms of computational efficiency and elegance. Thus, we seek to directly output an EDoF image from a separate network, which we termed \emph{EDoFNet}. \emph{EDoFNet} is composed of 20 convolutional layers, with PRelu as its activation function. The input of the \emph{EDoFNet} is the focal stack, the same as the input of \emph{FocusNet}, and the output is the EDoF image. With the kernel size of $3\times3$, a 20 layer convolutional network will produce a receptive field of $41\times41$, which is larger than the size of the largest blur kernel. Fig. \ref{} shows the architecture of \emph{EDoFNet}. Finally, we concatenate the depth image from \emph{FocusNet} and the EDoF image from the \emph{EDoFNet}, and fuse them by using another 10 layer convolutional network. We call the new network \emph{FocusNet-v2}. The architecture of \emph{FocusNet-v2} is illustrated in Fig. \ref{}. \subsection{StereoNet and BDfFNet for Depth from Binocular Focal Stack} \label{} Given the EDoF stereo pair from the \emph{EDoFNet}, we set out to estimate depth from stereo using another network, termed \emph{StereoNet}. For stereo matching, it is critical to consolidate both local and global cues to generate precise pixel-wise disparity. To this end, we propose \emph{StereoNet} by adopting the Hourglass network architecture~\cite{newell16}, as shown in Fig. \ref{}. The advantage of this network is that it can attentively evaluate the coherence of features across scales by utilizing large amount of residual modules~\cite{he2016deep}. The network composes of downsampling part and upsampling part. The downsampling part consists of a series of max pooling interleaved with residual modules while the upsampling part is a mirrored architecture of the downsampling part, with max pooling replaced by deconvolution layer for upsampling. Between any pair of corresponding max pooling and upsampling, there is a connection layer comprising of a residual module. Elementwise addition follows to add processed lower-level features to higher-level features. In this way, the network learns a more holistic representation of input images. Prediction is generated at the end of the upsampling part. One round of downsampling and upsampling part can be viewed as one iteration of predicting, whereas additional rounds can be stacked to refine initial estimates. For \emph{StereoNet}, we use two rounds of downsampling and upsampling parts as they already give good performance, while further rounds improve marginally at the cost of more training time. Note that the weights are not shared in the two rounds. Different from~\cite{newell16}, we do not downsample input images before the first downsampling part. This stems from the difference in problem settings: our solution aims for pixel-wise precision while~\cite{newell16} only requires structured understanding of images. Throughout the network, we use small convolution filters ($3\times3$ or $1\times1$). After each pair of downsampling and upsampling parts, supervision is applied using the same ground truth disparity map. The final output is of the same resolution as the input images. Finally, we construct \emph{BDfFNet} by concatenating the results from \emph{StereoNet}, \emph{FocusNet-v2} and \emph{EDoFNet}, and adding more convolutional layers. The convolutional layers serve to find the optimal combination from focus cue and disparity cue. \section{Implementation} \label{} \noindent\textbf{Optimization} Given the focal stack as input and ground truth color/depth image as label, we train all the networks end-to-end. In our implementation, we first train each network individually, then fine-tune the concatenated network with the pre-trained weights as initialization. Because \emph{FocusNet} and \emph{FocusNet-v2} contains multiple convolutional layers for downsampling, the input image needs to be cropped to the nearest number that is multiple of 8 for both height and width. We use the mean square error (MSE) with $l_2$-norm regularization as the loss for all models, which leads to the following objective function \begin{equation} \label{eqn: lossFunction} \min_{\theta}\frac{1}{N}\sum_{i=1}^N\left\| F(S^{i};\theta)- D^{i} \right\rVert_{2}^{2} + \frac{\lambda}{2}\left\| \theta \right\rVert_{2}^{2} \end{equation} where $S^i$ and $D^i$ are the $i$-th focal stack and depth image, $F(S^{i};\theta)$ is the function represented by the network and $\theta$ are the learned weights. Although there are works \cite{zhao17} suggesting the mean absolute error (MAE) might be a better loss function, our experiment shows that results from MAE are inferior to MSE. Following \cite{ioffe15}, we apply batch normalization after the convolution layer and before PRelu layer. We initialize the weights using the technique from \cite{he15}. We employ MXNET \cite{chen15} as the learning framework and train and test the networks on a NVIDIA K80 graphic card. We make use of the Adam optimizer \cite{kingma15} and set the weight decay = 0.002, $\beta1$ = 0.9, $\beta2$ = 0.999. The initial learning rate is set to be 0.001. We first train each sub-network of \emph{BDfFNet} separately and then combine them for further training. All the networks are trained for 80 epoches. \noindent\textbf{Data augmentation and preprocessing} For \emph{FocusNet} and \emph{EDoFNet}, the size of the analyzed patches determines the largest sensible blur kernel size. Therefore, we randomly crop a patch of size $64\times64$ from the image, which contains enough contextual information to extract the depth and EDoF image. For \emph{StereoNet}, a larger patch of size $256\times256$ is used to accommodate the large disparity between stereo images. To facilitate the generalization of the network, we augment the data by flipping the patches horizontally and vertically. All the data augmentations are performed on the fly at almost no extra cost. Finally, the range of all images are normalized to $0\sim1$. \begin{figure}[t] \begin{center} \includegraphics[width=1.0\linewidth]{DFF_result_EDoF.pdf} \end{center} \vspace{-8pt} \caption{Results of our \emph{EDoFNet}. First row shows two slices of the focal stack focusing at different depth. Second and third row show the EDoF and ground truth image respectively.} \label{} \end{figure} \section{Experiments} \label{} \begin{figure*}[h] \begin{center} \includegraphics[width=0.9\linewidth]{result_Focus_v3.pdf} \end{center} \vspace{-8pt} \caption{Comparisons on \emph{FocusNet} vs. \emph{FocusNet-v2}, i.e., without and with the guide of an all-focus image.} \label{} \end{figure*} \subsection{Extract the EDoF Image from Focal Stack} We train \emph{EDoFNet} on a single focal stack of 16 slices. Although the network has simple structure, the output EDoF image features high image quality. Our network also runs much faster than conventional methods based on global optimization: on the resolution of $960\times540$ it runs at 4 frames per second. Fig. \ref{} shows the result of \emph{EDoFNet}. Compared with ground truth image, the produced EDoF image is slightly blurry. However, given a very noisy focal stack as input, the resultant EDoF image gets rid of large part of the noise. Our experiments also show that it suffices to guide the refinement of depth image and be used as the input of \emph{StereoNet}. \subsection{Depth Estimation from Focal Stack} \label{} As mentioned in \ref{}, to construct \emph{FocusNet-v2}, we first train \emph{FocusNet} and \emph{EDoFNet} respectively, then concatenate their output with more fusion layers and train the combination. Fig. \ref{} shows the result of both \emph{FocusNet} and \emph{FocusNet-v2}. We observe that \emph{FocusNet} produces results with splotchy artifact, and depth bleeds across object's boundary. However, \emph{FocusNet-v2} utilizes the EDoF color image to assist depth refinement, alleviating the artifacts and leading to clearer depth boundary. It is worth noting that we also trained a network that has identical structure to \emph{FocusNet-v2} from scratch, but the result is of inferior quality. We suspect this is due to the good initialization provided by the pre-trained model. We compare our results with \cite{suwajanakorn15} and \cite{moeller15} using the data provided by the authors of \cite{suwajanakorn15}. We select 16 images from their focal stack for DfF. Fig. \ref{} illustrates the results. Our \emph{FocusNet-v2} is capable of predicting disparity value with higher quality, while using significantly less time (0.9 second) than \cite{suwajanakorn15} (10 mins) and \cite{moeller15} (4 seconds). We also train the \emph{FocusNet-v2} on a clean dataset without poisson noise. It performs better on synthetic data, but exhibits severe noise pattern on real images, as shown in Fig. \ref{}. The experiment confirms the necessity to add noise to the dataset for simulating real images. \begin{figure}[t] \begin{center} \includegraphics[width=0.99\linewidth]{result_compareVariational_v2.pdf} \end{center} \vspace{-8pt} \caption{Comparisons on depth estimation from a single focal stack using our \emph{FocusNet-v2} (last column) vs. \cite{suwajanakorn15} (second column) and \cite{moeller15} (third column). \emph{FocusNet-v2} is able to maintain smoothness on flat regions while preserving sharp occlusion boundaries.} \label{} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=0.5\linewidth]{compareDatasetWoNoise_v2.pdf} \end{center} \vspace{-8pt} \caption{Results from \emph{FocusNet-v2} trained by the clean dataset without poisson noise.} \label{} \end{figure} \subsection{Depth Estimation from Stereo and Binocular Focal Stack} \label{} Figure \ref{} shows the results from \emph{StereoNet} and \emph{BDfFNet}. Compared with \emph{FocusNet-v2}, \emph{StereoNet} gives better depth estimation. This is expected since \emph{StereoNet} requires binocular focal stacks as input, while \emph{FocusNet-v2} only use a single focal stack. However, \emph{StereoNet} exhibits blocky artifacts and overly smoothed boundary. In contrast, depth prediction from \emph{BDfFNet} features sharp edges. The depth in flat surface region is also smoother compared to \emph{FocusNet-v2}. Table \ref{} describes the mean absolute error (MAE) and running time of all models on $960\times540$ image. \begin{figure*}[t] \begin{center} \includegraphics[width=0.9\linewidth]{DFF_result_Stereo_BDfF_v2.pdf} \end{center} \vspace{-8pt} \caption{Comparisons on results only using \emph{StereoNet} vs. the composed \emph{BDfFNet}. BDfFNet produces much sharper boundaries while reducing blocky artifacts.} \label{} \end{figure*} \begin{table} [b] \begin{center} \begin{tabular}{|c|c |c |c |c|} \hline \ & \emph{FocusNet} & \emph{FocusNet-v2} & \emph{StereoNet} & \emph{BDfFNet} \\ [0.5ex] \hline MAE & 0.045 & 0.031 & 0.024 & 0.021 \\ Time(s) & 0.6 & 0.9 & 2.8 & 9.7 \\ \hline \end{tabular} \caption{MAE and running time of models.} \label{} \end{center} \end{table} \subsection{Real Scene Experiment} \label{} We further conduct tests on real scenes. To physically implement B-DfF, we construct a light field stereo pair by using two Lytro Illum cameras, as illustrated in Fig. \ref{}. Comparing with stereo focal sweeping, the Lytro pair can conduct high quality post-capture refocusing without the need for accurate synchronized mechanical control on focal length. In our experiment the two light field cameras share the same configuration including the zoom and focus settings. The raw images are preprocessed using Light Field Toolbox \cite{dansereau13}. Finally we conduct refocusing using shift-and-add algorithm \cite{ng05} to synthesize the focal stack. \begin{figure*}[h] \begin{center} \includegraphics[width=0.9\linewidth]{DFF_result_realScene_v2.pdf} \end{center} \vspace{-8pt} \caption{Comparisons of real scene results from \emph{FocusNet-v2}, \emph{StereoNet} and \emph{BDfFNet}.} \label{} \end{figure*} Figure \ref{} shows the predicted depth from \emph{FocusNet-v2}, \emph{StereoNet} and \emph{BDfFNet}. Results show that \emph{BDfFNet} benefits from both \emph{FocusNet-v2} and \emph{StereoNet} to offer smoother depth with sharp edges. The experiments also demonstrate that models learned from our dataset could be transferred to predict real scene depth. \begin{figure}[t] \begin{center} \includegraphics[width=0.7\linewidth]{dualLytro.pdf} \end{center} \vspace{-8pt} \caption{To emulate our B-DfF setup, we combine a pair of Lytro Illum cameras into a stereo setup.} \label{} \end{figure} \section{Discussions and Future Work} \label{} Our deepeye solution exploits efficient learning and computational light field imaging to infer depths from a focal stack pair. Our technique mimics human vision system that simultaneously employs binocular stereo matching and monocular depth-from-focus. Comprehensive experiments show that our technique is able to produce high quality depth estimation orders of magnitudes faster than the prior art. In addition, we have created a large dual focal stack database with ground truth disparity. Our current implementation limits the input size of our network to be focal stacks of 16 layers. In our experiments, we have shown that it is able to produce high fidelity depth estimation under our setup. To handle denser focal stacks, one possibility is to concatenate all images in the stack as a 3D $(XYS)$ focal cube or volume \cite{zhou12}, where $X$ and $Y$ are the width and height and $S$ is the index of a layer. We can then downsample the $XS$ slice along $S$ dimension to 16 slices using light field compression or simplification techniques such as tensor \cite{wanner12} and triangulation \cite{yu13}. Another important future direction we plan to explore is to replace one of the two focal stacks to be an all-focus image. This would further reduce the computational cost for constructing the network but would require adjusting the architecture. Finally, aside from computer vision, we hope our work will stimulate significant future work in human perception and the biological nature of human eyes. \section{Acknowledgments} This project was supported by the National Science Foundation under grant CBET-1706130. {\small \bibliographystyle{ieee}

\section{Introduction \label{sec-intro}}} \IEEEPARstart{I}{n} recent years, recommender systems, which help users discover items of interest from a large resource collection, have been playing an increasingly important role in various online services~\cite{dias2008value,koren2015advances}. Traditional recommendation methods (\emph{e.g.,}\xspace matrix factorization) mainly aim to learn an effective prediction function for characterizing user-item interaction records (\emph{e.g.,}\xspace user-item rating matrix). With the rapid development of web services, various kinds of auxiliary data (\emph{a.k.a.,}\xspace side information) become available in recommender systems. Although auxiliary data is likely to contain useful information for recommendation~\cite{schafer2007collaborative}, it is difficult to model and utilize these heterogeneous and complex information in recommender systems. Furthermore, it is more challenging to develop a relatively general approach to model these varying data in different systems or platforms. As a promising direction, heterogeneous information network (HIN), consisting of multiple types of nodes and links, has been proposed as a powerful information modeling method~\cite{sun2011pathsim,shi2017survey,shi2017heterogeneous}. Due to its flexility in modeling data heterogeneity, HIN has been adopted in recommender systems to characterize rich auxiliary data. In Fig.~\ref{fig_framework}(a), we present an example for movie recommendation characterized by HINs. We can see that the HIN contains multiple types of entities connected by different types of relations. Under the HIN based representation, the recommendation problem can be considered as a similarity search task over the HIN~\cite{sun2011pathsim}. Such a recommendation setting is called as \emph{HIN based recommendation}~\cite{yu2013collaborative}. HIN based recommendation has received much attention in the literature~\cite{feng2012incorporating,yu2013collaborative,yu2014personalized,shi2015semantic,shi2016integrating,zheng2017recommendation}. The basic idea of most existing HIN based recommendation methods is to leverage path based semantic relatedness between users and items over HINs, \emph{e.g.,}\xspace meta-path based similarities, for recommendation. \begin{figure*}[t \centering \includegraphics[width=15cm]{image/framework.pdf} \caption{\label{fig_framework}The schematic illustration of the proposed HERec approach.} \end{figure*} Although HIN based methods have achieved performance improvement to some extent, there are two major problems for these methods using meta-path based similarities. First, meta-path based similarities rely on explicit path reachability, and may not be reliable to be used for recommendation when path connections are sparse or noisy. It is likely that some links in HINs are accidentally formed which do not convey meaningful semantics. Second, meta-path based similarities mainly characterize semantic relations defined over HINs, and may not be directly applicable to recommender systems. It is likely that the derived path based similarities have no explicit impact on the recommendation performance in some cases. Existing methods mainly learn a linear weighting mechanism to combine the path based similarities~\cite{shi2016integrating} or latent factors~\cite{yu2013collaborative}, which cannot learn the complicated mapping mechanism of HIN information for recommendation. The two problems essentially reflect two fundamental issues for HIN based recommendation, namely effective information extraction and exploitation based on HINs for recommendation. For the first issue, it is challenging to develop a way to effectively extract and represent useful information for HINs due to data heterogeneity. Unlike previous studies using meta-path based similarities~\cite{sun2011pathsim,yu2013collaborative}, our idea is to learn effective heterogeneous network representations for summarizing important structural characteristics and properties of HINs. Following \cite{perozzi2014deepwalk,grover2016node2vec}, we characterize nodes from HINs with low-dimensional vectors, \emph{i.e.,}\xspace embeddings. Instead of relying on explicit path connection, we would like to encode useful information from HINs with latent vectors. Compared with meta-path based similarity, the learned embeddings are in a more compact form that is easy to use and integrate. Also, the network embedding approach itself is more resistant to sparse and noisy data. However, most existing network embedding methods focus on homogeneous networks only consisting of a single type of nodes and links, and cannot directly deal with heterogeneous networks consisting of multiple types of nodes and links. Hence, we propose a new heterogeneous network embedding method. Considering heterogeneous characteristics and rich semantics reflected by meta-paths, the proposed method first uses a random walk strategy guided by meta-paths to generate node sequences. For each meta-path, we learn a unique embedding representation for a node by maximizing its co-occurrence probability with neighboring nodes in the sequences sampled according to the given meta-path. We fuse the multiple embeddings \emph{w.r.t.} different meta-paths as the output of HIN embedding. After obtaining the embeddings from HINs, we study how to integrate and utilize such information in recommender systems. We don't assume the learned embeddings are naturally applicable in recommender systems. Instead, we propose and explore three fusion functions to integrate multiple embeddings of a node into a single representation for recommendation, including simple linear fusion, personalized linear fusion and non-linear fusion. These fusion functions provide flexible ways to transform HIN embeddings into useful information for recommendation. Specially, we emphasize that personalization and non-linearity are two key points to consider for information transformation in our setting. Finally, we extend the classic matrix factorization framework by incorporating the fused HIN embeddings. The prediction model and the fusion function are jointly optimized for the rating prediction task. By integrating the above two parts together, this work presents a novel HIN embedding based recommendation approach, called \emph{HERec} for short. HERec first extracts useful HIN based information using the proposed HIN embedding method, and then utilizes the extracted information for recommendation using the extended matrix factorization model. We present the overall illustration for the proposed approach in Fig.~\ref{fig_framework}. Extensive experiments on three real-world datasets demonstrate the effectiveness of the proposed approach. We also verify the ability of HERec to alleviate cold-start problem and examine the impact of meta-paths on performance. The key contributions of this paper can be summarized as follows: \textbullet ~We propose a heterogeneous network embedding method guided by meta-paths to uncover the semantic and structural information of heterogeneous information networks. Moreover, we propose a general embedding fusion approach to integerate different embeddings based on different meta-paths into a single representation. \textbullet ~We propose a novel heterogeneous information network embedding for recommendation model, called HERec for short. HERec can effectively integrate various kinds of embedding information in HIN to enhance the recommendation performance. In addition, we design a set of three flexible fusion functions to effectively transform HIN embeddings into useful information for recommendation. \textbullet ~Extensive experiments on three real-world datasets demonstrate the effectiveness of the proposed model. Moreover, we show the capability of the proposed model for the cold-start prediction problem, and reveal that the transformed embedding information from HINs can improve the recommendation performance. The remainder of this paper is organized as follows. Section~\ref{sec-rel} introduces the related works. Section~\ref{sec-def} describes notations used in the paper and presents some preliminary knowledge. Then, We propose the heterogeneous network embedding method and the HERec model in Section~\ref{sec-model}. Experiments and detailed analysis are reported in Section~\ref{sec-exp}. Finally, we conclude the paper in Section~\ref{sec-con}. \begin{figure*}[t] \centering \subfigure[Douban Movie]{ \begin{minipage}[b]{0.3\textwidth} \includegraphics[width=1\textwidth]{image/doubanmovie.pdf} \end{minipage} } \subfigure[Douban Book]{ \begin{minipage}[b]{0.3\textwidth} \includegraphics[width=1\textwidth]{image/doubanbook.pdf} \end{minipage} } \subfigure[Yelp]{ \begin{minipage}[b]{0.3\textwidth} \includegraphics[width=1\textwidth]{image/yelp.pdf} \end{minipage} } \caption{\label{fig_schema}Network schemas of heterogeneous information networks for the used three datasets. In our task, users and items are our major focus, denoted by large-sized circles, while the other attributes are denoted by small-sized circles. } \end{figure*} \section{Related Work \label{sec-rel}} In this section, we will review the related studies in three aspects, namely recommender systems, heterogeneous information networks and network embedding. In the literature of recommender systems, early works mainly adopt collaborative filtering (CF) methods to utilize historical interactions for recommendation~\cite{schafer2007collaborative}. Particularly, the matrix factorization approach~\cite{koren2009matrix,shi2012adaptive} has shown its effectiveness and efficiency in many applications, which factorizes user-item rating matrix into two low rank user-specific and item-specific matrices, and then utilizes the factorized matrices to make further predictions~\cite{koren2015advances}. Since CF methods usually suffer from cold-start problem, many works~\cite{yin2013lcars,feng2012incorporating,hong2013co} attempt to leverage additional information to improve recommendation performance. For example, Ma et al.~\cite{ma2011recommender} integrate social relations into matrix factorization in recommendation. Ling et al.~\cite{ling2014ratings} consider the information of both ratings and reviews and propose a unified model to combine content based filtering with collaborative filtering for rating prediction task. Ye et al.~\cite{ye2011exploiting} incorporate user preference, social influence and geographical influence in the recommendation and propose a unified POI recommendation framework. More recently, Sedhain et al.~\cite{sedhain2017low} explain drawbacks of three popular cold-start models~\cite{gantner2010learning,krohn2012multi,sedhain2014social} and further propose a learning based approach for the cold-start problem to leverage social data via randomised SVD. And many works begin to utilize deep models ($\emph{e.g.,}\xspace$ convolutional neural network, auto encoder) to exploit text information~\cite{zheng2017joint}, image information~\cite{he2016vbpr} and network structure information~\cite{zhang2016collaborative} for better recommendation. In addition, there are also some typical frameworks focusing on incorporating auxiliary information for recommendation. Chen et al.~\cite{chen2012svdfeature} propose a typical SVDFeature framework to efficiently solve the feature based matrix factorization. And Rendle~\cite{rendle2010factorization} proposes factorization machine, which is a generic approach to combine the generality of feature engineering. As a newly emerging direction, heterogeneous information network~\cite{shi2017survey} can naturally model complex objects and their rich relations in recommender systems, in which objects are of different types and links among objects represent different relations~\cite{sun2013mining}. And several path based similarity measures~\cite{lao2010relational,sun2011pathsim,shi2014hetesim} are proposed to evaluate the similarity of objects in heterogeneous information network. Therefore, some researchers have began to be aware of the importance of HIN based recommendation. Wang et al.~\cite{feng2012incorporating} propose the OptRank method to alleviate the cold-start problem by utilizing heterogeneous information contained in social tagging system. Furthermore, the concept of meta-path is introduced into hybrid recommender systems~\cite{yu2013recommendation}. Yu et al.~\cite{yu2013collaborative} utilize meta-path based similarities as regularization terms in the matrix factorization framework. Yu et al.~\cite{yu2014personalized} take advantage of different types of entity relationships in heterogeneous information network and propose a personalized recommendation framework for implicit feedback dataset. Luo et al.~\cite{luo2014hete} propose a collaborative filtering based social recommendation method using heterogeneous relations. More recently, Shi et al.~\cite{shi2015semantic} propose the concept of weighted heterogeneous information network and design a meta-path based collaborative filtering model to flexibly integrate heterogeneous information for personalized recommendation. In \cite{shi2016integrating,zheng2016dual,zheng2017recommendation}, the similarities of users and items are both evaluated by path based similarity measures under different semantic meta-paths and a matrix factorization based on dual regularization framework is proposed for rating prediction. Most of HIN based methods rely on the path based similarity, which may not fully mine latent features of users and items on HINs for recommendation. On the other hand, network embedding has shown its potential in structure feature extraction and has been successfully applied in many data mining tasks~\cite{hoff2002latent,yan2007graph}, such as classification~\cite{tu2016max}, clustering~\cite{wei2017cross,cao2016deep} and recommendation~\cite{liang2016factorization}. Deepwalk~\cite{perozzi2014deepwalk} combines random walk and skip-gram to learn network representations. Furthermore, Grover and Leskovec~\cite{grover2016node2vec} propose a more flexible network embedding framework based on a biased random walk procedure. In addition, LINE \cite{tang2015line} and SDNE~\cite{wang2016structural} characterize the second-order link proximity, as well as neighbor relations. Cao et al.~\cite{cao2015grarep} propose the GraRep model to capture higher-order graph proximity for network representations. Besides leaning network embedding from only the topology, there are also many works~\cite{pan2016tri,yang2015network,zhang2016homophily} leveraging node content information and other available graph information for the robust representations. Unfortunately, most of network embedding methods focus on homogeneous networks, and thus they cannot directly be applied for heterogeneous networks. Recently, several works~\cite{chang2015heterogeneous,tang2015pte,xu2017embedding,chen2017task,dong2017metapath2vec} attempt to analyze heterogeneous networks via embedding methods. Particularly, Chang et al.~\cite{chang2015heterogeneous} design a deep embedding model to capture the complex interaction between the heterogeneous data in the network. Xu et al.~\cite{xu2017embedding} propose a EOE method to encode the intra-network and inter-network edges for the coupled heterogeneous network. Dong et al.~\cite{dong2017metapath2vec} define the neighbor of nodes via meta-path and learn the heterogeneous embedding by skip-gram with negative sampling. Although these methods can learn network embeddings in various heterogeneous network, their representations of nodes and relations may not be optimum for recommendation. To our knowledge, it is the first attempt which adopts the network embedding approach to extract useful information from heterogeneous information network and leverage such information for rating prediction. The proposed approach utilizes the flexibility of HIN for modeling complex heterogeneous context information, and meanwhile borrows the capability of network embedding for learning effective information representation. The final rating prediction component further incorporates a transformation mechanism implemented by three flexible functions to utilize the learned information from network embedding. \section{Preliminary \label{sec-def}} A heterogeneous information network is a special kind of information network, which either contains multiple types of objects or multiple types of links. \begin{myDef} \textbf{Heterogeneous information network}~\cite{sun2012mining}. A HIN is denoted as $\mathcal{G} = \{\mathcal{V}, \mathcal{E}\}$ consisting of a object set $\mathcal{V}$ and a link set $\mathcal{E}$. A HIN is also associated with an object type mapping function $\phi: \mathcal{V} \rightarrow \mathcal{A}$ and a link type mapping function $\psi: \mathcal{E} \rightarrow \mathcal{R}$. $\mathcal{A}$ and $\mathcal{R}$ denote the sets of predefined object and link types, where $|\mathcal{A}| + |\mathcal{R}| > 2$. \end{myDef} The complexity of heterogeneous information network drives us to provide the meta level ($\emph{e.g.,}\xspace$ schema-level) description for understanding the object types and link types better in the network. Hence, the concept of network schema is proposed to describe the meta structure of a network. \begin{myDef} \textbf{Network schema}~\cite{sun2013mining,sun2009ranking}. The network schema is denoted as $\mathcal{S} = (\mathcal{A}, \mathcal{R})$. It is a meta template for an information network $\mathcal{G} = \{\mathcal{V}, \mathcal{E}\}$ with the object type mapping $\phi: \mathcal{V} \rightarrow \mathcal{A}$ and the link type mapping $\psi: \mathcal{E} \rightarrow \mathcal{R}$, which is a directed graph defined over object types $\mathcal{A}$, with edges as relations from $\mathcal{R}$. \end{myDef} \begin{exmp} As shown in Fig.~\ref{fig_framework}(a), we have represented the setting of movie recommender systems by HINs. We further present its corresponding network schema in Fig.~\ref{fig_schema}(a), consisting of multiple types of objects, including User ($U$), Movie ($M$), Director ($D$). There exist different types of links between objects to represent different relations. A user-user link indicates the friendship between two users, while a user-movie link indicates the rating relation. Similarly, we present the schematic network schemas for book and business recommender systems in Fig.~\ref{fig_schema}(b) and Fig.~\ref{fig_schema}(c) respectively. \end{exmp} In HINs, two objects can be connected via different semantic paths, which are called meta-paths. \begin{myDef} \textbf{Meta-path}~\cite{sun2011pathsim}. A meta-path $\rho$ is defined on a network schema $\mathcal{S} = (\mathcal{A}, \mathcal{R})$ and is denoted as a path in the form of $A_1 \xrightarrow{R_1} A_2 \xrightarrow{R_2} \cdots \xrightarrow{R_l} A_{l+1}$ (abbreviated as $A_1A_2 \cdots A_{l+1}$), which describes a composite relation $R = R_1 \circ R_2 \circ \cdots \circ R_l$ between object $A_1$ and $A_{l+1}$, where $\circ$ denotes the composition operator on relations. \end{myDef} \begin{exmp} Taking Fig.~\ref{fig_schema}(a) as an example, two objects can be connected via multiple meta-paths, $\emph{e.g.,}\xspace$ ``User - User" ($UU$) and ``User - Movie - User" ($UMU$). Different meta-paths usually convey different semantics. For example, the $UU$ path indicates friendship between two users, while the $UMU$ path indicates the co-watch relation between two users, \emph{i.e.,}\xspace they have watched the same movies. As will be seen later, the detailed meta-paths used in this work is summarized in Table~\ref{tab_metapath}. \end{exmp} Recently, HIN has become a mainstream approach to model various complex interaction systems \cite{shi2017survey}. Specially, it has been adopted in recommender systems for characterizing complex and heterogenous recommendation settings. \begin{myDef} \textbf{HIN based recommendation}. In a recommender system, various kinds of information can be modeled by a HIN $\mathcal{G} = \{\mathcal{V}, \mathcal{E}\}$. On recommendation-oriented HINs, two kinds of entities (\emph{i.e.,}\xspace users and items) together with the relations between them (\emph{i.e.,}\xspace rating relation) are our focus. Let $\mathcal{U}\subset \mathcal{V}$ and $\mathcal{I}\subset \mathcal{V}$ denote the sets of users and items respectively, a triplet $\langle u, i, r_{u,i}\rangle$ denotes a record that a user $u$ assigns a rating of $r_{u,i}$ to an item $i$, and $\mathcal{R}=\{\langle u, i, r_{u,i}\rangle\}$ denotes the set of rating records. We have $\mathcal{U}\subset \mathcal{V}$, $\mathcal{I}\subset \mathcal{V}$ and $\mathcal{R}\subset \mathcal{E}$. Given the HIN $\mathcal{G} = \{\mathcal{V}, \mathcal{E}\}$, the goal is to predict the rating score $r_{u,i'}$ of $u\in \mathcal{U}$ to a non-rated item $i'\in \mathcal{I}$. \end{myDef} Several efforts have been made for HIN based recommendation. Most of these works mainly leverage meta-path based similarities to enhance the recommendation performance~\cite{yu2013collaborative,yu2014personalized,shi2015semantic,shi2016integrating}. Next, we will present a new heterogeneous network embedding based approach to this task, which is able to effectively exploit the information reflected in HINs. The notations we will use throughout the article are summarized in Table~\ref{tabl_notations}. \begin{table}[htbp] \centering \caption{Notations and explanations.}\label{tabl_notations}{ \begin{tabular}{c||c} \hline {Notation} & {Explanation}\\ \hline \hline {$\mathcal{G}$}&{heterogeneous information network}\\ \hline {$\mathcal{V}$} & {object set} \\ \hline {$\mathcal{E}$} & {link set} \\ \hline {$\mathcal{S}$} & {network schema} \\ \hline {$\mathcal{A}$} & {object type set } \\ \hline {$\mathcal{R}$} & {link type set} \\ \hline {$\mathcal{U}$} & {user set} \\ \hline {$\mathcal{I}$} & {item set} \\ \hline {$\widehat{r_{u,i}}$} & {predicted rating user $u$ gives to item $i$}\\ \hline {$\bm{e}_v$} & {low-dimensional representation of node $v$} \\ \hline {$\mathcal{N}_u$} & {neighborhood of node $u$} \\ \hline {$\rho$} & {a meta-path} \\ \hline {$\mathcal{P}$} & {meta-path set} \\ \hline {$\bm{e}^{(U)}_u, \bm{e}^{(I)}_i$} & {final representations of user $u$, item $i$}\\ \hline {$d$} & {dimension of HIN embeddings} \\ \hline {$D$} & {dimension of latent factors} \\ \hline {$\mathbf{x}_u, \mathbf{y}_i$} & {latent factors of user $u$, item $i$} \\ \hline {$\bm{{\gamma}}^{(U)}_u$, $\bm{{\gamma}}_i^{(I)}$} & {latent factors for pairing HIN embedding of user $u$, item $i$} \\ \hline {$\alpha$, $\beta$} & {parameters for integrating HIN embeddings} \\ \hline {$\mathbf{M}^{(l)}$} & { transformation matrix w.r.t the $l$-th mete-path} \\ \hline {$\bm{b}^{(l)}$} & { bias vector w.r.t the $l$-th mete-path} \\ \hline {$w^{(l)}_u$} & {preference weight of user $u$ over the $l$-th meta-path} \\ \hline {$\bm{\Theta}^{(U)}$, $\bm{\Theta}^{(I)}$} & {parameters of fusion functions for users, items} \\ \hline {$\lambda$} & {regularization parameter} \\ \hline {$\eta$} & {learning rate} \\ \hline \end{tabular}} \end{table} \section{The Proposed Approach \label{sec-model}} In this section, we present a \emph{H}eterogeneous network \emph{E}mbedding based approach for \emph{Rec}emendation, called \emph{HERec}. For addressing the two issues introduced in Section 1, the proposed HERec approach consists of two major components. First, we propose a new heterogeneous network embedding method to learn the user/item embeddings from HINs. Then, we extend the classic matrix factorization framework by incorporating the learned embeddings using a flexible set of fusion functions. We present an overall schematic illustration of the proposed approach in Fig.~\ref{fig_framework}. After the construction of the HINs~(Fig.~\ref{fig_framework}(a)), two major steps are presented, namely HIN embedding~(Fig.~\ref{fig_framework}(b)) and recommendation~(Fig.~\ref{fig_framework}(c)). Next, we will present detailed illustration of the proposed approach. \subsection{Heterogeneous Network Embedding} Inspired by the recent progress on network embedding~\cite{perozzi2014deepwalk,grover2016node2vec}, we adopt the representation learning method to extract and represent useful information of HINs for recommendation. Given a HIN $\mathcal{G} = \{\mathcal{V}, \mathcal{E}\}$, our goal is to learn a low-dimensional representation $\bm{e}_v \in \mathbb{R}^d$ (\emph{a.k.a.,}\xspace embedding) for each node $v \in \mathcal{V}$. The learned embeddings are expected to highly summarize informative characteristics, which are likely to be useful in recommender systems represented on HINs. Compared with meta-path based similarity~\cite{sun2011pathsim,shi2015semantic}, it is much easier to use and integrate the learned representations in subsequent procedures. However, most of the existing network embedding methods mainly focus on homogeneous networks, which are not able to effectively model heterogeneous networks. For instance, the pioneering study \emph{deepwalk}~\cite{perozzi2014deepwalk} uses random walk to generate node sequences, which cannot discriminate nodes and edges with different types. Hence, it requires a more principled way to traverse the HINs and generate meaningful node sequences. \subsubsection{Meta-path based Random Walk} To generate meaningful node sequences, the key is to design an effective walking strategy that is able to capture the complex semantics reflected in HINs. In the literature of HINs, meta-path is an important concept to characterize the semantic patterns for HINs~\cite{sun2011pathsim}. Hence, we propose to use the meta-path based random walk method to generate node sequences. Giving a heterogeneous network $\mathcal{G} = \{\mathcal{V}, \mathcal{E}\}$ and a meta-path $\rho : A_1 \xrightarrow{R_1} \cdots A_t \xrightarrow{R_t} A_{t+1} \cdots \xrightarrow{R_l} A_{l+1}$, the walk path is generated according to the following distribution: \begin{eqnarray}\label{eq-rw} &&P({n_{t+1}=x |n_{t}=v, \rho})\\ &=&\begin{cases} \frac{1}{|\mathcal{N}^{A_{t+1}}(v)|}, &\text{($v, x$) $\in$ $\mathcal{E}$ and $\phi(x) = A_{t+1}$};\nonumber\\ 0,& \text{otherwise}, \end{cases} \end{eqnarray} where $n_t$ is the $t$-th node in the walk, $v$ has the type of $A_t$, and $\mathcal{N}^{(A_{t+1})}(v)$ is the first-order neighbor set for node $v$ with the type of $A_{t+1}$. A walk will follow the pattern of a meta-path repetitively until it reaches the pre-defined length. \begin{exmp} We still take Fig.~\ref{fig_framework}(a) as an example, which represents the heterogeneous information network of movie recommender systems. Given a meta-path $UMU$, we can generate two sample walks (\emph{i.e.,}\xspace node sequences) by starting with the user node of \emph{Tom}: (1) Tom$_{User}$ $\rightarrow$ The Terminator$_{Movie}$ $\rightarrow$ Mary$_{User}$, and (2) Tom$_{User}$ $\rightarrow$ Avater$_{Movie}$ $\rightarrow$ Bob$_{User}$ $\rightarrow$ The Terminator$_{Movie}$ $\rightarrow$ Mary$_{User}$. Similarly, given the meta-path $UMDMU$, we can also generate another node sequence: Tom$_{User}$ $\rightarrow$ The Terminator$_{Movie}$ $\rightarrow$ Cameron$_{Director}$ $\rightarrow$ Avater$_{Movie}$ $\rightarrow$ Mary$_{User}$. It is intuitive to see that these meta-paths can lead to meaningful node sequences corresponding to different semantic relations. \end{exmp} \begin{figure}[t \centering \includegraphics[width=9cm]{image/random_walk.pdf} \caption{\label{fig_randwalk}An illustrative example of the proposed meta-path based random walk. We first perform random walks guided by some selected meta-paths, and then filter the node sequences \emph{not with} the user type or item type.} \end{figure} \subsubsection{Type Constraint and Filtering} Since our goal is to improve the recommendation performance, the main focus is to learn effective representations for users and items, while objects with other types are of less interest in our task. Hence, we only select meta-paths starting with \emph{user type} or \emph{item type}. Once a node sequence has been generated using the above method, it is likely to contain nodes with different types. We further remove the nodes with a type different from the starting type. In this way, the final sequence will only consist of nodes with the starting type. Applying type filtering to node sequences has two benefits. First, although node sequences are constructed using meta-paths with heterogeneous types, the final representations are learned using homogeneous neighborhood. We embed the nodes with the same type in the same space, which relaxes the challenging goal of representing all the heterogeneous objects in a unified space. Second, given a fixed-length window, a node is able to utilize more homogeneous neighbors that are more likely to be relevant than others with different types. \begin{exmp} As shown in Fig.~\ref{fig_randwalk}, in order to learn effective representations for users and items, we only consider the meta-paths in which the starting type is \emph{user type} or \emph{item type}. In this way, we can derive some meta-paths, such as $UMU$, $UMDMU$ and $MUM$. Take the meta-path of $UMU$ as an instance. We can generate a sampled sequence ``$u_1 \rightarrow m_1 \rightarrow u_2 \rightarrow m_2 \rightarrow u_3 \rightarrow m_2 \rightarrow u_4$" according to Eq.~\ref{eq-rw}. Once a sequence has been constructed, we further remove the nodes with a different type compared with the starting node. In this way, we finally obtain a homogeneous node sequence ``$u_1 \rightarrow u_2 \rightarrow u_3 \rightarrow u_4$". \end{exmp} The connections between homogeneous nodes are essentially constructed via the heterogeneous neighborhood nodes. After this step, our next focus will be how to learn effective representations for homogeneous sequences. \subsubsection{Optimization Objective} Given a meta-path, we can construct the neighborhood $\mathcal{N}_u$ for node $u$ based on co-occurrence in a fixed-length window. Following node2vec~\cite{grover2016node2vec}, we can learn the representations of nodes to optimize the following objective: \begin{equation} \max_f \sum_{u \in \mathcal{V}}{\log Pr(\mathcal{N}_u | f(u))}, \end{equation} where $f: \mathcal{V} \rightarrow \mathbb{R}^d$ is a function (aiming to learn) mapping each node to $d$-dimensional feature space, and $\mathcal{N}_u \subset \mathcal{V}$ represents the neighborhood of node $u$, \emph{w.r.t.} a specific meta-path. We can learn the embedding mapping function $f(\cdot)$ by applying stochastic gradient descent (SGD) to optimize this objective. A major difference between previous methods and ours lies in the construction of $\mathcal{N}_u$. Our method selects homogeneous neighbors using meta-path based random walks. The whole algorithm framework is shown in Algorithm~\ref{alg_embedding}. \begin{algorithm}[htb] \caption{HIN embedding algorithm for a single meta-path.} \label{alg_embedding} \begin{algorithmic}[1] \Require the heterogeneous information network $\mathcal{G} = \{\mathcal{V}, \mathcal{E}\}$; the given meta-path $\rho$; the target node type $A_t$; the dimension of embedding $d$; the walk length $wl$; the neighborhood size $ns$; the number of walks per node $r$. \Ensure The embedding of target node type w.r.t the single meta-path, denoted by $\bm{e}$ \State Initialize $e$ by standard normal distribution; \State $paths = []$; \For {each $v$ $\in$ $\mathcal{V}$ and $\phi(v) == A_t$} \For {$i$ = 1 to $r$} \State $path = []$; \While {$wl > 0$} \State walk to node $x$ according to Eq.~\ref{eq-rw}; \If {$\phi(x) == A_t$} \State append node $x$ into $path$; \State $wl \leftarrow wl - 1$; \EndIf \EndWhile \State Add $path$ to $paths$; \EndFor \EndFor \State $\bm{e} = SGD(paths, d, ns)$; \State \Return {$\bm{e}$}. \end{algorithmic} \end{algorithm} \subsubsection{Embedding Fusion} Heterogeneous network embedding provides a general way to extract useful information from HINs. For our model, given a node $v \in \mathcal{V}$, we can obtain a set of representations $\{ \bm{e}^{(l)}_v \}_{l=1}^{|\mathcal{P}|}$, where $\mathcal{P}$ denotes the set of meta-paths, and $\bm{e}^{(l)}_v$ denotes the representation of $v$ \emph{w.r.t.} the $l$-th meta-path. It requires a principled fusion way to transform node embeddings into a more suitable form that is useful to improve recommendation performance. Existing studies usually adopt a linear weighting mechanism to combine the information mined from HINs (\emph{e.g.,}\xspace meta-path based similarities), which may not be capable of deriving effective information representations for recommendation. Hence, we propose to use a general function $g(\cdot)$, which aims to fuse the learned node embeddings for users and items: \begin{eqnarray}\label{eq-eui} \bm{e}^{(U)}_u &\leftarrow& g(\{\bm{e}^{(l)}_u\}),\\ \bm{e}^{(I)}_i &\leftarrow& g(\{\bm{e}^{(l)}_i\}),\nonumber \end{eqnarray} where $\bm{e}^{(U)}_u$ and $\bm{e}^{(I)}_i$ are the final representations for a user $u$ and an item $i$ respectively, called \emph{HIN embeddings}. Since users and items are our focus, we only learn the embeddings for users and items. At the current stage, we do not specify the form of function $g(\cdot)$. Instead, we believe that a good fusion function should be learned according to the specific task. Hence, we leave the formulation and optimization of the fusion function in our recommendation model. \subsection{Integrating Matrix Factorization with Fused HIN Embedding for Recommendation} Previously, we have studied how to extract and represent useful information from HINs for recommendation. With HIN embedding, we can obtain user embeddings $\{\bm{e}^{(U)}_u\}_{u \in \mathcal{U}}$ and item embeddings $\{\bm{e}^{(I)}_i\}_{i \in \mathcal{I}}$, which are further specified by a function $g(\cdot)$ that is to learn. Now we study how to utilize the learned embeddings for recommendation. \subsubsection{Rating Predictor} We build our rating predictor based on the classic matrix factorization (MF) model~\cite{mnih2008probabilistic}, which factorizes the user-item rating matrix into user-specific and item-specific matrices. In MF, the rating of a user $u$ on an item $i$ is simply defined as follows: \begin{equation}\label{eq-mf} \widehat{r_{u,i}} = \mathbf{x}_u^{\top}\cdot \mathbf{y}_i, \end{equation} where $\mathbf{x}_u \in \mathbb{R}^{D}$ and $\mathbf{y}_i \in \mathbb{R}^{D}$ denote the latent factors corresponding to user $u$ and item $i$. Since we have also obtained the representations for user $u$ and item $i$, we further incorporate them into the rating predictor as below \begin{equation}\label{eq-predictor} \widehat{r_{u,i}} = \mathbf{x}_u^{\top}\cdot \mathbf{y}_i + \alpha \cdot {\bm{e}_u^{(U)}}^{\top}\cdot\bm{{\gamma}}_i^{(I)} + \beta \cdot {\bm{{\gamma}}^{(U)}_u}^{\top}\cdot {\bm{e}_i^{(I)}}, \end{equation} where $\bm{e}_u^{(U)}$ and $\bm{e}_i^{(I)}$ are the fused embeddings, $\bm{{\gamma}}^{(U)}_u$ and $\bm{{\gamma}}_i^{(I)}$ are user-specific and item-specific latent factors to pair with the HIN embeddings $\bm{e}_u^{(U)}$ and $\bm{e}_i^{(I)}$ respectively, and $\alpha$ and $\beta$ are the tuning parameters to integrate the three terms. For Eq.~\ref{eq-predictor}, we need to note two points. First, $\bm{e}_u^{(U)}$ and $\bm{e}_i^{(I)}$ are the output of function $g(\cdot)$ in Eq.~\ref{eq-eui}. We assume that the derived embeddings after transformation by function $g(\cdot)$ are applicable in MF. Second, we don't directly pair $\bm{e}_u^{(U)}$ with $\bm{e}_i^{(I)}$. Recall the proposed embedding method indeed characterizes the relatedness between objects with the same type. We incorporate new latent factors $\bm{{\gamma}}^{(U)}_u$ and $\bm{{\gamma}}_i^{(I)}$ to relax the assumption that $\bm{e}_u^{(U)}$ and $\bm{e}_i^{(I)}$ have to be in the same space, which increases the flexility to the prediction model. \subsubsection{Setting the Fusion Function} Previously, we assume the function $g(\cdot)$ has been given in a general form. Now, we study how to set the fusion function, which transforms HIN embeddings into a form that is useful in recommender systems. We only discuss the function for fusing user embeddings, and it is similar to fuse item embeddings. We propose to use three fusion functions to integrate embeddings: \begin{itemize} \item Simple linear fusion. We assume each user has the same preference for each meta-path, and therefore, we assign each meta-path with a unified weight ($\emph{i.e.,}\xspace$ average value) for each user. Moreover, we linearly transform embeddings to target space. \begin{equation}\label{eq-slf} g(\{\bm{e}^{(l)}_u\}) = \frac{1}{|\mathcal{P}|}\sum_{l=1}^{|\mathcal{P}|}{(\mathbf{M}^{(l)} \bm{e}^{(l)}_u+\bm{b}^{(l)})}, \end{equation} where $\mathcal{P}$ is the set of considered meta-paths, $\mathbf{M}^{(l)} \in \mathbb{R}^{D\times d}$ and $\bm{b}^{(l)} \in \mathbb{R}^{D}$ are the transformation matrix and bias vector \emph{w.r.t.} the $l$-th meta-path. \item Personalized linear fusion. The simple linear fusion cannot model users' personalized preference over the meta-paths. So we further assign each user with a weight vector on meta-paths, representing user's personalized preference for each meta-path. It is more reasonable that each user has his/her personal interest preference in many real applications. \begin{equation}\label{eq-plf} g(\{\bm{e}^{(l)}_u\}) = \sum_{l=1}^{|\mathcal{P}|}{w^{(l)}_u (\mathbf{M}^{(l)} \bm{e}^{(l)}_u+\bm{b}^{(l)})}, \end{equation} where $w^{(l)}_u$ is the preference weight of user $u$ over the $l$-th meta-path. \item Personalized non-linear fusion. Linear fusion has limited expressive power in modeling complex data relations. Hence, we use non-linear function to enhance the fusion ability. \begin{equation}\label{eq-pnlf} g(\{\bm{e}^{(l)}_u\}) = \sigma\bigg( \sum_{l=1}^{|\mathcal{P}|} {w^{(l)}_u \sigma\big(\mathbf{M}^{(l)} \bm{e}^{(l)}_u+\bm{b}^{(l)}\big)}\bigg), \end{equation} where $\sigma(\cdot)$ is a non-linear function, \emph{i.e.,}\xspace sigmoid function in our work. Although we only use two non-linear transformations, it is flexible to extend to multiple non-linear layers, \emph{e.g.,}\xspace Multi-Layer Perceptrons. \end{itemize} \subsubsection{Model Learning} We blend the fusion function into matrix factorization framework for learning the parameters of the proposed model. The objective can be formulated as follows: \begin{align} \pounds &= \sum_{\langle u, i, r_{u,i}\rangle \in \mathcal{R}}{(r_{u,i} - \widehat{r_{u,i}})}^2 + \lambda \sum_u{(\|\mathbf{x}_u\|_2 + \|\mathbf{y}_i\|_2} \nonumber \\ &+ \|\bm{\gamma}^{(U)}_u\|_2 + \|\bm{\gamma}^{(I)}_i\|_2 + \|\bm{\Theta}^{(U)}\|_2 + \|\bm{\Theta}^{(I)}\|_2), \end{align} \noindent where $\widehat{r_{u,i}}$ is the predicted rating using Eq.~\ref{eq-predictor} by the proposed model, $\lambda$ is the regularization parameter, and $\bm{\Theta}^{(U)}$ and $\bm{\Theta}^{(I)}$ are the parameters of the function $g(\cdot)$ for users and items respectively. We adopt SGD to efficiently optimize the final objective. The update of original latent factors $\{\mathbf{x}_u\}$ and $\{\mathbf{y}_i\}$ are the same as that of standard MF in Eq.~\ref{eq-mf}. The parameters of the proposed model will be updated as follows: \begin{small} \begin{eqnarray} \bm{\Theta}_{u,l}^{(U)} &\leftarrow& \bm{\Theta}_{u,l}^{(U)} - \eta \cdot (-\alpha (r_{u,i}-\widehat{r_{u,i}})\bm{\gamma}_i^{(I)}\frac{\partial{\bm{e}_{u}^{(U)}}}{\partial{\bm{\Theta}_{u,l}^{(U)}}} + {\lambda_{\Theta}}\bm{\Theta}_{u,l}^{(U)}), \nonumber \\ \\ \bm{\gamma}_u^{(U)} &\leftarrow& \bm{\gamma}_u^{(U)} - \eta \cdot (-\beta (r_{u,i}-\widehat{r_{u,i}})\bm{e}_i^{(I)} + \lambda_{\gamma}\bm{\gamma}_u^{(U)}),\\ \nonumber\\ \bm{\Theta}_{i,l}^{(I)} &\leftarrow& \bm{\Theta}_{i,l}^{(I)} - \eta \cdot (-\beta (r_{u,i}-\widehat{r_{u,i}})\bm{\gamma}_u^{(U)}\frac{\partial{\bm{e}_{i}^{(I)}}}{\partial{\bm{\Theta}_{i,l}^{(I)}}} + {\lambda_{\Theta}}\bm{\Theta}_{i,l}^{(I)}), \nonumber \\ \\ \bm{\gamma}_i^{(I)} &\leftarrow& \bm{\gamma}_i^{(I)} - \eta \cdot (-\alpha (r_{u,i}-\widehat{r_{u,i}})\bm{e}_u^{(U)} + \lambda_{\gamma}\bm{\gamma}_i^{(I)}), \end{eqnarray} \end{small} where $\eta$ is the learning rate, $\lambda_{\Theta}$ is the regularization for parameters $\bm{\Theta}^{(U)}$ and $\bm{\Theta}^{(I)}$, and $\lambda_{\gamma}$ is the regularization for parameters $\bm{\gamma}^{(U)}$ and $\bm{\gamma}^{(I)}$. In our work, we utilize the sigmod function for non-linear transformation, and we can take advantage of the properties of sigmod function for ease of derivative calculation. It is worth noting that the symbol $\bm{\Theta}$ denotes all the parameters in the fusion function, and the calculation of $\frac{\partial{\bm{e}_{i}}}{\partial{\bm{\Theta}_{i,l}}}$ will be different for different parameters in $\bm{\Theta}$. Next, we present the detailed derivation of $\frac{\partial{\bm{e}_{i}}}{\partial{\bm{\Theta}_{i,l}}}$ for personalized non-linear fusion function \begin{eqnarray}\label{eq-gradient} \small &&\frac{\partial{\bm{e}_{i}}}{\partial{\bm{\Theta}_{i,l}}}=\\ &&\begin{cases} w^{(l)}_i\sigma(Z_s)\sigma(Z_f)(1-\sigma(Z_s))(1-\sigma(Z_f))e^{(l)}_i, &\text{$\bm{\Theta} = \bm{M}$};\nonumber\\ \\ w^{(l)}_i\sigma(Z_s)\sigma(Z_f)(1-\sigma(Z_s))(1-\sigma(Z_f)), &\text{$\bm{\Theta} = \bm{b}$}; \nonumber\\ \\ \sigma(Z_s)\sigma(Z_f)(1-\sigma(Z_s)), &\text{$\bm{\Theta} = w$}, \end{cases} \end{eqnarray} where $Z_s = \sum_{l=1}^{|\mathcal{P}|} {w^{(l)}_i \sigma\big(\mathbf{M}^{(l)} \bm{e}^{(l)}_i+\bm{b}^{(l)}\big)}$ and $Z_f = \mathbf{M}^{(l)} \bm{e}^{(l)}_i+\bm{b}^{(l)}$. The derivations of $\frac{\partial{\bm{e}_{i}}}{\partial{\bm{\Theta}_{i,l}}}$ can be calculated in the above way for both users and items. We omit the derivations for linear fusion functions, since it is relatively straightforward. The whole algorithm framework is shown in Algorithm~\ref{alg_herec}. In lines 1-6, we perform HIN embedding to obtain the representations of users and items. And in lines 8-19, we adopt the SGD algorithm to optimize the parameters in the fusion function and rating function. \begin{algorithm}[htb] \caption{The overall learning algorithm of HERec.} \label{alg_herec} \begin{algorithmic}[1] \Require the rating matrix $\mathcal{R}$; the learning rate $\eta$; the adjustable parameters $\alpha, \beta$; the regularization parameter $\lambda$; the meta-path sets for users and items, $\mathcal{P}^{(U)}$ and $\mathcal{P}^{(I)}$. \Ensure the latent factors for users and items, $\mathbf{x}$ and $\mathbf{y}$; the latent factors to pair HIN embedding of users and items, $\bm{{\gamma}}^{(U)}$ and $\bm{{\gamma}}^{(I)}$; the parameters of the fusion function for users and items, $\bm{\Theta}^{(U)}$ and $\bm{\Theta}^{(I)}$ \For {$l = 1$ to $|\mathcal{P}^{(U)}|$} \State Obtain users' embeddings $\{\bm{e}^{(l)}_u\}$ based on meta-path $\mathcal{P}^{(U)}_l$ according to Algorithm \ref{alg_embedding}; \EndFor \For {$l = 1$ to $|\mathcal{P}^{(I)}|$} \State Obtain items' embeddings $\{\bm{e}^{(l)}_i\}$ based on the meta-path set $\mathcal{P}^{(I)}_l$ according to Algorithm \ref{alg_embedding}; \EndFor \State Initialize $\mathbf{x}, \mathbf{y}, \bm{{\gamma}}^{(U)}, \bm{{\gamma}}^{(I)}, \bm{\Theta}^{(U)}, \bm{\Theta}^{(I)}$ by standard normal distribution; \While {not convergence} \State Randomly select a triple $\langle u, i, r_{u,i}\rangle \in \mathcal{R}$; \State Update $\mathbf{x}_u, \mathbf{y}_i$ by typical MF; \For {$l = 1$ to $|\mathcal{P}^{(U)}|$} \State Calculate $\frac{\partial{\bm{e}_{u}^{(U)}}}{\partial{\bm{\Theta}_{u,l}^{(U)}}}$ by Eq. 14; \State Update $\bm{\Theta}_{u,l}^{(U)}$ by Eq. 10; \EndFor \State Update $\bm{\gamma}_u^{(U)}$ by Eq. 11; \For {$l = 1$ to $|\mathcal{P}^{(I)}|$} \State Calculate $\frac{\partial{\bm{e}_{i}^{(I)}}}{\partial{\bm{\Theta}_{i,l}^{(I)}}}$ by Eq. 14; \State Update $\bm{\Theta}_{i,l}^{(I)}$ by Eq. 12; \EndFor \State Update $\bm{\gamma}_i^{(I)}$ by Eq. 13; \EndWhile \State \Return {$\mathbf{x}, \mathbf{y}, \bm{{\gamma}}^{(U)}, \bm{{\gamma}}^{(I)}, \bm{\Theta}^{(U)}, \bm{\Theta}^{(I)}$}. \end{algorithmic} \end{algorithm} \subsubsection{Complexity Analysis} HERec contains two major parts: (1) HIN embedding. The complexity of deepwalk is $\mathcal{O}(d \cdot |V|)$, where $d$ is the embedding dimension and $|V|$ is the number of nodes in the network. Therefore it takes $\mathcal{O}(d \cdot |\mathcal{U}|)$ and $\mathcal{O}(d \cdot |\mathcal{I}|)$ to learn users' and items' embeddings according to a single meta-path, respectively. And the total complexity of HIN embedding is $\mathcal{O}(|\mathcal{P}|\cdot d \cdot (|\mathcal{U}|+|\mathcal{I}|))$ since the number of selected meta-paths is $|\mathcal{P}|$. It is worth noting that HIN embedding can be easily trained in parallel and we will implement it using a multi-thread mode in order to improve the efficiency of model. (2) Matrix factorization. For each triplet $\langle u, i, r_{u,i}\rangle$, updating $\bm{x}_u$, $\bm{y}_i$, $\bm{\gamma}_u^{(U)}$, $\bm{\gamma}_i^{(I)}$ takes $\mathcal{O}(D)$ time, where $D$ is the number of latent factors. And updating $\bm{\Theta}_u^{(U)}$, $\bm{\Theta}_i^{(I)}$ takes $\mathcal{O}(|\mathcal{P}| \cdot D \cdot d)$ time to learn the transformation matrices $\mathbf{M}$ for all meta-paths. In the proposed approach, $|\mathcal{P}|$ is generally small, and $d$ and $D$ are at most several hundreds, which makes the proposed method efficient in large datasets. Specially, SGD has very good practice performance, and we have found that it has a fast convergence rate on our datasets. \section{Experiments \label{sec-exp}} In this section, we will demonstrate the effectiveness of HERec by performing experiments on three real datasets compared to the state-of-the-art recommendation methods. \begin{table*}[t \centering \caption{\label{tab_Data} Statistics of the three datasets.} { \begin{tabular}{|c||c|c|c|c|c|c|} \hline {Dataset} & {Relations} & {Number} & {Number} & {Number} & {Ave. degrees} & {Ave. degrees}\\ {(Density)} & {(A-B)} & {of A} & {of B} & {of (A-B)} & {of A} & {of B}\\ \hline \hline \multirow{6}{*}{Douban Movie} & {User-Movie} & {13,367} & {12,677} & {1,068,278} & {79.9} & {84.3} \\ \cline{2-7} \multirow{6}{*}{} & {User-User} & {2,440} & {2,294} & {4,085} & {1.7} & {1.8} \\ \cline{2-7} \multirow{6}{*}{(0.63\%)} &{User-Group} & {13,337} & {2,753} & {570,047} & {42.7} & {207.1}\\ \cline{2-7} \multirow{6}{*}{} & {Movie-Director} & {10,179} & {2,449} & {11,276} & {1.1} & {4.6} \\ \cline{2-7} \multirow{6}{*}{} & {Movie-Actor} & {11,718} & {6,311} & {33,587} & {2.9} & {5.3} \\ \cline{2-7} \multirow{6}{*}{} & {Movie-Type} & {12,678} & {38} & {27,668} & {2.2} & {728.1} \\ \hline \hline \multirow{5}{*}{Douban Book} & {User-Book} & {13,024} & {22,347} & {792,026} & {60.8} & {35.4} \\ \cline{2-7} \multirow{5}{*}{} & {User-User} & {12,748} & {12,748} & {169,150} & {13.3} & {13.3} \\ \cline{2-7} \multirow{5}{*}{(0.27\%)} & {Book-Author} & {21,907} & {10,805} & {21,905} & {1.0} & {2.0} \\ \cline{2-7} \multirow{5}{*}{} & {Book-Publisher} & {21,773} & {1,815} & {21,773} & {1.0} & {11.9} \\ \cline{2-7} \multirow{5}{*}{} & {Book-Year} & {21,192} & {64} & {21,192} & {1.0} & {331.1} \\ \hline \hline \multirow{5}{*}{Yelp} & {User-Business} & {16,239} & {14,284} & {198,397} & {12.2} & {13.9} \\ \cline{2-7} \multirow{5}{*}{(0.08\%)} & {User-User} & {10,580} & {10,580} & {158,590} & {15.0} & {15.0} \\ \cline{2-7} \multirow{5}{*}{} & {User-Compliment} & {14,411} & {11} & {76,875} & {5.3} & {6988.6} \\ \cline{2-7} \multirow{5}{*}{} & {Business-City} & {14,267} & {47} & {14,267} & {1.0} & {303.6} \\ \cline{2-7} \multirow{5}{*}{} & {Business-Category} & {14,180} & {511} & {40,009} & {2.8} & {78.3} \\ \hline \end{tabular}} \end{table*} \begin{table}[t \center \caption{\label{tab_metapath} The selected meta-paths for three datasets in our work.} {\begin{tabular}{|c||c|} \hline {Dataset} & {Meta-paths}\\ \hline\hline \multirow{2}{*}{Douban Movie} & {UMU, UMDMU, UMAMU, UMTMU}\\ \multirow{2}{*}{} & {MUM, MAM, MDM, MTM} \\ \hline \multirow{2}{*}{Douban Book} & {UBU, UBABU, UBPBU, UBYBU} \\ \multirow{2}{*}{} & {BUB, BPB, BYB} \\ \hline \multirow{2}{*}{Yelp} & {UBU, UBCiBU, UBCaBU} \\ \multirow{2}{*}{} & {BUB, BCiB, BCaB} \\ \hline \end{tabular}} \end{table} \begin{table*}[t] \centering \caption{\label{tab_Effectiveness} Results of effectiveness experiments on three datasets. A smaller MAE or RMSE value indicates a better performance. For ease of reading the results, we also report the improvement w.r.t. the PMF model for each other method. A larger improvement ratio indicates a better performance. } { \begin{tabular}{|c|c|c||c|c|c|c|c|c||c|c||c|} \hline {Dataset} & {Training} & {Metrics} & {PMF} &{SoMF} & {FM$_{HIN}$} & {HeteMF}& {SemRec} & {DSR} & {HERec$_{dw}$} & {HERec$_{mp}$}&{HERec}\\ \hline \hline \multirow{16}{*}{Douban}& \multirow{4}{*}{80\%} & {MAE} & {0.5741} & {0.5817} & {0.5696} & {0.5750} & {0.5695} & {0.5681} & {0.5703} & {\textbf{0.5515}} &{0.5519} \\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {-1.32\%} & {+0.78\%} & {-0.16\%} & {+0.80\%} & {+1.04\%} & {+0.66\%} &{+3.93\%} & {+3.86\%} \\ \cline{3-12} \multirow{16}{*}{Movie} &\multirow{4}{*}{}& {RMSE} & {0.7641} & {0.7680} & {0.7248} & {0.7556}& {0.7399} & {0.7225} & {0.7446} & {0.7121} & {\textbf{0.7053}} \\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {-0.07\%} & {+5.55\%} & {+1.53\%} & {+3.58\%} & {+5.85\%} & {+2.97\%} &{+7.20\%} & {+8.09\%} \\ \cline{2-12} \multirow{16}{*}{} &\multirow{4}{*}{}\multirow{4}{*}{60\%} & {MAE} & {0.5867} & {0.5991} & {0.5769} & {0.5894} & {0.5738} & {0.5831} & {0.5838} & {0.5611} &{\textbf{0.5587}} \\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {-2.11\%} & {+1.67\%} & {-0.46\%} & {+2.19\%} & {+0.61\%} & {+0.49\%} &{+4.36\%} & {+4.77\%} \\ \cline{3-12} \multirow{16}{*}{} &\multirow{4}{*}{}& {RMSE} & {0.7891} & {0.7950} & {0.7842} & {0.7785} & {0.7551} & {0.7408} & {0.7670} & {0.7264} &{\textbf{0.7148}} \\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {-0.75\%} & {+0.62\%} & {+1.34\%} & {+4.30\%} & {+6.12\%} & {+2.80\%} &{+7.94\%} & {+9.41\%} \\ \cline{2-12} \multirow{16}{*}{} &\multirow{4}{*}{}\multirow{4}{*}{40\%} & {MAE} & {0.6078} & {0.6328} & {0.5871} & {0.6165} & {0.5945} & {0.6170} & {0.6073} & {0.5747} &{\textbf{0.5699}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {-4.11\%} & {+3.40\%} & {-1.43\%} & {+2.18\%} & {-1.51\%} & {+0.08\%} &{+5.44\%} & {+6.23\%} \\ \cline{3-12} \multirow{16}{*}{} &\multirow{4}{*}{}& {RMSE} & {0.8321} & {0.8479} & {0.7563} & {0.8221} & {0.7836} & {0.7850} & {0.8057} & {0.7429} &{\textbf{0.7315}} \\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {-1.89\%} & {+9.10\%} & {+1.20\%} & {+5.82\%} & {+5.66\%} & {+3.17\%} &{+10.71\%} & {+12.09\%} \\ \cline{2-12} \multirow{16}{*}{} &\multirow{4}{*}{}\multirow{4}{*}{20\%} & {MAE}& {0.7247}& {0.6979} & {0.6080} & {0.6896} & {0.6392} & {0.6584} & {0.6699} & {0.6063} &{\textbf{0.5900}} \\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+3.69\%} & {+16.10\%} & {+4.84\%} & {+11.79\%} & {+9.14\%} & {+7.56\%} &{+16.33\%} & {+18.59\%} \\ \cline{3-12} \multirow{16}{*}{} &\multirow{4}{*}{}& {RMSE} & {0.9440} & {0.9852} & {0.7878} & {0.9357} & {0.8599} & {0.8345} & {0.9076} & {0.7877} &{\textbf{0.7660}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {-4.36\%} & {+16.55\%} & {+0.88\%} & {+8.91\%} & {+11.60\%} & {+3.86\%} &{+16.56\%}& {+18.86\%} \\ \hline \hline \multirow{16}{*}{Douban}& \multirow{4}{*}{80\%} & {MAE} & {0.5774} & {0.5756} & {0.5716} & {0.5740} & {0.5675} & {0.5740} & {0.5875} & {0.5591} &{\textbf{0.5502}} \\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+0.31\%} & {+1.00\%} & {+0.59\%} & {+1.71\%} & {+0.59\%} & {-1.75\%} &{+3.17\%} & {+4.71\%} \\ \cline{3-12} \multirow{16}{*}{Book} &\multirow{4}{*}{}& {RMSE} & {0.7414} & {0.7302} & {0.7199} & {0.7360} & {0.7283} & {0.7206} & {0.7450} & {0.7081} &{\textbf{0.6811}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+1.55\%} & {+2.94\%} & {+0.77\%} & {+1.81\%} & {+2.84\%} & {-0.44\%} &{+4.53\%} & {+8.17\%} \\ \cline{2-12} \multirow{16}{*}{} &\multirow{4}{*}{}\multirow{4}{*}{60\%} & {MAE} & {0.6065} & {0.5903} &{0.5812} & {0.5823} & {0.5833} & {0.6020} & {0.6203} & {0.5666} &{\textbf{0.5600}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+2.67\%} & {+4.17\%} & {+3.99\%} & {+3.83\%} & {+0.74\%} & {-2.28\%} &{+6.58\%} & {+7.67\%} \\ \cline{3-12} \multirow{16}{*}{} &\multirow{4}{*}{}& {RMSE} & {0.7908} & {0.7518} &{0.7319} & {0.7466} & {0.7505} & {0.7552} & {0.7905} & {0.7318} &{\textbf{0.7123}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+4.93\%} & {+7.45\%} & {+5.59\%} & {+5.10\%} & {+4.50\%} & {+0.04\%} &{+7.46\%} & {+9.93\%} \\ \cline{2-12} \multirow{16}{*}{} &\multirow{4}{*}{}\multirow{4}{*}{40\%} & {MAE} & {0.6800} & {0.6161} &{0.6028} & {0.5982} & {0.6025} & {0.6271} & {0.6976} & {0.5954} &{\textbf{0.5774}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+9.40\%} & {+11.35\%} & {+12.03\%} & {+11.40\%} & {+7.78\%} & {-2.59\%} &{+12.44\%}& {+15.09\%} \\ \cline{3-12} \multirow{16}{*}{} &\multirow{4}{*}{}& {RMSE} & {0.9203} & {0.7936} &{0.7617} & {0.7779} & {0.7751} & {0.7730} & {0.9022} & {0.7703} &{\textbf{0.7400}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+13.77\%} & {+17.23\%} & {+15.47\%} & {+15.78\%} & {+16.01\%} & {+1.97\%} &{+16.30\%} & {+19.59\%} \\ \cline{2-12} \multirow{16}{*}{} &\multirow{4}{*}{}\multirow{4}{*}{20\%} & {MAE}& {1.0344} & {0.6327} &{0.6396} & {0.6311} & {0.6481} & {0.6300} & {1.0166} & {0.6785} &{\textbf{0.6450}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+38.83\%} & {+38.17\%} & {+38.99\%} & {+37.35\%} & {+39.10\%} & {+1.72\%} &{+34.41\%} & {+37.65\%} \\ \cline{3-12} \multirow{16}{*}{} &\multirow{4}{*}{}& {RMSE} & {1.4414} & {0.8236} &{0.8188} & {0.8304} & {0.8350} & {0.8200} & {1.3205} & {0.8869} &{\textbf{0.8581}} \\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+42.86\%} & {+43.19\%} & {+42.39\%} & {+42.07\%} & {+43.11\%} & {+8.39\%} &{+38.47\%} & {+40.47\%} \\ \hline \hline \multirow{16}{*}{Yelp}& \multirow{4}{*}{90\%} & {MAE} & {1.0412} & {1.0095} &{0.9013} & {0.9487} & {0.9043} & {0.9054} & {1.0388} & {0.8822} &{\textbf{0.8395}} \\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+3.04\%} & {+13.44\%} & {+8.88\%} & {+13.15\%} & {+13.04\%} & {+0.23\%} &{+15.27\%} & {+19.37\%} \\ \cline{3-12} \multirow{16}{*}{} &\multirow{4}{*}{}& {RMSE} & {1.4268} & {1.3392} &{1.1417} & {1.2549} & {1.1637} & {1.1186} & {1.3581} & {1.1309} &{\textbf{1.0907}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+6.14\%} & {+19.98\%} & {+12.05\%} & {+18.44\%} & {+21.60\%} & {+4.81\%} &{+20.74\%} & {+23.56\%} \\ \cline{2-12} \multirow{16}{*}{} &\multirow{4}{*}{}\multirow{4}{*}{80\%} & {MAE} & {1.0791} & {1.0373} &{0.9038} & {0.9654} & {0.9176} & {0.9098} & {1.0750} & {0.8953} &{\textbf{0.8475}} \\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+3.87\%} & {+16.25\%} & {+10.54\%} & {+14.97\%} & {+15.69\%} & {+0.38\%} &{+17.03\%} & {+21.46\%} \\ \cline{3-12} \multirow{16}{*}{} &\multirow{4}{*}{}& {RMSE} & {1.4816} & {1.3782} &{1.1497} & {1.2799} & {1.1771} & {1.1208} & {1.4075} & {1.1516} &{\textbf{1.1117}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+6.98\%} & {+22.40\%} & {+13.61\%} & {+20.55\%} & {+24.35\%} & {+5.00\%} &{+22.27\%} & {+24.97\%} \\ \cline{2-12} \multirow{16}{*}{} &\multirow{4}{*}{}\multirow{4}{*}{70\%} & {MAE} & {1.1170} & {1.0694} &{0.9108} & {0.9975} & {0.9407} & {0.9429} & {1.1196} & {0.9043} &{\textbf{0.8580}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+4.26\%} & {+18.46\%} & {+10.70\%} & {+15.78\%} & {+15.59\%} & {-0.23\%} &{+19.04\%} & {+23.19\%} \\ \cline{3-12} \multirow{16}{*}{} &\multirow{4}{*}{}& {RMSE} & {1.5387} & {1.4201} &{1.1651} & {1.3229} & {1.2108} & {1.1582} & {1.4632} & {1.1639} &{\textbf{1.1256}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+7.71\%} & {+24.28\%} & {+14.02\%} & {+21.31\%} & {+24.73\%} & {+4.91\%} &{+24.36\%} & {+26.85\%} \\ \cline{2-12} \multirow{16}{*}{} &\multirow{4}{*}{}\multirow{4}{*}{60\%} & {MAE}& {1.1778} & {1.1135} &{0.9435} & {1.0368} & {0.9637} & {1.0043} & {1.1691} & {0.9257} &{\textbf{0.8759}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+5.46\%} & {+19.89\%} & {+11.97\%} & {+18.18\%} & {+14.73\%} & {+0.74\%} &{+21.40\%} & {+25.63\%} \\ \cline{3-12} \multirow{16}{*}{} &\multirow{4}{*}{}& {RMSE} & {1.6167} & {1.4748} &{1.2039} & {1.3713} & {1.2380} & {1.2257} & {1.5182} & {1.1887} &{\textbf{1.1488}}\\ \multirow{16}{*}{} &\multirow{4}{*}{}& {Improve} & {} & {+8.78\%} & {+25.53\%} & {+15.18\%} & {+23.42\%} & {+24.19\%} & {+6.09\%} &{+26.47\%} & {+28.94\%} \\ \hline \end{tabular}} \end{table*} \subsection{Evaluation Datasets} We adopt three widely used datasets from different domains, consisting of Douban Movie dataset \footnote{} from movie domain, Douban Book dataset \footnote{} from book domain, and Yelp dataset \footnote{} from business domain. Douban Movie dataset includes 13,367 users and 12,677 movies with 1,068,278 movie ratings ranging from 1 to 5. Moreover, the dataset also includes social relations and attribute information of users and movies. As for the Douban Book dataset, it includes 13,024 users and 22,347 books with 792,026 ratings ranging from 1 to 5. Yelp dataset records user ratings on local businesses and contains social relations and attribute information of businesses, which includes 16,239 users and 14,282 local businesses with 198,397 ratings raging from 1 to 5. The detailed descriptions of the three datasets are shown in Table~\ref{tab_Data}, and the network schemas of the three datasets are plotted in Fig.~\ref{fig_schema}. Besides the domain variation, these three datasets also have different rating sparsity degrees: the Yelp dataset is very sparse, while the Douban Movie dataset is much denser. \subsection{Metrics} We use the widely used mean absolute error (MAE) and root mean square error (RMSE) to measure the quality of recommendation of different models. The metrics MAE and RMSE are defined as follows: \begin{equation} MAE = \frac{1}{|\mathcal{D}_{test}|}\sum_{(i, j) \in \mathcal{D}_{test}}{|r_{i,j} - \widehat{r_{i,j}}|}, \end{equation} \begin{equation} RMSE = \sqrt{\frac{1}{|\mathcal{D}_{test}|}\sum_{(i, j) \in \mathcal{D}_{test}}{(r_{i,j} - \widehat{r_{i,j}})^2}}, \end{equation} where $r_{i,j}$ is the actual rating that user $i$ assigns to item $j$, $\widehat{r_{i,j}}$ is the predicted rating from a model, and $\mathcal{D}_{test}$ denotes the test set of rating records. From the definition, we can see that a smaller MAE or RMSE value indicates a better performance. \subsection{Methods to Compare} We consider the following methods to compare: \textbullet\ \textbf{PMF}~\cite{mnih2008probabilistic}: It is the classic probabilistic matrix factorization model by explicitly factorizing the rating matrix into two low-dimensional matrices. \textbullet\ \textbf{SoMF}~\cite{ma2011recommender}: It incorporates social network information into the recommendation model. The social relations are characterized by a social regularization term, and integrated into the basic matrix factorization model. \textbullet\ \textbf{FM$_{HIN}$}: It is the context-aware factorization machine~\cite{rendle2012factorization}, which is able to utilize various kinds of auxiliary information. In our experiments, we extract heterogeneous information as context features and incorporate them into the factorization machine for rating prediction. We implement the model with the code from the authors~\footnote{}. \textbullet\ \textbf{HeteMF}~\cite{yu2013collaborative}: It is a MF based recommendation method, which utilizes entity similarities calculated over HINs using meta-path based algorithms. \textbullet\ \textbf{SemRec}~\cite{shi2015semantic}: It is a collaborative filtering method on weighted heterogeneous information network, which is constructed by connecting users and items with the same ratings. It flexibly integrates heterogeneous information for recommendation by weighted meta-paths and a weight ensemble method. We implement the model with the code from the authors~\footnote{}. \textbullet\ \textbf{DSR}~\cite{zheng2017recommendation}: It is a MF based recommendation method with dual similarity regularization, which imposes the constraint on users and items with high and low similarities simultaneously. \textbullet \textbf{HERec$_{dw}$}: It is a variant of HERec, which adopts the homogeneous network embedding method deepwalk~\cite{perozzi2014deepwalk} and ignores the heterogeneity of nodes in HINs. We use the code of deepwalk provided by the authors~\footnote{}. \textbullet \textbf{HERec$_{mp}$}: It can be view as a variant of HERec, which combines the heterogeneous network embedding of metapath2vec++~\cite{dong2017metapath2vec} and still preserves the superiority of HINs. We use the code of metapath2vec++ provided by the authors~\footnote{}. \textbullet\ \textbf{HERec}: It is the proposed recommendation model based on heterogeneous information network embedding. In the effectiveness experiments, we utilize the personalized non-linear fusion function, which is formulated as Eq.~\ref{eq-pnlf}. And the performance of different fusion function will be reported in the later section. The selected baselines have a comprehensive coverage of existing rating prediction methods. PMF and SoMF are classic MF based rating prediction methods; FM$_{HIN}$, HeteMF, SemRec and DSR are HIN based recommendation methods using meta-path based similarities, which represent state-of-the-art HIN based methods. The proposed approach contains an HIN embedding component, which can be replaced by other network embedding methods. Hence, two variants HERec$_{dw}$ and HERec$_{mp}$ are adopted as baselines. Among these methods, HIN based methods need to specify the used meta-paths. We present the used meta-paths in Table~\ref{tab_metapath}. Following~\cite{sun2011pathsim}, we only select short meta-paths of at most four steps, since long meta-paths are likely to introduce noisy semantics. We set parameters for HERec as follows: the embedding dimension number $d = 64$ and the tuning coefficients $\alpha = \beta = 1.0$. Following~\cite{perozzi2014deepwalk}, the path length for random walk is set to 40. For HERec$_{dw}$ and HERec$_{mp}$, we set the embedding dimension number $d = 64$ for fairness. Following~\cite{shi2016integrating,zheng2017recommendation}, we set the number of latent factors for all the MF based methods to 10. Other baseline parameters either adopt the original optimal setting or are optimized by using a validation set of 10\% training data. \subsection{Effectiveness Experiments} For each dataset, we split the entire rating records into a training set and a test set. For Double Movie and Book datasets, we set four training ratios as in $\{80\%, 60\%, 40\%, 20\%\}$; while for Yelp dataset, following~\cite{shi2015semantic}, we set four larger training ratios as in $\{90\%, 80\%, 70\%, 60\%\}$, since it is too sparse. For each ratio, we randomly generate ten evaluation sets. We average the results as the final performance. The results are shown in Table \ref{tab_Effectiveness}. The major findings from the experimental results are summarized as follows: (1) Among these baselines, HIN based methods (HeteMF, SemRec, FM$_{HIN}$ and DSR) perform better than traditional MF based methods (PMF and SoMF), which indicates the usefulness of the heterogeneous information. It is worthwhile to note that the FM$_{HIN}$ model works very well among the three HIN based baselines. An intuitive explanation is that in our datasets, most of the original features are the attribute information of users or items, which are likely to contain useful evidence to improve the recommendation performance. (2) The proposed HERec method (corresponding to the last column) is consistently better than the baselines, ranging from PMF to DSR. Compared with other HIN based methods, HERec adopts a more principled way to leverage HINs for improving recommender systems, which provides better information extraction (a new HIN embedding model) and utilization (an extended MF model). Moreover, the superiority of the proposed HERec becomes more significant with less training data. In particular, the improvement ratio of HERec over the PMF is up to 40\% with 20\% training data on the Douban Book dataset, which indicates a significant performance boost. As mentioned previously, the Yelp dataset is very sparse. In this case, even with 60\% training data, the HERec model improves over PMF by about 26\%. As a comparison, with the same training ratio (\emph{i.e.,}\xspace 60\%), the improvement ratios over PMF are about 29\% in terms of RMSE scores. These results indicate the effectiveness of the proposed approach, especially in a sparse dataset. (3) Considering the two HERec variants HERec$_{dw}$ and HERec$_{mp}$, it is easy to see HERec$_{dw}$ performs much worse than HERec$_{mp}$. The major difference lies in the network embedding component. HERec$_{mp}$ adopts the recently proposed HIN embedding method \emph{metapath2vec}~\cite{dong2017metapath2vec}, while HERec$_{dw}$ ignores data heterogeneity and casts HINs as homogeneous networks for learning. It indicates HIN embedding methods are important to HIN based recommendation. The major difference between the proposed embedding method and metapath2vec (adopted by HERec$_{mp}$) lies in that we try to transform heterogeneous network information into homogeneous neighborhood, while metapath2vec tries to map all types of entities into the same representation space using heterogeneous sequences. Indeed, metapath2vec is a general-purpose HIN embedding method, while the proposed embedding method aims to improve the recommendation performance. Our focus is to learn the embeddings for users and items, while objects of other types are only used as the bridge to construct the homogeneous neighborhood. Based on the results (HERec $>$ HERec$_{mp}$), we argue that it is more effective to perform task-specific HIN embedding for improving the recommendation performance. \subsection{Detailed Analysis of The Proposed Approach} In this part, we perform a series of detailed analysis for the proposed approach. \begin{figure*} \centering \subfigure[Douban Movie]{ \begin{minipage}[b]{0.3\textwidth} \includegraphics[width=1\textwidth]{image/fusion_dm_rmse.pdf} \\ \includegraphics[width=1\textwidth]{image/fusion_dm_mae.pdf} \end{minipage} } \subfigure[Douban Book]{ \begin{minipage}[b]{0.3\textwidth} \includegraphics[width=1\textwidth]{image/fusion_db_rmse.pdf} \\ \includegraphics[width=1\textwidth]{image/fusion_db_mae.pdf} \end{minipage} } \subfigure[Yelp]{ \begin{minipage}[b]{0.3\textwidth} \includegraphics[width=1\textwidth]{image/fusion_yelp_rmse.pdf} \\ \includegraphics[width=1\textwidth]{image/fusion_yelp_mae.pdf} \end{minipage} } \caption{\label{fig_fusion}Performance comparison of different fusion functions on three datasets.} \end{figure*} \begin{figure*} \centering \subfigure[Douban Movie]{ \begin{minipage}[b]{0.3\textwidth} \includegraphics[width=1\textwidth]{image/cold_start_rmse_dm.pdf} \\ \includegraphics[width=1\textwidth]{image/cold_start_mae_dm.pdf} \end{minipage} } \subfigure[Douban Book]{ \begin{minipage}[b]{0.3\textwidth} \includegraphics[width=1\textwidth]{image/cold_start_rmse_db.pdf} \\ \includegraphics[width=1\textwidth]{image/cold_start_mae_db.pdf} \end{minipage} } \subfigure[Yelp]{ \begin{minipage}[b]{0.3\textwidth} \includegraphics[width=1\textwidth]{image/cold_start_rmse_yelp.pdf} \\ \includegraphics[width=1\textwidth]{image/cold_start_mae_yelp.pdf} \end{minipage} } \caption{\label{fig_cs}Performance comparison of different methods for cold-start prediction on three datasets. $y$-axis denotes the improvement ratio over PMF.} \end{figure*} \subsubsection{Selection of Different Fusion Functions} HERec requires a principled fusion way to transform node embeddings into a more suitable form that is useful to enhance recommendation performance. Therefore, we will discuss the impact of different fusion functions on the recommendation performance. For convenience, we call the HERec variant with the simple linear fusion function (Eq.~\ref{eq-slf}) as HERec$_{sl}$, the variant with personalized linear fusion function (Eq.~\ref{eq-plf}) as HERec$_{pl}$, and the variant with personalized non-linear fusion function (Eq.~\ref{eq-pnlf}) as HERec$_{pnl}$. We present the performance comparison of the three variants of HERec on the three datasets in Fig.~\ref{fig_fusion}. As shown in Fig.~\ref{fig_fusion}, we can find that overall performance ranking is as follows: HERec$_{pnl}$ $>$ HERec$_{pl}$ $>$ HERec$_{sl}$. Among the three variants, the simple linear fusion function performs the worst, as it ignores the personalization and non-linearity. Indeed, users are likely to have varying preferences over meta-paths~\cite{shi2015semantic}, which should be considered in meta-path based methods. The personalization factor improves the performance significantly. As a comparison, the performance improvement of HERec$_{pnl}$ over HERec$_{pl}$ is relatively small. A possible reason is that the incorporation of personalized combination parameters increases the capability of linear models. Nonetheless, HERec$_{pnl}$ still performs the best by considering personalization and non-linear transformation. In a complicated recommender setting, HIN embeddings may not be directly applicable in recommender systems, where a non-linear mapping function is preferred. Since HERec$_{pnl}$ is the best variant of the proposed model, in what follows, HERec will use the personalized non-linear fusion function, \emph{i.e.,}\xspace HERec$_{pnl}$ by default. \subsubsection{Cold-start Prediction} HINs are particularly useful to improve cold-start prediction, where there are fewer rating records but heterogeneous context information is available. We consider studying the performance \emph{w.r.t.} different cold-start degrees, \emph{i.e.,}\xspace the rating sparsity. To test it, we first categorize ``cold" users into three groups according to the numbers of their rating records, \emph{i.e.,}\xspace $(0, 5]$, $(5, 15]$ and $(15, 30]$. It is easy to see that the case for the first group is the most difficult, since users from this group have fewest rating records. Here, we only select the baselines which use HIN based information for recommendation, including SoMF, HeteMF, SemRec, DSR and FM$_{HIN}$. We present the performance comparison of different methods in Fig.~\ref{fig_cs}. For convenience, we report the improvement ratios \emph{w.r.t.} PMF. Overall, all the comparison methods are better than PMF (\emph{i.e.,}\xspace a positive $y$-axis value). The proposed method performs the best among all the methods, and the improvement over PMF becomes more significant for users with fewer rating records. The results indicate that HIN based information is effective to improve the recommendation performance, and the proposed HERec method can effectively utilize HIN information in a more principled way. \begin{figure*}[htbp] \centering \subfigure[Douban Movie]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1\textwidth]{image/metapath_dm.pdf} \end{minipage} } \subfigure[Douban Book]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1.05\textwidth]{image/metapath_db.pdf} \end{minipage} } \subfigure[Yelp]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1\textwidth]{image/metapath_yelp.pdf} \end{minipage} } \caption{\label{fig_metapath}Performance change of HERec when gradually incorporating meta-paths.} \end{figure*} \subsubsection{Impact of Different Meta-Paths} As shown in Table~\ref{tab_metapath}, the proposed approach uses a selected set of meta-paths. To further analyze the impact of different meta-paths, we gradually incorporate these meta-paths into the proposed approach and check the performance change. In Fig. \ref{fig_metapath}, we can observe that generally the performance improves (\emph{i.e.,}\xspace becoming smaller) with the incorporation of more meta-paths. Both meta-paths starting with the user type and item type are useful to improve the performance. However, it does not always yield the improvement with more meta-paths, and the performance slightly fluctuates. The reason is that some meta-paths may contain noisy or conflict information with existing ones. Another useful observation is that the performance quickly achieves a relatively good performance with the incorporation of only a few meta-paths. This confirms previous finding~\cite{shi2015semantic}: a small number of high-quality meta-paths are able to lead to large performance improvement. Hence, as mentioned before, we can effectively control the model complexity by selecting just a few meta-paths. \begin{figure*}[t \centering \subfigure[Douban Movie]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1\textwidth]{image/factor_dm.pdf} \end{minipage} } \subfigure[Douban Book]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1\textwidth]{image/factor_db.pdf} \end{minipage} } \subfigure[Yelp]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1\textwidth]{image/factor_yelp.pdf} \end{minipage} } \caption{\label{fig_factor}Performance with respect to the dimension of latent factors on three datasets.} \end{figure*} \begin{figure*}[t \centering \subfigure[Douban Movie]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1\textwidth]{image/reg_rmse_dm.pdf} \end{minipage} } \subfigure[Douban Book]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1\textwidth]{image/reg_rmse_db.pdf} \end{minipage} } \subfigure[Yelp]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1\textwidth]{image/reg_rmse_yelp.pdf} \end{minipage} } \caption{\label{fig_reg}Varying parameters $\alpha$ and $\beta$ on the three datasets.} \end{figure*} \begin{figure*}[t \centering \subfigure[Douban Movie]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1.06\textwidth]{image/iter_dm.pdf} \end{minipage} } \subfigure[Douban Book]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1\textwidth]{image/iter_db.pdf} \end{minipage} } \subfigure[Yelp]{ \begin{minipage}[t]{0.3\textwidth} \centering \includegraphics[width=1\textwidth]{image/iter_yelp.pdf} \end{minipage} } \caption{\label{fig_iter}Performance with respect to the number of iterations on three datasets.} \end{figure*} \subsubsection{Parameter Tuning} For matrix factorization based methods, an important parameter to tune is the number of latent factors. The proposed model also involves such a parameter. We vary it from 5 to 40 with a step of 5, and examine how the performance changes \emph{w.r.t} the number of latent factors. We present the tuning results in Fig.~\ref{fig_factor}. As we can see, using 10 latent factors yields the best performance, indicating that the number of latent factors should be set to a small number. Next, we fix the number of latent factors as ten, and tune two other parameters $\alpha$ and $\beta$ (Eq.~\ref{eq-predictor}), which are used to integrate different terms as weights. Now we examine how they influence the model performance. For both parameters, we vary them in the set of $\{0.1, 0.5, 1, 2\}$. As shown in Fig. \ref{fig_reg}, the optimal performance is obtained near $(1,1)$, \emph{i.e.,}\xspace both $\alpha$ and $\beta$ are around 1. The results show that the HIN embeddings from both the user and item sides are important to improve the prediction performance. Overall, the change trend is smooth, indicating that the proposed model is not very sensitive to the two parameters. Finally, we study the performance change \emph{w.r.t.} the number of iterations. As shown in Fig.~\ref{fig_iter}, we can see that the proposed model has a fast convergence rate, and about 40 to 60 iterations are required for dense datasets (\emph{i.e.,}\xspace Douban Movie and Book), while about 20 iterations are required for sparse datasets (\emph{i.e.,}\xspace Yelp). \section{Conclusion \label{sec-con}} In this paper, we proposed a novel heterogeneous information network embedding based approach (\emph{i.e.,}\xspace HERec) to effectively utilizing auxiliary information in HINs for recommendation. We designed a new random walk strategy based on meta-paths to derive more meaningful node sequences for network embedding. Since embeddings based on different meta-paths contain different semantic, the learned embeddings were further integrated into an extended matrix factorization model using a set of fusion functions. Finally, the extended matrix factorization model together with fusion functions were jointly optimized for the rating prediction task. HERec aimed to learn useful information representations from HINs guided by the specific recommendation task, which distinguished the proposed approach from existing HIN based recommendation methods. Extensive experiments on three real datasets demonstrated the effectiveness of HERec. We also verified the ability of HERec to alleviate cold-start problem and examine the impact of meta-paths on performance. As future work, we will investigate into how to apply deep learning methods (\emph{e.g.,}\xspace convolutional neural networks, auto encoder) to better fuse the embeddings of multiple meta-paths. In addition, we only use the meta-paths which have the same starting and ending types to effectively extract network structure features in this work. Therefore, it is interesting and natural to extend the proposed model to learn the embeddings of any nodes with arbitrary meta-paths. As a major issue of recommender systems, we will also consider how to enhance the explainablity of the recommendation method based on the semantics of meta-paths. \bibliographystyle{IEEEtran}

\section{Introduction} \label{} Solar jets are small scale eruptions observed at different heights of the solar atmosphere. They are observed in the different parts of the solar surface such as: in coronal hole regions \citep{Madjarska13,Zhelyazkov17}, active regions \citep{Schmieder13,Chandra15,Sterling17}, quiet regions (for example \citealp{Panesar16}). Since their first observations in X-rays, a very remarkable progress have been made to understand the physics of the solar jets including the high resolution data of \emph{Solar Dynamics Observatory\/} (\emph{SDO}). The solar jets can be divided into two categories, that is, ``standard'' and ``blow-out'' jets (\citealp{Moore13,Moore15,Raouafi16,Chandra17}). Initially this classification was based on their morphology. In the case of standard jets their spires are thin and narrow during their entire lifetime and their bases remained relatively dim, except for the commonly observed compact jet bright point (JBP) on one side of the jet's base. In contrary to this the blow-out jets can be defined by broad spires. A blow-out jet initiated as standard jet starts as a emerging bipole reconnecting with ambient open field. But sometime during this reconnection process, the emerging bipole is triggered unstable and erupts outward and it becomes the ``blow-out'' jet. The ``blow-out'' jet is associated with a small flux rope eruption. Sometime this erupting flux rope can become coronal mass ejection (CME) (for example see \citealp{Liu15,Chandra17}). The interpretation of ``blow-out'' jets was first proposed by \cite{Moore10,Moore13}. Later on, these were reported and confirmed in several observations \citep{Sterling16,Panesar16,Sterling17}. The accepted mechanism for the generation of solar jets is the magnetic reconnection. Conditions for the magnetic reconnection can be magnetic flux emergence \citep{Heyvaerts77,Shibata92,Guo13}, also flux cancellation \citep{Priest94,Longcope98,Innes10,Adams14}, or can be both \citep{Young14,Chandra15}. Several numerical simulations have been done in this direction. Most of these includes the magnetic flux emergence. Moreover, in case of Pariat and co-workers \citep{Pariat17} one makes the condition for magnetic reconnection imposing horizontal photospheric twisting motions. Very recently it is proposed that the jets are small scale phenomena and this physical mechanism is similar to the large scale eruptions \citep{Wyper17}. Therefore to understand these small scale phenomena is very crucial to explain the large scale solar eruptions. Apart from this, the jets in the solar atmosphere can support the propagation of magnetohydrodynamic (MHD) waves, which becoming unstable can be one of the possible ways for the coronal heating. Several models have been proposed for the heating of solar corona. It is believed that there are two possible candidates for the coronal heating namely magnetic reconnection and the dissipation of energy by MHD waves. However, it is still not clear which one is the favorable mechanism. Now people thinks it may be the combination of these two mechanisms. It is known since long time that when the fluids of different speeds flow in the same direction, there will be a strong velocity shear near the interface region of two different speeds. This velocity shear produced the vortex sheet at the boundaries. As a result of this Kelvin--Helmholtz (KH) instability occurs. The KHI MHD modeling have been done in several studies (for reviews see \cite{Zhelyazkov15, Nakariakov16}). In addition to MHD modelings, there are few cases where the KHI vortex were observed \citep{Foullon11, Zhelyazkov15, Kuridze16, Zhelyazkov17}. The vortex sheet, which is due to KHI can become unstable like spiral perturbations. The vortex sheet causes the conversion of the directed flow energy into the turbulent energy (e.g., \cite{Maslowe85}). And this energy can be one of the possible source of coronal heating. Therefore, to study the KH instability in the solar features is very crucial and useful to understand the solar coronal heating problem. Keeping this in mind, we aim here the modeling of KH instability of one solar jet observed by the Atmospheric Imaging Assembly (AIA) on board \emph{SDO\/} on 2014 April 16. The organization of the article is as follows: we present the observations and the selection criteria in Section 2. The jet and magnetic field geometry, as well as the governing MHD equations for the modeling are given in Section 3. Section 4 deals with the derivation of wave dispersion relations obtained from our model. Finally, in Sections 5 and 6, we discuss the numerical results and conclude our study. \section{Observations} \label{} During 2014 April 15--16, the active region NOAA AR 12035 produced several jets. These jets are well observed by the Atmospheric Imaging Assembly (AIA) on board \emph{SDO\/} satellite. The AIA observe the Sun in different UV and EUV wavelengths. The pixel size and temporal evolution of the AIA data is 0.6~arcsec and 12~s respectively. The dynamics and the kinematics of these jets were described by \cite{Joshi17}. They found two jet sites close to each other and noticed the slippage of jets from one site to other site. Using the Heliospheric Magnetic Imager (HMI) photospheric magnetic field data, they found that the jets site were associated with the flux emergence as well the flux cancellation and presence of several null points (see their Figure 10). In this article, we have selected one jet, notably J\ensuremath{'}${_6}$ from \citealp{Joshi17} Table 1 and model it for the Kelvin--Helmholtz instability. Our selection criterion is based on its high speed, large size, longer life-time and visibility in all EUV channels relative to other jets. In addition to this criterion, this jet indicates eruption of blobs type structures at its boundary. We believe that these blobs like structures could be the result of KHI. Therefore, the chosen jet is a good candidate for KHI modeling. The jet starts ${\approx}$14:47 UT and reached its maximum length {$\approx$}14:57 UT and finally finish around 15:00 UT. The jet's maximum speed was ${\approx}343$~km\,s$^{-1}$ at AIA 211~\AA\ and the average speed of all the EUV channels was ${\approx}332$~km\,s$^{-1}$. \begin{figure}[!h] \centering \includegraphics[width=1.0\textwidth,clip=]{na-izh-fg1.eps} \caption{Evolution of the jet in AIA 171 and 193~\AA. The field-of-view is $130\times180$~arcsec.} \label{} \end{figure} \begin{figure}[!h] \centering \includegraphics[width=0.45\textwidth,clip=]{na-izh-fg2.eps} \vspace*{-1cm} \caption{AIA 171~\AA\ images of the jet. The white arrows indicate different visible structures of blobs, which could be due to KHI. The field-of-view of the images is $40\times80$~arcsec.} \label{} \end{figure} For a better image quality for \emph{SDO}/AIA data, we have used the Multi-Gaussian Normalization (MGN) techniques \citep{Morgan14} to enhanced the image quality. The evolution of the jet in AIA/EUV 171 and 193~\AA\ is displayed in Figure~\ref{}. During jet's evolution, we have noticed the propagation of plasma blobs along the jet boundary in AIA 171~\AA. To investigate these structures in more detail, we have enlarged the jet image in smaller field-of-view and this is shown in Figure~\ref{}. An inspection of these images evidenced the movement of the plasma blobs at the jet's boundary. These structures are shown by the white arrows. These blob structures evidenced the KHI. \begin{figure}[!h] \centering \includegraphics[width=1.0\textwidth,clip=]{na-izh-fg3.eps} \vspace*{-1cm} \caption{Maps of temperature and emission measure inside and outside the jet derived by the \emph{SDO\/} data.} \label{} \end{figure} To model the jet for the KHI, we have computed the temperature and the density inside and outside of the jet. For this calculation, we have used the technique established by \cite{Aschwanden13}. This techniques uses the observations of six AIA EUV channels, i.e., 94, 131, 171, 193, 211 and 335 \AA. The estimated values for the jet are given in Table 1. The subscripts `i' and `e' stamp for \emph{interior\/} and \emph{exterior}, respectively. An example of temperature and emission measure maps are presented in Figure~\ref{} during the peak phase of the jet. \begin{figure}[!t] \centerline{\includegraphics[width=0.65\textwidth,clip=]{na-izh-fg4.eps}} \caption{Magnetic fields and flow velocity configurations in axially moving solar jet flux tubes.} \label{} \end{figure} \begin{table} \caption{Different jet's physical parameters derived from \emph{SDO}/AIA data.} \label{parameter} \centering \begin{tabular}{ccccc} \hline Maximum & Average speed & Temperature & Electron density\\ speed & (EUV wavelengths) & (MK) & (${\times}10^9$ cm$^{-3}$) \\ (km\,s$^{-1}$) & (km\,s$^{-1}$) & $T_\mathrm{i}$~~~~~$T_\mathrm{e}$ & $\rho_\mathrm{i}$~~~~$\rho_\mathrm{e}$ \\ \hline 343 at 304 \AA\ & 332&1.61~~1.81 & 9.45~~6.38 \\ \hline \end{tabular} \end{table} \section{Jet's geometry and the governing magnetohydrodynamic equations} \label{} We model the jet as a moving cylindrical magnetic flux tube of radius $a$ surrounded by a static coronal plasma. We assume that the jet and ambient plasma densities, $\rho_\mathrm{i}$ and $\rho_\mathrm{e}$, respectively, are homogeneous. Concerning magnetic field topology we consider three cases illustrated in Figure~\ref{}. In the first case (the left column in the picture) the magnetic fields inside and outside the tube are homogeneous accordingly with magnitudes $B_\mathrm{i}$ and $B_\mathrm{e}$. In the second case (the middle column in Figure~\ref{}) we assume that the magnetic field inside the jet/magnetic flux tube, $\boldsymbol{B}_\mathrm{i}$, is slightly twisted, while the one of the environment, $\boldsymbol{B}_\mathrm{e}$, is still homogeneous. In the latter case (the right column in the picture) both magnetic fields are twisted. The jet velocity $\boldsymbol{v}_0$ and wavevector $\boldsymbol{k}$ are directed along the central line of each flux tube, which (the line) is the $z$ axis of a cylindrical coordinate system that will be used in deriving the wave dispersion relations of propagating normal MHD modes in the jet. The primary reason for considering three types of magnetic flux tubes is to see, first, how the magnetic field twist of the internal magnetic field $\boldsymbol{B}_\mathrm{i}$ will change/affect the condition for emerging of KHI of given MHD mode, and second, how an additionally involved twisted external magnetic field $\boldsymbol{B}_\mathrm{e}$ will change the picture. In cylindrical coordinates all perturbations associated with the wave propagation are proportional to $\exp[\mathrm{i}(-\omega t + m\phi + k_z z)]$, where $\omega$ is the angular wave frequency, $m$ is the azimuthal mode number, and $k_z$ is the axial wavenumber, that is, $\boldsymbol{k} = (0, 0, k_z)$. Among the various MHD modes the kink one with $m = 1$ turns out to be the most unstable against the KHI---that is why we will focus our study on this mode only. The KHI occurs when the jet velocity $\boldsymbol{v}_0$ exceeds some critical value, $v_0^{\mathrm{cr}}$. That critical flow velocity depends upon two important parameters of the jet--environment system, notably the density contrast between the two media and the ratio of external to internal magnetic fields (more correctly the ratio of their axial components) \citep{Zaqarashvili14}. The density contrast, $\eta$, can be defined in two ways: (i) as the ratio $\rho_\mathrm{e}/\rho_\mathrm{i}$ (the standard definition), or (ii) as suggested by \cite{Paraschiv15}, as the ratio $\rho_\mathrm{e}/(\rho_\mathrm{i} + \rho_\mathrm{e})$. For the magnetic fields ratio, $b$, we have, according to Figure~\ref{}, three definitions of $b$, namely equal to $B_\mathrm{e}/B_\mathrm{i}$ for a untwisted magnetic flux tube, $B_\mathrm{e}/B_{\mathrm{i}z}$ for a slightly twisted tube, and $B_{\mathrm{e}z}/B_{\mathrm{i}z}$ for the case when both magnetic field are twisted. According to Table~1, the standard density contrast $\eta$ is equal to $0.675$. The estimated temperatures inside and outside the jet define the following sound speeds in both media: $c_\mathrm{si} = 151$~km\,s$^{-1}$ and $c_\mathrm{se} = 158$~km\,s$^{-1}$, respectively. Regarding the coronal magnetic field value many people tried to find it through observations. The methods include Helioseismology, Radio observations as well as magnetic field modeling. According to very recent study done by \cite {Luna17} the coronal magnetic field value ranges from $7$ to $30$~G. They have estimated these values using the Helioseismology of the filament and magnetic field modeling and the result from both these methods are consistent. Keeping these observational facts in mind, we assume that the background magnetic field is $B_\mathrm{e} = 7$~G, then the Alfv\'en speed in the environment, calculated from the standard formula $v_\mathrm{A} = B/\!\!\sqrt{\mu \rho}$, where $\mu$ being the vacuum permeability, is $v_\mathrm{Ae} = 191$~km\,s$^{-1}$. The Alfv\'en speed inside the flux tube can be derived from the pressure balance equation (the sum of thermal and magnetic pressure to be constant, that is, to have the same values in both media) and it yields $v_\mathrm{Ai} = 132.5$~km\,s$^{-1}$. This implies an internal magnetic field $B_\mathrm{i} = 5.9$~G; thus the magnetic fields ratio is $b = 1.18$. Both plasma betas accordingly are $\beta_\mathrm{i} = 1.55$ and $\beta_\mathrm{e} = 0.83$. Using the \cite{Paraschiv15} definition of the density contrast, we get a lower value for $\eta$, notably $0.403$. The sound speeds are the same as listed above. At this value of $\eta$, to satisfy the pressure balance equation, one needs to increase the background magnetic field: $B_\mathrm{e} = 8$~G fits the bill giving $v_\mathrm{Ae} = 218.3$~km\,s$^{-1}$, $v_\mathrm{Ai} = 63.2$~km\,s$^{-1}$, $b = 2.192$, $\beta_\mathrm{i} = 6.83$ and $\beta_\mathrm{e} = 0.63$, respectively. With aforementioned density contrasts, sound and Alfv\'en speeds, as well as magnetic fields ratios, prior to starting numerical calculation of wave dispersion relations of the kink mode ($m = 1$), we can make some predictions first for the nature of the propagating normal mode (be it pure surface, pseudo surface/body, or leaky wave) that depends on the speeds ordering \citep{Cally86}; second, to evaluate the so called `kink speed $c_\mathrm{k}$' in a static magnetic flux tube expressed in terms of plasma densities and Alfv\'en speeds in both media \citep{Edwin83}: \begin{equation} \label{} c_\mathrm{k} = \left( \frac{\rho_\mathrm{i} v_\mathrm{Ai}^2 + \rho_\mathrm{e} v_\mathrm{Ae}^2}{\rho_\mathrm{i} + \rho_\mathrm{e}} \right)^{1/2} = \left( \frac{1 + b^2}{1 + \eta} \right)^{1/2}v_\mathrm{Ai}, \end{equation} which, as seen, is independent of sound speeds and characterizes the propagation of transverse perturbations; and third, to find the expected threshold/critical Alfv\'en Mach number $M_\mathrm{A}$ at which KHI would start--the later is determined by the inequality \citep{Zaqarashvili14}: \begin{equation} \label{} M_\mathrm{A}^2 > \left( 1 + 1/\eta \right)\left( b^2 + 1 \right). \end{equation} Here, the Alfv\'en Mach number is defined as $v_0/v_\mathrm{Ai}$. These two formulas are written down for the kink mode propagating on a untwisted magnetic flux tube. When the magnetic field $\boldsymbol{B}_\mathrm{i}$ is twisted, we should use modified values for the ratio $b$, and also for $v_\mathrm{Ai}$---for computational reasons, the Alfv\'en speed inside the jet should be defined as $B_{\mathrm{i}z}/\!\!\sqrt{\mu \rho_\mathrm{i}}$. Dispersion relations of MHD modes are generally derived from the basic linearized equations of ideal magnetohydrodynamics. Linearization implies that each variable (plasma density, fluid velocity, magnetic field, thermal and magnetic pressures) consists of an equilibrium value and its perturbation, and each term in the linearized MHD equations should contain only one perturbation. We note the equilibrium quantities with subscript `$0$' and their perturbations with subscript `$1$'. Dispersion relation of normal MHD modes in a moving untwisted magnetic flux tube are well known and their solution in complex variables is more or less straightforward. With a twisted flux tube situation is more complicated. Up to now we have no appropriate dispersion relation of normal MHD modes for compressible plasma magnetic flux tube. That is why we are forced to do some approximations. In the case $\eta = 0.675$ plasma betas, as we already computed, are $\beta_\mathrm{i} = 1.55$ and $\beta_\mathrm{e} = 0.83$, respectively, that is both are close to $1$. Then, according to \cite{Zank93} both media can be considered as incompressible plasmas and the wave dispersion relation can be obtained from the linearized equation of the incompressible magnetohydrodynamics. In the second case with $\eta = 0.403$ where $\beta_\mathrm{i} = 6.83$ and $\beta_\mathrm{e} = 0.63$, obviously the internal medium can be treated as incompressible plasma, but an adequate approximation for its environment would be a cool plasma. As \cite{Zank93} claim, for the $\beta \ll 1$ regime, there is a strong tendency for nearly incompressible perturbations to propagate in a 1D direction parallel to the magnetic field. Thus, in our study for KHI development of Alfv\'en-like perturbations/waves in the latter twisted solar jet it is appropriate to consider the jet as incompressible plasma and its environment as a cool, also incompressible, medium. In summary, we have to obtain five different dispersion relations depending upon the nature of the jet medium and its environment, namely one for the case of untwisted flux tube when both plasmas are treated as compressible fluids; two equations for the case when the internal and external magnetic fields are twisted and the two media are considered as incompressible plasmas, and finally two equations for magnetically twisted flux tubes when the internal medium is assumed to be incompressible plasma while the external one is cool plasma. The derivation of the MHD modes dispersion relation for an untwisted flux tube on using the basic equation of ideal magnetohydrodynamics is more or less straightforward and we will take its final form. The same is the case when the internal magnetic field is twisted only. Dispersion relation of MHD waves traveling on a double-twisted magnetic flux tube assuming incompressible plasma approximation in both media was obtained by \cite{Zaqarashvili14} on using the operator coefficient techniques developed by \cite{Goossens92}. Here, we will re-derive that dispersion relation starting from the basic MHD equations governing the incompressible plasma dynamics. Moreover, assuming a cool plasma environment, we will obtain a new form of the dispersion relation applicable in the case when internal plasma is incompressible and the surrounding medium is a cool plasma. Fortunately, the governing equations for space and time evolution of fluid velocity $\vec{v}$, magnetic field $\vec{B}$, and pressure $p$ perturbations for incompressible and cool plasmas have practically the same form: \begin{equation} \label{} \rho_0 \frac{\partial}{\partial t}\boldsymbol{v}_1 + \rho_0 (\vec{v}_0 \cdot \nabla)\vec{v}_1 + \nabla \left( p_1 + \frac{\vec{B}_0 \cdot \vec{B}_1}{\mu} \right) - \frac{1}{\mu}(\vec{B}_0 \cdot \nabla)\vec{B}_1 - \frac{1}{\mu}(\vec{B}_1 \cdot \nabla)\vec{B}_0 = 0, \end{equation} \begin{equation} \label{} \frac{\partial}{\partial t}\vec{B}_1 - \nabla (\vec{v}_1 \times \vec{B}_0) - \nabla (\vec{v}_0 \times \vec{B}_1) = 0, \end{equation} \begin{equation} \label{} \nabla \cdot \vec{v}_1 = 0, \end{equation} \begin{equation} \label{} \nabla \cdot \vec{B}_1 = 0. \end{equation} We note that in the case of cool plasma, the thermal pressure $p$ and its perturbation $p_1$ are equal to zero, and the fluid velocity perturbation, $\vec{v}_1$, in cylindrical coordinates, is presented as $\vec{v}_1 = (v_{1r},v_{1\phi}, 0)$, while the magnetic field perturbation has its three non-zero components, notably $\vec{B}_1 = (B_{1r}, B_{1\phi}, B_{1z})$. Notice also that for a cool plasma the total pressure perturbation $p_\mathrm{tot} = p_1 + \vec{B}_0 \cdot \vec{B}_1/\mu$ reduces to magnetic pressure perturbation only. \section{Dispersion Relations} \label{} \subsection{Dispersion relation in untwisted moving flux tube} \label{} We start with the simplest magnetic fields configuration pictured by the left column in Figure~\ref{}. Dispersion relation of normal MHD modes propagating in a flowing compressible jet surrounded by a static compressible plasma reads \citep{Terra-Homem03,Nakariakov07,Zhelyazkov12} \begin{eqnarray} \label{} \frac{\rho_\mathrm{e}}{\rho_\mathrm{i}}\left( \omega^2 - k_z^2 v_\mathrm{Ae}^2 \right) \kappa_\mathrm{i}\frac{I_m^{\prime}(\kappa_\mathrm{i}a)}{I_m(\kappa_\mathrm{i}a)} - \left[ \left( \omega - \vec{k} \cdot \vec{v}_0 \right)^2 - k_z^2 v_\mathrm{Ai}^2 \right] \kappa_\mathrm{e}\frac{K_m^{\prime}(\kappa_\mathrm{e}a)}{K_m(\kappa_\mathrm{e}a)} = 0. \end{eqnarray} Here, $\kappa$ is the wave attenuation coefficient, which characterizes the space structure of the wave in the corresponding medium and whose squared magnitude is given by the expression \begin{equation} \label{} \kappa^2 = -\frac{\left[ \left( \omega - \vec{k} \cdot \vec{v}_0 \right)^2 - k_z^2 c_{\rm s}^2 \right]\left[ \left( \omega - \vec{k} \cdot \vec{v}_0 \right)^2 - k_z^2 v_\mathrm{A}^2 \right]}{\left( c_\mathrm{ s}^2 + v_\mathrm{A}^2 \right)\left[ \left( \omega - \vec{k} \cdot \vec{v}_0 \right)^2 - k_z^2 c_\mathrm{T}^2 \right]}, \end{equation} where \begin{equation} \label{} c_\mathrm{T} = c_\mathrm{s} v_\mathrm{A}/\left( c_\mathrm{s}^2 + v_\mathrm{A}^2 \right)^{1/2} \end{equation} is the so-called \emph{tube velocity\/} \citep{Edwin83}. As seen from the expressions for the attenuation coefficient and tube velocity, both quantities have different values inside and outside the flux magnetic tube owing to the different magnitudes of sound and Alfv\'en speeds in both media. In dispersion equation~(\ref{}), $I_m$ and $K_m$ are the modified Bessel functions, and the prime means differentiation with respect to the function argument. It is worth noticing that the mode frequency inside the flux tube is Doppler-shifted. \subsection{Dispersion relations in twisted moving flux tube} \label{} If the magnetic field is twisted, we assume that in cylindrical coordinate system the magnetic field has the following form: $\vec{B} = \left( 0, B_\phi(r), B_z(r) \right)$. The unperturbed magnetic field and thermal pressure satisfy the pressure balance equation \begin{equation} \label{} \frac{\mathrm{d}}{\mathrm{d}r}\left( p + \frac{B_\phi^2 + B_z^2}{\mu} \right) = -\frac{B_\phi^2}{\mu r}. \end{equation} For simplicity we consider magnetic flux tube with uniform twist, that is \begin{equation} \label{} \vec{B}_\mathrm{i} = (0, Ar, B_{\mathrm{i}z}), \end{equation} where $A$ and $B_{\mathrm{i}z}$ are constants. The magnetic field in the environment is homogeneous, that is, $\vec{B}_\mathrm{e} = (0, 0, B_\mathrm{e})$. The flow velocity of the moving flux tube is also homogeneous: $\vec{v}_0 = (0, 0, v_0)$. When considering that both media are incompressible plasmas, from aforementioned set of equations (\ref{})--(\ref{}) and appropriate boundary conditions for continuity of total Lagrangian pressure perturbation and the transfer Lagrangian displacement $\xi_r$ (obtainable from the relation $v_{1r} = \partial \xi_r/\partial t$), one can get the following dispersion relation \citep{Zhelyazkov12a}: \begin{equation} \label{} \frac{\left( \Omega^2 - \omega_\mathrm{Ai}^2 \right)F_m(\kappa_\mathrm{i}a) - 2mA \omega_\mathrm{Ai}/\sqrt{\mu \rho_\mathrm{i}}} {\left( \Omega^2 - \omega_\mathrm{Ai}^2 \right)^2 - 4A^2\omega_\mathrm{Ai}^2/\mu \rho_\mathrm{i} } = \frac{P_m(k_z a)} { \left(\rho_\mathrm{e}/\rho_\mathrm{i}\right) \left( \omega^2 - \omega_\mathrm{Ae}^2 \right) + A^2 P_m(k_z a)/\mu \rho_\mathrm{i}}, \end{equation} where $\Omega = \omega - \vec{k}\cdot \vec{v}_0$ is the Doppler-shifted wave frequency in the moving medium, \[ F_m(\kappa_\mathrm{i}a) = \frac{\kappa_\mathrm{i}a I_m^{\prime}(\kappa_\mathrm{i}a)}{I_m(\kappa_\mathrm{i}a)}, \qquad P_m(k_z a) = \frac{k_z a K_m^{\prime}(k_z a)}{K_m(k_z a)}, \] \[ \omega_\mathrm{Ai} = \frac{\vec{k}\cdot \vec{B}_\mathrm{i}}{\sqrt{\mu \rho_\mathrm{i}}} = \frac{1}{\sqrt{\mu \rho_\mathrm{i}}} \left( \frac{m}{r}B_{\mathrm{i}\phi} + k_z B_{\mathrm{i}z} \right) \quad \mbox{and} \quad \omega_\mathrm{Ae} = \frac{k_z B_\mathrm{e}}{\sqrt{\mu \rho_\mathrm{e}}} \] are the local Alfv\'en frequencies inside the moving flux tube and outside its static environment. It is easy to see that $\omega_\mathrm{Ae} = k_z v_\mathrm{Ae}$. The wave attenuation coefficient inside the tube, $\kappa_\mathrm{i}$, is given by \begin{equation} \label{} \kappa_\mathrm{i} = k_z \left[ 1 - 4A^2 \omega_\mathrm{Ai}^2/\mu \rho_\mathrm{i} \left( \Omega^2 - \omega_\mathrm{Ai}^2 \right)^2 \right]^{1/2}, \end{equation} while that in the surrounding medium, $\kappa_\mathrm{e}$, is simply equal to $k_z$. In the case when the environment is treated as a cool medium, the form of the wave dispersion relation is the same as in Equation~(\ref{}), but the wave attenuation coefficient $\kappa_\mathrm{e} = k_z$ should be replaced by \begin{equation} \label{} \kappa_\mathrm{e} = \left( k_z^2 v_\mathrm{Ae}^2 - \omega^2 \right)^{1/2}v_\mathrm{Ae}. \end{equation} \subsection{Dispersion relations in twisted moving flux tube and twisted external magnetic field} \label{} The most complicated case is that pictured by the right column in Figure~\ref{}. The basic idea in deriving the wave dispersion relation is to find out solutions for the total pressure perturbation, $p_\mathrm{tot}$, and perturbed interface, $\xi_r$, and merge them via appropriate boundary conditions. The equilibrium magnetic field inside the moving flux tube is the same as in the previous subsection~\ref{}, namely $\vec{B}_\mathrm{i} = (0, Ar, B_{\mathrm{i}z})$, where, recall, $A$ and $B_{\mathrm{i}z}$ are constants. The jet's flow velocity has only $z$ component, equal to $v_0$. As unperturbed parameters depend on the $r$ coordinate only, all the perturbations can be Fourier analyzed with $\exp[\mathrm{i}(-\omega t + m\phi + k_z z)]$ and the governing MHD equations (\ref{}) and (\ref{}) are: \begin{equation} \label{} -\mathrm{i}(\omega - \vec{k}\cdot \vec{v}_0)v_{1r} + \frac{1}{\rho_\mathrm{i}} \frac{\mathrm{d}}{\mathrm{d}r} p_\mathrm{tot} - \mathrm{i}\frac{1}{\sqrt{\mu \rho_\mathrm{i}}}\omega_\mathrm{Ai}B_{1r} + \frac{2A}{\mu \rho_\mathrm{i}}B_{1\phi} = 0, \end{equation} \begin{equation} \label{} -\mathrm{i}(\omega - \vec{k}\cdot \vec{v}_0)v_{1\phi} + \mathrm{i}\frac{1}{\rho_\mathrm{i}}\frac{m}{r}p_\mathrm{tot} - \mathrm{i}\frac{1}{\sqrt{\mu \rho_\mathrm{i}}}\omega_\mathrm{Ai}B_{1\phi} - \frac{2A}{\mu \rho_\mathrm{i}}B_{1r} = 0, \end{equation} \begin{equation} \label{} -(\omega - \vec{k}\cdot \vec{v}_0)v_{1z} + \frac{1}{\rho_\mathrm{i}}k_z p_\mathrm{tot} - \frac{1}{\sqrt{\mu \rho_\mathrm{i}}}\omega_\mathrm{Ai}B_{1z} = 0, \end{equation} \begin{equation} \label{} (\omega - \vec{k}\cdot \vec{v}_0)B_{1r} + (mA + k_z B_{\mathrm{i}z})v_{1r} = 0, \end{equation} \begin{equation} \label{} (\omega - \vec{k}\cdot \vec{v}_0)B_{1\phi} + (mA + k_z B_{\mathrm{i}z})v_{1\phi} = 0, \end{equation} \begin{equation} \label{} (\omega - \vec{k}\cdot \vec{v}_0)B_{1z} + (mA + k_z B_{\mathrm{i}z})v_{1z} = 0, \end{equation} where \begin{equation} \label{} \omega_\mathrm{Ai} = \frac{1}{\sqrt{\mu \rho_\mathrm{i}}}\left( mA + k_z B_{\mathrm{i}z} \right) \end{equation} is the local Alfv\'en frequency inside the moving flux tube. From the above equations one can obtain expressions for the fluid velocity components in terms of the total pressure perturbation, $p_\mathrm{tot}$, and its derivative with respect to $r$, namely \begin{equation} \label{} v_{1r} = -\mathrm{i}\frac{1}{Y_\mathrm{i}}\frac{1}{\Omega_\mathrm{i}\rho_\mathrm{i}}\left( \frac{\mathrm{d}}{\mathrm{d}r} - Z_\mathrm{i} \frac{m}{r} \right)p_\mathrm{tot}, \end{equation} \begin{equation} \label{} v_{1\phi} = \frac{1}{Y_\mathrm{i}}\frac{1}{\Omega_\mathrm{i}\rho_\mathrm{i}}\left( \frac{m}{r} - Z_\mathrm{i} \frac{\mathrm{d}}{\mathrm{d}r} \right)p_\mathrm{tot}, \end{equation} \begin{equation} \label{} v_{1z} = \frac{1}{\Omega_\mathrm{i}\rho_\mathrm{i}}k_z p_\mathrm{tot}. \end{equation} Here, \[ \frac{1}{\Omega_\mathrm{i}} \equiv \frac{\omega - \vec{k}\cdot \vec{v}_0}{(\omega - \vec{k}\cdot \vec{v}_0)^2 - \omega_\mathrm{Ai}^2}, \quad Z_\mathrm{i} \equiv \frac{2A\omega_\mathrm{Ai}}{\sqrt{\mu \rho_\mathrm{i}}\left[ (\omega - \vec{k}\cdot \vec{v}_0)^2 - \omega_\mathrm{Ai}^2 \right]}, \quad \mbox{and} \quad Y_\mathrm{i} = 1 - Z_\mathrm{i}^2. \] On using these expressions for fluid velocity perturbation in the constraint equation (\ref{}), after some algebra we get that the total pressure perturbation obeys the equation \begin{equation} \label{} \left[ \frac{\mathrm{d}^2}{\mathrm{d}r^2} + \frac{1}{r}\frac{\mathrm{d}}{\mathrm{d}r} - \left( \frac{m^2}{r^2} + \kappa_\mathrm{i}^2\right) \right]p_\mathrm{tot} = 0, \end{equation} where $\kappa_\mathrm{i}$, not surprisingly, coincides with expression (\ref{}), that is, \[ \kappa_\mathrm{i}^2 = k_z^2 \left[ 1 - \frac{4A^2 \omega_\mathrm{Ai}^2}{\mu \rho_\mathrm{i}\left( \Omega^2 - \omega_\mathrm{Ai}^2 \right)} \right], \quad \mbox{where as before} \quad \Omega \equiv \omega - \vec{k}\cdot \vec{v}_0. \] Equation~(\ref{}) is the Bessel equation whose solution bounded at the tube axis is \begin{equation} \label{} p_\mathrm{i\,tot} = \alpha_\mathrm{i}I_m(\kappa_\mathrm{i}r). \end{equation} Here, $I_m$ is the modified Bessel function of order $m$ and $\alpha_\mathrm{i}$ is a constant. Transverse displacement, $\xi_{\mathrm{i}r}$, can be obtained from expression~(\ref{}) and has the form: \begin{equation} \label{} \xi_{\mathrm{i}r} = \frac{\alpha_\mathrm{i}}{r} \frac{\left( \Omega^2 - \omega_\mathrm{Ai}^2 \right)\kappa_\mathrm{i}rI_m^{\prime}(\kappa_\mathrm{i}r) - 2mA\omega_\mathrm{Ai} I_m(\kappa_\mathrm{i}r)/\!\!\sqrt{\mu \rho_\mathrm{i}}}{\rho_\mathrm{i}\left( \Omega^2 - \omega_\mathrm{Ai}^2 \right)^2 - 4A^2 \omega_\mathrm{Ai}^2/\mu}, \end{equation} where the prime, $^{\prime}$, implies a differentiation by the Bessel function argument. Let us now go to the environment and find similar expressions for $p_\mathrm{e\,tot}$ and $\xi_{\mathrm{e}r}$, respectively. We consider the external magnetic field of the form \begin{equation} \label{} \vec{B}_\mathrm{e} = \left( 0, B_{\mathrm{e}\phi}(a)\frac{a}{r}, B_{\mathrm{e}z}(a)\frac{a^2}{r^2} \right) \end{equation} (where for convenience we denote $B_{\mathrm{e}\phi}(a) \equiv B_\phi$ and $B_{\mathrm{e}z}(a) \equiv B_z$) and the plasma density in the form $\rho_0 = \rho_\mathrm{e}(a/r)^4$ so that the Alfv\'en frequency \begin{eqnarray} \label{} \omega_\mathrm{Ae} = \frac{1}{\sqrt{\mu \rho_0(r)}}\left[ \frac{m}{r}B_{\mathrm{e}\phi}(r) + k_z B_{\mathrm{e}z}(r) \right] = \frac{r^2}{\sqrt{\mu \rho_\mathrm{e}a^4}}\left( \frac{maB_\phi}{r^2} + \frac{k_z a^2 B_z}{r^2} \right) \nonumber \\ \nonumber \\ {}= \frac{mB_\phi + k_z aB_z}{\sqrt{\mu \rho_\mathrm{e} a^2}} \end{eqnarray} is constant. This circumstance allows us to find analytical solution to the governing equations. Now momentum and Faraday equations are displayed in the form: \begin{equation} \label{} -\mathrm{i}\omega v_{1r} + \frac{1}{\rho_\mathrm{e}}\frac{\mathrm{d}}{\mathrm{d}r}p_\mathrm{tot} - \mathrm{i}\frac{1}{\mu \rho_\mathrm{e}}\left( \frac{maB_\phi}{r^2} + \frac{k_z a^2 B_z}{r^2} \right)B_{1r} + 2\frac{a}{r^2} \frac{B_\phi}{\mu \rho_\mathrm{e}}B_{1\phi} = 0, \end{equation} \begin{equation} \label{} -\mathrm{i}\omega v_{1\phi} + \frac{1}{\rho_\mathrm{e}}\frac{m}{r}p_\mathrm{tot} - \mathrm{i}\frac{1}{\mu \rho_\mathrm{e}}\left( \frac{maB_\phi}{r^2} + \frac{k_z a^2 B_z}{r^2} \right)B_{1\phi} = 0, \end{equation} \begin{equation} \label{} -\mathrm{i}\omega v_{1z} + \frac{1}{\rho_\mathrm{e}}k_z p_\mathrm{tot} - \mathrm{i}\frac{1}{\mu \rho_\mathrm{e}}\left( \frac{maB_\phi}{r^2} + \frac{k_z a^2 B_z}{r^2} \right)B_{1z} + 2\frac{a^2}{r^3} \frac{B_z}{\mu \rho_\mathrm{e}}B_{1r} = 0, \end{equation} \begin{equation} \label{} \omega B_{1r} + \left( \frac{maB_\phi}{r^2} + \frac{k_z a^2 B_z}{r^2} \right)v_{1r} = 0, \end{equation} \begin{equation} \label{} \mathrm{i}\omega B_{1\phi} + \mathrm{i}\left( \frac{maB_\phi}{r^2} + \frac{k_z a^2 B_z}{r^2} \right)v_{1\phi} + 2 \frac{a}{r^2}B_\phi v_{1r} = 0, \end{equation} \begin{equation} \label{} \mathrm{i}\omega B_{1z} + \mathrm{i}\left( \frac{maB_\phi}{r^2} + \frac{k_z a^2 B_z}{r^2} \right)v_{1z} + 2 \frac{a^2}{r^3}B_z v_{1r} = 0. \end{equation} By using expression~(\ref{}), from above equations, after a lengthy algebra one obtains that \begin{equation} \label{} v_{1r} = -\mathrm{i}\frac{1}{Y_\mathrm{e}}\frac{1}{\Omega_\mathrm{e}\rho_\mathrm{e}}\left( \frac{\mathrm{d}}{\mathrm{d}r} - Z_\mathrm{e}\frac{m}{r} \right)p_\mathrm{tot}, \end{equation} \begin{equation} \label{} v_{1\phi} = \frac{1}{\Omega_\mathrm{e}\rho_\mathrm{e}}\left[ \left( 1 + \frac{Z_\mathrm{e}^2}{Y_\mathrm{e}} \right) \frac{m}{r} - \frac{Z_\mathrm{e}}{Y_\mathrm{e}} \frac{\mathrm{d}}{\mathrm{d}r} \right]p_\mathrm{tot}, \end{equation} \begin{equation} \label{} v_{1z} = \frac{1}{\Omega_\mathrm{e}\rho_\mathrm{e}}k_z p_\mathrm{tot}, \end{equation} where now \[ \frac{1}{\Omega_\mathrm{e}} \equiv \frac{\omega}{\omega^2 - \omega_\mathrm{Ae}^2}, \quad Z_\mathrm{e} \equiv \frac{2B_\phi \omega_\mathrm{Ae}}{\sqrt{\mu \rho_\mathrm{e}a^2}\left(\omega^2 - \omega_\mathrm{Ae}^2 \right)}, \quad \mbox{and} \quad Y_\mathrm{e} = 1 - \frac{\omega^2}{\omega_\mathrm{Ae}^2}Z_\mathrm{e}^2. \] With these expressions for fluid velocity perturbations, on using again the equation $\nabla \cdot \vec{v}_1 = 0$, we obtain that the total pressure perturbation obeys the equation \begin{equation} \label{} \left[ \frac{\mathrm{d}^2}{\mathrm{d}r^2} + \frac{5}{r}\frac{\mathrm{d}}{\mathrm{d}r} - \left( \frac{n^2}{r^2} + \kappa_\mathrm{e}^2\right) \right]p_\mathrm{tot} = 0, \end{equation} which is an equation for Bessel function with complex order $\nu = \sqrt{4 + n^2}$, with a solution bounded at infinity in the form \begin{equation} \label{} p_\mathrm{e\,tot} = \alpha_\mathrm{e} \frac{a^2}{r^2} K_\nu(\kappa_\mathrm{e}r), \end{equation} where $\alpha_\mathrm{e}$ is a constant, the wave attenuation coefficient is equal to \begin{equation} \label{} \kappa_\mathrm{e} = k_z \left( 1 - \frac{4B_\phi^2 \omega^2}{\mu \rho_\mathrm{e} \left( \omega^2 - \omega_\mathrm{Ae}^2 \right)^2 a^2} \right)^{1/2}, \end{equation} and the term $n^2$ is given by the expression \[ n^2 = m^2 - \frac{4m^2 B_\phi^2}{\mu \rho_\mathrm{e}a^2 \left( \omega^2 - \omega_\mathrm{Ae}^2 \right)} + \frac{8mB_\phi \omega_\mathrm{Ae}}{\sqrt{\mu \rho_\mathrm{e}}a \left( \omega^2 - \omega_\mathrm{Ae}^2 \right)}. \] With the help of the relation $\xi_r = \mathrm{i}v_{1r}/\omega$, on using Equation~(\ref{}), one obtains that \begin{eqnarray} \label{} \xi_{\mathrm{e}r} = \alpha_\mathrm{e} \frac{r\left( \omega^2 - \omega_\mathrm{Ae}^2 \right)\kappa_\mathrm{e}r K_\nu^{\prime}(\kappa_\mathrm{e}r)}{a^2 \rho_\mathrm{e}\left( \omega^2 - \omega_\mathrm{Ae}^2 \right)^2 - 4B_\phi^2 \omega^2/\mu} \nonumber \\ \nonumber \\ {}- \alpha_\mathrm{e} \frac{r}{a} \left( \frac{2a \left( \omega^2 - \omega_\mathrm{Ae}^2 \right) + 2mB_\phi \omega_\mathrm{Ae}/\!\!\sqrt{\mu \rho_\mathrm{e}}}{a^2 \rho_\mathrm{e}\left( \omega^2 - \omega_\mathrm{Ae}^2 \right)^2 - 4B_\phi^2 \omega^2/\mu} \right) K_\nu(\kappa_\mathrm{e}r). \end{eqnarray} Merging the solutions for $p_\mathrm{tot}$ and $\xi_r$ in both media at the tube surface, $r = a$, one obtains the dispersion equation of the normal MHD modes propagating on the moving magnetic flux tube. As we already said, the boundary conditions at the tube surface are the continuity of the Lagrangian displacement \begin{equation} \label{} \xi_{\mathrm{i}r}|_{r=a} = \xi_{\mathrm{e}r}|_{r=a} \end{equation} (where $\xi_{\mathrm{i}r}$ and $\xi_{\mathrm{e}r}$ are given by Equations~(\ref{}) and (\ref{})), and the total Lagrangian pressure \citep{Bennett99} \begin{equation} \label{} \left.p_\mathrm{i\,tot} - \frac{B_{\mathrm{i}\phi}^2}{\mu a}\xi_{\mathrm{i}r}\right\vert_{r=a} = \left.p_\mathrm{e\,tot} - \frac{B_{\mathrm{e}\phi}^2}{\mu a}\xi_{\mathrm{e}r}\right\vert_{r=a}, \end{equation} where the total pressure perturbations are given by Equations~(\ref{}) and (\ref{}), respectively. Using these boundary conditions we recover the dispersion equation governing the oscillations in moving twisted flux tube surrounded by twisted incompressible magnetized plasma derived in \citealp{Zaqarashvili14}: \begin{eqnarray} \label{} \frac{\left( \left[ \omega - \vec{k} \cdot \vec{v}_0 \right]^2 - \omega_\mathrm{Ai}^2 \right)F_m(\kappa_\mathrm{i}a) - 2mA \omega_\mathrm{Ai}/\sqrt{\mu \rho_\mathrm{i}}} {\rho_\mathrm{i}\left( \left[ \omega - \vec{k} \cdot \vec{v}_0 \right]^2 - \omega_\mathrm{Ai}^2 \right)^2 - 4A^2\omega_\mathrm{Ai}^2/\mu} \nonumber \\ \nonumber \\ {}= \frac{a^2 \left( \omega^2 - \omega_\mathrm{Ae}^2 \right)Q_\nu(\kappa_\mathrm{e}a) - G} {L - H \left[a^2 \left( \omega^2 - \omega_\mathrm{Ae}^2 \right)Q_\nu(\kappa_\mathrm{e}a) - G \right]}, \end{eqnarray} where \[ F_m(\kappa_\mathrm{i}a) = \frac{\kappa_\mathrm{i}a I_m^{\prime}(\kappa_\mathrm{i}a)}{I_m(\kappa_\mathrm{i}a)}, \quad Q_\nu(\kappa_\mathrm{e}a) = \frac{\kappa_\mathrm{e}a K_\nu^{\prime}(\kappa_\mathrm{e}a)}{K_\nu(\kappa_\mathrm{e}a)}, \quad L = a^2 \rho_\mathrm{e} \left( \omega^2 - \omega_\mathrm{Ae}^2 \right)^2 - 4B_{\mathrm{e}\phi}^2 \omega^2/\mu, \] \[ H = B_{\mathrm{e}\phi}^2/\mu a^2 - A^2/\mu, \quad G = 2a^2 \left( \omega^2 - \omega_\mathrm{Ae}^2 \right) + 2maB_{\mathrm{e}\phi} \omega_\mathrm{Ae}/\!\!\sqrt{\mu \rho_\mathrm{e}}. \] When surrounding plasma is treated as a cool medium, then, recall, the plasma pressure perturbation $p_1 = 0$ and the total pressure perturbation, $p_\mathrm{tot}$ reduces to magnetic pressure perturbation only. Moreover, the axial component of fluid velocity perturbation also is zero, that is, $v_{1z} = 0$. Under these circumstances, among the six governing MHD equations (\ref{})--(\ref{}), only Equations~(\ref{}) and (\ref{}) are slightly changed; we will denote the magnetic pressure perturbation by $p_\mathrm{mag1}$. By using Equation~(\ref{}), we express $B_{1z}$ in terms of $B_{1r}$ and $B_{1\phi}$ and obtain \[ B_{1z} = \mathrm{i}\frac{1}{k_z}\left( \frac{\mathrm{d}}{\mathrm{d}r} + \frac{1}{r} \right)B_{1r} - \frac{1}{k_z}\frac{m}{r}B_{1\phi}. \] Inserting this $B_{1z}$ alongside with $B_{1r}$ and $B_{1\phi}$ expressed via $v_{1r}$ and $v_{1\phi}$ (by using Equations~(\ref{}) and (\ref{})) into Equation~(\ref{}), we get \[ -\mathrm{i}\frac{1}{\rho_\mathrm{i}}k_z^2 p_\mathrm{mag1} - \frac{\omega_\mathrm{Ae}^2}{\omega}\left( \frac{\mathrm{d}}{\mathrm{d}r} + 3\frac{1}{r} \right)v_{1r} - \mathrm{i}\frac{\omega_\mathrm{Ae}^2}{\omega} \frac{m}{r} v_{1\phi} = 0. \] Now, on using Equations~(\ref{}) and (\ref{}) giving us fluid velocity perturbation components in terms of total (that is, magnetic) pressure perturbation, above equation takes the form \begin{equation} \label{} \left[ \frac{\mathrm{d}^2}{\mathrm{d}r^2} + \frac{3}{r}\frac{\mathrm{d}}{\mathrm{d}r} - \left( \frac{n_\mathrm{c}^2}{r^2} + \kappa_\mathrm{ec}^2\right) \right]p_\mathrm{mag1} = 0, \end{equation} which like before is the equation for Bessel function with complex order $\nu_\mathrm{c} = \sqrt{2 + n_\mathrm{c}^2}$, where \[ n_\mathrm{c}^2 = m^2 - \frac{4m^2 B_\phi^2}{\mu \rho_\mathrm{e}a^2 \left( \omega^2 - \omega_\mathrm{Ae}^2 \right)} + \frac{4mB_\phi \omega_\mathrm{Ae}}{\sqrt{\mu \rho_\mathrm{e}}a \left( \omega^2 - \omega_\mathrm{Ae}^2 \right)}. \] Here, the label `c' stamps for \emph{cool}. Note that in the cool environment the wave attenuation coefficient is given by \begin{equation} \label{} \kappa_\mathrm{ec} = k_z \left( 1 - \frac{4B_\phi^2 \omega^2}{\mu \rho_\mathrm{e} \left( \omega^2 - \omega_\mathrm{Ae}^2 \right)^2 a^2} \right)^{1/2}\left( \frac{\omega^2}{\omega_\mathrm{Ae}^2} - 1 \right)^{1/2}. \end{equation} Further on, following the standard steps for deriving the wave dispersion relation, we arrive at Equation~(\ref{}) in which have to replace $\nu$ by $\nu_\mathrm{c}$ and $\kappa_\mathrm{e}$ by $\kappa_\mathrm{ec}$, respectively. Having derived all necessary wave dispersion equations, we can now apply them in studying the propagation characteristics primarily of the kink ($m = 1$) mode running on the observed EUV jet, that will be done in the next section. \section{Numerical results and discussion} \label{} Since we are looking for unstable solutions to wave dispersion relations, assume the MHD modes are running along the flux tube having a real axial wavenumber $k_z$ and a complex angular wave frequency $\omega \equiv \mathrm{Re}\omega + \mathrm{i}\,\mathrm{Im}\omega$. Numerical results are usually presented as dependence of the complex wave phase velocity $v_\mathrm{ph} = \omega/k_z$ on the $k_z$. For convenience, we normalize all velocities with respect to the Alfv\'en speed inside the tube, $v_\mathrm{Ai}$, and the wavelength $\lambda = 2\pi/k_z$ with respect to the tube radius, $a$. Thus we have a complex dimensionless wave phase velocity $v_\mathrm{ph}/v_\mathrm{Ai}$ and dimensionless axial wavenumber $k_z a$. The normalization of sound, $c_\mathrm{s}$ and tube, $c_\mathrm{T}$, speeds in the attenuating coefficients (\ref{}), contained in dispersion equation (\ref{}), requires the values of both the reduced plasma betas $\tilde{\beta}_\mathrm{i,e} = c_\mathrm{i,e}^2/v_\mathrm{Ai,e}^2$ and the magnetic fields ratio $b = B_\mathrm{e}/B_\mathrm{i}$. In the case of a twisted magnetic flux tube, for normalizing the local Alfv\'en speed (\ref{}), along with the fixed input parameters $\eta = \rho_\mathrm{e}/\rho_\mathrm{i}$ and $b_\mathrm{twist} = B_\mathrm{e}/B_{\mathrm{i}z}$, one needs to define the parameter $\varepsilon \equiv B_{\mathrm{i}\phi}(a)/B_{\mathrm{i}z}(a) = Aa/B_{\mathrm{i}z}$ that characterizes the twisted magnetic field inside the tube. When the external magnetic field is twisted too, its normalization requires an additional $\varepsilon$, equal to $B_{\mathrm{e}\phi}(a)/B_{\mathrm{e}z}(a) = B_\phi/B_z$, which will have a subscript `2', while the similar parameter characterizing the twist of the internal magnetic field will possesses the subscript `1'. After these preliminary notes, we can report the results of numerical computations, carried out for the two values of the density contrast, $\eta$, equal to $0.675$ and $0.403$, respectively. \subsection{Kink mode propagation characteristics at a density contrast of $0.675$} \label{} We begin with the simplest jet's model as an untwisted axially moving flux tube (the left column in Figure~\ref{}). The ordering of basic speeds (sound, Alfv\'en, and tube one) in both media is as follows: \[ c_\mathrm{Ti} < c_\mathrm{Te} < v_\mathrm{Ai} < c_\mathrm{si} < c_\mathrm{se} < v_\mathrm{Ae}. \] According to \cite{Cally86} (see TABLE I there), at this ordering the kink ($m = 1$) mode, propagating in a static ($v_0 = 0$) tube, would be a pseudo surface/body wave of B$^{-}_{-}$ type. In addition to $\eta = 0.675$, the other input parameters for the numerical task are: $\tilde{\beta}_\mathrm{i} = 1.2946$, $\tilde{\beta}_\mathrm{e} = 0.6836$, and $b = 1.184$. At Alfv\'en Mach number $M_\mathrm{A} = 0$ (static flux tube) the anticipated normalized kink speed (\ref{}) has a magnitude of $1.1977$. With these input parameters the numerical computations of dispersion relation (\ref{}) confirm that the kink mode traveling on the tube is a pseudo surface/body wave possessing, at $k_z a \ll 1$, a normalized phase velocity equal to the normalized kink speed within 4 places after the decimal point. In moving flux tube, at relatively small $M_\mathrm{A}$s, the kink speed splits into a pair of phase velocities \citep{Zhelyazkov12}, whose dispersion curves initially go almost parallel to each \begin{figure}[!ht] \centerline{\hspace*{0.015\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg5a.eps} \hspace*{-0.03\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg5b.eps} } \caption{(Left panel) Dispersion curves of stable and unstable kink ($m = 1$) mode propagating in a moving untwisted magnetic flux tube at $\eta = 0.675$, $b = 1.184$, and at various values of Alfv\'en Mach number $M_\mathrm{A}$. The threshold Alfv\'en Mach number for KHI occurring is equal to $2.4425$ (red curve). (Right panel) The normalized growth rates of the unstable kink mode for the same values of the input parameters. The purple curve has been calculated at $M_\mathrm{A} = 2.4475$.} \label{} \end{figure} other, but for higher Alfv\'en Mach numbers, when one reaches the region of expected $M_\mathrm{A} > 2.44$ for occurring of KHI according to the criterion (\ref{}), their behavior become completely different. Look, for example, at the green and purple curves labeled by $M_\mathrm{A} = 2.4$ and $M_\mathrm{A} = 2.42$ in the left panel of Figure~\ref{}. The low-speed curves have more or less normal move while the high-speed ones turn over at some $k_z a$-values forming semi-closed dispersion curves. The instability arises at the threshold Alfv\'en Mach number equal to $2.4425$---indeed rather close to the predicted value. This value of $M_\mathrm{A}$ tells us that the KHI should emerge at a critical speed of the jet equal to $323.6$~km\,s$^{-1}$, which is less than the average jet speed of $332$~km\,s$^{-1}$ (see Table 1). The marginal red curves divides the 1D $M_\mathrm{A}$-space into two regions: for all $M_\mathrm{A} < 2.4425$ the kink wave is a stable MHD mode, while in the opposite case it becomes unstable; alongside the marginal red curve in the left panel of Figure~\ref{}, one can see other three dispersion curves (in green, purple, and blue colors) presenting the unstable kink mode at the corresponding Alfv\'en Mach numbers. The normalized growth rates of the same clutch of unstable kink waves are presented in the right panel of Figure~\ref{}. For the computation of unstable kink ($m = 1$) mode propagating in axially moving twisted flux tube (the middle column in Figure~\ref{}), our choice for the magnetic field twist parameter is $\varepsilon = 0.025$. The other two input \begin{figure}[!ht] \centerline{\hspace*{0.015\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg6a.eps} \hspace*{-0.03\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg6b.eps} } \caption{(Left panel) Dispersion curves of unstable kink ($m = 1$) mode propagating in a moving twisted magnetic flux tube at $\eta = 0.675$, $b = 1$, and $\varepsilon = 0.025$ and at four various values of Alfv\'en Mach number $M_\mathrm{A}$. The threshold Alfv\'en Mach number for KHI occurring is equal to $2.30839$ (red curve). (Right panel) The normalized growth rates of the unstable kink mode for the same values of the input parameters. The purple curve has been calculated at $M_\mathrm{A} = 2.33$.} \label{} \end{figure} parameters are $\eta = 0.675$ and $b = 1$. The dispersion curves and corresponding normalized growth rates of the unstable mode for four values of the Alfv\'en Mach number are displayed in Figure~\ref{}. There is an unexpected peculiarity---the threshold Alfv\'en Mach number for instability arising turns out to be lower than that in untwisted flux tube, notably equal to $2.30839$, which yields a critical flow speed of ${\approx}308$~km\,s$^{-1}$---$24$~km\,s$^{-1}$ less than the average jet speed. An extensive study of KHI in an EUV jet situated on the west side of NOAA AR 10938 and observed on board \emph{Hinode\/} on 2007 January 15/16 showed just the opposite inequality: $2.4068$ vs $2.354327$, or $114.4$~km\,s$^{-1}$ vs $112$~km\,s$^{-1}$ \citep{Zhelyazkov16}. It is curious to see what will be the wave growth rate, $\gamma_\mathrm{KH}$, instability developing time, $\tau_\mathrm{KH} = 2\pi/\gamma_\mathrm{KH}$, and the kink wave phase velocity, $v_\mathrm{ph}$, for a given wavelength. Bearing in mind that our jet has a width of $4$~Mm and height of $152$~Mm, if we chose $\lambda_\mathrm{KH} = 4$~Mm to be a reasonable wavelength of the unstable mode, then the aforementioned instability parameters, determined by the crossed points of cyan lines and red marginally dispersion and growth rate curves at $k_z a = 3.141592$ in Figure~\ref{}, are: \[ \gamma_\mathrm{KH} = 31.7 \times 10^{-3}~\mathrm{s}^{-1}, \quad \tau_\mathrm{KH} = 198.4~\mathrm{s} = 3.3~\mathrm{min}, \quad \mbox{and} \quad v_\mathrm{ph} = 202.3~\mathrm{km}\,\mathrm{s}^{-1}, \] \begin{figure}[!ht] \centerline{\hspace*{0.015\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg7a.eps} \hspace*{-0.03\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg7b.eps} } \caption{(Left panel) Dispersion curves of unstable kink ($m = 1$) mode propagating in a moving twisted magnetic flux tube surrounded by a twisted magnetized plasma at $\eta = 0.675$, $b = 1$, $\varepsilon_1 = 0.025$, $\varepsilon_2 = 0.001$ and at five various values of Alfv\'en Mach number $M_\mathrm{A}$. The two cyan curves labeled by $M_\mathrm{A} = 3.4$ are associated with the propagation of spurious unstable kink ($m = 1$) mode. The threshold Alfv\'en Mach number for KHI occurring is equal to $3.6075$ (red curve). (Right panel) The normalized growth rates of the unstable and spurious unstable kink mode for the same values of the input parameters. The green and purple curves have been calculated at $M_\mathrm{A} = 3.7$ and $3.8$, respectively.} \label{} \end{figure} respectively. As seen, the KHI evolution time of $3.3$~min is much less than the jet lifetime of $16$~min (see Table 1 in \citealp{Joshi17}). Note that the two cross points in Figure~\ref{} can be considered as a ``phase portrait'' of KHI in the dimensionless phase velocity--wavenumber-plane. In the most complicated case when the external magnetic field is also twisted (the right column in Figure~\ref{}), as we have mentioned, it is necessary to introduce two magnetic field twist parameters, $\varepsilon_1$ for the internal field, and $\varepsilon_2$ for the external one, respectively. We take $\varepsilon_1$, as in the previous case, to be equal to $0.025$; our choice for $\varepsilon_2$ is $0.001$. Thus, with $\eta = 0.403$, $b = 1$, $\varepsilon_1 = 0.025$, and $\varepsilon_2 = 0.001$, the solutions to the wave dispersion relation (\ref{}) of the kink ($m = 1$) mode at five values of the Alfv\'en Mach number are graphically presented in Figure~\ref{}. Here, we are faced with a distinctly different issue, namely at relatively low Alfv\'en Mach numbers, in the very long wavelength limit, $k_z a \ll 1$, one appears two branches of the dispersion curve separated by a gap (see the cyan curves in Figure~\ref{}). With increasing the magnitude of the Alfv\'en Mach number, that gap becomes narrower and at some $M_\mathrm{A}$ the two branches merge forming a continuous dispersion curve---this is the marginal dispersion curve and the corresponding Alfv\'en Mach number is the threshold one for appearance of the KHI---in our case its value is $3.6075$. We would like to notice that the red normalized growth rate curve has no cusp at $k_z a = 0.00168$---it is a normal smooth curve. For larger values of $k_z a$ both the dispersion and growth rate curves go with gradually increasing magnitudes. But at such high threshold Alfv\'en Mach number $M_\mathrm{A} = 3.6075$ and Alfv\'en speed of $132.46$~km\,s$^{-1}$, the required flow speed for instability onset is equal to $478$~km\,s$^{-1}$---a value which is inaccessible by the jet under consideration. Hence, even at a very small twisted external magnetic field suppresses the KHI occurrence. \subsection{Kink mode propagation characteristics at a density contrast of $0.403$} \label{} With the lower value of the density contrast, $\eta = 0.403$, the magnitudes of sound speeds in both media are unchanged, but those of the Alfv\'en speeds are changed and now the ordering of four basic speeds is \[ c_\mathrm{Ti} < v_\mathrm{Ai} < c_\mathrm{Te} < c_\mathrm{si} < c_\mathrm{se} < v_\mathrm{Ae}, \] which implies (see TABLE I in \citealp{Cally86}) that the kink mode propagating in a rest untwisted flux tube should be a surface wave of S$^{-}_{+}$ type. The normalized magnitude of the kink speed is $c_\mathrm{k}/v_\mathrm{Ai} = 2.03457$. The mode will become unstable against KHI if $M_\mathrm{A} > 4.496$. The input parameters for the numerical solution of dispersion equation (\ref{}) now are: $\eta = 0.403$, $\tilde{\beta}_\mathrm{i} = 5.6899$, $\tilde{\beta}_\mathrm{e} = 0.6282$, and $b = 2.192$. With these input values, at $M_\mathrm{A} = 0$ (static magnetic flux tube) numerical computations confirm that the kink ($m = 1$) mode is a pure surface wave and in the long wavelength limit, $k_z a \ll 1$, recover the normalized kink speed of $2.03457$ within three places behind the decimal point. In searching that Alfv\'en Mach number at which the KHI will merge, we use another strategy, notably starting the computations with a higher than the predicted $4.496$ value of $M_\mathrm{A}$, and latter on decreasing it (with small steps) until reach the marginal dispersion and normalized growth rate curves---the results are \begin{figure}[!ht] \centerline{\hspace*{0.015\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg8a.eps} \hspace*{-0.03\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg8b.eps} } \caption{(Left panel) Dispersion curves of unstable kink ($m = 1$) mode propagating in a moving untwisted magnetic flux tube at $\eta = 0.403$, $b = 2.19$, and at various values of $M_\mathrm{A}$. The threshold Alfv\'en Mach number for KHI occurring is equal to $4.651$ (red curve). (Right panel) The normalized growth rates of the unstable kink mode for the same values of the input parameters. The green and purple curve have been calculated at $M_\mathrm{A} = 4.66$ and $4.67$, respectively.} \label{} \end{figure} \begin{figure}[!ht] \centerline{\hspace*{0.015\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg9a.eps} \hspace*{-0.03\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg9b.eps} } \caption{(Left panel) The same as in Figure~\ref{}, but for $\eta = 0.403$ (cool environment) and $b = 2.19$. The threshold $M_\mathrm{A}$ for KHI occurring is equal to $4.67$ (red curve). (Right panel) The green and purple curves have been calculated at $M_\mathrm{A} = 4.68$ and $4.69$, respectively.} \label{} \end{figure} shown in Figure~\ref{}. The threshold Alfv\'en Mach number is equal to $4.651$ which means that with $v_\mathrm{Ai} = 63.2$~km\,s$^{-1}$ the critical flow velocity for KHI onset is $294$~km\,s$^{-1}$, which is lower than the average jet speed of $332$~km\,s$^{-1}$. When the internal magnetic field is twisted and the external medium is cool magnetized plasma, we have to numerically solve dispersion Equation~(\ref{}), but calculating the external wave attenuation coefficient $\kappa_\mathrm{e}$ by using expression~(\ref{}). With the input parameters $\eta = 0.403$, $b = 2.19$, and $\varepsilon = 0.025$, the numerical solutions for the unstable ($m = 1$) kink mode are displayed in Figure~\ref{}. The threshold Alfv\'en Mach number for instability onset being equal to $4.67$ is higher than that for untwisted moving flux tube, but critical flow velocity (calculated with $v_\mathrm{Ai} = 63.18$~km\,s$^{-1}$) of $295$~km\,s$^{-1}$ is still lower that the average jet speed. This means that the KHI should arise and at an instability wavelength $\lambda_\mathrm{KH} = 4$~Mm (that is, at $k_z a = 3.141592$) the instability \begin{figure}[!ht] \centerline{\hspace*{0.015\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg10a.eps} \hspace*{-0.03\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg10b.eps} } \caption{(Left panel) Dispersion curves of unstable kink ($m = 1$) mode propagating in a moving twisted magnetic flux tube surrounded by a cool twisted magnetized plasma at $\eta = 0.675$, $b = 2.19$, $\varepsilon_1 = 0.025$, $\varepsilon_2 = 0.001$ and at five various values of $M_\mathrm{A}$. The cyan curve labeled by $M_\mathrm{A} = 4.9$ simulates spurious unstable kink ($m = 1$) mode. The threshold $M_\mathrm{A}$ for KHI occurring is equal to $5.0375034$ (red curve). (Right panel) The normalized growth rates of the unstable and spurious unstable kink mode for the same values of the input parameters. The green and purple curves have been calculated at $M_\mathrm{A} = 5.08$ and $5.11$, respectively.} \label{} \end{figure} wave growth rate, instability developing/evolution time, and mode phase velocity are: \[ \gamma_\mathrm{KH} = 17.0 \times 10^{-3}~\mathrm{s}^{-1}, \quad \tau_\mathrm{KH} = 369.5~\mathrm{s} = 6.2~\mathrm{min}, \quad \mbox{and} \quad v_\mathrm{ph} = 204.6~\mathrm{km}\,\mathrm{s}^{-1}, \] respectively. It is immediately seen that now the KHI instability evolution time is $1.88$ times longer than that in the case of $\eta = 0.675$, while the wave phase velocities are of the same order. In this case, when the environment is a cool medium, looking at the right panel in Figure~\ref{}, one recognizes that there exists a peculiarity, notably the instability ceases at some dimensionless wavenumber---in our case is it equal to $4.017$. If one needs a little bit shorter instability wavelength $\lambda_\mathrm{KH}$, that is bigger $k_z a$, one has to increase the magnitude of the threshold Alfv\'en Mach number. This circumstance is not too dangerous because there is room for KHI onset at threshold $M_\mathrm{A}$ up to $5.25$. Twisted external magnetic field, as in the previous case of $\eta = 0.675$ when the environment was considered as incompressible medium, dramatically change dispersion characteristics of the kink mode. Solving dispersion relation (\ref{}) with $\eta = 0.403$, $b = 2.19$, $\varepsilon_1 = 0.025$, and $\varepsilon_2 = 0.001$, we obtain families of dispersion and normalized growth rate curves pictured in Figure~\ref{}. The distinctly peculiar shape of the dispersion and growth rate curves at the very beginning of the $k_z a$-axis is seen in Figure~\ref{}. That figure is similar to Figure~\ref{}, but \begin{figure}[!ht] \centerline{\hspace*{0.015\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg11a.eps} \hspace*{-0.03\textwidth} \includegraphics[width=0.515\textwidth,clip=]{na-izh-fg11b.eps} } \caption{(Left panel) A zoom of dispersion and growth rate curves of the kink ($m = 1$) mode as displayed in Figure~\ref{} at the very beginning of the horizontal $k_z a$-axis.} \label{} \end{figure} note that the cusp-like growth rate curves are shifted on one order to the right. Cyan-colored curves in both figures (\ref{} and \ref{}) calculated at $M_\mathrm{A} = 4.9$ are associated with a spurious unstable kink mode---a really unstable kink wave arises at $M_\mathrm{A} = 5.0375034$ (red curves in the two figures). By contrast to the previous case of $\eta = 0.675$, now the KHI instability of the kink mode can merge at a flow velocity of ${\approx}318$~km\,s$^{-1}$ which is lower that the average jet speed of $332$~km\,s$^{-1}$. Going back to Figure~\ref{}, it is intriguing to see how the twist of the external magnetic filed changes the KHI characteristics of the kink mode at the instability wavelength $\lambda_\mathrm{KH} = 4$~Mm compared with those calculated for the previous case, when the internal magnetic field is twisted only. The cross points of vertical cyan lines, plotted at $k_z a = 3.141592$, with corresponding marginal dispersion and normalized growth rate curves yield the following values: \[ \gamma_\mathrm{KH} = 109.7 \times 10^{-3}~\mathrm{s}^{-1}, \quad \tau_\mathrm{KH} = 57.3~\mathrm{s} \cong 1~\mathrm{min}, \quad \mbox{and} \quad v_\mathrm{ph} = 271~\mathrm{km}\,\mathrm{s}^{-1}. \] We see that now the instability evolution time, $\tau_\mathrm{KH}$, is much shorter (${\cong}1$~min), that is KHI is faster, and the phase velocity of the unstable kink ($m = 1$) mode is higher, $271$ vs $205$~km\,s$^{-1}$. \section{Conclusion} \label{} We have explored the possibilities for arising the KHI of the kink ($m = 1$) mode in a fast EUV jet observed by \emph{SDO}/AIA in the NOAA active region 12035 on April 16, 2014. We have modeled the jet as a moving cylindrical magnetic flux tube of radius $a$ at three different magnetic fields' topologies, namely untwisted homogeneous magnetic fields inside and outside the tube, twisted internal magnetic field and untwisted one in the environment, and both twisted magnetic fields. Depending on the value of the density contrast (be it $\eta = 0.675$ or $0.403$), the magnitude of the background magnetic field $B_\mathrm{e}$, and magnetic field topology, we treat the internal and external media as compressible plasmas (when the moving tube is untwisted), or as incompressible internal and external media (at $\eta = 0.675$), or as incompressible internal and cool environment media (at $\eta = 0.403$) when one or both magnetic fields are twisted. Studying the kink ($m = 1$) MHD mode propagation in an untwisted moving magnetic flux tube allows us to determine depending on the ordering of sound, Alfv\'en, and tube velocities in both media the nature of the propagating mode (pure surface or pseudo surface wave \citep{Cally86}) along with the value of the characteristic kink speed $c_\mathrm{k}$ (\ref{})---both issues in a rest flux tube. The normalized with respect $v_\mathrm{Ai}$ kink speed must be reproduced in the $k_z a \ll 1$ limit during the numerical solution of dispersion equation (\ref{}) if the numerical code is correct. The inclusion of jet speed via the Alfv\'en Mach number does not change the wave nature until the kink mode is stable, but splits the kink speed dispersion curves into a pair of curves whose behavior might be generally rather complex as illustrated in the left panel of Figure~\ref{}. The threshold $M_\mathrm{A}$ can be predicted on using criterion~(\ref{}), but its actual value is given by the computed unstable normalized wave phase velocity and growth rate curves (termed marginal ones and plotted in red color). Numerically derived threshold Alfv\'en Mach numbers turn out to be very close to the predicted values for the two sets of input data of both density contrasts. If the critical jet velocity for KHI onset is accessible, one can, in principle, evaluate at a fixed wavelength the instability developing time and the corresponding wave phase velocity. We note that when the kink mode becomes unstable both attenuation coefficients determining its spatial structure in the two media are complex quantities and it is logical such a wave to be called \emph{generalized surface mode}, that is, being neither pure surface, pseudo surface/body, or leaky wave. Observations show that in the most cases solar atmosphere jets are more or less twisted. That is why the focus of our study was to see how the magnetic field twist changes the picture. At $\eta = 0.675$ with $b = 1$ and $\varepsilon = 0.025$, the threshold $M_\mathrm{A}$, unexpectedly was found to be lower than the one for untwisted magnetic flux tube, namely $2.30839$ vs $2.4425$. At the other $\eta = 0.403$, with $b = 2.19$ and $\varepsilon = 0.025$, we have just the opposite relation: KHI instability onset of the kink ($m = 1$) mode requires a little bit higher critical jet speed than that for untwisted tube: $295$ vs $294$~km\,s$^{-1}$. It is instructive to note that the difference between the corresponding threshold Alfv\'en Mach numbers is more emphatic ($4.67$ vs $4.651$), but the slightly different Alfv\'en speeds ($63.18$ and $63.2$~km\,s$^{-1}$, respectively) make the critical flow velocities difference rather small---$1$~km\,s$^{-1}$ only. Another important observation is the different shapes of normalized growth rate curves, more specifically, in the case when the environment is treated as a cool medium ($\eta = 0.403$), the growth rate curve is limited on the right-hand side of the horizontal axis in the Im($v_\mathrm{ph}/v_\mathrm{Ai})$--$k_z a$-plane. In such a case, if one needs a wider $k_z a$ instability range, one must increase the threshold $M_\mathrm{A}$. It is worth mentioning that irrespective of slightly different background magnetic fields ($7$ and $8$~G, respectively), the KHI developing times of the kink mode at $\lambda_\mathrm{KH} = 4$~Mm are not too distinct though at $\eta = 0.304$ the evolution time of $6.2$~min is approximately two times longer than the one at $\eta = 0.675$ being equal to $3.3$~min. On the other side, the wave phase velocities of the unstable kink mode are very close: $202.3$~km\,s$^{-1}$ for $\eta = 0.675$ and $204.6$~km\,s$^{-1}$ at $\eta = 0.403$. A definitely new achievement in our paper is the study of how the external twisted magnetic field affects the development of KHI. In that case, in addition to the internal magnetic field twist parameter, now subscribed by `1', that is, $\varepsilon_1 = 0.025$, we had to introduce a second twist parameter characterizing the external magnetic field, whose value is relatively small: $\varepsilon_2 = 0.001$. The influence of the twisted external magnetic field $\vec{B}_\mathrm{e}$ on the instability onset critically depends upon the nature of surrounding coronal plasma. If the environment is treated as incompressible plasma, that is, $\eta = 0.675$, the external twisted magnetic field suppresses the instability arising: with a threshold Alfv\'en Mach number equal to $3.6075$ (see Figure~\ref{}) and Alfv\'en speed $v_\mathrm{Ai} = 232.46$~km\,s$^{-1}$, the required critical flow velocity of ${\cong}478$~km\,s$^{-1}$ for the instability triggering is far beyond the average jet speed of $332$~km\,s$^{-1}$. In the case when the surrounding medium is considered as cool plasma ($\eta = 0.403$), it turns out that the kink ($m = 1$) mode can become unstable at an accessible critical jet speed of ${\approx}318$~km\,s$^{-1}$. Comparing the KHI evolution times of the kink mode propagating in single-twisted and double-twisted moving magnetic flux tube we establish that in the second case $\tau_\mathrm{KH}$ is much shorter---it is below $1$~min, more precisely $57$~s vs $370$~s in a single-twisted flux tube. The wave phase velocity of the unstable kink mode is, however, with $66$~km\,s$^{-1}$ higher than that velocity in a single-twisted tube been equal to ${\cong}205$~km\,s$^{-1}$. Our study shows that the KHI occurrence of the kink ($m =1$) mode primarily depends on the density contrast between the two media: the jet and its environment. Here comes the big question: ``Which is the adequate approach of defining the density contrast: the standard one or that suggested by \cite{Paraschiv15}?'' The answer can only be obtained when we have more information about the parameters of an observationally recorded KHI, for example like that in a coronal mass ejection registered by \cite{Foullon11} and \cite{Ofman11}, namely observationally deduced wavelength $\lambda_\mathrm{KH}$, instability growth rate $\gamma_\mathrm{KH}$ and wave phase velocity $v_\mathrm{ph}$ to be compared with their values obtained from numerically derived growth rate and dispersion curves. The Kelvin--Helmholtz instability is an important instability as we have mentioned in the Introduction section because it is capable of converting well-ordered flows, such solar EUV and X-ray jets, into more disordered, even turbulent, flows, which can lead to a heating of the solar atmosphere. \begin{center} Acknowledgments \end{center} The work of M.B.\ and I.Zh.\ was supported by the Bulgarian Science Fund under Indo-Bulgarian bilateral project DNTS/INDIA 01/7, and that of R.J.\ and R.C.\ by the Department of Science \& Technology, Government of India Fund under the project /Int/Bulgaria/P-2/12. The authors are thankful to the \emph{Solar Dynamic Observatory}, the data from which are used in the present investigation. We are also indebted to Dr.~Snezhana Yordanova for plotting one figure.

\section{Introduction} Image-to-image translation has been made much progress~\cite{Isola_2017_CVPR,Zhu_2017_ICCV,Yi_2017_ICCV,Chen_2017_ICCV,kim2017learning,mao2018semantic} because many tasks~\cite{zhou2016similarity,wang2016super,ma2017unsupervised,shi2017end} in image processing, computer graphics, and computer vision can be posed as translating an input image into a corresponding output image~\cite{zhu2016generative,li2016precomputed,wang2016generative,Ledig_2017_CVPR,taigman2017unsupervised,Sela_2017_ICCV,Tung_2017_ICCV}. And its achievements mainly owes to the success of Generative Adversarial Networks (GANs)~\cite{goodfellow2014generative}, especially conditional GANs (cGANs)~\cite{mirza2014conditional,radford2016unsupervised,Isola_2017_CVPR}. However, the current studies mainly concern image-to-image translation tasks with low-level visual information conversion, \emph{e.g.}, photo-to-sketch~\cite{liu2018auto}. A caricature is a rendered image showing the features of its subject in an exaggerated way and usually used to describe a politician or movie star for political or entertainment purpose. Creating caricatures can be considered as artistic creation tracing back to the 17th century with the profession of caricaturist. Then some efforts have been made to produce caricatures semi-automatically using computer graphics techniques~\cite{akleman2000making}, which intend to provide warping tools specifically designed toward rapidly producing caricatures. But there are very few software programs designed specifically for automatically creating caricatures, and to the best of our knowledge, none can work to be comparable with caricaturist. Nowadays, besides the political and public-figure satire, caricatures are also used as gifts or souvenirs, and more and more museums dedicated to caricature throughout the world were opened. So it would be very useful and meaningful if computers can create caricatures from photos automatically and intelligently. Photo-to-caricature is a typical high-level image-to-image translation problem but with bigger challenge than other low-level translation problems such as photo-to-label, photo-to-map, or photo-to-sketch~\cite{Isola_2017_CVPR}, because caricatures \begin{itemize} \item require \textbf{satire} and \textbf{exaggeration} of photos; \item need \textbf{artistry} with different styles; \item must be \textbf{lifelike}, especially the expression of a face photo. \end{itemize} Specifically, for a face photo, we want to create the face caricatures with different styles, which exaggerate the face shape or facial sense organs (\emph{i.e.}, ears, mouth, nose, eyes and eyebrow) but keep the vivid expression while producing artistry. In this paper, we propose a GAN-based method for learning to translate faces in the wild from the source photo domain to the target caricature domain (see Figure~\ref{} for translating examples and Figure~\ref{} for the architecture of our proposed method). Although deep convolutional neural networks with adversarial training~\cite{radford2015unsupervised} can generate images with enough precise facial features~\cite{berthelot2017began}, these images sometimes still have wrong relationships between facial features or mismatch among facial organs such as nose and eyes, \emph{e.g.}, a face with more than two eyes or crooked nose. Traditional GANs can produce correct facial organs but wrong relationships between them, and also it's very challenging to abstract and exaggerate face and facial organs. We attribute this problem to the deficient capacity of the discriminator of GAN to distinguish real-fake images. Therefore, our motivation is to design an adversarial training with multiple discriminators to improve the ability of GAN's discriminator for feature representation. \begin{figure}[!ht] \centering \includegraphics[width=\linewidth]{demo.pdf} \caption{Translating faces in the wild from photo to caricature with different styles by our proposed method. (a) Example results on IIIT-CFW dataset~\cite{mishra2016iiit}; (b) Example results on PHOTO-SKETCH dataset~\cite{zhang2011coupled,wang2009face}.} \label{} \end{figure} Based on the model of CycleGAN~\cite{Zhu_2017_ICCV}, we design a dual pathway model of GAN for high-level image-to-image translation tasks, where one pathway of coarse discriminator is in charge of abstracting the global structure information, while another pathway of fine discriminator is responsible for concerning the local statistics information. For generator, besides the adversarial loss, we provide one more extra perceptual similarity loss to constrain consistency for generated output itself and with the unpaired target domain image. By using our proposed method, the photos of faces in the wild can be translated to caricatures with learned general-purpose exaggerated artistic styles while still keeping the original lifelike expression (see Figure~\ref{} and \ref{} for references). Considering that traditional GANs are not robust and easily attracted by noise, we design a noise-added training procedure to improve the robustness of our model. Inspired by InfoGAN~\cite{chen2016infogan}, we find that auxiliary noise can help model learning the caricature style information while translating images in our task. We have extensively evaluated our method on IIIT-CFW dataset~\cite{mishra2016iiit}, PHOTO-SKETCH dataset~\cite{zhang2011coupled,wang2009face}, Caricature~\cite{abaci2015matching}, FEI dataset~\cite{thomaz2010new}, Yale dataset~\cite{georghiades1997yale}, KDEF dataset~\cite{lundqvist1998karolinska} and CelebA dataset~\cite{liu2015faceattributes}. The experimental results show that our method can create acceptable caricatures from face photos while current state-of-the-art image-to-image translation methods can't. Also the designed experiments indicate the effectiveness of our proposed dual pathway of discriminators, additional noise input and extra perceptual loss, respectively. Besides, we tested our photo-to-caricature translation method for producing caricatures with adding different proportions of noise to show the translating robustness and style diversity. Furthermore, the proposed method can create caricatures for arbitrary face photos without pre-training on extra face datasets. Another prominent performance of our methods is that our model can capture the expression information and make some abstraction and exaggeration. This might be helpful to fill aforementioned gap of automatic and intelligent caricature creation. \section{Related Work} \textbf{Image-to-image translation.} Owing to the success of GANs~\cite{goodfellow2014generative}, especially various conditional GANs~\cite{mirza2014conditional,mathieu2015deep,reed2016generative,yan2016attribute2image,wang2016generative,zhang2017image}, image-to-image translation problems have been made much progress recently, which aims to translate an input image from one domain to another domain given input-output images pairs~\cite{Isola_2017_CVPR}. Earlier image-conditional models for specific applications have achieved impressive results on inpainting~\cite{Pathak_2016_CVPR}, de-raining~\cite{zhang2017image}, texture synthesis~\cite{li2016precomputed}, style transfer~\cite{wang2016generative}, video prediction~\cite{mathieu2015deep} and super-resolution~\cite{Ledig_2017_CVPR}. The general-purpose solution for image-to-image translation developed by Isola \emph{et al.}~\cite{Isola_2017_CVPR} with the released \verb|Pix2pix| software has achieved reasonable results on many translation tasks by using paired images for training such as photo-to-label, photo-to-map and photo-to-sketch. Then CycleGAN~\cite{Zhu_2017_ICCV}, DualGAN~\cite{Yi_2017_ICCV}, and DiscoGAN~\cite{kim2017learning} were proposed for unpaired or unsupervised image-to-image translation with almost comparable results to paired or supervised methods. However, the translation tasks that these work can tackle usually concern the conversion of low-level visual information such as line (photo$\rightarrow$sketch), color (summer$\rightarrow$winter), and texture (horse$\rightarrow$zebra), but it's still very challenging for some translation tasks with the requirement of high-level visual information conversion, \emph{e.g.}, abstraction and exaggeration (photo$\rightarrow$caricature). Recently, Iizuka \emph{et al.}~\cite{iizuka2017globally} combined two discriminators called global discriminator and local discriminator to improve the adversarial training for image completion task. Experimental results have proven that the global discriminator can distinguish the images based on the global parts while the local discriminator pays attention to the details of parts. We exploit the design of two discriminators for high-level image-to-image translation tasks in this paper. \textbf{Photo-to-cartoon translation.} Translating photo to cartoon by computer algorithms has been studied for a long time because of the corresponding interesting and meaningful applications. The earlier work mainly relied on the analysis of facial features~\cite{luo2002exaggeration,chen2004automatic,liao2004automatic}, which is hard to be applicable for large-scale faces in the wild. Thanks to the invention of GANs~\cite{goodfellow2014generative}, automatic photo-to-cartoon translation becomes feasible. But most of the current related work mainly focused on the generation of anime~\cite{zhang2017style} or emoji~\cite{taigman2017unsupervised} with specific style\footnote{We consider that anime, emoji, and caricature are three types of cartoon or three subsets of cartoon.}. And these works have nothing to do with caricature creation that needs to be exaggerated, lifelike and artistic. Unlike the prior works for image-to-image translation dealing with low-level visual information conversion, our study mainly focuses on the translation problems of high-level visual information conversion, \emph{e.g.}, photo-to-caricature. For this purpose, our method differs from the past work in network architecture as well as the layers and losses. We design a dual pathway model of GAN with two discriminators named coarse discriminator and fine discriminator to capture global structure and local statistics respectively, and we apply the perceptual similarity loss for generator. For learning the style information and improving the robustness of our model, we provide the input with auxiliary noise. Here we show that our approach is effective on the face photo-to-caricature translation task, which typically requires high-level visual information conversion. \section{Method} Conditional GANs are generative models that learn a mapping from observed input image $x$, to target image $y$, $G:\left\{x\right\}\rightarrow y$. The objective of our conditional GAN can be expressed as \begin{equation} \begin{split} \mathcal{L}_{c}(G,D)=\mathbb{E}_{x,y\sim P_{data}(x,y)}\left[\log D(x,y)\right]+\\ \mathbb{E}_{x\sim P_{data}(x)}\left[\log (1-D(x,G(x)))\right]. \end{split} \label{} \end{equation} The goal is to learn a generator distribution over data $y$ that matches the real data distribution $P_{data}$ by transforming an observed input image $x\sim P_{data}(x)$ into a sample $G(x)$. This generator is trained by playing against an adversarial discriminator $D$ that aims to distinguish between samples from the true data distribution and the generator's distribution. To deal with the high-level visual information conversion for some more challenging image-to-image translation tasks, we exploit one discriminator to dual discriminators and add two more losses (perceptual loss and cycle consistency loss) to adversarial loss, so our method can tackle conversions of both the low-level (line, color, texture, \emph{etc.}) and the high-level (expression, exaggeration, artistry, \emph{etc.}) visual information. And in order to improve the robustness of our model, we provide our model a noise-added training and use auxiliary noise to learn the style information while translating. The overall network architecture and data flow are illustrated in Figure~\ref{}. \begin{figure}[!ht] \centering \includegraphics[width=\textwidth]{architecture.pdf} \caption{Network architecture and data flow chart of our proposed method for face photo-to-caricature translation.} \label{} \end{figure} \subsection{Cycle consistency loss} Researchers apply adversarial training to learn the mapping function between two different image domains. Here we use two same generators (named \verb|G1| and \verb|G2| in Figure~\ref{}) for observing the sample distribution of two domains. And $X$ denotes the photo domain while $Y$ denotes the caricature domain in our task. We use our generators to emulate the translation between $X$ and $Y$. However, only using adversarial loss can't guarantee a plausible generation. So researchers use Cycle Consistency Loss ($\mathcal{L}_{cyc}$) to help establishing mapping function between domains~\cite{Zhu_2017_ICCV}. And $\mathcal{L}_{cyc}$ is expressed as \begin{equation} \begin{split} \mathcal{L}_{cyc}(G1,G2)=\mathbb{E}_{x\sim P_{data}(x)}\left[||G_{2}(G_{1}(x))-x||_{1}\right]\\ +\mathbb{E}_{y\sim P_{data}(y)}\left[||G_{1}(G_{2}(y))-y||_{1}\right], \end{split} \end{equation} where $G_{1}$ and $G_{2}$ represent the two generators, and $x$ and $y$ are samples from $X$ and $Y$ domain respectively. We use $L1$ loss for the cycle consistency loss following Zhu \emph{et al.}~\cite{Zhu_2017_ICCV}. \subsection{Perceptual loss} \label{} To further reduce the space gap of possible mapping functions between domains, we apply the perceptual loss $\mathcal{L}_{p}$ for our model. For a constrained translation problem, finding an appropriate loss function is critical. We adopt the content loss of Gatys \emph{et al.}~\cite{gatys2016image}, which is also referred as a perceptual similarity loss or feature matching~\cite{bruna2015super,dosovitskiy2016generating,johnson2016perceptual,ledig2016photo}. We apply the perceptual loss to our model followed by the cycle consistence loss, and compute the perceptual loss between unpaired images from different domains to push generator to capture the feature representations. Let $\phi$ denotes a pre-trained visual perception network (we use pre-trained \verb|VGG19| in our experiments) and $n$ denotes the number of feature maps. Different layers in the networks represent low-to-high level information: from edges and color to object and semantic representation. Matching both low and high layers in the perception network can help achieving fantastic translation. And the perceptual loss can be expressed as \begin{equation} \mathcal{L}_{p}=\sum^{N}_{n}{\lambda}_{n}||{\Phi}_{n}(y)-{\Phi}_{n}(G(x);\theta))||_{1}, \end{equation} where $y$ denotes the image from caricature domain in our task and $G(x)$ denotes the synthesized caricature image using $x$. Note that these two images are unpaired. \subsection{Auxiliary noise input} Previous researches have fully proved that we can get plausible image results from noise input~\cite{radford2016unsupervised,chen2016infogan,zhao2016energy,berthelot2017began}. In order to improve the robustness and enrich the diversity of image translation between domains, we design a noise-added training procedure before the translation shown in Figure~\ref{}. First we obtain a random noise input from random uniform distribution (range from $0$ to $255$), then we merge the noise input and the raw image input to acquire the final input using approximate weights. Here we define $\alpha$ to denote the proportion of the raw image accounting for the final image. This can be expressed as \begin{equation} \label{} x=x_{i}*\alpha+(1-\alpha)*n, \quad 0\leq\alpha\leq1, \end{equation} where $x_{i}$ denotes the raw input from photo domain and $n$ denotes the noise that has a uniform distribution $P_{noise}$. Figure~\ref{} shows one sample with adding noise. With auxiliary noise input, the GAN object is expressed as: \begin{equation} \begin{split} L_{c}(G,D)=\mathbb{E}_{x,y\sim P_{data}(x,y)}\left[\log D_{p}\left(x,y\right)\right] \\ + \mathbb{E}_{x\sim P_{data}(x),n\sim P_{noise}(n)}\left[\log (1-D(x,G(x,n)) \right]. \end{split} \end{equation} \begin{figure}[!ht] \centering \includegraphics[width=0.8\textwidth]{add_noise.pdf} \caption{Sample for adding noise, here we use $\alpha=0.5$.} \label{} \end{figure} We add the auxiliary noise to improve the robustness of our model. And we can get different styles of output through adjusting the noise input (see Section~\ref{}). Besides, we add the noise from a uniform distribution for making the style information more representable and matching the style space while translating. Furthermore, we also hope to control the style by adding different noise and make the translating conditional, which will be our future work. \subsection{Dual discriminators} Traditional GANs usually have one generator and one discriminator to leverage adversarial training. Different from them, we design two different discriminators to capture different level information. In our method, we propose two different discriminators called coarse discriminator and fine discriminator respectively. The coarse discriminator aims to encourage generator to synthesize images based on global style and structure information for domain translation. In our high level image-to-image task, the coarse discriminator is capable for capturing the structure information and abstracting the representative information of face photo such as emotion and style. While the fine discriminator aims to achieve the feature matching and help generating more plausible and precise images, and the fine discriminator builds the adversarial training for the face details with generator such as lip and eyes. Different from Satoshi \emph{et al.}~\cite{iizuka2017globally}, we are not using the image patches as input of local discriminator, we provide both the two discriminators with the whole image as input while the outputs of the two discriminators are different (see Figure~\ref{}). The output of coarse discriminator is a $4\times 4$ patch matrix after the sigmoid activity function while the output of fine discriminator is $16\times 16$. Note that both the two discriminators are using sigmoid function at the last layer. We have tried using different output size combinations to get different results, and the experiments show that the combination of $16\times 16$ for fine discriminator and $4\times 4$ for coarse discriminator can obtain the best result for translation. The coarse discriminator has a smaller feature map with more abstractive representation compared to the fine discriminator. The \verb|D1| and \verb|D2| in Figure~\ref{} represent the dual discriminators for translating two different domains $X$ and $Y$. \subsection{Generator} \label{} Previous studies have found it beneficial to mix the GAN objective with a more traditional norm, such as $L2$~\cite{Pathak_2016_CVPR} and $L1$~\cite{Isola_2017_CVPR} distance. We explore this opinion to cycle consistency loss by applying $L1$ distance to compute $L_{cyc}$. Our final objective is \begin{align} \label{} G^{*} = \arg\underset{G}{\min}\;\underset{D}{\max}\mathcal{L}(G,D) + \gamma \mathcal{L}_{cyc} + \sigma \mathcal{L}_{p}, \\ \text{w.r.t}\quad \mathcal{L}(G,D)= \begin{cases} \mathcal{L}_{p}(G,D_{c}) \quad \text{for}\ D_{c},\\ \mathcal{L}_{g}(G,D_{f}) \quad \text{for}\ D_{f}, \end{cases} \end{align} where $\mathcal{L}_{cyc}$ means cycle consistency loss, $\mathcal{L}_{p}$ indicates perceptual loss, and $\gamma$ and $\sigma$ are hyper parameters to balance the contribution of each loss to the objective. We use greedy search to optimize the hyper parameters with $\gamma=10$ and $\sigma=2.0$ for all the experiments in this paper. As shown in Figure~\ref{}, we use Conv-Residual blocks-Deconv~\cite{he2016deep} as the generator to share the low-level and high-level information between the input and output directly across the net. \section{Experiments} As a typical image-to-image translation task, photo-to-caricature requires high-level visual information conversion, which is very challenging for the state-of-the-art general-purpose solutions. To explore the effect of our proposed model, we tested the method on a variety of datasets for translating faces in the wild from photo to caricature with different styles, and the qualitative results are shown in Figs.~\ref{} and~\ref{}. \subsection{Dataset and training} Our proposed model is trained in a supervised unpaired fashion on a paired face photo-caricature dataset, named IIIT-CFW-P2C dataset, which was rebuilt based on IIIT-CFW~\cite{mishra2016iiit}. The IIIT-CFW is a dataset for the cartoon faces in the wild and contains $8928$ annotated cartoon faces of famous personalities in the world with varying profession. Also it provides $1000$ real faces of the public figure for cross modal retrieval tasks. However, it's not suitable for the training of photo-to-caricature translation task using some paired methods (such as Pix2pix) because the face photos and face cartoons\footnote{We consider that caricature is a type of a cartoon or a subset of a cartoon.} are not paired, \emph{e.g.}, the facial orientation and expression of the photo and caricature for the same person are varying a lot. So we rebuild a photo-caricature dataset with $1171$ paired images by searching the IIIT-CFW dataset and Internet as the training set for compared experiments. Here we use $800$ for training and the left for testing. At inference time, we run the generator in exactly the same manner as during the training phase. Besides, we also extensively evaluated our method on a variety of datasets with faces in the wild, including Caricature~\cite{abaci2015matching}, FEI~\cite{thomaz2010new}, IIIT-CFW~\cite{mishra2016iiit}, Yale~\cite{georghiades1997yale}, KDEF~\cite{lundqvist1998karolinska} and CelebA~\cite{liu2015faceattributes}. Besides, we also consider photo-to-sketch as a photo-to-caricature task for experiments using PHOTO-SKETCH dataset~\cite{zhang2011coupled,wang2009face}, which has $1194$ paired images and hence can be directly used for supervised training of some compared paired methods. And following DualGAN, we use $995$ unpaired images for training and $199$ for testing. Note that we train CycleGAN, DualGAN and our model using unpaired images and train Pix2pix using paired images of the two datasets. \subsection{Comparison with state-of-the-arts} Using IIIT-CFW-P2C dataset, we first compare our proposed method with Pix2pix~\cite{Isola_2017_CVPR}, DualGAN~\cite{Yi_2017_ICCV}, DiscoGAN~\cite{kim2017learning} and CycleGAN~\cite{Zhu_2017_ICCV} on photo-to-caricature translation task. All the four methods were trained on the same training dataset and tested on novel data from IIIT-CFW-P2C dataset that does not overlap those for training. \paragraph{Qualitative evaluation} Figure~\ref{} shows the experimental results of the comparison, it can be seen that, DualGAN only learned the color and edge translation rather than structure information, Pix2pix makes structure error in almost all cases, while CycleGAN can keep the structure information of the input image but without enough conversion for caricature creation task. For DiscoGAN, it's really hard to generate plausible and meaningful results due to the lack data of training ($800$ pairs in our experiments vs. tens of thousands of pairs in DiscoGAN's experiments) and the big challenge of task (photo-to-caricature vs. edge-to-photo). Although it's still not good enough for the results of our method compared to human caricaturists, the experiments on photo-to-caricature translation of faces show considerable performance gain of our proposed method over state-of-the-art image-to-image translation methods, especially the encouraging ability of exaggeration and abstraction. However, due to the very challenging task with less training data but on various styles, our method also messes some details while translation (see the mouth of the first row and the eyes of the fourth row on our results in Figure~\ref{}). \begin{figure}[!ht] \centering \includegraphics[width=\textwidth]{compare.pdf} \caption{Comparison of state-of-the-art image-to-image translation methods with our proposed method for face photo-to-caricature translation on IIIT-CFW-P2C dataset.} \label{} \end{figure} This experiment also expresses that high-level image-to-image translation tasks like photo-to-caricature are generally more difficult than those low-level translation tasks such as photo-to-sketch, because it not only needs to abstract the facial features, but also requires to exaggerate the emotional expression. So that the pixel-level methods (like Pix2pix) might fail as they force the generator to concentrate on local information rather than whole structure. Besides, we also evaluate our methods on PHOTO-SKETCH dataset. Figure~\ref{} shows the compared results. Comparing with Pix2pix, our method can reduce the effect of being blurry and artifact. Although DualGAN and CycleGAN can also reserve the structure information of input faces, they are not good at achieving abstraction and artistry. Similarly, DiscoGAN collapses when facing insufficient training data and big challenging task. \begin{figure}[!ht] \centering \includegraphics[width=\textwidth]{compare_sketch.pdf} \caption{Comparison of state-of-the-art image-to-image translation methods with our proposed method for face photo-to-caricature translation on PHOTO-SKETCH dataset.} \label{} \end{figure} \paragraph{Quantitative evaluation} Beyond the visually qualitative evaluation, we also evaluated the translated cartoon results of different methods quantitatively on IIIT-CFW-P2C dataset and PHOTO-SKETCH dataset in terms of both human judge and machine grade, and the average results are shown in Table~\ref{} and Table~\ref{} respectively. For human score, we invited 40 volunteers to evaluate the generated image quality of different methods in terms of satire, exaggeration, lifelikeness and artistry compared with the given original photo, by grading from 1 to 5 (1 represents the worst and 5 represents the best). For inception score, we used a pre-trained classifier network and sampled images for evaluation followed~\cite{salimans2016improved}. The results shows that our proposed method outperforms state-of-the-art image-to-image translation methods with the highest human and inception scores. \begin{table} \centering \caption{The generated image equality evaluation results of different methods on IIIT-CFW-P2C dataset.} \label{} \begin{tabular}{|c|c|c|} \hline Method & Human score & Inception score \\ \hline\hline Pix2pix & 2.0106 & 1.5069$\pm$0.1090\\ DualGAN & 1.9946 & 1.4843$\pm$0.1049\\ DiscoGAN & 1.5014 & 1.3366$\pm$0.0714\\ CycleGAN & 3.5001 & 1.5684$\pm$0.1331\\ Ours & 4.0120 & 1.6043$\pm$0.0918 \\ \hline \end{tabular} \end{table} \begin{table} \centering \caption{The generated image equality evaluation results of different methods on PHOTO-SKETCH dataset.} \label{} \begin{tabular}{|c|c|c|} \hline Method & AMT & Inception score \\ \hline\hline Pix2pix & 2.0971 & 1.3625$\pm$0.0706\\ DualGAN & 2.3810 & 1.4063$\pm$0.0822\\ DiscoGAN & 1.0858 & 1.3142$\pm$0.0206\\ CycleGAN & 3.2272 & 1.3980$\pm$0.1130\\ Ours & 4.0750 & 1.4298$\pm$0.0818 \\ \hline \end{tabular} \end{table} \subsection{Dual discriminators} \label{} In this experiment, we verify the effectiveness of our proposed dual pathway of discriminators. We first use only one coarse discriminator (Coarse D) and one fine discriminator (Fine D) separately, and then use dual discriminators with one coarse discriminator plus one fine discriminator (Fine D + Coarse D), while keeping all other architectures and settings fixed for training and testing. Some example results are shown in Figure~\ref{}, and one Fine D model almost misses the key structure information on faces, but our dual Coarse D + Fine D model can render the structure of facial features well. It further proves that the Fine D model only concerns the local statistics for tackling low-level image-to-image translation tasks (\emph{e.g.}, photo-to-sketch). \begin{figure}[!ht] \centering \includegraphics[width=0.5\textwidth]{candf.pdf} \caption{Comparison of only one discriminator (Coarse D and Fine D respectively) with our dual discriminators (Coarse D + Fine D, C \& F).} \label{} \end{figure} And we also took some experiments to greedy search the best combination size of output patches for coarse discriminator and fine discriminator. Figure~\ref{} shows results of different combinations, which indicates that large Fine D patches (\emph{e.g.}, F$32$) fails to abstract and exaggerate faces while small Fine D patches (\emph{e.g.}, F$8$) abstracts and exaggerates faces too much. \begin{figure}[!ht] \centering \includegraphics[width=\textwidth]{combinations.pdf} \caption{Examples of different combination sizes of output patches for coarse discriminator and fine discriminator. C$k$ denotes that the output of coarse discriminator is a $k\times k$ patch, and F$k$ denotes that the output of fine discriminator is a $k\times k$ patch. It can be seen that large F$k$ such as F$32$ for fine discriminator fails to abstract faces and achieve exaggerations, while small F$k$ such as F$8$ abstracts and exaggerates faces too much.} \label{} \end{figure} \subsection{Loss selection} \label{} We first consider to check if the cycle consistency loss $\mathcal{L}_{cyc}$ should be provided to the GAN objective (Equation~\ref{}), and the second column in Figure~\ref{} shows the results without cycle consistency loss. It's easy to see that without cycle consistency loss, although the adversarial system with adversarial loss can capture the facial features, it's hard to generate caricature images with plausible objects and meaningful relationship between facial organs. The third column in Figure~\ref{} shows the results by adding cycle consistency loss $\mathcal{L}_{cyc}$. Therefore, the normal adversarial training can lead to some kind of caricature style, but it fails to be lifelike without meaningful components. \begin{figure}[!ht] \centering \includegraphics[width=0.5\textwidth]{cycle.pdf} \caption{Comparison of extra loss for final objective of generator: without ($w/o$) $\mathcal{L}_{cyc}$ and with ($w/$) $\mathcal{L}_{cyc}$.} \label{} \end{figure} Based on the cycle consistency loss, we provide the perceptual loss $\mathcal{L}_{p}$ for generator in our system. Figure~\ref{} shows the compared results of without and with $\mathcal{L}_{p}$. It can be seen that perceptual loss can produce images with the exaggerated facial features such as eyes, nose, and mouth. The perceptual loss, which expresses some perceptual errors on facial features such as head, eyes, and mouth, could improve the artistic expression of image generation and show better abstraction ability. And it can also reduce the effect of being blurry. The second column of Figure~\ref{} without using perceptual loss illustrates the indistinguishable facial expressions with distorted facial organs while translation, and the third column with adding perceptual loss improves the performance on facial expression and organ translation with caricature effect, \emph{e.g.}, the smile woman of last row. \begin{figure}[t] \centering \includegraphics[width=0.5\linewidth]{perceptual.pdf} \caption{Comparison of extra loss for final objective of generator: without ($w/o$) $\mathcal{L}_{p}$ and with ($w/$) $\mathcal{L}_{p}$.} \label{} \end{figure} \subsection{Auxiliary noise input} \label{} By adding auxiliary noise to our photo-to-caricature system, we can improve the robustness and diversity of synthesized facial caricatures. Figure~\ref{} shows the example results of adding auxiliary noise for photo-to-caricature translation, and the output results for adding different proportions ($\alpha$) of noise indicate that our system can still synthesize meaningful facial caricatures with even more than a half noise ($1-\alpha$, see Equation~\ref{} for reference) as inputs, besides, the added different proportions of noise also lead to different styles of output results, which indicates that it might be used as a factor for tuning different synthesized styles. \begin{figure}[!ht] \centering \includegraphics[width=\textwidth]{info.pdf} \caption{Examples of adding auxiliary noise for robustness and diversity of our photo-to-caricature translation. Different proportions ($\alpha$ in Equation~\ref{}) of noise inputs can also lead to meaningful different styles of output caricatures.} \label{} \end{figure} \subsection{Freestyle face caricature creation} To evaluate the ability for real applications in the daily life, we tested our method on a variety of face datasets, including Caricature~\cite{abaci2015matching}, FEI~\cite{thomaz2010new}, IIIT-CFW~\cite{mishra2016iiit}, Yale~\cite{georghiades1997yale}, KDEF~\cite{lundqvist1998karolinska} and CelebA~\cite{liu2015faceattributes}, to illustrate the photo-to-caricature translation on faces in the wild, and the results are shown in Figure~\ref{}. These freestyle face caricature creation results validate that our model works not bad on arbitrary faces and show the potential value for the related applications. And we can see that the translated results in KDEF, FEI and Yale datasets also have different facial expression corresponding the input faces. Our methods successfully reserve the emotion information and emulate the facial organs with caricature style. So we can conclude that the more abstracted information such as facial emotion and expression with global structure information are reserved. Besides, our model can also enlarge or narrow the facial organs such as chin, lips, eyes and so on, which is required for high-level image-to-image translation tasks. \begin{figure}[!ht] \centering \includegraphics[width=\linewidth]{free.pdf} \caption{Example translated caricatures of facial photos from several datasets (Caricature~\cite{abaci2015matching}, FEI~\cite{thomaz2010new}, IIIT-CFW~\cite{mishra2016iiit}, Yale~\cite{georghiades1997yale}, KDEF~\cite{lundqvist1998karolinska} and CelebA~\cite{liu2015faceattributes}) using our trained model on IIIT-CFW-P2C.} \label{} \end{figure} \section{Conclusion and Future Work} We present a novel GAN-based method to deal with high-level image-to-image translation task, \emph{i.e.}, photo-to-caricature translation. The proposed method uses dual discriminators for capturing global structure and local statistics information with abstraction ability, also provides extra perceptual loss on GAN objective to constrain the consistency under exaggeration. Besides, the style information can be learned and representative by adding auxiliary noise input. And the robustness can be improved by the noise-added training. Experimental results show that our method not only outperforms other state-of-the-art image-to-image translation methods, but also works well on a variety of datasets for photo-to-caricature translation of faces in the wild. \textbf{Limitations.} Translating photo to caricature is a very challenging high-level image-to-image translation task. Thus, our model also fails in some cases, \emph{e.g.}, some generated images of Yale and CelebA dataset in Figure~\ref{}. Although our method can keep the structure information of faces, it is still hard to render the details for providing high-quality caricatures and some tiny organs (such as eyes) are lack of details in Figure~\ref{}. Besides, it's also sensitive to side face with complex background, \emph{e.g.}, some cases on CelebA and KDEF datasets in Figure~\ref{}. \textbf{Future work.} With regard to future work, first, it would be interesting to investigate our method on other tasks of high-level image-to-image translation (\emph{e.g.}, human-to-cartoon translation for cartoon movies); second, for the proposed method, the model still needs to be improved to provide high-quality rendered translated results; third, we intend to apply our method on the real applications of automatic and intelligent photo-to-caricature translation; fourth, we hope that we can control caricature style while translating images between domains by tuning input noise. \section*{Acknowledgement} We thanks the volunteers for grading human scores of translation results from different methods. This work was supported by the National Natural Science Foundation of China [61771440, 41776113], and Qingdao Municipal Science and Technology Program [17-1-1-5-jch].

\section{Introduction} Data for the $\Lambda^0_b\to p J/\psi K^-$ decay provide a definite argument for the existence of a pentaquark, a baryon system in $ p J/\psi$ spectra which has the following quark content: $P_{\bar c cuud}=\bar c(cuud)$.\cite{LHCbB} In terms of the antiquark-diquark-diquark states it can be presented as a three-body system: \bea \label{2mn} && P=a\;(\bar c^{\alpha}\cdot D_{cq_1}^{\beta}\cdot D_{q_2q_3}^{\gamma}\epsilon_{\alpha\beta\gamma}) + b\;(-\bar c^{\alpha}\cdot D_{q_2q_3}^{\beta}\cdot D_{cq_1}^{\gamma}\epsilon_{\alpha\beta\gamma}), \nn \\ && D_{cq_1}^{\beta} = c_{\beta'}q_{1\beta''}\epsilon^{\beta\beta'\beta''}, \qquad D_{q_2q_3}^{\gamma} = q_{2\gamma'}q_{3\gamma''}\epsilon^{\gamma\gamma'\gamma''} \eea where $\alpha,\beta,\gamma$ refer to color indices. Let us stress that both terms in (\ref{2mn}) have the same weight in color space. The right-hand side diquarks are forming baryons (it is convential condition). The second term in Eq.~(\ref{2mn}) gives a zero result for the states under consideration. Taking into account coordinate wavefunction lead us to nonvanishing second term in Eq.~(\ref{2mn}), however such states are radial excited ones and they are out of our consideration (see Appendix A). A diquark is a color triplet member, similar to a quark, and the right-hand side of Eq.~(\ref{2mn}) presents a three-body system with a color structure similar to that in low-lying baryons. It was supposed in Refs.~\citen{maia-polo,lebe,ala-2} that it results in the formation of a diquark-diquark-antiquark system. Following this idea it is possible to perform a classification of such baryon states and give estimations of their masses. Estimation of the charmed diquark masses is given in Ref.~\citen{cq-cq} where diquark-antidiquark states with hidden charm are studied. The notion of the diquark was introduced by Gell-Mann.\cite{gell-mann} Diquarks were discussed for baryon states during a long time, see the pioneering papers \citen{ida,licht,ono,vva75,schm} and recent studies \citen{santo1,santo2,santo3,santo4,santo5,santo6}. The systematization of baryons in terms of the quark-diquark states is presented in Refs.~\citen{lightdiq,book4}. Pentaquarks built of light-light and light-heavy diquarks present a natural extension of multiquark schemes studied in the last decade for mesons.\cite{wein,brod,maiani,voloshin,alii,jaff,ross} The LHCb collaboration observed two resonance states in the $J/\psi \,p$ -spectrum: \cite{LHCbB} the broad state $P(4380)$ and the narrow one $P(4450)$. These resonances generated a wide discussion.\cite{wang,Petrov,espo,os,pol,ali,skw,wo,st,pdg1,pdg2,pdg5,pdg6,pdg7,pdg8,pdg9,pdg10}. We will concentrate on discussing the narrow state because there are different versions of interpretation for a broad state (for example see Ref.~\citen{odnakartinka}). We suppose that narrow pentaquark states are formed by decays which are not $s$-wave recombination ones.\cite{ournew} Such a narrow state is $P(4450)$ with the width $\sim 40\pm 20$ MeV. In the present paper we calculate the recombination transitions for multiplet $(\bar ccuud)$. We see that the $P(4450)$ state does not decay in $s$-wave recombination mode. (The $s$-wave transition $P(4450)\to J/\psi + N(J=\frac 32)$ is impossible because of the large mass of $N(J=\frac 32)$ state). Tere is not such a problem for all other multiplet members. The narrow $P(4450)$ recombinates in $d$-vawe $P(4450)\to J/\psi + p$, so the $d$-vawe transition gives a comparatively small value of width. \section{Pentaquarks as composite states of light diquark ($qq'$), charmed diquark ($qc$) and antiquark ($\bar c$)} In Ref.~\citen{ala-2} it was supposed that the genuine antiquark-diquark-diquark state in the LHCb experiment is the narrow $P_c^{\frac52-}(4450)$ while the broad $P_c^{\frac32+}(4380)$ is the result of rescatterings of final state hadrons. \subsection{Color-spin-isospin structure of the pentaquark} Generally, in color space the wave function of the antiquark-diquark-diquark pentaquark is represented by Eq.~(\ref{2mn}). As it was mentioned above, in the case under consideration we have the first term of Eq.~(\ref{2mn}) only. It is written as follows: \bea\label{3m} && \bar c^{\alpha}\cdot D_{cq_1}^{\beta}\cdot D_{q_2q_3}^{\gamma} \epsilon_{\alpha\beta\gamma}= {\bar c}^{\alpha}(cq_1)^{\beta}(q_2q_3)^{\gamma}\epsilon_{\alpha\beta\gamma} \\ &=&-({\bar c}^{\alpha}c_{\alpha}) (q_{1\gamma}q_{2\gamma'}q_{3\gamma''}\epsilon^{\gamma\gamma'\gamma''})+ ({\bar c}^{\alpha} q_{1\alpha})(c_{\gamma}q_{2\gamma'}q_{3\gamma''}\epsilon^{\gamma\gamma'\gamma''}) \nn \\ &=& -({\bar c}c)(q_{1}q_{2}q_{3})+({\bar c}q_{1})(cq_{2}q_{3}) \nn \eea The last line presents the result in a short form, the brackets separate the wite meson and baryon states. To take into account the isospin structure we write: \be\label{1} P_{\bar c cuud}= -\sqrt{\frac{1}{3}}\bar c\cdot(cu)\cdot(ud) +\sqrt{\frac{2}{3}}\bar c \cdot(cd)\cdot(uu). \ee We discuss a scheme in which the exotic baryon states are formed by standard QCD-motivated interactions (gluonic exchanges, confinement forces) but in addition with diquarks as constituents. We work with four diquarks, two scalars $S^{II_z,JJ_z}$ and two axial-vectors $A^{II_z,JJ_z}$: \bea \label{5} && S_{cq}^{\frac 12 I_z,00}(I=1/2,\,J=0),\qquad A_{cq}^{\frac 12 I_z,1J_z}(I=1/2,\,J=1), \\ && S_{q'q''}^{00,00}(I=0,\,J=0),\qquad\,\, A_{q'q''}^{1I_z,1J_z}(I=1,\,J=1), \nn \eea where $I$ and $J$ refer to isospin and spins of the diquarks. In terms of these diquarks the color-flavor wave funcion of pentaquark reads: \be \label{6} P_{\bar c\cdot cq\cdot q'q''}=\bar c^\alpha\cdot \epsilon_{\alpha\beta\gamma}\, \begin{tabular}{|l|} $S^\beta_{cq}$\\ $A^\beta_{cq}$ \end{tabular} \cdot \begin{tabular}{|l|} $S^\gamma_{q'q''}$\\ $A^\gamma_{q'q''}$ \end{tabular} \ee We have six diquark-diquark states: \be \label{51} P_{\bar c\cdot cq\cdot q'q''}= \bar c^\alpha\cdot \begin{tabular}{|l|} $(S_{cq}S_{q'q''})^\alpha(\frac12; 0^{+})$\\ $(S_{cq}A_{q'q''})^\alpha(\frac12,\frac32; 1^{+})$\\ $(A_{cq}S_{q'q''})^\alpha(\frac12; 1^{+})$\\ $(A_{cq}A_{q'q''})^\alpha(\frac12,\frac32; 0^{+})$\\ $(A_{cq}A_{q'q''})^\alpha(\frac12,\frac32; 1^{+})$\\ $(A_{cq}A_{q'q''})^\alpha(\frac12,\frac32; 2^{+})$ \end{tabular} \ee with the isospin and spin-parity numbers of the diquark-diquark pair: $I=\frac12, \frac32$ and $J^P= 0^+, 1^+,2^+$. So, the low-laing $s$-wave pentaquark multiplet $P(I,J^P)$ reads: \be\label{7} P_{\bar c\cdot (cq)\cdot(q'q'')}= \begin{tabular}{l} $P(\frac12,\frac12^{-})$\\ $P(\frac12,\frac12^{-})$, $P(\frac12,\frac32^{-})$, $P(\frac32,\frac12^{-})$, $P(\frac32,\frac32^{-})$, \\ $P(\frac12,\frac12^{-})$, $P(\frac12,\frac32^{-})$\\ $P(\frac12,\frac12^{-})$, $P(\frac32,\frac 12^{-})$\\ $P(\frac12,\frac12^{-})$, $P(\frac12,\frac32^{-})$, $P(\frac32,\frac12^{-})$, $P(\frac32,\frac32^{-})$ \\ $P(\frac12,\frac32^{-})$, $P(\frac12,\frac52^{-})$, $P(\frac32,\frac32^{-})$, $P(\frac32,\frac52^{-})$ \end{tabular} \ee \subsection{Masses of the $s$-wave pentaquarks} It was undersood already relatively long ago that the mass splitting of hadrons can be well described in the framework of the quark model by the short-ranged spin-spin interactions of the constituents.\cite{ZS,RGG,glashow} For mesons and baryons the mass formulae discussed by Glashow read:\cite{glashow} \bea \label{8a} && M_M=\sum\limits_{j=1,2}m_{q(j)}+a \frac{\vec{s}_1\vec{s}_2}{m_{q(1)}m_{q(2)}}, \\ && M_B=\sum\limits_{j=1,2,3}m_{q(j)}+b \sum\limits_{j>\ell}\frac{\vec{s}_j\vec{s}_\ell}{m_{q(j)}m_{q(\ell)}}, \nn \eea where $\vec{s}_j$ and $m_{q(j)}$ refer to spins and masses of the constituents. Mass splitting parameters in Eq.~(\ref{8a}), $a$ and $b$, are characterized by the size of the colour-magnetic interaction and the size of the discussed hadron, the short-range interaction is supposed in Ref.~\citen{GL}. For the 36-plet mesons ($q\bar q$) and 56-plet baryons ($qqq$) formulae of Eq.~(\ref{8a}) work well. The only exception is the pion, its calculated mass is $\sim$ 350 MeV that point out the existence of additional forces in the pseudoscalar channel (possibly, istanton-induced forces). It seems natural to apply modified formulae (\ref{8a}) to pentaquark systems. For multiquark states from Ref.~\citen{LHCbB} we have: \bea \label{8}&& M_{q_1q_2\cdot q_3c\cdot \bar c}=3m_{D(q_1q_2)D(q_3c)\bar c} +4\Delta\Big(\vec{\mu}_{D(q_1q_2)}\vec{\mu}_{D(q_3c)}+ \vec{\mu}_{D(q_1q_2)}\vec{\mu}_{\bar c}+\vec{ \mu}_{\bar c}\vec{\mu}_{D(q_3c)}\Big) \nn \\ &&3m_{D(q_1q_2)D(q_3c)\bar c}= m_{D(q_1q_2)}+m_{D(q_3c)}+m_{\bar c} \eea where $\vec \mu_D$ and $\vec \mu_{\bar c}$ are color-magnetic moments of diquarks and $c$-quark, $\Delta$ is the parameter of spin splitting. Diquarks are considered as composite systems of quarks analogous to light nuclei. The magnetic moments are written as sums of quark magnetic moments: \bea \label{10e} && \vec{\mu}_{D(q_1q_2)}=\frac{\vec{s}_{q_1} }{m_{q_1}}+ \frac{\vec{s}_{q_2}}{m_{q_2}}, \qquad \vec{\mu}_{D(q_3c)} = \frac{\vec{s}_{q_3} }{m_{q_3}} +\frac{\vec{s}_c }{m_{c}}\simeq \frac{\vec{s}_{q_3} }{m_{q_3}},\nn\\ && \vec{\mu}_c=\frac{\vec{s}_c }{m_{c}}\simeq 0. \eea In our estimations we take into account that $m_{c}>>m_{q} $. Estimation of diquark masses is the most problematic issue in the study of diquarks (see for example Ref.~\citen{santopinto}). Basing on Refs.~\citen{lightdiq,QQQQ} we estimate the masses of scalar $S$ ($J^P=0^+$) and axial $A$ ($J^P=1^+$) diquarks as follows (in MeV units): \be \begin{tabular}{ll} $m_q=330,$ & $m_c=1450$,\\ $m_{S(q_1q_2)}=700$, & $m_{S(q_3c)}=2000$,\\ $m_{A(q_1q_2)}=800$, & $m_{A(q_3c)}=2100$.\\ \end{tabular} \ee Correspondingly we write a set of the low-laying pentaquark states: \be \label{13} \begin{tabular}{l|l} $\qquad I=1/2$& $\qquad I=3/2 $ \\ \hline $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac12,\frac32^{-})}(4450)$, $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac12,\frac52^{-})}(4450)$, & $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac32,\frac32^{-})}(4450)$, $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac32,\frac52^{-})}(4450)$\,,\\ $P_{\bar c\,S_{(cq)}S_{(q'q'')}}^{(\frac12,\frac12^{-})}(4300)$, & \\ $P_{\bar c\,S_{(cq)}A_{(q'q'')}}^{(\frac12,\frac12^{-})}(4300)$, $P_{\bar c\,S_{(cq)}A_{(q'q'')}}^{(\frac12,\frac32^{-})}(4300)$, & $P_{\bar c\,S_{(cq)}A_{(q'q'')}}^{(\frac32,\frac12^{-})}(4300)$, $P_{\bar c\,S_{(cq)}A_{(q'q'')}}^{(\frac32,\frac32^{-})}(4300)$, \\ $P_{\bar c\,A_{(cq)}S_{(q'q'')}}^{(\frac12,\frac12^{-})}(4300)$, $P_{\bar c\,A_{(cq)}S_{(q'q'')}}^{(\frac12,\frac32^{-})}(4300)$, &\\ $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac12,\frac12^{-})}(4150)$, $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac12,\frac32^{-})}(4150)$, & $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac32,\frac12^{-})}(4150)$, $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac32,\frac32^{-})}(4150)$, \\ $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac12,\frac12^{-})}(4000)$, & $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac32,\frac12^{-})}(4000)$, \\ \end{tabular} \ee Masses are given in MeV units, an uncertainty in the determination of masses is of the order of $\pm$150 MeV. It should be noted that the process $P\to meson+baryon\to P$ can shift mass values in Eq.~(\ref{13}). Such a rescattering was taken into account for tetraquark states in Ref.~\citen{QQQQ}. Following Ref.~\citen{QQQQ} we estimate this shift to be of the order of 100 MeV. \section{Spin-isospin structure of pentaquarks } \subsection {Pentaquark $P_{\bar c A_{cu}\cdot A_{ud} }^{\frac12 \frac12,\frac52 \frac52}(4450)$} In the antiquark-diquark-diquark scheme the isospin structure of the $P_{\bar ccuud }^{\frac12\frac12,\frac52 \frac52}(4450)$ with spin and isospin projections can be written as follows: \be\label{n6} P_{\bar c\uparrow A_{cu}A_{ud}}^{\frac12\frac12,\frac52 \frac52 }(4450)= -\frac{1}{\sqrt3}\;\bar c^{\uparrow}\cdot A_{cu}^{\frac12\frac12,11}\cdot A_{ud}^{10,11} +\sqrt{\frac23}\;\bar c^{\uparrow}\cdot A_{cd}^{\frac12-\frac12,11}\cdot A_{uu}^{11,11} \ee The wave functions for axial diquarks read: \bea\label{n14} &&A_{cu}^{\frac12\frac12,11}=c^{\uparrow}u^{\uparrow}, \qquad A_{ud}^{10,11}=\frac{1}{\sqrt2}(d^{\uparrow}u^{\uparrow}+u^{\uparrow}d^{\uparrow}),\nn \\ &&A_{cd}^{\frac12-\frac12,11}=c^{\uparrow}d^{\uparrow}, \qquad A_{uu}^{11,11}=u^{\uparrow}u^{\uparrow}. \eea Taking into account color structure (\ref{3m}), we can rewrite expresion (\ref{n6}) as follows: \bea P_{\bar c\uparrow A_{cu}A_{ud}}^{\frac12\frac12,\frac52 \frac52 }(4450)&=& -\frac{1}{\sqrt3}\;\bar c^{\uparrow}\cdot c^{\uparrow}u^{\uparrow}\cdot A_{ud}^{10,11} +\sqrt{\frac23}\;\bar c^{\uparrow}\cdot c^{\uparrow}d^{\uparrow}\cdot A_{uu}^{11,11}=\nn \\ &=&-\frac{1}{\sqrt3}\;\left(-\bar c^{\uparrow}c^{\uparrow}\cdot u^{\uparrow}A_{ud}^{10,11}+\bar c^{\uparrow}u^{\uparrow}\cdot c^{\uparrow}A_{ud}^{10,11}\right)+\nn \\ &+&\sqrt{\frac23}\;\left(-\bar c^{\uparrow}c^{\uparrow}\cdot d^{\uparrow}A_{uu}^{11,11}+\bar c^{\uparrow}d^{\uparrow}\cdot c^{\uparrow}A_{uu}^{11,11}\right). \eea The white quark-antiquark and three-quark states should be projected on states of real mesons and baryons. We need mesons: \bea\label{n15} && J/\psi^{\Uparrow} =\bar c^{\uparrow}c^{\uparrow},\quad \bar D^{*0\Uparrow} =\bar c^{\uparrow}u^{\uparrow},\quad D^{*-\Uparrow} =\bar c^{\uparrow}d^{\uparrow}, \eea and baryons: \bea\label{n16} && N^{\frac 12\frac 12,\frac 32\frac 32} = \sqrt\frac23 d^{\uparrow}A_{uu}^{11,11}-\sqrt\frac13 u^{\uparrow}A_{ud}^{10,11},\nn \\ && \Sigma_c^{+(10,\frac32\frac32)}= c^{\uparrow}A^{10,11}_{ud}\,,\nn \\ && \Sigma_c^{++(11,\frac32\frac32)}= c^{\uparrow}A^{11,11}_{uu}. \eea Therefore we write for the pentaquark wave function: \bea\label{5252} && P_{\bar c\uparrow A_{cu}A_{ud}}^{\frac12\frac12,\frac52 \frac52 }=- J/\psi^{\Uparrow}N^{\frac 12\frac 12,\frac 32\frac 32}-\sqrt{\frac13}\bar D^{*0\Uparrow}\Sigma^{+(10,\frac32\frac32)}_{c} +\sqrt{\frac23}D^{*-\Uparrow}\Sigma^{++(11,\frac32\frac32)}_{c}\,. \eea The state $P_{\bar c\,A_{(cq)}A_{(q'q'')}}^{(\frac12,\frac52^{-})}(4520\pm 150)$ is a good candidate to be a state which was observed by LHCb \cite{LHCbB}: $5/2^?(4450\pm 4)$ with a width of $\Gamma=39\pm 24$ MeV. Recombination to $\bar D^{*0}\Sigma^{*+}_{c}$ and $D^{*-\Uparrow}\Sigma^{*++}_{c}$ channels is strongly suppressed because their tresholdls (4460 and 4464 MeV correspondingly) are larger than the pentaquark mass. So the leading recombination channel is $P\to J/\psi + p$ which is possible in $d$-wave only. $s$-wave transition $P\to J/\psi + N(J=\frac 32)$ has the treshold 4817 MeV. That explains the relatively small width of $P(4450)$. \subsubsection{Propagator} Resonance propagator: \bea && \frac{1}{m^2-s +G^2\cos^2\gamma |k_{J/\psi N^*}| -iG^2\sin^2\gamma k^5_{J/\psi p}-iG^2 k_{D^*\Sigma^*}} \\ &=& \frac{1}{m^2_{ P(\frac12 \frac52^-)}-s -iG^2_{J/\psi p} k^5_{J/\psi p}-iG^2 k_{D^*\Sigma^*}} \nn \eea Here: \bea && m^2_{ P(\frac12 \frac52^-)}=m^2+G^2\cos^2\gamma |k_{J/\psi N^*}|, \qquad G^2_{J/\psi p}=G^2\sin^2\gamma \eea Thresholds under discussion: \bea && M_{J/\psi}+M_{N^*}=(3100+1720) MeV=4820\; MeV, \\ && M_{D^*}+M_{\Sigma^*}=(2010+2520) MeV= 4520\; MeV \nn\\ && M_{J/\psi}+M_{p}=(3100+940) MeV= 4040\; MeV \nn \eea Hadron momenta in the pentaquark region ($\sqrt s\sim 4450 $ MeV): \bea && |k_{J/\psi N^*}|=\sqrt{\frac{[-s+(M_{J/\psi}+M_{N^*})^2][s-(M_{J/\psi}-M_{N^*})^2]}{4s}} \\ && k_{D^*\Sigma^*}=i\sqrt{\frac{[-s+(M_{D^*}+M_{\Sigma^*})^2][s-(M_{D^*}-M_{\Sigma^*})^2]}{4s}} ,\quad s< (M_{D^*}+M_{\Sigma^*})^2 \nn\\ && k_{D^*\Sigma^*}=\sqrt{\frac{[s-(M_{D^*}+M_{\Sigma^*})^2][s-(M_{D^*}-M_{\Sigma^*})^2]}{4s}} ,\quad s> (M_{D^*}+M_{\Sigma^*})^2 \nn\\ \nn \\ && k_{J/\psi p}=\sqrt{\frac{[s-(M_{J/\psi }+M_{ p})^2][s-(M_{J/\psi }-M_{ p})^2]}{4s}} ,\quad s> (M_{J/\psi}+M_{ p})^2 \nn \eea Amplitude of production of pentaquark with background contribution reads: \bea A&=& A_{background}+ \frac{A_{ production}}{m^2_{ P(\frac12 \frac52^-)}-s -iG^2_{J/\psi p} k^5_{J/\psi p}-iG^2 k_{D^*\Sigma^*}} \\ &=& A_{background}\bigg[1+ \frac{\mu^2_{prod}}{m^2_{ P(\frac12 \frac52^-)}-s -iG^2_{J/\psi p} k^5_{J/\psi p}-iG^2 k_{D^*\Sigma^*}} \bigg] \nn \eea At complex $\mu^2_{prod}$ and $m_{ P(\frac12 \frac52^-)}\sim 4450 \, MeV$ we have resonance-cusp interplay: \be \sigma\sim |A|^2 \ee with \bea && \mu^2_{prod}=|\mu^2_{prod}|e^{i\Phi_{prod}} \,,\quad \sqrt{|\mu^2_{prod}|}\la m_{ P(\frac12 \frac52^-)} \\ && \Big[\frac{G^2_{J/\psi p} k^5_{J/\psi p}}{m_{ P(\frac12 \frac52^-)}} \Big]_{\sqrt s \sim 4500}\quad \la 50\,MeV,\qquad \Big[ \frac{G^2 k_{D^*\Sigma^*}}{m_{ P(\frac12 \frac52^-)}} \Big]_{\sqrt s \sim 4500} \la 50\,MeV \nn \eea \subsection {Pentaquark $ $ $P^{\frac12\frac12,\frac12\frac12}_{\bar c S_{cu}\cdot S_{ud}}$ } Combination $P=\bar c S_{cu}\cdot S_{ud}$ contains scalar diquarks $S_{cu}^{\frac12\frac12,00}=[c^{\uparrow}u^{\downarrow}-c^{\downarrow}u^{\uparrow}]\frac{1}{\sqrt 2}$ and $S_{ud}^{00,00}=[u^{\uparrow}d^{\downarrow}-u^{\downarrow}d^{\uparrow}-d^{\uparrow}u^{\downarrow}+d^{\downarrow}u^{\uparrow}]\frac{1}{2}$: \bea\label{n7} P^{\frac12\frac12,\frac12\frac12}_{\bar c\uparrow(cuud)}&&=\bar c^{\uparrow} S_{cu}^{\frac12\frac12,00}S_{ud}^{10,00}. \eea Using (\ref{3m}) expression (\ref{n7}) can be represented as follows: \bea \label{n1} P^{\frac12\frac12,\frac12\frac12}_{\bar c\uparrow S_{cu}S_{ud}}&&= \bar c^{\uparrow}[c^{\uparrow}u^{\downarrow}-c^{\downarrow}u^{\uparrow}]\frac{1}{\sqrt 2}S_{ud}= \\ &&=\frac{1}{\sqrt 2}\left\{-\bar c^{\uparrow}c^{\uparrow}\cdot u^{\downarrow}S_{ud}+\bar c^{\uparrow}u^{\downarrow}\cdot c^{\uparrow}S_{ud}+\bar c^{\uparrow}c^{\downarrow}\cdot u^{\uparrow}S_{ud}-\bar c^{\uparrow}u^{\uparrow}\cdot c^{\downarrow}S_{ud}\right\}.\nn \eea Now we need expressions for the meson parts: \bea\label{n2} &&\bar c^\uparrow c^{\uparrow}=J/\psi^{\Uparrow},\quad \bar c^\uparrow c^{\downarrow}=\frac{1}{\sqrt2}(J/\psi^{(0)}-\eta_c),\nn \\ &&\bar c^\uparrow u^{\uparrow}=\bar D^{*0\Uparrow},\quad \bar c^\uparrow u^{\downarrow}=\frac{1}{\sqrt2}(\bar D^{*0(0)}-\bar D^{0}), \eea and for baryon parts ($\Lambda_c^{+}(cud)$ and $p(uud)$): \bea\label{n3} &&\Lambda_c^{+(00,\frac 12\frac 12)}=c^{\uparrow}S^{00,00}_{ud},\\ &&N^{\frac12\frac12,\frac12\frac12}= \frac{1}{\sqrt{2}} [\frac{2}{3} d^{\downarrow} A^{11,11}_{uu} - \frac{\sqrt2}{3} d^{\uparrow} A^{11,10}_{uu} - \frac{\sqrt2}{3} u^{\downarrow} A^{10,11}_{ud} + \frac{1}{3} u^{\uparrow} A^{10,10}_{ud}] +\frac{1}{\sqrt{2}}u^{\uparrow} S^{00,00}_{ud},\nn\\ &&N^{'\frac12\frac12,\frac12\frac12}=- \frac{1}{\sqrt{2}} [\frac{2}{3} d^{\downarrow} A^{11,11}_{uu} - \frac{\sqrt2}{3} d^{\uparrow} A^{11,10}_{uu} - \frac{\sqrt2}{3} u^{\downarrow} A^{10,11}_{ud} + \frac{1}{3} u^{\uparrow} A^{10,10}_{ud}] +\frac{1}{\sqrt{2}}u^{\uparrow} S^{00,00}_{ud},\nn \eea where $N^{'\frac12\frac12,\frac12\frac12}$ is some radial excitation of proton $N^{\frac12\frac12,\frac12\frac12}$. Taking into account expressions (\ref{n2}) and (\ref{n3}) we can write: \bea\label{n4} P^{\frac12\frac12,\frac12\frac12}_{\bar c\uparrow S_{cu}S_{ud}}&&=\frac{1}{\sqrt 2}\left\{J/\psi^{\Uparrow}(p^{\uparrow}+p^{'\uparrow})\frac{1}{\sqrt2}+\frac{1}{\sqrt2}(\bar D^{*0(0)}-\bar D^{0})\Lambda_c^{+\uparrow}+\right.\nn \\ &&\left.+\frac{1}{\sqrt2}(J/\psi^{(0)}-\eta_c)(p^{\uparrow}+p^{'\uparrow})\frac{1}{\sqrt2}-\bar D^{*0\Uparrow}\Lambda_c^{+\downarrow}\right\}. \eea Combination $P=-\bar c S_{ud}\cdot S_{cu}$ (see Eq.~(\ref{2mn})) gives zero result without taking into account the coordinate wave function. \subsection {Pentaquark $P_{\bar c(A_{cu}\cdot A_{ud})_{j=0} }^{\frac12\frac12,\frac12 \frac12 }(\sim 4100 )$} Isospin and spin structure of pentaquark state $P_{\bar c(A_{cu}\cdot A_{ud})_{j=0} }^{\frac12\frac12,\frac12 \frac12}$ is: \bea\label{n8} P_{\bar cA_{cu}\cdot A_{ud}}^{\frac12\frac12,\frac12 \frac12} &=& -\frac{1}{\sqrt3}\;\bar c^{\uparrow}\cdot \frac{1}{\sqrt 3} [A_{cu}^{\frac12\frac12,11} A_{ud}^{10,1-1} -A_{cu}^{\frac12\frac12,10} A_{ud}^{10,10} +A_{cu}^{\frac12\frac12,1-1} A_{ud}^{10,11}] \nn \\ && +\sqrt\frac23\;\bar c^{\uparrow}\cdot \frac{1}{\sqrt 3} [A_{cd}^{\frac12-\frac12,11} A_{uu}^{11,1-1} -A_{cd}^{\frac12-\frac12,10} A_{uu}^{11,10} +A_{cd}^{\frac12-\frac12,1-1} A_{uu}^{11,11}] \nonumber \\ &=& -\frac{1}{3}\;\bar c^{\uparrow}\cdot [c^{\uparrow}u^{\uparrow}A_{ud}^{10,1-1} -\sqrt\frac12(c^{\uparrow}u^{\downarrow}+c^{\downarrow}u^{\uparrow})A_{ud}^{10,10} +c^{\downarrow}u^{\downarrow}A_{ud}^{10,11}] \\ && +\frac{\sqrt2}{3}\;\bar c^{\uparrow}\cdot [c^{\uparrow}d^{\uparrow}A_{uu}^{11,1-1} -\sqrt\frac12(c^{\uparrow}d^{\downarrow}+c^{\downarrow}d^{\uparrow})A_{uu}^{11,10} +c^{\downarrow}d^{\downarrow}A_{uu}^{11,11}]. \nonumber \eea Using expressions for mesons: \bea\label{n11} \bar c^{\uparrow}d^{\uparrow}=D^{*-\Uparrow},\quad \bar c^{\uparrow}d^{\downarrow}=\frac{1}{\sqrt 2}(D^{*-(0)}-D^-), \eea and baryons: \bea\label{n12} \Sigma_c^{+(10,\frac12\frac12)}=&=&\sqrt{\frac23}c^{\downarrow}A^{10,11}_{ud}-\sqrt{\frac13}c^{\uparrow}A^{10,10}_{ud},\nn \\ \Sigma_c^{++(11,\frac12\frac12)}&=&\sqrt{\frac23}c^{\downarrow}A^{11,11}_{uu}-\sqrt{\frac13}c^{\uparrow}A^{11,10}_{uu}, \eea we can write for the state $P_{\bar c(A_{cu}\cdot A_{ud})_{j=0} }^{\frac12\frac12,\frac12 \frac12}$: \bea\label{n13} P_{\bar cA_{cu}\cdot A_{ud}}^{\frac12\frac12,\frac12 \frac12} &=&\frac12 J/\psi^{\Uparrow}(-N^{\frac12\frac12,\frac12-\frac12}+N^{'\frac12\frac12,\frac12-\frac12})+\nn\\ &+&\frac{1}{2\sqrt 2}(J/\psi^{(0)}-\eta_c)(-N^{\frac12\frac12,\frac12\frac12}+N^{'\frac12\frac12,\frac12\frac12})-\nn\\ &-&\frac{1}{\sqrt 6}\bar D^{*0\Uparrow}\Sigma_c^{+(10,\frac12-\frac12)}-\frac{1}{2\sqrt 3}(\bar D^{*0(0)}-\bar D^0)\Sigma_c^{+(10,\frac12\frac12)}+\nn \\ &+&\frac{1}{\sqrt 3}D^{*-\Uparrow}\Sigma_c^{++(11,\frac12-\frac12)}+\frac{1}{\sqrt 6}(D^{*-(0)}-D^-)\Sigma_c^{++(11,\frac12\frac12)}. \eea \subsection {Pentaquark $P_{\bar c A(cu)\cdot A(ud)_{j=1} }^{\frac12\frac12,\frac12\frac12}(\sim 4240 )$} In the case of diquark-diquark spin $j_{AA}=1$, pentaquark state $P_{\bar c A(cu)\cdot A(ud)}^{\frac12\frac12,\frac12\frac12}$ can be written as follows (upper indecies in right part of the expression represent spin and its projection): \be P_{\bar c A(cu)\cdot A(ud)}^{\frac12\frac12,\frac12\frac12}=-\frac{1}{\sqrt3}[\bar c(A_{cu}A_{ud})^{1J_z}]^{\frac12\frac12}+\sqrt{\frac23}[\bar c(A_{cd}A_{uu})^{1J_z}]^{\frac12\frac12}. \ee In terms of mesons and baryons it can be represented in form: \bea &&P_{\bar c A(cu)\cdot A(ud)}^{\frac12\frac12,\frac12\frac12}=J/\Psi^{\Uparrow}\left(-\frac{1}{3\sqrt2}p^{\downarrow}+\frac{1}{3\sqrt2}p^{'\downarrow}+\frac{1}{3\sqrt2}N^{\frac12\frac12,\frac32-\frac12}\right)+\frac{1}{\sqrt6}J/\Psi^{\Downarrow}N^{\frac12\frac12,\frac32\frac32}+\nn\\ &&+J/\Psi^{(0)}\left(\frac16 p^{\uparrow}-\frac16 p^{'\uparrow}-\frac13 N^{\frac12\frac12,\frac32\frac12}\right)+\eta_c\left(\frac12 p^{\uparrow}-\frac12 p^{'\uparrow}\right)+\nn\\ &&+\bar D^{*0(0)}\left(\frac{1}{3\sqrt6}\Sigma_c^{+10,\frac12\frac12}-\frac{1}{3\sqrt3}\Sigma_c^{+10,\frac32\frac12}\right)+\frac{1}{\sqrt6}\bar D^0\Sigma_c^{+10,\frac12\frac12}+\nn\\ &&+D^{*-(0)}\left(\frac{1}{3\sqrt3}\Sigma_c^{++11,\frac12\frac12}+\frac{2\sqrt2}{3\sqrt3}\Sigma_c^{++11,\frac32\frac12}\right)+D^-\left(-\frac{1}{3\sqrt3}\Sigma_c^{++11,\frac12\frac12}+\frac{\sqrt2}{3\sqrt3}\Sigma_c^{++11,\frac32\frac12}\right)+\nn\\ &&+\bar D^{*0\Uparrow}\left(\frac{1}{3\sqrt6}\Sigma_c^{+10,\frac32-\frac12}-\frac{1}{3\sqrt3}\Sigma_c^{+10,\frac12-\frac12}\right)+\frac{1}{3\sqrt2}\bar D^{*0\Downarrow}\Sigma_c^{+10,\frac32\frac32}-\nn\\ &&-D^{*-\Uparrow}\left(-\frac{\sqrt2}{3\sqrt3}\Sigma_c^{++11,\frac12-\frac12}+\frac{1}{3\sqrt3}\Sigma_c^{++11,\frac32-\frac12}\right)-\frac13 D^{*-\Downarrow}\Sigma_c^{++11,\frac32\frac32}. \eea \subsection {Pentaquark $P_{\bar ccuud }^{\frac32 I_z,\frac52 J_z}$ near 4450 MeV} It is the isotopic counterpart of the observed state $P_{\bar ccuud }^{\frac12 I_z,\frac52 J_z}(4450)$. We consider here the case with $I_z=\frac12$, namely, the state with the same quark content as $P_{\bar ccuud }^{\frac12 \frac12 ,\frac52 J_z}(4450)$. The $P_{\bar ccuud }^{\frac32\frac12,\frac52 J_z}$ can be seen in the decay $\Sigma^0_b\to K^-J/\psi\, N\pi$, in the spectrum $J/\psi\, \Delta^+(1240)\to J/\psi\, N\pi $. In the quark-diquark model the spin-isospin wave function of the $P_{\bar ccuud }^{II_z,JJ_z}$ with spin-isospin projections $I_z=\frac12\,,\;J_z=\frac 52$ is written as: \be P^{\frac32\frac12, \frac52\frac52}_{\bar ccuud } = \sqrt{\frac23}\;\bar c^{\uparrow}\cdot A_{cu}^{\frac12\frac12,11}\cdot A_{ud}^{10,11} +\frac{1}{\sqrt3}\;\bar c^{\uparrow}\cdot A_{cd}^{\frac12-\frac12,11}\cdot A_{uu}^{11,11} \ee As previously we use axial-vector diquarks (\ref{n14}) and write: \bea P^{\frac32\frac12, \frac52\frac52}_{\bar ccuud } &=&\sqrt{\frac23}[-\bar c^{\uparrow}c^{\uparrow}\cdot u^{\uparrow}A_{ud}^{10,11}+\bar c^{\uparrow}u^{\uparrow}\cdot c^{\uparrow}A_{ud}^{10,11}]+\nn\\ &+&\frac{1}{\sqrt 3}[-\bar c^{\uparrow}c^{\uparrow}\cdot d^{\uparrow}A_{uu}^{11,11}+\bar c^{\uparrow}d^{\uparrow}\cdot c^{\uparrow}A_{uu}^{11,11}]. \eea The white quark-antiquark and three-quark states should be projected on states of real mesons (\ref{n15}) and baryons (\ref{n16}), (\ref{26}). Therefore we write for the pentaquark spin-flavor wave function: \bea \label{25u} && P^{\frac32\frac12, \frac52\frac52}_{\bar ccuud } = -J/\psi^{\Uparrow}\Delta^{+(\frac 32\frac 12,\frac 32\frac 32)} +\sqrt{\frac23}D^{*0\Uparrow}\Sigma^{+(10,\frac 32\frac 32)}_{c} +\sqrt{\frac13}D^{*-\Uparrow}\Sigma^{++(11,\frac 32\frac 32)}_{c}\,. \eea Let us emphasise, that states under consideration are $s$-wave states, we consider them as dominant ones. \subsection {Pentaquark $P_{\bar cA(cu)A(ud)_{j=1} } ^{\frac12\frac12,\frac32\frac32}$} Pentaquark state $P_{\bar cA(cu)A(ud)} ^{\frac12\frac12,\frac32\frac32}$ for the case $j_{AA}=1$ can be written as follows: \bea &&P_{\bar cA(cu)A(ud)} ^{\frac12\frac12,\frac32\frac32}=-\frac{1}{\sqrt3}\;\Big[\bar c\cdot A_{cu}^{\frac12\frac12,1J_z}\cdot A_{ud}^{10,1J_z}\Big]^{\frac32\frac32}+\sqrt{\frac23}\;\Big[\bar c\cdot A_{cd}^{\frac12-\frac12,1J_z}\cdot A_{uu}^{11,1J_z}\Big]^{\frac32\frac32}=\nn\\ &&=-\frac{1}{\sqrt3}[\bar c^{\uparrow}\frac{1}{\sqrt2}(A_{cu}^{\frac12\frac12,11}A_{ud}^{10,10}-A_{cu}^{\frac12\frac12,10}A_{ud}^{10,11})]+\nn\\ &&+\sqrt{\frac23}[\bar c^{\uparrow}\frac{1}{\sqrt2}(A_{cd}^{\frac12-\frac12,11}A_{uu}^{11,10}-A_{cd}^{\frac12-\frac12,10}A_{uu}^{11,11})]. \eea Using formulae from \ref{appB} we can rewrite this expression in terms of mesons and baryons: \bea &&P_{\bar cA(cu)A(ud)}^{\frac12\frac12,\frac32\frac32}=\frac{1}{\sqrt3}J/\Psi^{\Uparrow}p^{\uparrow}-\frac{1}{\sqrt3}J/\Psi^{\Uparrow}p^{'\uparrow}-\frac{1}{2\sqrt3}J/\Psi^{\Uparrow}N^{\frac12\frac12,\frac32\frac12}+\nn\\ &&+\frac{1}{2\sqrt2}J/\Psi^{(0)}N^{\frac12\frac12,\frac32\frac32}+\frac{1}{2\sqrt2}\eta_cN^{\frac12\frac12,\frac32\frac32}+\frac{\sqrt2}{3}\bar D^{*0\Uparrow}\Sigma_c^{+10,\frac12\frac12}-\frac16\bar D^{*0\Uparrow}\Sigma_c^{+10,\frac32\frac12}-\nn\\ &&-\frac23 D^{*-\Uparrow}\Sigma_c^{+11,\frac12\frac12}+\frac{1}{3\sqrt2}D^{*-\Uparrow}\Sigma_c^{+11,\frac32\frac12}+\frac{1}{2\sqrt6}\bar D^{*0(0)}\Sigma_c^{+10,\frac32\frac32}-\nn\\ &&-\frac{1}{2\sqrt6}\bar D^0\Sigma_c^{+10,\frac32\frac32}-\frac{1}{2\sqrt3}D^{*-(0)}\Sigma_c^{+11,\frac32\frac32}+\frac{1}{2\sqrt3}D^-\Sigma_c^{+11,\frac32\frac32}. \eea \subsection {Pentaquark $P_{\bar cA(cu)A(ud) _{j=2}} ^{\frac12\frac12,\frac32\frac32}$} Pentaquark state $P_{\bar cA(cu)A(ud)} ^{\frac12\frac12,\frac32\frac32}$ for the case $j_{AA}=2$ can be written as follows: \bea && P_{\bar cA(cu)A(ud)}^{\frac12\frac12,\frac32\frac32}=-\frac{1}{\sqrt3}\; \Big[\bar c\cdot A_{cu}^{\frac12\frac12,1J_z}\cdot A_{ud}^{10,1J_z}\Big]^{\frac32\frac32}+\sqrt{\frac23}\;\Big[\bar c\cdot A_{cd}^{\frac12-\frac12,1J_z}\cdot A_{uu}^{11,1J_z}\Big]^{\frac32\frac32}. \eea In terms of mesons and baryons this state is expressed in the form: \bea &&P_{\bar cA(cu)A(ud)}^{\frac12\frac12,\frac32\frac32}=-\frac{\sqrt 5}{2\sqrt 2}\eta_cN^{\frac12\frac12,\frac32\frac32}-\frac{3\sqrt 2}{4\sqrt 5}J/\Psi^{(0)}N^{\frac12\frac12,\frac32\frac32}+\frac{\sqrt 3}{2\sqrt 5}J/\Psi^{\Uparrow}N^{\frac12\frac12,\frac32\frac12}+ \nn\\ &&+\frac{1}{2\sqrt 5}\bar D^{*0\Uparrow}\Sigma_c^{+10,\frac32\frac12}-\frac{\sqrt 6}{4\sqrt 5}\bar D^{*0(0)}\Sigma_c^{+10,\frac32\frac32}-\frac{\sqrt 5}{2\sqrt 6}\bar D^0\Sigma_c^{+10,\frac32\frac32}- \nn\\ &&-\frac{\sqrt 2}{2\sqrt 5}D^{*-\Uparrow}\Sigma_c^{++11,\frac32\frac12}+\frac{\sqrt 5}{2\sqrt 3}D^-\Sigma_c^{++11,\frac32\frac32}+\frac{\sqrt 3}{2\sqrt 5}D^{*-(0)}\Sigma_c^{++11,\frac32\frac32}. \eea \subsection {Pentaquark $P_{\bar cS(cu)A(ud) } ^{\frac12\frac12,\frac12\frac12}$} Using diquarks this state can be represented as follows: \bea\label{p1} &&P_{\bar cS(cu)A(ud)} ^{\frac12\frac12,\frac12\frac12}= -\frac{1}{\sqrt3}\; \Big[\bar c\cdot S_{cu}^{\frac12\frac12,00}\cdot A_{ud}^{10,1J_z}\Big]^{\frac12\frac12} +\sqrt{\frac23}\;\Big[\bar c\cdot S_{cd}^{\frac12-\frac12,00}\cdot A_{uu}^{11,1J_z}\Big]^{\frac12\frac12}=\nn\\ &&=-\frac{1}{\sqrt 3}[\sqrt{\frac23}\bar c^{\downarrow}S_{cu}^{\frac12\frac12,00}A_{ud}^{10,11}-\frac{1}{\sqrt 3}c^{\uparrow}S_{cu}^{\frac12\frac12,00}A_{ud}^{10,10}]+\nn\\ &&+\sqrt{\frac23}[\sqrt{\frac23}\bar c^{\downarrow}S_{cd}^{\frac12-\frac12,00}A_{uu}^{11,11}-\frac{1}{\sqrt 3}c^{\uparrow}S_{cd}^{\frac12-\frac12,00}A_{uu}^{11,10}]. \eea Using formulae from \ref{appB}, we can rewrite this expression in terms of mesons and baryons: \bea\label{p2} &&P_{\bar cS(cu)A(ud)} ^{\frac12\frac12,\frac12\frac12}=-\frac{\sqrt 2}{3}J/\Psi^{(0)}N^{\frac12\frac12,\frac32\frac12}-\frac{\sqrt 2}{12}J/\Psi^{(0)}p^{\uparrow}+\frac{\sqrt 2}{12}J/\Psi^{(0)}p^{'\uparrow}+\frac{\sqrt 2}{4}\eta_c p^{'\uparrow}+\nn\\ &&+\frac{\sqrt 3}{3}J/\Psi^{\Downarrow}N^{\frac12\frac12,\frac32\frac32}+\frac13 J/\Psi^{\Uparrow}N^{\frac12\frac12,\frac32-\frac12}+\frac16 J/\Psi^{\Uparrow}p^{\downarrow}-\frac16 J/\Psi^{\Uparrow}p^{'\downarrow}-\frac{\sqrt 2}{4}\eta_c p^{\uparrow}-\\ &&-\frac{\sqrt 6}{18}\bar D^{*0\Uparrow}\Sigma_c^{+10,\frac12-\frac12}-\frac13\bar D^{*0\Downarrow}\Sigma_c^{+10,\frac32\frac32}+\frac{\sqrt 3}{18}\bar D^{*0(0)}\Sigma_c^{+10,\frac12\frac12}+\frac{\sqrt 3}{6}\bar D^0\Sigma_c^{+10,\frac12\frac12}-\nn\\ &&-\frac{\sqrt 3}{9}\bar D^{*0\Uparrow}\Sigma_c^{+10,\frac32-\frac12}+\frac{\sqrt 6}{9}\bar D^{*0(0)}\Sigma_c^{+10,\frac32\frac12}+\frac{\sqrt 3}{9}D^{*-\Uparrow}\Sigma_c^{++11,\frac12-\frac12}-\frac{\sqrt 6}{6}D^-\Sigma_c^{++11,\frac12\frac12}-\nn\\ &&-\frac{\sqrt 6}{18}D^{*-(0)}\Sigma_c^{++11,\frac12\frac12}+\frac{\sqrt 6}{9}D^{*-\Downarrow}\Sigma_c^{++11,\frac32-\frac12}-\frac{2\sqrt 3}{9}D^{*-(0)}\Sigma_c^{++11,\frac32\frac12}+\frac{\sqrt 2}{3}D^{*-\Downarrow}\Sigma_c^{++11,\frac32\frac32}.\nn \eea \subsection {Pentaquark $P_{\bar cA(cu)S(ud) } ^{\frac12\frac12,\frac12\frac12}$} Here we use axial-vector diquarks $A_{cu}^{\frac12\frac12,11}=c^{\uparrow}u^{\uparrow}$, $A_{cu}^{\frac12\frac12,10}=[c^{\uparrow}u^{\downarrow}+c^{\downarrow}u^{\uparrow}]\frac{1}{\sqrt 2}$ and scalar diquark $S_{ud}^{00,00}=\frac12[u^{\uparrow}d^{\downarrow}-u^{\downarrow}d^{\uparrow}-d^{\uparrow}u^{\downarrow}+d^{\downarrow}u^{\uparrow}]$: \bea && P_{\bar cA(cu)S(ud) }^{\frac12\frac12,\frac12\frac12}= \sqrt{\frac23}\bar c^{\downarrow}A_{cu}^{\frac12\frac12,11}S_{ud}^{00,00}-\frac{1}{\sqrt 3}\bar c^{\uparrow}A_{cu}^{\frac12\frac12,10}S_{ud}^{00,00}=\nn\\ && \sqrt{\frac23}[-\bar c^{\downarrow}c^{\uparrow}\cdot u^{\uparrow}S_{ud}^{00,00}+\bar c^{\downarrow}u^{\uparrow}\cdot c^{\uparrow}S_{ud}^{00,00}]-\nn\\ && \frac{1}{\sqrt 6}[-\bar c^{\uparrow}c^{\uparrow}\cdot u^{\downarrow}S_{ud}^{00,00}+\bar c^{\uparrow}u^{\downarrow}\cdot S_{ud}^{00,00}]- \frac{1}{\sqrt 6}[-\bar c^{\uparrow}c^{\downarrow}\cdot u^{\uparrow}S_{ud}^{00,00}+\bar c^{\uparrow}u^{\uparrow}\cdot c^{\downarrow}S_{ud}^{00,00}]. \eea Using meson and baryon states we write: \bea P_{\bar cA(cu)S(ud)}^{\frac12\frac12,\frac12\frac12}&=&-\frac{1}{\sqrt 6}(J/\psi^{(0)}+\eta_c)(p^{\uparrow}+p^{'\uparrow})+\frac{1}{\sqrt 3}(\bar D^{*0(0)}+D^0)\Lambda_c^{+\uparrow}+\nn\\ &+&\frac{1}{\sqrt 6}J/\psi^{\Uparrow}(p^{\downarrow}+p^{'\downarrow})-\frac{1}{\sqrt 6}(\bar D^{*0(0)}-D^0)\Lambda_c^{+\uparrow}+\nn\\ &+&\frac{1}{\sqrt 6}(J/\psi^{(0)}-\eta_c)(p^{\uparrow}+p^{'\uparrow})-\frac{1}{\sqrt 6}\bar D^{*0\Uparrow}\Lambda_c^{+\downarrow}. \eea \section{Conclusion} We considered the recombination process for pentaquark transition to the meson-baryon channel. With the fixed set of quarks we have two transitions in color space. Both color channels have the same weight. However, different meson-baryon channels have different weights due to different phase space and different tresholds. We analized in detail all possible transitions for the system $\bar ccuud$. We suppose that the $s$-wave recombination leads to large decay widths. We see that the $J=\frac 52$ channel only can not decay in the $s$-wave mode; $d$-wave recombination is possible and naturally leads to dumping of width. There is no such dumping of width for other members of pentaquark multiplet with $J=\frac 12,\frac 32$. So we can expect that $P^{\frac 12}$, $P^{\frac 32}$ states are wide and corresponding poles are drowned in complex energy plane. So it is natural that $P(4450)$ is seen experimentally as a narrow peak only.\cite{LHCbB} \section*{Acknowledgments} The paper was supported by grant RSF 16-12-10267. We thank A.V. Anisovich for useful discussion.

\section{Introduction} Given a closed, irreducible 3--manifold, its complexity is the minimum number of tetrahedra in a (pseudo-simplicial) triangulation of the manifold. This number agrees with the complexity defined by Matveev~\cite{Matveev-complexity-1990} unless the manifold is $S^3,$ $\mathbb{R} P^3$ or $L(3,1).$ We denote the complexity of $M$ by $c(M).$ The only known infinite families of closed 3--manifolds for which an exact value of the complexity is known are spherical space forms \cite{Jaco-minimal-2009, Jaco-coverings-2011, Jaco-Z2-2013}. Asymptotic bounds for an infinite family of double branched coverings of certain alternating closed 3--braids are given by Ni and Wu~\cite{Ni-correction-2015} using the results of \cite{Jaco-Z2-2013}. New lower bounds on the complexity of closed 3--manifolds were recently given with different methods by Cha~\cite{Cha-complexity-2016, Cha-topological-2016}. In this paper, we continue to be interested in determining the exact complexity and minimal triangulations of infinite classes of closed 3--manifolds. In order to describe our results, we need the following definition. Let $M$ be a closed, orientable, irreducible, connected 3--manifold, and let $S$ be an embedded closed surface dual to a given $\varphi \in H^1(M;\mathbb{Z}_2).$ An analogue of Thurston's norm \cite{Thurston-norm-1986} is defined in \cite{Jaco-Z2-2013} as follows. If $S$ is connected, let $\chi_{-}(S) = \max \{ 0,-\chi(S)\},$ and otherwise let $$\chi_{-}(S) = \sum_{S_i\subset S} \max \{ 0,-\chi(S_i)\},$$ where the sum is taken over all connected components of $S.$ Note that $S_i$ is not necessarily orientable. Define: $$|| \ \varphi \ || = \min \{ \chi_{-}(S) \mid S \text{ dual to } \varphi\}.$$ The surface $S$ dual to $\varphi \in H^1(M;\mathbb{Z}_2)$ is said to be \emph{$\mathbb{Z}_2$--taut} if no component of $S$ is a sphere and $\chi(S) = -|| \ \varphi \ ||.$ As in \cite{Thurston-norm-1986}, one observes that (after possibly deleting compressible tori) every component of a $\mathbb{Z}_2$--taut surface is non-separating and geometrically incompressible. This follows from the fact that the Klein bottle does not embed in $\mathbb{R} P^3.$ Our previous papers \cite{Jaco-minimal-2009} and \cite{Jaco-Z2-2013} use this norm on $H^1(M;\mathbb{Z}_2)$ to obtain lower bounds on the complexity of 3--manifolds. The first main result (Theorem~\ref{ 1}) in this paper generalises the work from \cite{Jaco-minimal-2009}, and the second main result (Theorem~\ref{ 3}) improves the main result of \cite{Jaco-Z2-2013}. As applications of both results, we give the first infinite families of minimal triangulations of Seifert fibred spaces modelled on Thurston's geometry $\widetilde{\SL_2(\mathbb{R})}.$ \subsection{Rank one} In \cite{Jaco-minimal-2009} we exhibit an infinite family of lens spaces each member $L$ of which has even order fundamental group which satisfy the relationship $c(L) = 1 + 2 || \ \varphi \ ||,$ where $\varphi$ is the generator for $H^1(L;\mathbb{Z}_2).$ We call these lens spaces \emph{balanced} (since another characterisation, using the terminology of \S\ref{}, is that the number of $\varphi$--even edges equals the number of $\varphi$--odd edges). An example is $L = L(2n, 1),$ where $|| \ \varphi \ || = n-2$ and $c(L) = 2n-3.$ Our first main result puts the main result of \cite{Jaco-minimal-2009} in a greater context. It can be summarised with the slogan: \emph{If $M$ contains a non-orientable $\mathbb{Z}_2$--taut surface of genus $n$, then $M$ is at least as complex as $L(2n, 1).$} An analogous statement can also be made for orientable $\mathbb{Z}_2$--taut surfaces. The proof of the below theorem does not use the main result from \cite{Jaco-minimal-2009}, but borrows heavily from its proof. \begin{theorem}\label{ 1} Let $M$ be a closed orientable, irreducible, connected 3--manifold not homeomorphic with $\mathbb{R} P^3,$ and suppose that $0 \neq \varphi \in H^1(M;\mathbb{Z}_2).$ Then $c(M) \ge 1 + 2 || \ \varphi \ ||.$ Moreover, if equality holds, then $M$ is a balanced lens space. \end{theorem} \begin{corollary} \label{} Let $M$ be a closed orientable, irreducible, connected 3--manifold not homeomorphic with a balanced lens space and suppose that $0 \neq \varphi \in H^1(M;\mathbb{Z}_2).$ Then $c(M) \ge 2 + 2 || \ \varphi \ ||.$ \end{corollary} We now describe a number of infinite families of triangulations that are minimal due to Corollary~\ref{}. \textbf{Lens spaces.} This infinite family of triangulations contains lens spaces with fundamental group of even order, whose minimal layered triangulation has all $\varphi$--even edges of degree 4 except for either two $\varphi$--even edges or three $\varphi$--even edges. In the former family of triangulations, we have one edge of degree 3 and one of degree 5, and in the latter we have two edges of degree 3 and one of degree 6. To see that these triangulations are minimal, one checks that the minimal layered triangulation of such a lens space $L$ contains a unique (hence $\mathbb{Z}_2$--taut) non-orientable normal surface $S_\varphi$ representing the non-trivial class in $H^1(L;\mathbb{Z}_2).$ Denoting $\mathfrak{e}$ the total number of even edges, then the number of tetrahedra in the minimal layered triangulation is $2\mathfrak{e},$ and the Euler characteristic of the taut surface is $1-\mathfrak{e}.$ Hence $c(L) \le 2 \mathfrak{e} = 2 + 2\mathfrak{e} -2 = 2 + 2 || \ \varphi \ ||,$ which forces equality by Corollary~\ref{}. See \S\ref{ spaces} for details and a description of these lens spaces. \textbf{Small Seifert fibred spaces with a horizontal taut surface.} The Seifert fibred spaces $$M_{k,m,n} = S^2( (1,1), (2k+1,1), (2m+1,1), (2n+1,1)),$$ where $k, m$ and $n$ are positive integers, are triangulated by \emph{augmented solid tori} having $2(k+m+n+1)$ tetrahedra. These triangulations were first described by Burton~\cite{regina}. We show that this is a minimal triangulation satisfying $ c(M_{k,m,n}) = 2 + 2|| \ \varphi \ ||$. The manifolds $M_{k,m,n} $ are modelled on Thurston's geometry $\widetilde{\SL_2(\mathbb{R})}$, except for the $8$-tetrahedra case $k=m=n=1,$ which has a Nil structure. The minimality of this three-parameter family is explained in \S\ref{}. \subsection{Rank two} Under the additional hypothesis that $M$ is atoroidal, it is shown in \cite{Jaco-Z2-2013} that if $H \le H^1(M;\mathbb{Z}_2)$ is a subgroup of rank two, then $$c(M) \ge 2 + \sum_{0 \neq \varphi \in H} ||\;\varphi\;||.$$ The improved approach to the rank one case given in this paper, namely taking into account the existence of compression discs, similarly applies in the rank two case to show that the hypothesis that $M$ be atoroidal is superfluous. This has also been observed independently by Nakamura~\cite{Nakamura-complexity-2017}. Our next main result adds an important observation to this, namely a characterisation of the minimal triangulations realising the above lower bound. This again allows us to give a simple corollary which determines further infinite families of minimal triangulations of small Seifert fibred spaces. In this introduction, we give the following simplification of the technical Theorem~\ref{ 3} stated in \S\ref{ 2}. \begin{theorem}\label{ 2} Let $M$ be a closed, orientable, irreducible, connected 3--manifold with the property that $H^1(M;\mathbb{Z}_2)$ has rank two. If $$c(M) = 2 + \sum_{0 \neq \varphi \in H^1(M;\mathbb{Z}_2)} ||\;\varphi\;||,$$ then $M$ is a generalised quaternionic space and the minimal triangulation is a twisted layered loop triangulation. \end{theorem} Every generalised quaternionic space with $H^1(M;\mathbb{Z}_2)$ of rank two satisfies the equation in the above theorem. These spaces and their triangulations are described in \cite{Jaco-coverings-2011}. As a consequence, we show that in the Pachner graph of all triangulations of a manifold, there may be triangulations with one tetrahedron more than the minimal triangulation but requiring arbitrarily many 2--3 and 3--2 moves to be reduced to a minimal triangulation. (See \S\ref{}.) \begin{corollary} \label{cor:3+} Let $M$ be a closed, orientable, irreducible, connected 3--manifold such that $H^1(M;\mathbb{Z}_2)$ has rank two. If $M$ is not a generalised quaternionic space, then $$c(M) \ge 3 + \sum_{0 \neq \varphi \in H^1(M;\mathbb{Z}_2)} ||\;\varphi\;||.$$ \end{corollary} \textbf{Small Seifert fibred spaces with three vertical taut surfaces.} The Seifert fibred space $$M'_{k,m,n} = S^2( (1,-1), (2k+2,1), (2m+2,1), (2n+2,1)),$$ where $k, m$ and $n$ are positive integers, is triangulated by an augmented solid torus having $2k+2m+2n+3$ tetrahedra. This has an $\widetilde{\SL_2(\mathbb{R})}$ structure and satisfies $c(M'_{k,m,n}) = 3 + \sum || \ \varphi \ ||.$ See \S\ref{}. In \S\ref{} we also describe an infinite family of minimal triangulations of prism manifolds realising the lower bound in Corollary~\ref{cor:3+}. \textbf{Acknowledgements.} This research was supported through the programme ``Research in Pairs" by the Mathematisches Forschungsinstitut Oberwolfach in 2017. The authors would like to thank the staff at MFO for an excellent collaboration environment. Jaco is partially supported by NSF grant DMS-1308767 and the Grayce B. Kerr Foundation. Spreer is supported by the Einstein Foundation (project ``Einstein Visiting Fellow Santos''). Research of Rubinstein and Tillmann is supported in part under the Australian Research Council's Discovery funding scheme (project number DP160104502). \section{Preliminaries} \label{} This section summarises the results and definitions we need from \cite{Jaco-minimal-2009} in order to carry out the new proofs. The material is repeated here for the benefit of the reader since only a small amount from each subsection in \cite{Jaco-minimal-2009} is needed. The only change is that we have improved the notation in \S\ref{}. \subsection{Triangulations} The notation of \cite{Jaco-0-efficient-2003, Jaco-layered-2006} is used in this paper. A triangulation, $\mathcal{T},$ of a 3--manifold $M$ consists of a union of pairwise disjoint 3--simplices, $\widetilde{\Delta},$ a set of face pairings, $\Phi,$ and a natural quotient map $p\colon\thinspace \widetilde{\Delta} \to \widetilde{\Delta} / \Phi = M.$ Since the quotient map is injective on the interior of each 3--simplex, we refer to the image in $M$ of a 3--simplex in $\mathcal{T}$ as a \emph{tetrahedron} and to its faces, edges and vertices with respect to the pre-image. Similarly for images of 2--simplices, 1--simplices and vertices, which are referred to as \emph{faces,} \emph{edges} and \emph{vertices} in $M,$ respectively. For an edge $e,$ the number of pairwise distinct 1--simplices in $p^{-1}(e)$ is referred to as its \emph{degree}, denoted $d(e).$ If an edge is contained in $\partial M,$ then it is termed a \emph{boundary edge}; otherwise it is referred to as an \emph{interior edge}. \subsection{Minimal and 0--efficient triangulations} A triangulation of the closed, orientable, connected 3--manifold $M$ is termed \emph{minimal} if $M$ has no triangulation with fewer tetrahedra. A triangulation of $M$ is termed \emph{0--efficient} if the only embedded, normal 2--spheres are vertex linking. It is shown by the first two authors in \cite{Jaco-0-efficient-2003} that (1) the existence of a 0--efficient triangulation implies that $M$ is irreducible and $M\neq \mathbb{R} P^3,$ (2) a minimal triangulation is 0--efficient unless $M=\mathbb{R} P^3$ or $L(3,1),$ and (3) a 0--efficient triangulation has a single vertex unless $M=S^3$ and the triangulation has precisely two vertices. These facts are used implicitly throughout the article. \subsection{Layered triangulations of the solid torus and lens spaces} \label{ triangulations of the solid torus and lens spaces} Layering along a boundary edge is defined in \cite{Jaco-layered-2006} and illustrated in Figure \ref{} on the left. We only need this in the restricted setting where $M= D^2 \times S^1$ is a solid torus and $\mathcal{T}_\partial$ is a triangulation of $\partial M$ with one vertex. Given a solid torus with a one-vertex triangulation on its boundary, it follows from an Euler characteristic argument that this triangulation has two faces on the boundary meeting along three edges. Suppose $e$ is an edge in $\partial M.$ This is incident to two distinct faces. We say the 3--simplex $\Delta$ is \emph{layered along the (boundary) edge} $e$ if two faces of $\Delta$ are paired, ``without a twist," with the two faces of $\mathcal{T}_\partial$ incident with $e$. The resulting 3--manifold is homeomorphic with $M$. If $\mathcal{T}_\partial$ is the restriction of a triangulation $\mathcal{T}$ of $M$ to $\partial M$, then we obtain a new triangulation of $M.$ \begin{figure}[htb] \centering{\includegraphics[width=\textwidth]{layering}} \caption{Left: Layering on a boundary edge. Right: Layered triangulation of the solid torus \label{}} \end{figure} Starting point for a layered triangulation of a solid torus is the one-triangle triangulation of the M\"obius band shown in Figure~\ref{} on the right and considered as a degenerate solid torus and labeled $\mathbf{T}(1,1,2).$ The first layer is a single tetrahedron layered on the edge labelled $e$ of the M\"obius band in Figure~\ref{}, with the two back faces of the tetrahedron identified as indicated by the edge labels. One can then \emph{layer} on any of the three boundary edges (see \cite{Jaco-layered-2006}), giving a layered triangulation of the solid torus with two tetrahedra. Inductively, a layered triangulation of a solid torus with $k$ tetrahedra is obtained by layering on a boundary edge of a layered triangulation with $k-1$ tetrahedra, $k \ge 1.$ See also Figure~\ref{} for an encoding of these structures by the so-called \emph{L--graph.} A layered triangulation of a solid torus with $k$ tetrahedra has one vertex, $2k+1$ faces, and $k+2$ edges. Except in the degenerate case, the vertex, three edges, and two faces are contained in the boundary, and there is a special edge in the boundary having degree one; it is called the \emph{univalent edge}. There is a unique set of unordered nonnegative integers, $\{p,q,p+q\}$, associated with the triangulation on the boundary and determined from the geometric intersection of the meridional slope of the solid torus with the three edges of the triangulation. Two such triangulations with numbers $\{p,q,p+q\}$ and $\{p',q',p'+q'\}$ can be carried to each other via a homeomorphism of the solid torus if and only if the two sets of numbers are identical. There are three possible ways to identify the two faces in the boundary, each producing a lens space. These identifications can be thought of as ``folding along an edge in the boundary", which is also referred to as ``closing the book" with an edge as the binding. Folding along the edge $p$ gives the lens space $L(2q+p,q)$; along the edge $q$ the lens space $L(2p+q,p)$; and along the edge $p+q$ the lens space $L(\abs{p-q},p)$, see \cite{Jaco-layered-2006}. If the solid torus is triangulated, then after identification one obtains a triangulation of the lens space. If the triangulation of the solid torus is a layered triangulation of the solid torus, then the induced triangulation of the lens space is termed a \emph{layered triangulation} of the lens space. \subsection{Edges of low degree in minimal triangulations} \label{ degree edges} Given a 1--vertex triangulation of a closed 3--manifold, let $E$ denote the number of edges, $E_i$ denote the number of edges of degree $i,$ and $T$ denote the number of tetrahedra. Then $E = T+1,$ and hence $6 = \sum (6-i)E_i$ since the Euler characteristic of a closed 3--manifold is zero and that of the vertex link is two. It follows that there exist at least two edges of degree at most five. We summarise the results we use; a discussion can be found in \cite{Jaco-minimal-2009}. \begin{proposition}[Edges of degree one or two, \cite{Jaco-0-efficient-2003, Jaco-minimal-2009}]\label{ one and two edges} A minimal and 0--efficient triangulation $\mathcal{T}$ of the closed, orientable, connected and irreducible 3--manifold $M$ has \begin{enumerate} \item[(1)] no edge of order one unless $M=S^3,$ and \item[(2)] no edge of order two unless $M=L(3,1)$ or $L(4,1).$ \end{enumerate} \end{proposition} \begin{proposition}[Edges of degree three, \cite{Jaco-0-efficient-2003, Jaco-minimal-2009}]\label{ three edges} A minimal and 0--efficient triangulation $\mathcal{T}$ of the closed, orientable, connected and irreducible 3--manifold $M$ has no edge of order three unless either \begin{enumerate} \item[(3a)] $\mathcal{T}$ contains a single tetrahedron and $M=L(5,2);$ or \item[(3b)] $\mathcal{T}$ contains two tetrahedra and $M=L(5,1)$ or $L(7,2);$ or \item[(3c)] $\mathcal{T}$ contains, as an embedded subcomplex, the two tetrahedron, layered triangulation $\mathbf{T} (1,3,4)$ of the solid torus. Moreover, $\mathcal{T}$ contains at least three tetrahedra and every edge of degree three is an interior edge of such a subcomplex. \end{enumerate} Moreover, the triangulations in (3b) are obtained by identifying the boundary faces of $\mathbf{T} (1,3,4)$ appropriately. \end{proposition} \subsection{Intersections of maximal layered solid tori} \label{ of maximal layered solid tori} Throughout this section, assume that $M$ is an irreducible, orientable, connected 3--manifold with a fixed triangulation, $\mathcal{T}.$ Further assumptions are stated. \begin{definition}(Layered solid torus) A \emph{layered solid torus with respect to $\mathcal{T}$} in $M$ is a subcomplex in $M$ which is combinatorially equivalent to a layered solid torus. Any reference to $\mathcal{T}$ is suppressed when $\mathcal{T}$ is fixed. The edges of the layered solid torus $\mathbf{T}$ are termed as follows: If there exists more than one tetrahedron, then there is a unique edge which has been layered on first, called the \emph{base-edge (of $\mathbf{T}$).} The edges in the boundary of $\mathbf{T}$ are called \emph{boundary edges (of $\mathbf{T}$)} and all other edges (including the base-edge) are termed \emph{interior edges (of $\mathbf{T}$).} \end{definition} For every edge in $M,$ we refer to its degree as its $M$--degree, and its degree with respect to the layered solid torus $\mathbf{T}$ in $M$ is called its $\mathbf{T}$--degree. The $\mathbf{T}$--degree of an edge not contained in $\mathbf{T}$ is zero. The unique edge of $\mathbf{T}$--degree one is termed a \emph{univalent edge (for $\mathbf{T}$)}. Clearly, $\mathbf{T}$--degree and $M$--degree agree for all interior edges of $\mathbf{T}.$ \begin{definition}(Maximal layered solid torus) A layered solid torus is a \emph{maximal layered solid torus with respect to $\mathcal{T}$} in $M$ if it is not strictly contained in any other layered solid torus in $M.$ \end{definition} A layered triangulation of a lens space with at least three tetrahedra contains precisely two maximal layered solid tori, and they meet in all but two of the tetrahedra in the triangulation. In general, the intersection of maximal layered solid tori is small: \begin{lemma}\cite{Jaco-0-efficient-2003, Jaco-minimal-2009}\label{ of maximal layered solid tori} Assume that the triangulation $\mathcal{T}$ is minimal and 0--efficient. If $M$ is not a lens space with layered triangulation, then the intersection of two distinct maximal layered solid tori in $M$ consists of at most a single edge. \end{lemma} \begin{lemma}\cite{Jaco-0-efficient-2003, Jaco-minimal-2009}\label{ lens char 2} Assume that the triangulation $\mathcal{T}$ is minimal and 0--efficient, and suppose that $M$ contains a layered solid torus, $\mathbf{T},$ made up of at least two tetrahedra and having a boundary edge, $e,$ which has degree four in $M.$ Then either \begin{enumerate} \item $\mathbf{T}$ is not a maximal layered solid torus in $M;$ or \item $e$ is the univalent edge for $\mathbf{T}$ and it is contained in four distinct tetrahedra in $M;$ or \item $M$ is a lens space with minimal layered triangulation. \end{enumerate} \end{lemma} \subsection{Colouring of edges and dual normal surface} \label{} Throughout this section, let $\mathcal{T}$ be an arbitrary 1--vertex triangulation of a closed 3--manifold $M,$ and let $\varphi \colon\thinspace \pi_1(M) \to \mathbb{Z}_2$ be a non--trivial homomorphism. Each edge, $e,$ is given a fixed orientation, and hence represents an element $[e]\in \pi_1(M).$ If $\varphi[e]=0,$ the edge is termed $\varphi$--even, otherwise it is termed $\varphi$--odd. This terminology is independent of the chosen orientation for $e.$ Faces in the triangulation give relations between loops represented by edges. It follows that a tetrahedron falls into one of the following categories, see Figure~\ref{ class}: \begin{itemize} \item[] Type $\Delta_{\mathfrak{q}}$: A pair of opposite edges are $\varphi$--even, all others are $\varphi$--odd. \item[] Type $\Delta_{\mathfrak{t}}$: The three edges incident to a vertex are $\varphi$--odd, all others are $\varphi$--even. \item[] Type $\Delta_{\emptyset}$: All edges are $\varphi$--even. \end{itemize} \begin{figure}[htb] \begin{center} \includegraphics[width=.6\textwidth]{tetrahedra_1} \end{center} \caption{Types of tetrahedra and normal discs of the dual surface} \label{ class} \end{figure} It follows from the classification of the tetrahedra in $\mathcal{T}$ that, if $\varphi$ is non-trivial, then one obtains a unique normal surface, $S_\varphi(\mathcal{T}),$ with respect to $\mathcal{T}$ by introducing a single vertex on each $\varphi$--odd edge. This surface is disjoint from the tetrahedra of type $\Delta_{\emptyset}$; it meets each tetrahedron of type $\Delta_{\mathfrak{t}}$ in a single triangle meeting all $\varphi$--odd edges; and each tetrahedron of type $\Delta_{\mathfrak{q}}$ in a single quadrilateral dual to the $\varphi$--even edges. Moreover, $S_\varphi(\mathcal{T})$ is dual to the $\mathbb{Z}_2$--cohomology class represented by $\varphi.$ \begin{lemma}\cite{Jaco-minimal-2009} Assume that the triangulation $\mathcal{T}$ contains a single vertex and that all edge loops are coloured using a homomorphism $\varphi \colon\thinspace \pi_1(M) \to \mathbb{Z}_2.$ Then all tetrahedra in a layered solid torus in $M$ are either of type $\Delta_{\mathfrak{q}}$ or type $\Delta_{\emptyset}$, but not both. \end{lemma} \begin{definition}[(Types of layered solid tori)] Assume that the triangulation $\mathcal{T}$ contains a single vertex and that all edge loops are coloured using a homomorphism $\varphi \colon\thinspace \pi_1(M) \to \mathbb{Z}_2.$ A layered solid torus containing a tetrahedron of type $\Delta_{\mathfrak{q}}$ (respectively $\Delta_{\emptyset}$) is accordingly termed of type $\Delta_{\mathfrak{q}}$ (respectively $\Delta_{\emptyset}$). \end{definition} \section{Rank one} \label{ one} In this section, we give the proof of Theorem~\ref{ 1}. Suppose $M$ is an orientable, irreducible, closed 3--manifold not homeomorphic with $\mathbb{R} P^3,$ and assume $\varphi$ is a non-trivial element of $H^1(M;\mathbb{Z}_2).$ By hypothesis, $\varphi$ is dual to a surface of non-negative Euler characteristic. Suppose that $c(M) \le 1 + 2 || \ \varphi \ ||.$ To prove Theorem~\ref{ 1} we need to show that $M$ is a balanced lens space. To simplify notation, we write $g-2=|| \ \varphi \ ||.$ We point out that \emph{if} $\varphi$ can be represented by a connected non-orientable surface, \emph{then} $g$ is its genus. Otherwise it is merely a useful symbol for us. Using this definition of $g,$ this gives the slogan from the introduction: $$c(M) \le 1 + 2 || \ \varphi \ || = 2g - 3 = c(L(2g,1)).$$ Suppose we have an arbitrary minimal triangulation of $M$ and colour the edges using $\varphi.$ \subsection{Promoting triangulations with edge flips} \label{} We first modify the triangulation of $M$ using 4--4 bistellar flips to rule out maximal layered solid tori of special types. This follows the argument in \cite{Jaco-minimal-2009} on pages 173--175. We note that this does not change the number of tetrahedra, hence producing other minimal triangulations. A maximal layered solid torus $\mathbf{T}$ of type $\Delta_{\mathfrak{q}}$ containing a $\varphi$--even edge of degree three and having the property that all other $\varphi$--even edges in its interior or boundary have degree four is called \emph{supportive}. It follows from Lemma~\ref{ lens char 2} that the $\varphi$--even boundary edge $e$ of $\mathbf{T}$ is its univalent edge and contained in four pairwise distinct tetrahedra in the triangulation. As in \cite{Jaco-minimal-2009}, the complex formed by the four tetrahedra around $e$ is termed an octahedron (even though it may not be embedded). There are four cases to consider for the types of tetrahedra ordered cyclically around $e,$ and we start the labelling with the tetrahedron in $\mathbf{T}$ of type $\Delta_{\mathfrak{q}}$ containing $e.$ The proofs of the following statements are in \cite{Jaco-minimal-2009}, where the notation for the types is 1 instead of $\Delta_{\mathfrak{q}}$; 2 instead of $\Delta_{\mathfrak{t}}$, and 3 instead of $\Delta_{\emptyset}.$ \begin{itemize} \item[] $(\Delta_{\mathfrak{q}},\Delta_{\mathfrak{q}},\Delta_{\mathfrak{q}},\Delta_{\mathfrak{q}})$: Any 4--4 bistellar flip on the octahedron results in four tetrahedra of type $\Delta_{\mathfrak{t}},$ and gives a triangulation with fewer supportive solid tori and the same number of tetrahedra of type $\Delta_{\emptyset}.$ \item[] $(\Delta_{\mathfrak{q}},\Delta_{\mathfrak{t}},\Delta_{\mathfrak{t}},\Delta_{\mathfrak{q}})$ or $(\Delta_{\mathfrak{q}},\Delta_{\mathfrak{q}},\Delta_{\mathfrak{t}},\Delta_{\mathfrak{t}})$: Performing a specific 4--4 bistellar flip on the octahedron results in a triangulation with fewer supportive solid tori and the same number of tetrahedra of type $\Delta_{\emptyset}.$ \item[] $(\Delta_{\mathfrak{q}},\Delta_{\mathfrak{t}},\Delta_{\emptyset},\Delta_{\mathfrak{t}})$: Any 4--4 bistellar flip on the octahedron results in tetrahedra of types $\Delta_{\mathfrak{q}},\Delta_{\mathfrak{q}},\Delta_{\mathfrak{t}},\Delta_{\mathfrak{t}},$ and hence reduces the number of tetrahedra of type $\Delta_{\emptyset}.$ \end{itemize} It follows by induction that after finitely many edge flips, we may assume that the triangulation does not contain any supportive solid tori. \subsection{The fundamental equation} Let \begin{itemize} \item[] $n_{\mathfrak{q}}=$ number of tetrahedra of type $\Delta_{\mathfrak{q}}$, \item[] $n_{\mathfrak{t}}=$ number of tetrahedra of type $\Delta_{\mathfrak{t}}$, \item[] $n_{\emptyset}=$ number of tetrahedra of type $\Delta_{\emptyset}$, \item[] $\mathfrak{e}(\mathcal{T})=$ number of $\varphi$--even edges, item[] $\tilde{\mathfrak{e}}(\mathcal{T})=$ number of pre-images of $\varphi$--even edges in $\widetilde{\Delta}.$ \end{itemize} Since the triangulation is 0--efficient, no component of $S_\varphi=S_\varphi(\mathcal{T})$ is a sphere or a projective plane. Hence $$2-g = - || \ \varphi \ || \ge \chi (S_\varphi)$$ with equality if $S_\varphi$ is taut. If $S_\varphi$ is not geometrically incompressible but admits $k_\varphi$ pairwise disjoint independent compression discs, then this inequality may be written as: $$2-g = - || \ \varphi \ || \ge \chi (S_\varphi) + 2k_\varphi$$ Consider the complex $K$ spanned by the $\varphi$--even edges. Then taking the normal surface corresponding to twice the normal coordinate of $S_\varphi$ gives the boundary of a small regular neighbourhood of $K.$ Hence $$2 \chi(S_\varphi) = \chi(\partial K) = 2 \chi(K) = 2 \big( 1 - \mathfrak{e} + \frac{n_{\mathfrak{t}} + 4 n_{\emptyset}}{2} - n_{\emptyset} \big) = 2 - 2 \mathfrak{e} + {n_{\mathfrak{t}}} + 2n_{\emptyset} $$ This gives: $$2(2-g) \ge 2\chi (S_\varphi) + 4k_\varphi= 2 - 2\mathfrak{e} + n_{\mathfrak{t}} + 2n_{\emptyset}+ 4k_\varphi.$$ The number $c(M)$ of all tetrahedra satisfies $$2g-3 \ge c(M) = n_{\mathfrak{q}} + n_{\mathfrak{t}} + n_{\emptyset}.$$ We also have $\sum d \mathfrak{e}_d = \tilde{\mathfrak{e}} = 2 n_{\mathfrak{q}} + 3 n_{\mathfrak{t}} + 6 n_{\emptyset}.$ This gives: \begin{align*} 2 = 4(2-g) + 2(2g-3) \ge& 4 \chi(S_\varphi) + 8 k_\varphi+ 2 c(M)\\ =& 4 + 8 k_\varphi + \sum (d-4) \mathfrak{e}_d + n_{\mathfrak{t}}. \end{align*} We may assume that $M$ is not the lens space $L(4,1)$ since in this case the conclusion follows by inspection of the triangulation with a single tetrahedron of $L(4,1).$ Hence the smallest degree of an edge in $M$ is 3. Rearranging the inequality gives: \begin{equation}\label{inequ for min tri} \mathfrak{e}_3 \ge 2 + \sum_{j=5}^{\infty} (j-4) \mathfrak{e}_j + 8 k_\varphi + n_{\mathfrak{t}}. \end{equation} We first remark that if $M$ is a lens space \emph{and} the minimal triangulation is a layered triangulation, then a direct argument shows that $M$ is a balanced lens space. This comes from the fact that in a minimal layered triangulation there are at most two edges of degree three. Hence inequality (\ref{inequ for min tri}) must be an equality, $\mathfrak{e}_3=2$ and all other terms must vanish. In particular, apart from the two $\varphi$--even edges of degree 3 all other $\varphi$--even edges have degree 4. The lens spaces satisfying this condition can now be constructed and identified (see \S2.4, Corollary 30 and Figure 6 in \cite{Jaco-minimal-2009}). Hence we assume that it is not the case that $M$ is a lens space and the minimal triangulation is a layered triangulation. We aim to show that this gives a contradiction. Since the triangulation has at least two even edges of degree three, Proposition~\ref{ three edges} and Lemma~\ref{ of maximal layered solid tori} imply that there are at least two maximal layered solid tori, and that any two maximal layered solid tori meet in at most one edge. \subsection{Deficit from maximal layered solid tori of type $\Delta_{\emptyset}$} As in \cite{Jaco-minimal-2009}, one sees that the collection of maximal layered solid tori of type $\Delta_{\emptyset}$ containing a degree 3 edge contribute a \emph{deficit} to the inequality (\ref{inequ for min tri}), in the sense that the contribution to the right hand side from $\sum_{j=5}^{\infty} (j-4) \mathfrak{e}_j$ exceeds the contribution to the left hand side $\mathfrak{e}_3.$ The argument is as follows. Any maximal layered solid tori of type $\Delta_{\emptyset}$ having an internal edge of degree at least 5, either \emph{balances} out or contributes a \emph{deficit} to the inequality (\ref{inequ for min tri}). Hence assume we are left with all such tori having one internal edge of degree 3 and all others of degree 4. The edges on the boundary have internal degrees $1, 3, 3+2m,$ where $m \ge 1.$ The edge $e_0$ of internal degree $3+2m$ is the longitude of the layered solid torus. Now consider the collection of all maximal layered solid tori of type $\Delta_{\emptyset}$ meeting in this longitude $e_0$ (some may meet it in their longitude, some in another edge). Whence the degree in $M$ of the longitude is at least 6 (in fact, at least 7). Write this as $5 + k,$ where $k\ge 1.$ Now at most $\frac{k+1}{2}$ layered solid tori can meet in this edge, since no two meet in more than one edge. Hence the contribution to the left hand side is at most $\frac{k+1}{2}$ (and this may be double counting some degree 3 edges), and the contribution to the right hand side is $d(e_0) - 4 = 1+k,$ where $k \ge 0,$ Whence one obtains a deficit since $\frac{k+1}{2} < k+1.$ It follows that there must be at least two maximal layered solid tori of type $\Delta_{\mathfrak{q}},$ and that from the collection of all such maximal layered solid tori there must be a \emph{gain} of at least two. \subsection{No gain from maximal layered solid tori of type $\Delta_{\mathfrak{q}}$} First note that given any maximal layered solid tori of type $\Delta_{\mathfrak{q}}$ having an internal $\varphi$--even edge of degree at least five, then it either \emph{balances} out or contributes a \emph{deficit} to the inequality (\ref{inequ for min tri}). Hence we now restrict to the collection of all maximal layered solid tori of type $\Delta_{\mathfrak{q}}$ containing a $\varphi$--even edge of degree three and having all other interior $\varphi$--even edges of degree four. Since there are no supportive tori (cf. \S\ref{}), we know that the even boundary edge $e$ of such a torus $\mathbf{T}$ has degree at least five in $M.$ We call $\mathbf{T}$ an \emph{almost supportive torus}. If $e$ is not contained in any other almost supportive torus, then again we have either a balance or a deficit. This shows that there is some almost supportive torus $\mathbf{T}$ with the property that it shares its even boundary edge with other almost supportive tori. If $k$ almost supportive tori meet in an even degree edge $e$ of degree $d,$ then $d \ge 2k$ by Lemma~\ref{ of maximal layered solid tori}. For at least one collection of such tori, there must be a positive contribution. Hence $k \ge 1 + d-4 = d-3$ and so $k+3 \ge d \ge 2k$ for at least one collection. This gives $k\le 3.$ Since we also have $d\ge 5,$ we have the following cases: \begin{enumerate} \item $d=6$ and $k=3;$ \item $d=5$ and $k=2.$ \end{enumerate} Assume that $d=6$ and $k=3.$ It follows that $e$ is contained in six pairwise distinct tetrahedra, with supportive tori alternating around $e.$ All these tetrahedra have to be of type $\Delta_{\mathfrak{q}}.$ Thus, there is a compression disc for $S_\varphi$ associated with $e.$ This compression disc has boundary on the six quadrilateral discs in the tetrahedra containing $e;$ see Figure~\ref{}. This contributes $8$ to the right hand side of the equation. We now adjust the contribution of $e$ and the even edges in the boundary of the abstract neighbourhood as follows. Each $\varphi$--even edge not in the interior of an almost supportive torus is given a contribution of 1 (leaving us with at least 4 to still balance). We then perform the compression, resulting in a new normal surface. However, we do not change the types of the tetrahedra. Note that in each tetrahedron associated with the compression, one normal quadrilateral is replaced with two normal triangles and the two even edges of the tetrahedron are adjusted with an extra weight of $1,$ i.e. increasing the contribution of its edge degree by one. The resulting surface is normal and since the triangulation is 0--efficient and $M$ is not $\mathbb{R} P^3,$ no resulting component from the compression is a sphere or projective plane. We also note that if one of the other even edges involved in this process is of the type ($d=6$ and $k=3$) or ($d=5$ and $k=2$), then it now is counted as ($d+1=6+1=7$ and $k=3$) or ($d+1=5+1=6$ and $k=2$) and hence gives a balance or deficit. See Figure~\ref{} for details. \begin{figure}[htb] \centerline{\includegraphics[width=.7\textwidth]{compression}} \caption{Compression in the case $d=6$ and $k=3$. A similar situation occurs in the case $d=5$ and $k=2$. \label{}} \end{figure} Thus, for each edge of type ($d=6$ and $k=3$) we either perform a compression of $S_\varphi$ or, if any of the tetrahedra in its link have already been used in a compression of $S_\varphi$, we do not obtain a gain from it. Also note that this process never splits off spheres or projective planes. Hence there must be an edge $e$ of type ($d=5$ and $k=2$) that we have not considered yet. The abstract neighbourhood of $e$ contains three tetrahedra of type $\Delta_q,$ namely two from the two almost supportive tori $\mathbf{T}_1$ and $\mathbf{T}_2$, and one meeting each of these in a face. The remaining tetrahedra are either of both of type $\Delta_q,$ or both of type $\Delta_t.$ We first show that they must be distinct. First consider the case where all tetrahedra are of type $\Delta_q.$ Then the tetrahedra are pairwise distinct since otherwise $\mathbf{T}_1$ and $\mathbf{T}_2$ meet in more than one edge. Hence there is a compression disc for $S_\varphi$ associated with the edge $e$, see Figure~\ref{}. As in the case of ($d=6$ and $k=3$) this allows us to distribute an extra weight to each of the edges $e$ and the $\varphi$--even edges not in the interior of $\mathbf{T}_1$ or $\mathbf{T}_2$ but incident with the collection of five tetrahedra around $e.$ Next, consider the case where three tetrahedra in the abstract edge neighbourhood of $e$ are of type $\Delta_q,$ and the remaining two of type $\Delta_t.$ The tetrahedra of type $\Delta_q$ are pairwise distinct since each of $\mathbf{T}_1$ and $\mathbf{T}_2$ contains exactly one of them and the remaining shares a face with each. The tetrahedra of type $\Delta_t$ are distinct since each has only three $\varphi$--odd edges, and the union of their $\varphi$--odd edges contains the four $\varphi$--odd edges in the boundaries of $\mathbf{T}_1$ and $\mathbf{T}_2.$ Hence there is a contribution of 2 to the right hand side of inequality (\ref{inequ for min tri}) from the two tetrahedra of type $\Delta_t.$ The two tetrahedra of type $\Delta_t$ meet in a face containing three $\varphi$--even edges. If not all of these are of the type ($d=5$ and $k=2$) and have not been visited before, then we obtain a total balance or deficit. Hence assume all three are of type ($d=5$ and $k=2$) and have not been visited before. In this case, we have two normal triangles (from the two tetrahedra of type $\Delta_t$ and nine normal quadrilaterals (three each from the three $\varphi$--even edges of degree 5) forming a subsurface of $S_\varphi$ that is a thrice punctured sphere. This comes from the fact that no two of the $\varphi$--even edge can now be identified since otherwise two maximal layered solid tori meet in two edges. We replace this thrice punctured sphere by three discs---this corresponds to two compressions and an isotopy on $S_\varphi.$ The resulting surface is again normal: in each tetrahedron of type $\Delta_t$ we replace the unique normal triangles by the other three normal triangles; and in each tetrahedron of type $\Delta_q$ involved in the links of the three $\varphi$--even edges we replace the normal quadrilateral by the two normal triangles meeting the $\varphi$--even edge on the boundary of the layered solid torus. On the level of normal surface theory, this amounts to adding to the normal coordinate of $S_\varphi$ the \emph{edge solutions} corresponding to the three $\varphi$--even edges and subtracting the \emph{tetrahedral solutions} corresponding to the tetrahedra of type $\Delta_t$ (see \cite[\S2.5-2.6]{Luo-angle-2008}). Here, our convention is that edge and tetrahedral solutions have negative quadrilateral coordinates. Then the resulting normal coordinate has non-negative coordinates, the same edge weights modulo 2 as $S_\varphi$ (hence representing the same class) and its Euler characteristic is $\chi(S_\varphi)+4,$ since tetrahedral solutions have formal Euler characteristic 1 and edge solutions have formal Euler characteristic 2. See Figure~\ref{} for an illustration of this process. Now the overall contribution to the left hand side is 6, and the overall contribution to the right hand side is 3 from the degree 5 edges, plus 2 from the type $\Delta_t$ tetrahedra and 16 from the two compressions. Hence we have an overall deficit associated with the three type ($d=5$ and $k=2$) edges. \begin{figure}[htb] \centerline{\includegraphics[width=\textwidth]{doubleCompression}} \caption{The double compression in the case $d=5$ and $k=2$. \label{}} \end{figure} It follows that no type ($d=5$ and $k=2$) edge can result in a gain, and hence we have a contradiction to our assumptions. This completes the proof of Theorem~\ref{ 1}. \section{Rank two} \label{ 2} Before we state our second main result, we describe the \emph{twisted squares} from \cite{Luo-combinatorial-2017}. Two pairs of opposite edges in a tetrahedron are the boundary of a quadrilateral properly embedded in the tetrahedron. If opposite pairs of these edges are identified, then one either obtains a pinched projective plane, an embedded Klein bottle or an embedded torus. These are shown in Figure~\ref{} and the latter two are called, respectively, a \emph{Klein square} and a \emph{torus square}. \begin{figure}[htb] \centering{\includegraphics[width=.7\textwidth]{cases}} \caption{Tetrahedron with two pairs of opposite edges identified forming a twisted square of type \emph{real projective plane} (left), \emph{Klein bottle} (center) or \emph{torus} (right). \label{}} \end{figure} As discussed in the introduction, replacing the use of the lemmata in \cite{Jaco-Z2-2013} that assume $M$ atoroidal with an improvement of the counting argument analogous to \S\ref{ one} by taking into account compression discs shows that if $H \le H^1(M;\mathbb{Z}_2)$ is a subgroup of rank two, then $$c(M) \ge 2 + \sum_{0 \neq \varphi \in H} ||\;\varphi\;||.$$ A detailed argument is given by Nakamura~\cite{Nakamura-complexity-2017}. Our next main result characterises precisely those manifolds that realise the lower bound. Theorem~\ref{ 2} is a direct corollary of this result. \begin{theorem}\label{ 3} Let $M$ be a closed orientable, irreducible, connected 3--manifold. Suppose that $H \le H^1(M;\mathbb{Z}_2)$ is a subgroup of rank two. If $c(M) = 2 + \sum_{0 \neq \varphi \in H} ||\;\varphi\;||,$ then every minimal triangulation contains a tetrahedron with a Klein square or a torus square. Moreover, either \begin{enumerate} \item $M$ is a generalised quaternionic space and $H = H^1(M;\mathbb{Z}_2)$; or \item the rank of $H^1(M;\mathbb{Z}_2)$ is at least three and a Klein square is an incompressible surface in $M;$ or \item the rank of $H^1(M;\mathbb{Z}_2)$ is at least three and a torus square is an incompressible non-separating surface in $M.$ \end{enumerate} \end{theorem} \begin{proof} Suppose we have a triangulation $M$ realising the bound $c(M) = 2 + \sum_{0 \neq \varphi \in H} ||\;\varphi\;||$ for $H \le H^1(M;\mathbb{Z}_2)$ subgroup of rank $2$. It follows from \cite{Jaco-Z2-2013} (see also \cite{Nakamura-complexity-2017}) that, given $0 \neq \varphi \in H$, then all tetrahedra of $M$ are of type $\Delta_{\mathfrak{q}}$ and the canonical normal representative $S_\varphi$ is taut. Denote the edges in the kernel of $\varphi$ by $e_1, \dots , e_k$. We identify vectors in $\mathbb{Z}_2^{c(M)+1}$ with subsets of edges of $M$ in the natural way. It follows that for each $b \in \mathbb{Z}_2^k \subset \mathbb{Z}_2^{c(M)+1}$ we can modify $S_\varphi$ in the following way: for every $b_i = 1$, increase the edge weight of $e_i$ from $0$ to $2$ and leave all other edge weights the same. This gives a normal surface $S_\varphi^b$ still representing $\varphi$. The new surface intersects the tetrahedra of $M$ in three possible ways illustrated in Figure~\ref{}. We have for the Euler characteristic of $S_\varphi^b$: $$ \chi (S_\varphi^b) = \chi(S_\varphi) -2 \mathfrak{o} (b) + 2 |b|$$ where $\mathfrak{o} (b)$ denotes the number of octagons in $S_\varphi^b$ and $|b|$ is the $L^1$--norm. \begin{figure}[htb] \centering{\includegraphics[width=.7\textwidth]{modifyS}} \caption{A quadrilateral disc in $S$ changes to two normal triangles or one normal octagon, depending on whether only one edge opposite the quadrilateral has its weight increased to two or both. \label{}} \end{figure} Since $S_\varphi$ is taut we have $\chi (S_\varphi) \geq \chi (S_\varphi^b) $ and hence $\mathfrak{o} (b) \geq |b|$. Since each edge in the triangulation $M$ is in the kernel for some non-trivial $\varphi \in H$, we know that each unit vector in $\mathbb{Z}_2^{c(M)+1}$ gives at least one octagon. Each unit vector in $\mathbb{Z}_2^{c(M)+1}$ corresponds to one edge and hence there exists at least one tetrahedron with a pair of opposite edges identified with this edge. Since there are strictly more edges than tetrahedra, the pigeonhole principle implies that there exists at least one tetrahedron in which this happens for at least two pairs of opposite edges. This leaves us with three cases illustrated in Figure~\ref{} giving rise to three types of \emph{twisted squares} in the respective tetrahedra. In the first case, the twisted square forms a pinched projective plane in $M$. As in Lemma~15 of \cite{Luo-combinatorial-2017} this can be surgered to give an embedded projective plane which implies that $M \cong \mathbb{R} P^3$, contradicting our hypothesis on the rank of $H$. In the second case the twisted square forms an embedded Klein bottle in $M$. This must be non-separating and incompressible and thus there exists a class in $H^1 (M, \mathbb{Z}_2)$ with norm zero. If this class is in $H,$ then one of the quadrilateral surfaces is a Klein bottle and $M$ has a Klein bottle as a 1--sided Heegaard splitting surface, and hence is a prism manifold by \cite{Rubinstein-one-sided-1978}. In this case we apply Proposition~\ref{} given below to obtain the first conclusion. If this class is not in $H,$ then we have the second conclusion. In the third case, the twisted square forms an embedded torus $T^2$ in $M.$ Denote $\Gamma = \text{im}(\pi_1(T^2) \to \pi_1(M)).$ This is either equal to $\mathbb{Z}\oplus\mathbb{Z},$ $\mathbb{Z},$ or the trivial group. However, two of the edges are contained in the torus and hence $\Gamma$ has a surjection onto $\mathbb{Z}_2\oplus\mathbb{Z}_2$ when mapping $\pi_1(M)$ onto $H_1(M;\mathbb{Z}_2).$ This implies that the torus is incompressible. Moreover, application of the standard half-lives half-dies argument to a component of $M\setminus T^2$ shows that $T^2$ must be non-separating. Hence there is a loop in $M$ transverse to $T^2$ and meeting it in a single point showing that the rank of $H_1(M;\mathbb{Z}_2)$, and hence the rank of $H^1(M;\mathbb{Z}_2)$, is at least 3. \end{proof} \section{Examples} \label{} In this section, we give the examples described in the introduction, and discuss methods to obtain more families. \subsection{Lens spaces} \label{ spaces} We use the equations given in \cite{Jaco-minimal-2009}. Suppose that $c(L) = 2 + 2 ||\;\varphi\;||.$ For the canonical surface, we have $\chi(S_\varphi) = 1-\mathfrak{e}$ and the number of tetrahedra is $\mathfrak{e} + \mathfrak{o} - 1.$ It also follows by inspection that $||\;\varphi\;|| = - \chi(S_\varphi) = \mathfrak{e} -1.$ Whence $c(L) = 2 \mathfrak{e}.$ In our fundamental Equation~(\ref{inequ for min tri}), this gives $\sum(d-4)\mathfrak{e}_d = 0$ and since $\mathfrak{e}_3\le 2,$ we have four possibilities, which are summarised as follows: \begin{enumerate} \item $\mathfrak{e}_3 =0.$ This implies that only $\mathfrak{e}_4 \neq 0.$ This cannot happen. \item $\mathfrak{e}_3 = 1.$ This forces $\mathfrak{e}_5=1$ and all other $\varphi$--even edges are of degree 4. This gives the first infinite family of lens spaces. \item $\mathfrak{e}_3 = 2.$ We have ($\mathfrak{e}_5=2$ or $\mathfrak{e}_6=1$) and all other $\varphi$--even edges are of degree 4. The former case is again not possible. The latter case results in the second infinite family of lens spaces. \end{enumerate} We indicate the families of lens spaces satisfying these conditions by adding these new families to the L-Graph in Figure 7 of \cite{Jaco-minimal-2009} and given here in Figure~\ref{}. The fraction $p/q, q > p > 0$, at a vertex of the L-graph corresponds to the one-vertex triangulation on the boundary of the solid torus for the triple $(p,q,p+q)$. The unique minimal path from $p/q$ to $1/1$ describes the unique minimal layered triangulation of the solid torus extending the $(p,q,p+q)$ triangulation on the boundary. At the vertex $p/q$ there are two possible lens spaces determined with minimal layered triangulations (see Figure~\ref{} on how to extend the L-graph at $p/q$). We obtain the triangulation of a lens space by what we call folding along either the edge labeled $p$ or the edge labeled $q$ in the boundary of the layered solid torus. In our situation, we require either $p$ or $q$ to be even and fold along the even edge getting the lens space $L(2q+p,q), p$ even, and $L(2p+q,p)$, $q$ even. \begin{figure}[htp] \centering{\includegraphics[width=.4\textwidth]{lgraph}} \caption{Extending the L-graph at vertex $p/q$. \label{}} \end{figure} Let $\overline{\mathfrak{e}}$ and $\overline{\mathfrak{o}}$ denote the number of even and the number of odd labeled edges in the layered triangulation of the layered solid torus corresponding to the triple $(p,q,p+q),$ respectively. When viewing the L-graph we define the \emph{deficiency} at the vertex $p/q$ as $d = \overline{\mathfrak{e}} - \overline{\mathfrak{o}}$. Note that when we have either $p$ or $q$ even and we fold along the even edge at $p/q$, the resulting lens space has the same number of even edges as the layered solid torus but one less odd edge. Hence, we satisfy the conditions of Theorem 5 of \cite{Jaco-minimal-2009}, for the lens space formed by folding along an even edge at $p/q$ with $d=1$. Similarly, we satisfy the conditions for the minimal layered triangulation of a lens space to be its minimal triangulation in the two cases described above when we fold along an even edge at $p/q$ with $d=2$. The new entries from the results in this paper lie along the infinite subgraph in Figure~\ref{} designated by solid edges and obtained by folding along the even edge at a vertex of valence 2 in this solid subgraph. \begin{sidewaysfigure}[p] \centering{\includegraphics[width=\textwidth]{tree}} \caption{L-graph of minimal layered solid torus triangulations. \label{}} \end{sidewaysfigure} \subsection{Prism manifolds} \label{} It is observed in \cite{Jaco-Z2-2013} that if a triangulation of $M$ has three taut quadrilateral surfaces representing the non-trivial elements of a subgroup $H$ of $H^1(M;\mathbb{Z}_2)$ of rank 2, then the number of tetrahedra in that triangulation equals $2 + \sum_{0 \neq \varphi \in H} ||\;\varphi\;||,$ and hence is minimal. This characterisation is used in the next result. \begin{proposition}\label{} Suppose $M$ is a prism manifold with with $H^1(M;\mathbb{Z}_2)$ of rank two. Then a triangulation of $M$ has three taut quadrilateral surfaces if and only if $M$ is the generalised quaternionic space $$M_k = S^3/Q_{4k} = S^2(\ (1,-1), (2,1), (2,1), (k,1) \ ),$$ for some $k\in 2\mathbb{N}.$ \end{proposition} \begin{proof} Since $M$ is a prism manifold, one of the quadrilateral surfaces is a Klein bottle, and the triangulation is dual to the 1--sided Heegaard splitting defined by this Klein bottle. Writing the fundamental group of the Klein bottle as $b^{-1} a b = a^{-1},$ a presentation for the fundamental group of the prism manifold is $$\langle a, b \ |\ b^{-1}ab = a^{-1}, (b^2)^q a^{2p} = 1 \rangle.$$ Here, the boundary of the singular compression disc for the Klein bottle is given by the curve $(b^2)^q a^{2p}$ on the Klein bottle. The number of tetrahedra in the dual triangulation is $2pq.$ We know that $$2pq = 2 + \sum_{0 \neq \varphi \in H} ||\;\varphi\;|| = 2 + 0 + 2n = 2(1+n),$$ where $n$ is the norm of the two surfaces not equal to the Klein bottle. The slopes on the singular fiberes are $(1,-1), (2,1), (2,1), (2p,q)$. Let $S$ be a pair of pants, i.e. a compact surface with three boundary components and Euler characteristic $-1.$ Then $M$ is obtained from $S \times S^1$ by gluing in three solid tori $D^2 \times S^1.$ The gluing maps are uniquely determined by giving the slopes of the curves on $D^2 \times S^1$ identified with curves $\star \times S^1$ on the boundary of $S \times S^1.$ They are $(2,1)$, $(2,1)$ and $(2p,q-2p)$. A standard calculation with Seifert invariants shows that the resulting manifold is homeomorphic by a fibre preserving map to $M.$ Then the horizontal incompressible surface described by Frohman~\cite{Frohman-one-sided-1986} is obtained by taking $S \times \star$ and extending it over each solid torus by adding the unique one-sided surface of slope $(2,1)$, $(2,1)$ and $(2p,q-2p)$ respectively. We seek two vertical surfaces with negative Euler characteristic equal to $pq-1$. It follows by inspection that there are such surfaces if $q=1.$ Now if $q$ increases in the slope $(2p,q-2p),$ then the norm decreases, hence we only have the desired equality when $q=1.$ This gives precisely the family of generalised quaternionic spaces. This proves the forwards direction. The converse is shown in \cite[\S6]{Jaco-Z2-2013}. \end{proof} \begin{remark} The generalised quaternionic space $M_k$, $k$ even, is also triangulated by the augmented solid torus $A(2,1 \mid 2,1 \mid k, 1-k).$ This has one more tetrahedron than $c(M_k).$ In the Pachner graph of all 1--vertex triangulations of $M_k$ connected by 2--3 and 3--2 moves any path from this triangulation to the minimal triangulation is of length at least $2k-11.$ The reason is that the augmented solid torus triangulation contains a layered solid torus consisting of $k-5$ tetrahedra. To remove the self-identification of the core tetrahedron requires at least $k-5$ Pachner moves of type 2--3 in any connecting sequence. Since this increases the number of tetrahedra, and the target triangulation has one tetrahedron fewer, one also needs at least $k-4$ moves of type 3--2 in any connecting sequence. In total, this amounts to at least $2k-9$ Pachner moves. \end{remark} \begin{figure} \centerline{\includegraphics[height=4cm]{fpg_1} \hspace{2cm} \raisebox{.3cm}{\includegraphics[height=3.2cm]{fpg_2}}} \caption{Left: Dual graph of layered loop triangulation of $M_{10}$. Right: Dual graph of augmented solid torus triangulation of $M_{10}$.} \end{figure} \textbf{Constructions.} We note that the generalised quaternionic spaces are obtained from an augmented solid torus by folding two pairs of its faces to obtain the Seifert invariants $(2,1),$ $(2,1)$ and then adding a layered solid torus that belongs to the subfamily $\mathbf{T}(1, 1-k, k)$ of deficiency zero from \S\ref{ spaces}. Adding a layered solid torus that belongs to the subfamily $\mathbf{T}(3, 1+6k, 4+6k)$ gives the prism manifolds $$P_k = S^2( (1,-1), (2,1), (2,1), (4+6k,1+6k))$$ and these satisfy $c(P_k) = 3 + \sum_{0 \neq \varphi \in H} ||\;\varphi\;||,$ and hence are minimal. More families can be found this way, but we wish to emphasise that not all layered solid tori of deficiency zero result in such prism manifolds with this property. We explain in Remark~\ref{} how to build such examples with \emph{Regina}~\cite{regina}. \subsection{Small Seifert fibred spaces} \label{} In this section we determine minimal triangulations for two infinite classes of small Seifert fibred spaces admitting $\widetilde{\SL_2(\mathbb{R})}$ structures. As in \cite{Jaco-Z2-2013} we rely on work of Waldhausen~\cite{Waldhausen-klasse-1967}, Rubinstein~\cite{Rubinstein-one-sided-1978} and Frohman~\cite{Frohman-one-sided-1986} to determine the norms of the $\mathbb{Z}_2$--homology classes. The discussion in this paper is less detailed than in \cite{Jaco-Z2-2013}. It is a pleasant exercise, using the structure of the triangulations, to give a combinatorial proof of our claims using normal surface theory. \textbf{Small Seifert fibred spaces with a horizontal taut surface.} Let $$M = M_{k,m,n} = S^2( (1,1), (2k+1,1), (2m+1,1), (2n+1,1)),$$ where $k, m$ and $n$ are positive integers. Now $H^1(M; \mathbb{Z}_2) \cong \mathbb{Z}_2$ and we show that the generator $\varphi$ is represented by an incompressible non-orientable horizontal surface of Euler characteristic $-(k+m+n).$ Hence $c(M) \ge 2(k+m+n+1).$ The horizontal surface is obtained as follows. On a 2--sphere choose three open discs with pairwise disjoint closures, and denote $S$ the complement of the interiors of the discs. Then $M$ is obtained from $S \times S^1$ by gluing in three solid tori $D^2 \times S^1.$ The gluing maps are uniquely determined by giving the slopes of the curves on $D^2 \times S^1$ identified with curves $\star \times S^1$ on the boundary of $S \times S^1.$ They are $(2k+1, 2k+2),$ $(2n+1, 2n+2)$ and $(2m+1, -2m).$ A standard calculation with Seifert invariants shows that the resulting manifold is homeomorphic by a fibre preserving map to $M.$ Then the horizontal incompressible surface described by Frohman~\cite{Frohman-one-sided-1986} is obtained by taking $S \times \star$ and extending it over each solid torus by adding the unique one-sided surface of slope $(2k+2, 2k+1),$ $(2n+2, 2n+1)$ and $(-2m, 2m+1)$ respectively. These surfaces have respective Euler characteristics $-k,$ $-n$ and $-m+1.$ Adding the Euler characteristic of $S$ gives a total of $-(k+m+n).$ It follows from Frohman~\cite{Frohman-one-sided-1986} that this is the unique incompressible surface representing the nontrivial class in $H^1(M; \mathbb{Z}_2).$ To obtain the equality $c(M) = 2 + 2 ||\;\varphi\;||$ it suffices to exhibit a triangulation with $2(k+m+n+1)$ tetrahedra. The augmented solid torus $A(2k+1,-2k-2 \mid 2m+1,-2m-2 \mid 2n+1, -1)$ is built by first triangulating a prism with three tetrahedra and then gluing the top triangle to the bottom triangle. We denote this $3$-tetrahedron $3$-vertex solid torus by $\tilde{P}$. Every vertex in $\tilde{P}$ is of degree four. See Figure~\ref{} for a detailed picture of $\tilde{P}$ and its boundary. We then pinch the vertical edges to obtain a complex with three boundary tori with two triangles each, and only one vertex, denoted by $P$. \begin{figure}[htb] \centering{\includegraphics[width=\textwidth]{prism}} \caption{Left: The solid torus $\tilde{P}$. Right: Boundary complex of $\tilde{P}$, top and bottom vertices become identified in pairs. \label{}} \end{figure} The construction of the augmented solid torus is finished by gluing layered solid tori of type $(1,2k+1,2k+2)$, $(1,2m+1,2m+2)$ and $(1,2n,2n+1)$ along the boundary components of $P$. For each of the boundary components, the even-order edge is glued to the horizontal edge, the edge of order one is glued to a vertical edge, and the remaining edge is glued to the remaining diagonal edge. See Figure~\ref{} for details and see \cite{regina} for a more elaborate discussion of augmented layered solid tori. From the description of $M$ above, this shows that the resulting identification space is homeomorphic with $M$. Since the triangulation has $2(k+m+n+1)$ tetrahedra, we obtain $c(M) \le 2(k+m+n+1) = 2 + 2 ||\;\varphi\;||,$ which forces equality. We conclude this example by describing a normal surface approach to determining the norm. Any normal representative of the unique non-zero homology class $\varphi \in H_2 (M; \mathbb{Z}_2)$ intersects the horizontal edges of $P$ (i.e., the three edges belonging to the top and bottom triangle of the prism) an even number of times, and all other edges of $P$ an odd number of times. It can be shown that the only least weight least genus normal surface representative of $\varphi$ is the canonical normal representative $S_\varphi.$ The intersection of $S_\varphi$ with the boundary faces of $P$ os shown as a dashed line in the induced boundary of $\partial P$ in Figure~\ref{}. \textbf{Small Seifert fibred spaces with three vertical taut surfaces.} Let $$M= M'_{k,m,n} = S^2( (1,-1), (2k+2,1), (2m+2,1), (2n+2,1)),$$ where $k, m$ and $n$ are positive integers. As above, one can show that this is triangulated by the \emph{augmented solid torus} $A(2k+2,-1 \mid 2m+2,-1 \mid 2n+2, -1)$, which has $2k+2m+2n+3$ tetrahedra. Hence $c(M) \le 2k+2m+2n+3.$ One has $H^1(M, \mathbb{Z}_2) \cong \mathbb{Z}_2\oplus \mathbb{Z}_2.$ The manifold is obtained similar to above by gluing solid tori to $S\times S^1,$ this time using slopes $(2k+2, 1),$ $(2m+2, 1)$ and $(2n+2, -2n-1).$ A vertical one-sided surface is obtained by taking an arc $\alpha$ in $S$ connecting distinct boundary component of $S$ and attaching to the annulus $\alpha \times S^1$ one-sided incompressible surfaces in the tori met by the boundary of the annulus. This gives surfaces of respective Euler characteristics $-k-m,$ $-m-n$ and $-n-k$ representing the non-trivial $\mathbb{Z}_2$--classes. As in \cite{Jaco-Z2-2013} one argues that these are norm minimising, hence giving $c(M) = 3 + \sum_{0\neq \varphi} ||\;\varphi\;||.$ \begin{remark} \label{} To construct a minimal triangulation of $M_{k,m,n}$ (resp. $M'_{k,m,n}$) with \emph{Regina} \cite{regina}, one creates a new 3--manifold by selecting type of triangulation ``Seifert fibred space over $2$-sphere'' and passing the (differently normalised) parameters $(2k+1,1) (2m+1,1) (2n+1,2n+2)$ (resp. $(2k+2,1) (2m+2,1) (2n+2,-2n-1)$). Regina then generates an augmented solid torus of type $A(2k+1,-1 \mid 2m+1,-1 \mid 2n+1, -4n-3)$ (resp. $A(2k+2,-1 \mid 2m+2,-1 \mid 2n+2, -1)$. These minimal triangulations are in general not unique, e.g. another minimal triangulation of $M_{k,m,n}$ is obtained as $A(2k+1,-2k-2 \mid 2m+1,-2m-2 \mid 2n+1, -1).$ \end{remark} \bibliographystyle{plain}

\section{Introduction} \IEEEPARstart{M}{assive} multiple-input multiple-output (MIMO) has been considered as one of the key enabling technologies for 5G wireless communications systems, and thus has recently drawn lots of research interest. It could bring significant spectral efficiency (SE) and energy efficiency (EE) gains when the base station (BS) has a large number of antennas \cite{Andrews}, \cite{Ngo}. Due to the large number of antennas at BS, a simple linear receiver can be employed to achieve a high SE, and the impact of imperfect hardware can also be mitigated. Most of the current works concerning massive MIMO systems assume each antenna at BS is equipped with a high-resolution ADC. However, high-resolution ADCs can incur substantial power assumption and high-cost hardware implementation \cite{Le}, which may be not affordable in practical massive MIMO systems. This motives us to study the use of low-resolution ADCs to understand their impact on the performance of the system. Several recent works have investigated the effects of low-resolution ADCs on the performance of massive MIMO systems \cite{Bai}-\cite{Zhang2016}. In these works, the effect of low-resolution ADCs is described by using the additive quantization noise model (AQNM) \cite{Bai}. Using AQNM, Fan {\it et al.} \cite{Fan} derive an approximate analytical expression for the uplink achievable rate of massive MIMO systems over Rayleigh fading channels when low-resolution ADCs and the maximal-ratio combining (MRC) technique are used at the receiver; Zhang {\it et al.} \cite{Zhang2016} derive tractable and exact approximation expressions for the uplink SE of massive MIMO systems with low-resolution ADCs and MRC over Rician fading channels, where both perfect and imperfect channel state information (CSI) are considered. However, only uniform resolution ADCs and MRC receiver are considered in \cite{Fan} and \cite{Zhang2016}, which cannot be applied to analyze the uplink SE performance of massive MIMO systems with non-uniform resolution ADCs and MMSE receiver. In \cite{Liang}, a mixed-ADC architecture for massive MIMO systems is proposed, in which the BS antennas are equipped with both high-resolution ADCs and low-resolution ADCs. Under mixed-ADC massive MIMO with MRC receiver, the closed-form approximate expressions for the uplink achievable rate are derived for both perfect and imperfect CSI when the Rician fading channels are considered \cite{Zhang2017}. In \cite{Dong}, the asymptotic equivalent of the uplink SE is obtained when non-uniform resolution ADCs and MMSE receivers are used at the BS of massive MIMO systems. The MMSE receiver designed in \cite{Dong} could cause significant SE loss for massive MIMO systems with non-uniform resolution ADCs. This is because it does not consider non-uniform quantization noise incurred by non-uniform resolution ADCs. In this letter, with AQNM, we first design a proper MMSE receiver for the uplink massive MIMO system with low-resolution ADCs. The proposed MMSE receiver takes both additive white Gaussian noise (AWGN) and quantization noise into account. It can work well for both cases of uniform resolution ADCs and non-uniform resolution ADCs. We then derive a tight asymptotic equivalent of the uplink SE using random matrix theory. Finally, numerical results are obtained to verify our theoretical analysis. Notations: Boldface lower and upper-case symbols represent vectors and matrices, respectively. The ${N \times N}$ identity matrix reads as ${{\bf{I}}_N}$. let ${{\bf{X}}^H}$ and ${\mathop{\rm tr}\nolimits} \left( {\bf{X}} \right)$ be the conjugate transpose and trace of a matrix ${\bf{X}}$. ${\mathop{\rm diag}\nolimits} \left( {\bf{X}} \right)$ keeps only the diagonal entries of ${\bf{X}}$. A random vector ${\bf{x}} \sim {\cal C}{\cal N}\left( {{\bf{m}},{\bf{V}}} \right)$ is complex Gaussian distributed with mean vector ${\bf{m}}$ and covariance matrix ${\bf{V}}$. $\overset{a.s.}{\longrightarrow}$ denotes almost sure convergence. \section{System Model} Consider an uplink massive MIMO system with low-resolution ADCs, where the BS equipped with $M$ antennas receives signal from $K$ single-antenna users using the same time-frequency resources. The received signal ${\bf{y}} \in {{\mathbb{C}}^{M \times 1}} $ at BS can be written as \begin{align} \label{EQ1} {\bf{y}} = \sqrt {{p_u}} {\bf{Gx}} + {\bf{n}}, \end{align} where ${p_u}$ is the average transmit power of each user, ${\bf{x}} \in {{\mathbb{C}}^{K \times 1}}$ is the transmit vector for all $K$ users, and ${\bf{n}} \sim {\cal C}{\cal N} \left( {0,{{\bf{I}}_M}} \right)$ is the normalized AWGN vector. The input covariance matrix ${{\bf{R}}_{\bf{x}}} = {{\bf{I}}_K}$ is assumed. The channel matrix ${\bf{G}} \in {{\mathbb{C}}^{M \times K}}$ between the BS and all users is given as \begin{align} \label{EQ2} {\bf{G}} = {\bf{H}}{{\bf{D}}^{{1 \mathord{\left/ {\vphantom {1 2}} \right. \kern-\nulldelimiterspace} 2}}}, \end{align} where $\bf{H}$ denotes the small-scale fading channel matrix and ${\bf{D}}$ is the $K \times K$ diagonal matrix with diagonal entries $\left\{ {{\beta _k}} \right\}$ to model the large-scale fading. Rayleigh fading is assumed here, thus without loss of generality, the elements of ${\bf{H}}$ can be considered to follow distribution ${\cal C}{\cal N} \left( {0,1} \right)$. As the quantization error can be well approximated as a linear gain with AQNM \cite{Bai}, the output vector of low-resolution ADCs can be expressed as \begin{align} \label{EQ3} {{\bf{y}}_q} = {\bf{Q}}\left( {\bf{y}} \right) \approx \Lambda {\bf{y}} + {{\bf{n}}_q} = \sqrt {{p_u}} \Lambda {\bf{Gx}} + \Lambda {\bf{n}} + {{\bf{n}}_q}, \end{align} where ${\bf{Q}}\left( \cdot \right)$ is the quantizer function which applies component-wise and separately to the real and imaginary parts, and ${{\bf{n}}_q}$ is the additive Gaussian quantization noise vector which is uncorrelated with ${\bf{y}}$. The linear gain matrix $\Lambda$ is an $M \times M$ diagonal matrix with entries $\left\{ {{\alpha _m}} \right\}$, and the linear gain coefficient ${\alpha _m} = 1 - {\rho _m}$, where ${\rho _m}$ is the inverse of the signal-to-quantization-noise ratio. Given a fixed channel realization ${\bf{G}}$, the covariance matrix of ${{\bf{n}}_q}$ can be written as \cite{Fan} \begin{align} \label{EQ4} \begin{array}{l} {{\bf{R}}_{{{\bf{n}}_q}}} = {\rm E}\left( {{{\bf{n}}_q}{\bf{n}}_q^H\left| {\bf{G}} \right.} \right)\\ {\rm{~~~~~= }}\Lambda \left( {{{\bf{I}}_M} - \Lambda } \right){\bf{diag}}\left( {{p_u}{\bf{G}}{{\bf{G}}^H} + {{\bf{I}}_M}} \right). \end{array} \end{align} \section{Asymptotic Equivalent of the Uplink Spectral Efficiency} \label{Section 3} In this section, using random matrix theory, we derive the asymptotic equivalent of the uplink SE for massive MIMO systems with low-resolution ADCs, when the MMSE receiver is considered. Here the perfect CSI is known at the BS. First, the proposed high-performance MMSE receiver is designed as \begin{align} \label{EQ5} {\bf{W}} = {{\bf{G}}^H}{\Lambda ^H}{\left( {\Lambda {\bf{G}}{{\bf{G}}^H}{\Lambda ^H} + {\bf{Z}} + \theta {{\bf{I}}_M}} \right)^{ - 1}}, \end{align} where ${\bf{Z}} = \frac{1}{{{p_u}}}{{\bf{\bar R}}_{{{\bf{n}}_q}}} = \frac{1}{{{p_u}}}\Lambda \left( {{{\bf{I}}_M} - \Lambda } \right)\left( {{p_u}{\bf{tr}}\left( {\bf{D}} \right) + 1} \right)$, and $\theta = {1 \mathord{\left/ {\vphantom {1 {{p_u}}}} \right. \kern-\nulldelimiterspace} {{p_u}}}$. ${{\bf{\bar R}}_{{{\bf{n}}_q}}}$ is the expectation of ${{\bf{R}}_{{{\bf{n}}_q}}}$ with respect to ${\bf{G}}$. Thus, we can see that the proposed MMSE receiver can suppress both AWGN and quantization noise at the same time. This proposed MMSE receiver will be used in the following analysis. By multiplying ${\bf{W}}$ with ${{\bf{y}}_q}$, the quantized output vector is processed as \begin{align} \label{EQ6} {\bf{r}} = {\bf{W}}{{\bf{y}}_q} = \sqrt {{p_u}} {\bf{W}}\Lambda {\bf{Gx}} + {\bf{W}}\Lambda {\bf{n}} + {\bf{W}}{{\bf{n}}_q}. \end{align} The output signal for the $k{\rm{th}}$ user can be expressed as \begin{align} \label{EQ7} \begin{array}{l} {{\bf{r}}_k} = \sqrt {{p_u}} {\bf{g}}_k^H{\Lambda ^H}{\bf{V}}\Lambda {{\bf{g}}_k}{x_k} + \sqrt {{p_u}} \sum\limits_{i \ne k} {{\bf{g}}_k^H{\Lambda ^H}{\bf{V}}\Lambda {{\bf{g}}_i}{x_i}} \\ {\rm{~~~~~~+ }}{\bf{g}}_k^H{\Lambda ^H}{\bf{V}}\Lambda {\bf{n}} + {\bf{g}}_k^H{\Lambda ^H}{\bf{V}}{{\bf{n}}_q}, \end{array} \end{align} where ${\bf{V}} = {\left( {\Lambda {\bf{G}}{{\bf{G}}^H}{\Lambda ^H} + {\bf{Z}} + \theta {{\bf{I}}_M}} \right)^{ - 1}}$, ${{\bf{g}}_i}$ is the $i{\bf{th}}$ column of ${\bf{G}}$ and ${x_i}$ is the $i{\bf{th}}$ element of ${\bf{x}}$. The achievable uplink rate of the $k{\rm{th}}$ user is thus expressed as \begin{align} \label{EQ8} {R_k} = {\log _2}\left( {1 + \frac{{{p_u}{{\left| {{\bf{g}}_k^H{\Lambda ^H}{\bf{V}}\Lambda {{\bf{g}}_k}} \right|}^2}}}{{{I_k}}}} \right), \end{align} where the variance of the interference-plus-noise term is \begin{align} \label{EQ9} \begin{array}{l} {I_k} = {p_u}\sum\limits_{i \ne k} {{{\left| {{\bf{g}}_k^H{\Lambda ^H}{\bf{V}}\Lambda {{\bf{g}}_i}} \right|}^2}} + {\bf{g}}_k^H{\Lambda ^H}{\bf{V}}\Lambda {\Lambda ^H}{\bf{V}}\Lambda {{\bf{g}}_k}\\ {\rm{~~~~~+ }}{\bf{g}}_k^H{\Lambda ^H}{\bf{V}}{{\bf{R}}_{{{\bf{n}}_q}}}{\bf{V}}\Lambda {{\bf{g}}_k} \end{array}. \end{align} Now the sum uplink SE of the whole system is obtained as \begin{align} \label{EQ10} {R_{sum}} = \sum\limits_k {{R_k}}. \end{align} To obtain the asymptotic equivalent of ${R_k}$, we first derive the asymptotic equivalent for the power of the signal of interest and the three parts of ${I_k}$. Then, the asymptotic equivalent of ${R_k}$ is given as follows: \emph{Lemma 1:} When $M,K \to \infty $ with fixed ratio, the asymptotic equivalent of the uplink SE in Eq. (\ref{EQ8}) is written as \begin{align} \label{EQ11} {\bar R_k} = \frac{{{p_u}{\beta _k}{{\left| {{\Gamma _1}} \right|}^2}}}{{{p_u}\sum\limits_{i \ne k} {{{\left| {\frac{1}{{1 + {{{\beta _i}{\Gamma _1}} \mathord{\left/ {\vphantom {{{\beta _i}{\Gamma _1}} M}} \right. \kern-\nulldelimiterspace} M}}}} \right|}^2}{\beta _i}{\Gamma _2} + {\Gamma _2} + {\Gamma _3}} }}, \end{align} where ${\Gamma _1}$, ${\Gamma _2}$ and ${\Gamma _3}$ are calculated as Eqs. (14), (18) and (23), respectively. \emph{Proof:} For the signal of interest, using Sherman-Morrison formula, we can obtain \begin{align} \label{EQ12} {\bf{g}}_k^H{\Lambda ^H}{\bf{V}}\Lambda {{\bf{g}}_k} = \frac{{{\bf{g}}_k^H{\Lambda ^H}{{\left( {\Lambda {{\bf{G}}_{ - k}}{\bf{G}}_{ - k}^H{\Lambda ^H} + {\bf{Z}} + \theta {{\bf{I}}_M}} \right)}^{ - 1}}\Lambda {{\bf{g}}_k}}}{{1 + {\bf{g}}_k^H{\Lambda ^H}{{\left( {\Lambda {{\bf{G}}_{ - k}}{\bf{G}}_{ - k}^H{\Lambda ^H} + {\bf{Z}} + \theta {{\bf{I}}_M}} \right)}^{ - 1}}\Lambda {{\bf{g}}_k}}}, \end{align} where ${{\bf{G}}_{ - k}}$ denotes that the $k{\bf{th}}$ column of ${\bf{G}}$ is removed from ${\bf{G}}$. Through \emph{Theorem 1} in {\bf{Appendix A}}, we have \begin{equation} {\bf{g}}_k^H{\Lambda ^H}{\left( {\Lambda {{\bf{G}}_{ - k}}{\bf{G}}_{ - k}^H{\Lambda ^H} + {\bf{Z}} + \theta {{\bf{I}}_M}} \right)^{ - 1}}\Lambda {{\bf{g}}_k} \nonumber \\ \end{equation} \begin{equation} \overset{(a)}{\longrightarrow} {\bf{tr}}\left( {{\beta _k}\Lambda {\Lambda ^H}{{\bf{V}}_{\left( k \right)}}} \right) \nonumber \\ {\longrightarrow} {\beta _k}{\bf{tr}}\left( {\Lambda {\Lambda ^H}{\bf{V}}} \right) \end{equation} \begin{equation} \label{EQ13} \overset{a.s.}{\longrightarrow} {\beta _k}{\bf{tr}}\left( {\Lambda {\Lambda ^H}{\bf{S}}\left( {\frac{\theta }{M}} \right)} \right) = {\beta _k}{\Gamma _1}, \end{equation} where (a) follows from \emph{Lemma 4} in \cite{Wagner}, ${{\bf{V}}_{\left( k \right)}} = {\left( {\Lambda {{\bf{G}}_{ - k}}{\bf{G}}_{ - k}^H{\Lambda ^H} + {\bf{Z}} + \theta {{\bf{I}}_M}} \right)^{ - 1}}$ and \begin{equation} \label{EQ14} {\Gamma _1} = {\bf{tr}}\left( {\Lambda {\Lambda ^H}{\bf{S}}\left( {\frac{\theta }{M}} \right)} \right). \end{equation} With Eq. (\ref{EQ13}), the asymptotic equivalent for the power of the signal of interest can be given as \begin{equation} \label{EQ15} {\left| {{\bf{g}}_k^H{\Lambda ^H}{{\bf{V}}^{ - 1}}\Lambda {{\bf{g}}_k}} \right|^2} \overset{a.s.}{\longrightarrow} \left| {\frac{{{\beta _k}{\Gamma _1}}}{{1 + {\beta _k}{\Gamma _1}}}} \right|. \end{equation} For the inter-user interference part, by using Sherman-Morrison formula twice, we have \begin{equation} {\left| {{\bf{g}}_k^H{\Lambda ^H}{{\bf{V}}^{ - 1}}\Lambda {{\bf{g}}_i}} \right|^2} = {\left| {\frac{1}{{1 + {\bf{g}}_k^H{\Lambda ^H}{{\bf{V}}_{\left( k \right)}}\Lambda {{\bf{g}}_k}}}} \right|^2} \cdot {\left| {\frac{1}{{1 + {\bf{g}}_i^H{\Lambda ^H}{{\bf{V}}_{\left( {ki} \right)}}\Lambda {{\bf{g}}_i}}}} \right|^2} \nonumber \\ \end{equation} \begin{equation} \label{EQ16} {\rm{~~~~~~~~~}} \cdot {\bf{g}}_k^H{\Lambda ^H}{{\bf{V}}_{\left( {ki} \right)}}\Lambda {{\bf{g}}_i}{\bf{g}}_i^H{\Lambda ^H}{{\bf{V}}_{\left( {ki} \right)}}\Lambda {{\bf{g}}_k}. \end{equation} Through \emph{Theorem 2} in {\bf{Appendix A}}, we have \begin{equation} {\bf{g}}_k^H{\Lambda ^H}{{\bf{V}}_{\left( {ki} \right)}}\Lambda {{\bf{g}}_i}{\bf{g}}_i^H{\Lambda ^H}{{\bf{V}}_{\left( {ki} \right)}}\Lambda {{\bf{g}}_k} \nonumber \\ \end{equation} \begin{equation} \overset{(a)}{\longrightarrow} {\bf{tr}}\left( {{\beta _k}{\beta _i}{\Lambda ^H}{{\bf{V}}_{\left( {ki} \right)}}\Lambda {\Lambda ^H}{{\bf{V}}_{\left( {ki} \right)}}\Lambda } \right) {\longrightarrow} {\beta _k}{\beta _i}{\bf{tr}}\left( {\Lambda {\Lambda ^H}{\bf{V}}\Lambda {\Lambda ^H}{\bf{V}}} \right) \nonumber \\ \end{equation} \begin{equation} \label{EQ17} \overset{a.s.}{\longrightarrow} {\beta _k}{\beta _i}{\bf{tr}}\left( {\Lambda {\Lambda ^H}{\bf{S'}}\left( {\frac{\theta }{M}} \right)} \right) = {\beta _k}{\beta _i}{\Gamma _2}, \end{equation} where \begin{equation} \label{EQ18} {\Gamma _2} = {\bf{tr}} \left( {\Lambda {\Lambda ^H}{\bf{S'}}\left( {\frac{\theta }{M}} \right)} \right). \end{equation} Thus as $M,K \to \infty $, the asymptotic equivalent of the inter-user interference part is expressed as \begin{equation} \label{EQ19} {\left| {{\bf{g}}_k^H{\Lambda ^H}{{\bf{V}}^{ - 1}}\Lambda {{\bf{g}}_i}} \right|^2}\ \overset{a.s.}{\longrightarrow} {\left| {\frac{1}{{1 + {\beta _k}{\Gamma _1}}}} \right|^2}{\left| {\frac{1}{{1 + {\beta _i}{\Gamma _1}}}} \right|^2}{\beta _k}{\beta _i}{\Gamma _2}. \end{equation} For the Gaussian noise part, again by using Sherman-Morrison formula, we have \begin{align} \label{EQ20} {\bf{g}}_k^H{\Lambda ^H}{\bf{V}}\Lambda {\Lambda ^H}{\bf{V}}\Lambda {{\bf{g}}_i} = {~~~~~~~~~~~~~~~~~~~~~~~~~~~~~} \nonumber \\ {\rm{ }}{\left| {\frac{1}{{1 + {\bf{g}}_k^H{\Lambda ^H}{{\bf{V}}_{\left( k \right)}}\Lambda {{\bf{g}}_k}}}} \right|^2}{\bf{g}}_k^H{\Lambda ^H}{{\bf{V}}_{\left( k \right)}}\Lambda {\Lambda ^H}{{\bf{V}}_{\left( k \right)}}\Lambda {{\bf{g}}_k}. \end{align} Then, using Eqs. (\ref{EQ13}) and (\ref{EQ17}) , the asymptotic equivalent of the Gaussian noise part is expressed as \begin{equation} \label{EQ21} {\bf{g}}_k^H{\Lambda ^H}{\bf{V}}\Lambda {\Lambda ^H}{\bf{V}}\Lambda {{\bf{g}}_i} \overset{a.s.}{\longrightarrow} {\left| {\frac{1}{{1 + {\beta _k}{\Gamma _1}}}} \right|^2}{\beta _k}{\Gamma _2}. \end{equation} The quantization noise part contains ${{\bf{R}}_{{{\bf{n}}_q}}}$ that is dependent on the channel matrix ${\bf{G}}$. We use ${{\bf{\bar R}}_{{{\bf{n}}_q}}}$ to replace ${{\bf{R}}_{{{\bf{n}}_q}}}$ in ${\bf{g}}_k^H{\Lambda ^H}{\bf{V}}{{\bf{R}}_{{{\bf{n}}_q}}}{\bf{V}}\Lambda {{\bf{g}}_k}$ and the effect of this replacement on the asymptotic equivalent is negligible. Similar to the derivation of the Gaussian noise part, the asymptotic equivalent of the quantization noise part is obtained as \begin{align} \label{EQ22} {\bf{g}}_k^H{\Lambda ^H}{\bf{V}}{{\bf{R}}_{{{\bf{n}}_q}}}{\bf{V}}\Lambda {{\bf{g}}_k} \approx {~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~} \nonumber \\ {\left| {\frac{1}{{1 + {\bf{g}}_k^H{\Lambda ^H}{{\bf{V}}_{\left( k \right)}}\Lambda {{\bf{g}}_k}}}} \right|^2}{\bf{g}}_k^H{\Lambda ^H}{{\bf{V}}_{\left( k \right)}}{{{\bf{\bar R}}}_{{{\bf{n}}_q}}}{{\bf{V}}_{\left( k \right)}}\Lambda {{\bf{g}}_k} \nonumber \\ \overset{a.s.}{\longrightarrow} {\left| {\frac{1}{{1 + {\beta _k}{\Gamma _1}}}} \right|^2}{\beta _k}{\bf{tr}}\left( {{{{\bf{\bar R}}}_{{{\bf{n}}_q}}}{\bf{S'}}\left( {\frac{\theta }{M}} \right)} \right) = {\left| {\frac{1}{{1 + {\beta _k}{\Gamma _1}}}} \right|^2}{\beta _k}{\Gamma _3}, \end{align} where \begin{equation} \label{EQ23} {\Gamma _3} = {\bf{tr}}\left( {{{{\bf{\bar R}}}_{{{\bf{n}}_q}}}{\bf{S'}}\left( {\frac{\theta }{M}} \right)} \right). \end{equation} Until now, substituting Eqs. (\ref{EQ15}), (\ref{EQ19}), (\ref{EQ21}) and (\ref{EQ22}) into Eq. (\ref{EQ8}) and after some mathematics manipulations, we can get the \emph{Lemma 1}. {~~~~~~~~~~~}$\blacksquare$ Note that \emph{Lemma 1} is a general result and an exact asymptotic equivalent of the uplink SE for both uniform resolution ADCs and non-uniform resolution ADCs in massive MIMO systems. In addition, our proposed MMSE receiver shows better performances than that in [9], which is verified in simulation. Furthermore, the closed-form asymptotic equivalent of the uplink SE only depends on the statistics parameters of the channel that keeps static within a coherence time block; therefore it can be readily used to analytically evaluate the effects of low-resolution ADCs, the number of antennas $M$, and the transmit power ${p_u}$ on the considered system's performance. \section{Numerical results} This section provides numerical results to verify \emph{Lemma 1} in Sec. \ref{Section 3}. Consider a hexagonal cell with a radius of 1000 meters. User locations are generated independently and randomly in the cell by following uniform distribution, but the distance between each user and the BS is at least 100 meters. The path loss is expressed as $r_n^{ - v}$, where ${r_n}$ is the distance from the $n{\bf{th}}$ user to the BS and $v = 3.8$ is the path loss exponent. Shadowing is modeled as a log-normal random variable ${s_n}$ with a standard deviation of 8dB. Thus the large-scale fading can be expressed as ${\beta _n} = {s_n}{\left( {{{{r_n}} \mathord{\left/ {\vphantom {{{r_n}} {{r_{\min }}}}} \right. \kern-\nulldelimiterspace} {{r_{\min }}}}} \right)^{ - v}}$. The resolution of ADCs is $b = 1 \sim 3$ bits, and the values of $\rho$ are listed in Table I for $b \le 5$ bits. \begin{table} \caption{$\rho$ for different resolution bits $b$} \centering \begin{tabular}{c c c c c c} \hline $b$ &1 &2 &3 &4 &5 \\ \hline $\rho$ &0.3634 &0.1175 &0.03454 &0.009497 &0.002499 \\ \hline \end{tabular} \end{table} In Fig. \ref{fig1}, the simulated uplink SE along with the derived asymptotic equivalent in \emph{Lemma 1} are plotted against the user power ${p_u}$. The number of antennas $M$ is set as 60 and 120. The number of users $K$ is 8. The resolution bits of each ADC is selected with equal probability over $b = 1\sim3$ bits. The simulated uplink SE from [9] is included for comparison. It is observed that the derived asymptotic equivalent of the uplink SE matches very well with the simulation results for the two different numbers of antennas $M$ and for all cases of the user power ${p_u}$. As the user power ${p_u}$ increases, our proposed MMSE receiver shows a much better performance than that proposed in \cite{Dong}. This is because, when the user power ${p_u}$ is high, the quantization noise becomes dominant than AWGN; our proposed MMSE receiver can suppress the quantization noise and AWGN at the same time, while the proposed MMSE receiver in \cite{Dong} can just suppress AWGN. It is also observed that as the user power ${p_u}$ increases, the uplink SE is saturated, because the power of quantization noise scales up with the user power ${p_u}$, as shown in Eq. (\ref{EQ4}). \begin{figure}[!h] \centering \includegraphics[width=3.5in, height=2.5in]{Fig_1.eps} \caption{Simulated and analytical SEs of massive MIMO with non-uniform resolution ADCs versus the transmit user power ${p_u}$.} \label{fig1} \end{figure} \begin{figure}[!h] \centering \includegraphics[width=3.5in, height=2.5in]{Fig_2.eps} \hfill \includegraphics[width=3.5in, height=2.5in]{Fig_3.eps} \caption{Simulated and analytical SEs of massive MIMO with uniform-resolution ADCs versus the number of antennas $M$.} \label{fig2} \end{figure} Fig. \ref{fig2} shows the uplink SE versus the number of antennas with uniform-resolution ADCs as $1, 2,$ infinite bits. The number of users is 8. We can see that, in all cases of the number of antennas, the simulation results match very well the analytical results. With only $2$ bits, the SE nearly approaches that with ideal ADCs (infinite resolution). For fixed values of ${p_u}$, all curves grow without bound as $M$ increases, while for variable values of ${p_u} = {{{E_u}} \mathord{\left/ {\vphantom {{{E_u}} M}} \right. \kern-\nulldelimiterspace} M}$, the corresponding curves eventually saturate as $M$ increases. Note that ${E_u}$ is set as 30dB. \section{Conclusion} In this paper, AQNM is used to model the effect of the low-resolution ADCs. Based on the AQNM, we propose a high-performance MMSE receiver for massive MIMO systems with low-resolution ADCs, which can suppress both AWGN and quantization noise at the same time. Using the proposed MMSE receiver and random matrix theory, we derive the asymptotic equivalent of the uplink SE. Numerical results show that the asymptotic equivalent is very tight. Moreover, the loss of the uplink SE brought by the low-resolution ADCs can be compensated by increasing the number of antennas at BS. \appendices \label{Appendix} \section{} {\emph{Theorem 1 ([10])}}: Let $ {\bf{U}} \in {{\mathbb{C}}^{M \times M}}$ and $ {\bf{Z}} \in {{\mathbb{C}}^{M \times M}}$ be Hermitian nonnegative definite and $ {\bf{H}} \in {{\mathbb{C}}^{M \times K}}$ be random with independent columns ${{\bf{h}}_k} \sim {\cal{C}}{\cal{N}} \left( {0,\frac{1}{M}{{\bf{R}}_k}} \right)$. Assume that ${\bf{U}}$ and ${{\bf{R}}_k}\left( {k = 1,2, \ldots ,K} \right)$, have uniformly bounded spectral norms (with respect to $M$). Then when $M,K \to \infty $ with fixed ratio and any $\rho > 0$, \begin{equation} \frac{1}{M}{\bf{tr}}\left( {{\bf{U}}{{\left( {{\bf{H}}{{\bf{H}}^H} + {\bf{Z}} + \rho {{\bf{I}}_M}} \right)}^{ - 1}}} \right) - \frac{1}{M}{\bf{tr}}\left( {{\bf{US}}\left( \rho \right)} \right) \overset{a.s.}{\longrightarrow}{0}, \nonumber \end{equation} where ${\bf{S}}\left( \rho \right) = {\left( {\frac{1}{M}\sum\nolimits_{k = 1}^K {\frac{{{{\bf{R}}_k}}}{{1 + {\delta _k}\left( \rho \right)}}} + {\bf{Z}} + \rho {{\bf{I}}_M}} \right)^{ - 1}}$ and ${\delta _k}\left( \rho \right)$ is defined as ${\delta _k}\left( \rho \right) = {\lim _{t \to \infty }}\delta _k^t\left( \rho \right)$, $k = 1, \ldots ,K$, where \begin{equation} \delta _k^{\left( t \right)}\left( \rho \right) = \frac{1}{M}{\bf{tr}}\left( {{{\bf{R}}_k}{{\left( {\frac{1}{M}\sum\limits_{i = 1}^K {\frac{{{{\bf{R}}_i}}}{{1 + \delta _i^{\left( {t - 1} \right)}\left( \rho \right)}}} + {\bf{Z}} + \rho {{\bf{I}}_M}} \right)}^{ - 1}}} \right), \nonumber \end{equation} for $t=1,2, \ldots ,$ with initial value $\delta _k^{\left( 0 \right)} = {1 \mathord{\left/ {\vphantom {1 \rho }} \right. \kern-\nulldelimiterspace} \rho }$ for all $k$. {\emph{Theorem 2 ([10])}}: Let $ {\Theta} \in {{\mathbb{C}}^{M \times M}}$ be Hermitian nonnegative definite with uniformly bounded spectral norm (with respect to $M$). Under the same conditions as in {\emph{Theorem 1}}, \begin{equation} \frac{1}{M}{\bf{tr}}\left( {{\bf{U}}{{\bf{A}}^{ - 1}}\Theta {{\bf{A}}^{ - 1}}} \right) - \frac{1}{M}{\bf{tr}}\left( {{\bf{US'}}\left( \rho \right)} \right) \overset{a.s.}{\longrightarrow}{0}, \nonumber \end{equation} where ${\bf{A}} = {\bf{H}}{{\bf{H}}^H} + {\bf{Z}} + \rho {{\bf{I}}_M}$ and ${\bf{S'}}\left( \rho \right) \in {{\mathbb{C}}^{M \times M}}$ is \begin{equation} {\bf{S'}}\left( \rho \right) = {\bf{S}}\left( \rho \right)\Theta {\bf{S}}\left( \rho \right) + {\bf{S}}\left( \rho \right)\frac{1}{M}\sum\limits_{k = 1}^K {\frac{{{{\bf{R}}_k}{{\delta '}_k}\left( \rho \right)}}{{{{\left( {1 + {\delta _k}\left( \rho \right)} \right)}^2}}}{\bf{S}}\left( \rho \right)} . \nonumber \end{equation} ${\bf{S}}\left( \rho \right)$ and ${\delta _k}\left( \rho \right)$ are defined in {\emph{Theorem 1}}. And ${\bf{\delta '}}\left( \rho \right) = {\left[ {{{\delta '}_1}\left( \rho \right), \ldots ,{{\delta '}_K}\left( \rho \right)} \right]^T}$ is calculated as ${\bf{\delta '}}\left( \rho \right) = {\left( {{{\bf{I}}_M} - {\bf{J}}\left( \rho \right)} \right)^{ - 1}}{\bf{v}}\left( \rho \right)$, where \begin{equation} \begin{array}{l} {\left[ {{\bf{J}}\left( \rho \right)} \right]_{kl}} = \frac{{\frac{1}{M}{\bf{tr}}\left( {{{\bf{R}}_k}{\bf{S}}\left( \rho \right){{\bf{R}}_l}{\bf{S}}\left( \rho \right)} \right)}}{{M{{\left( {1 + {\delta _l}\left( \rho \right)} \right)}^2}}},{\rm{ 1}} \le k,l \le K\\ {\left[ {{\bf{v}}\left( \rho \right)} \right]_k} = \frac{1}{M}{\bf{tr}}\left( {{{\bf{R}}_k}{\bf{S}}\left( \rho \right)\Theta {\bf{S}}\left( \rho \right)} \right),1 \le k \le K. \end{array} \nonumber \end{equation} \ifCLASSOPTIONcaptionsoff \newpage \fi

\section{Introduction} Since Generative Adversarial Network (GAN) was presented by Goodfellow in 2014\cite{goodfellow2014generative}, GAN is proved to be very efficient in image generation tasks in many research fields\cite{zhu2017unpaired}\cite{zhao2016energy}\cite{mirza2014conditional}\cite{press2017language}. Specially, many researchers use GAN for face images generation. Researchers show that the extension of GAN can not only generate face images with different poses\cite{yin2017towards}\cite{tran2017representation}, but also with different accessories\cite{chen2016infogan}. However, when the target images are too complex, GAN is still insufficient to achieve objects generation precisely. In the other hand, different facial expression face images are always required to provide samples and labels for expression recognition\cite{elfenbein2002universality}. And multi-view face images can also be used for both pose and facial expression analysis\cite{lee2006generating}\cite{calvo2008facial}. However, insufficient facial images with different pose and facial expression bring us difficulty when using supervised machine learning methods. Although human face image datasets are not uncommon, a lot of datasets only include one frontal neutral face image for each subject\cite{LFWTech}. In order to solve this problem, we aim to synthesize different facial expression images through limited neutral images\cite{antipov2017face}. Besides, image with different pose can enrich the diversity of the face samples and could be applied to the pose recognition. It is not easy to consider both pose and facial expressions simultaneously during image translation. We need a desired generative model to comprehend the current image as well as capturing the underlying data distribution. This is often a very difficult task, since a collection of image samples on one attribute may lie on a very complex manifold. Images with multiple attributes would be further more sophisticated to be modeled. As described in Figure~\ref{}, the image on the left is the input image and all the right images are the target images with 2 kinds of attributes. In order to handle this task, the model should be able to generate the target images based on the input image with the appointed attributes. \begin{figure}[t] \begin{center} \includegraphics[width=1.0\linewidth]{figure0.pdf} \end{center} \caption{The images in the first column are the neutral images and the other column images mean the different facial expression images synthesized by our model.} \label{} \end{figure} \begin{figure*} \begin{center} \includegraphics[width=1.0\linewidth]{figure1.pdf} \end{center} \caption{The framework of our model.} \label{} \end{figure*} Many research show that the extension of GAN is able to disentangle image information encoded by vector z and by condition y making them independent \cite{isola2016image}\cite{yi2017dualgan}. We observe that vector z is able to encode persons' facial feature such as facial pose and hair styles with traditional GAN technologies. However, the traditional GAN based image-to-image methods meet difficulty when we try to translate a facial image into an image with appointed expression and pose. Inspired by other image translation concepts, we proposed an end-to-end image translation framework that includes two generators and two discriminators for facial images generation with multiple attributes. These generators and discriminators make up a pipeline architecture. The detail of our work is shown in Figure.~\ref{}. The overview of our model can be found in Figure~\ref{ consideration of the difficulty of generating images with multiple attributes at the same time, we divided the synthesizing process into two stages. We assign the two different attributes: facial expression and head pose, at the two different stages respectively. We denote two image generators as the expression generator $G_{e}$ and the pose generator $G_{p}$ for the two stages separately. Each generator has its respective discriminator. We denote the expression discriminator $D_{e}$ and the pose discriminator $D_{p}$. Similar to the idea of \cite{isola2016image}, the $G_{e}$ and the $D_{e}$ is a simple image-to-image generative adversarial network and so do $G_{p}$ and $G_{e}$. And the $X_{e}$ means the expression images that are synthesized by the expression generator and so do $X_{p}$. At the first stage. We use one generators for the pose image synthesizing. The output image $X_{p}$ and real pose images are evaluated by the pose discriminator to improve the image output quality. And then we take $X_{p}$ as the input of the expression generator $G_{e}$ for the next stage. The output images from the expression generator $G_{e}$ are the final outputs of our model. Through this pipeline operation, we can break a difficult problem into several easy steps. It's proved that the conditional vector can represent the extension information when synthesizing images using GANs\cite{lu2017conditional}\cite{mirza2014conditional}\cite{odena2016conditional}. Zhang et al. and Reed et al.'s work also show that GANs with information are able to synthesize images conditioned on text descriptions and on spatial constraints such as bounding boxes or key points\cite{reed2016generative}\cite{zhang2016stackgan}. Following this idea, we added two conditional vectors on the coded layer of two GANs respectively. It's believed that the coded layer includes high dimensional information. By adding conditional vectors on the coded layer, we can control the facial attribute and generate the appointed synthesized images. These additional information from latent space\cite{wu2016learning} can provides the generator additional constrains and force it generates the appointed images rather than generating casually. In order to make the encoder network focus on how to select the useful features, we used an additional discriminator on the bottom neck of the encoder-decoder network for parallel classfication task to enhance the feature extraction ability of the encoder network. The detail of this conditional adversarial architecture is described in section \ref{section3.1} and \ref{section3.2}. Inspired by Chen et al.'s work\cite{chen2017photographic} and Gulrajani et al.'s work\cite{gulrajani2017improved}, we include the cascade loss and gradient penalty to improve the image generation quality \ref{section3.3} and \ref{section3.4}. The full loss function of our framework is described in section \ref{section3.5}. Please note that there is another possible pipeline sequence of Figure~\ref{}. We define the sequence described in Figure~\ref{} as PE (pose generator at first), and we define EP (expression generator at first) as its inverse sequence structure. We evaluated these two structures in section \ref{section4}. Our main contributions are summarized as follows: \begin{itemize} \item We present a new pipeline architecture that can synthesize human face images with multiple attributes. With this framework, we are able to break a tough image generation task into several easy ones. \item We propose a parallel classification task in the bottom neck of the encoder-decoder structure to force the encoder network focus on image feature selection. \item We merge the cascade loss into the discriminator of our model to improve the image generation performance. An independent pre-trained cascade network is deployed in the discriminator to control the generated image quality layer by layer. \item We combine the gradient penalty with our model for the training process to make the training more stable. \end{itemize} Experimental results demonstrate the combination of these methods lead to convincing performance. \section{Related work} Recently, there have been remarkable progress in this field with the development of generative models that do not explicitly require this integration and can be trained using back-propagation algorithm. Two famous examples of such models are Generative Adversarial Networks (GANs)\cite{goodfellow2014generative} and Variational Autoencoders(VAE)\cite{KingmaW13}. These models are able to produce convincing image samples but not flexible enough to handle an image-to-image translation task with multiple attributes. Recent research related with GANs are mostly based on the work of DCGAN (deep convolutional generative adversarial network)\cite{radford2015unsupervised}. DCGAN has been proved to learn good feature representation from image pixels in many research\cite{zhu2017unpaired}\cite{isola2016image}. And the deep architectures have also shown the effectiveness of synthesizing photorealistic images in the adversarial networks\cite{chen2017photographic}. Image generation, especially face image generation, is still a hot research topic. Some early research focus on how to recover face images using incomplete face images with stacked Autoencoders and Boltzmann Machines\cite{Hinton504}. Nowadays, many researchers focus on image generation research after GANs are proposed by Goodfellow et al \cite{goodfellow2014generative}. GAN simultaneously train two models: a generative models for generating images and a discriminative model to differentiate between natural images and generated images. The key to the success of GANs is that they proposed an adversarial loss that can both optimize the generative model and discriminative model. The two-player game aims to force the generative model to synthesize more realistic images and discriminative model to differentiate between synthesized images and natural images. In the original work of Goodfellow et al. \cite{goodfellow2014generative}, GANs were used to synthesize MNIST digits and 32*32 images that aimed to reproduce the appearance of different classes in the CIFAR-10 dataset. Denton et al. \cite{denton2015deep} proposed training multiple separate GANs, one for each level in a Laplacian pyramid. Each model is trained independently to synthesize details at its scale. Assembling separately trained models in this fashion enabled the authors to synthesize smoother images and to push resolution up to 96*96. Zhao et al. \cite{mirza2014conditional} propose the conditional GANs and combine the conditional vectors with GAN to produce the images of appointed classes. In these works, researchers focus on how to convert a noise vector $z$ into an image $G(z|y)$ with a condition $y$ based on the adversarial frameworks. Reed et al.'s work further explore this field \cite{reed2016generative}. They proposed a text-to-image framework and try to convert a noise vector $z$ under a text condition $\varphi(t)$ to generate an image $G(z,\varphi(t))$ according to the text meaning. The recent work called InfoGAN\cite{chen2016infogan} proposed an information-theoretic extension to GANs to explore the potential of the noise vector . In their work, they decompose the input noise vector into two parts: incompressible noise $z$ and latent code $c$. And they try to make use of latent code $c$ based on the mutual information. Meanwhile, Isola et al. use a composite loss and provide an image-to-image translation solution that combines a GAN and a regression term\cite{isola2016image}. Instead of a one-dimensional conditional vector, they use a two-dimensional image as the input condition to control the input and the output performance. The authors applied their model on various datasets and demonstrate that their model is able to synthesize 256*256 images for given semantic layouts. Due to the development of GANs, image translation successfully drawn researchers' attention. The key research of this topic is to find the mapping function from one image domain to another image domain\cite{kim2017learning}\cite{yi2017dualgan}, which can be used for image painting and the image style conversion. Zhu et al. proposed a method which contains two mapping functions, and relatively there are two adversarial discriminators\cite{zhu2017unpaired}. They introduced two cycle consistency losses that can capture the image domain distribution and the translation from one image domain to another image domain on unpaired image samples. Although we can collect unpaired images more easily than paired samples, we need paired samples with multiple attributes to design a precise image translation system. That's why we choose Isola et al.'s model for our research\cite{isola2016image}. \section{Architecture\label{section3}} Similar to Isola et al.'s work, our network includes six convolutional layers with stride-2 and a residual block. Both generator and discriminator use modules of the form convolution-BatchNorm-Relu which is mentioned in\cite{ronneberger2015u}. All Relus in the encoder are leaky. \subsection{Conditional Adversarial Loss\label{section3.1}} This objective loss with different attributes can be expressed as \begin{equation} \label{eq1} \begin{split} L_{CGAN}(G,D)=E_{x,y\sim p_{data}(x,y)}\left[ logD\left(x,y\right)\right] \\ + E_{x\sim p_{data}(x),z\sim p_{z}(z)}\left[log(1-D(x,G(x,z)) \right] \end{split} \end{equation} As an extension of Conditional GAN, the adversarial loss is a necessary term to our model. In Eq.~\ref{eq1}, the vector $z$ is a random noise vector following a prior noise distribution $p_{z}(z)$ and $x$ presents the observed image following the distribution $p_{data}(x)$. Here $x$ is also a 2-dimensional condition to the adversarial network. The adversarial framework try to learn a mapping function from $\{x,z\}$ to the target image $y$. The binary discriminator D try to classify between the real image pair $\{x,y\}$ and the synthesized pair $\{x, G(x,z)\}$. The detail of the conditional adversarial loss is shown in Figure.~\ref{}. Following Isola et al. and Mirza et al.'s work\cite{isola2016image}\cite{mirza2014conditional}, the condition $x$, which is required in both generator and discriminator sides, can force the generator to generate image $G(x,z)$ according to the observed x. The conditional advesarial loss is a basic requirement in a conditional adversarial network. We inherit this term as a part of our hybrid discriminator in our adversarial system. \begin{figure}[t] \begin{center} \includegraphics[width=1.0\linewidth]{figure333.pdf} \end{center} \caption{The conditional adversarial loss.} \label{} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=1.0\linewidth]{figure2.pdf} \end{center} \caption{The parallel classification task and the conditional biases.} \label{} \end{figure} \subsection{Parallel Classification and Conditional Biases\label{section3.2}} We apply a parallel classification (PC) task in the coded layer for the generator as shown in Figure. \ref{}. We use the cross entropy as the objective function for the classification task: \begin{equation} L_{PC}=\sum_{k}^{K}y_{k}log{P(y=k)} \end{equation} Suppose we have $K$ classes ($K$ poses for example), the generator need to generate facial images with $K$ different poses based on the conditional biases $C$ from the coded layer 2. We introduce an additional classification task in the coded layer 1 to help the encoder network focus on how to extract the useful features from the raw images for encoding. After that, our model includes a conditional vector which indicates the appointed label we want. Please note that the conditional vector is inserted after we calculate the classification outputs and cross entropy error propagation to avoid short circuit. The pseudo code for training is listed in Algorithm \ref{algorithm1}. \begin{algorithm}[t] \caption{Pseudo code for parallel classification training.} \label{algorithm1} \begin{algorithmic}[1] \STATE $i=p$ \FOR{$i=1$ to $N steps$} \STATE Use $x'$ to get $D'(x')$ \STATE Calculate loss for encoder network \STATE Use $x$ and $C$ to get $G(x,C)$ \STATE Calculate loss for encoder and decoder \STATE Update the weights of the generator \ENDFOR \end{algorithmic} \end{algorithm} The classification task is implemented alternates with the image generation task. During one step, weights are updated after all the loss calcualtion of the generator are finished. There is a copy operation between encoder and decoder networks. The classification targets are the same as the conditional biases. But the weights $w_{c1}$ from the coded layer of the encoder network to parallel classification is different with the weights $w_{c2}$ from the conditional vector to the coded layer of the decoder network. The activation value of coded layer 1 and 2 are designed as following: \begin{equation} y_{c1}=f(x_{c1}+b_{c1}) \end{equation} \begin{equation} y_{c2}=f(x_{c1}+w_{c2}C+b_{c2}) \end{equation} where $y_{c1}$ denotes the outputs of coded layer 1 and $y_{c2}$ means the outputs of coded layer 2; $x_{c1}$ is the input of coded layer 1; $b_{c1}$ and $b_{c2}$ are the layer biases; $C$ is the conditional vector which uses one-hot encoding and $f( )$ is the activation function. Please note that the conditional vector $C$ is different with the conditional image $x$ mentioned in Eq. \ref{eq1} of section \ref{section3.1}. We use $C$ to control the expression and the pose information and make our model to synthesize the facial images with appointed pose and expression class. \subsection{Cascade Loss\label{section3.3}} \begin{figure}[t] \begin{center} \includegraphics[width=1.0\linewidth]{figure33.pdf} \end{center} \caption{The framework of cascade loss.} \label{} \end{figure} In order to make the synthesized images clearer and perform better, we consider Chen et al.'s work and include an extra pre-trained VGG-19 network to provide the cascade loss in our hybrid discriminator\cite{chen2017photographic}. Figure.~\ref{} shows the detail of cascade loss. The feature maps of convolutional networks can express some important information of images. To achieve lower bound between the real data distribution and the sample distributions, the cascade refinement network can provide some important guidance for our model. The cascade loss is considered as a measurement of similarity between the target and the output images. We initialize the network with the weights and bias of the pre-trained VGG network on the Imagenet dataset\cite{deng2009imagenet}. Each layer of the network will provide cascade loss between the real target images and the synthesized images. We used the first five convolutional layers for the cascade refine network. \begin{equation} \label{eq5} {L}_{Cascade}(\theta)=\sum^{N}_{n}{\lambda}_{n}||{\Phi}_{n}(x)-{\Phi}_{n}(G((x,z);\theta))||_{1} \end{equation} Following the definition mentioned in Eq. \ref{eq1}, here $y$ denotes the target image and $(G(x,z))$ is the image produced by the generator $G$; $\theta$ is the parameters of the generator G; $ {\Phi}_{n}$ is the cascade response in the $n_{th}$ level in the Cascade Refinement Network (CRN). We choose the first 5 convolutional layers in VGG-19 to calculate the cascade loss. So we have $N=5$. Please note the loss ${L}_{cascade}(\theta)$ mentioned in Eq. \ref{eq5} is only used to train the parameter $\theta$ of the generator G. The CRN is a pre-trained network. The weights of CRN will not be changed during we train the generator G. The parameter ${\lambda}_{n}$ controls the influence of cascade loss in the $n_{th}$ layer of CRN. The cascade loss can provide measure the similarity between the output and the target images under $N$ different scales and enhance the ability of discriminator. Cascade loss from the higher layer controls the global structure; the loss from the low layer controls the local detail during generation. Thus, the generator should provide better synthesized images to cheat the hybrid discriminator and finally improve the synthesized image quality. The loss constrains the generator and aims to seek the boundary between the real data distribution and sample data distribution. \subsection{Gradient Penalty\label{section3.4}} To further improve the performance of GANs, we consider Gulrajani et al.'s work and include the gradient penalty (GP) to our model\cite{gulrajani2017improved}. The gradient penalty can be expressed as \begin{equation} L_{GP} = E_{\hat{x}\sim P_{\hat{x}}} \left[(\left \| \bigtriangledown_{\hat{x}}D(\hat{x})\right \| _{2}- 1)^{2}\right] \end{equation} \begin{equation} \hat{x}=(1-\alpha)x+\alpha \tilde{x}\: \: \: \: \: where\: \alpha\sim U[0 , 1] \end{equation} Here $x$ represents the real sample and $\tilde{x}$ means the synthesized sample; $\alpha$ is the random parameter that controls the balance between real and fake and follows the average distribution between 0 and 1; $\bigtriangledown_{\hat{x}}$ is the gradient of $D(\hat{x})$. Please note that $\hat{x}$ is only used for calculating $L_{gp}$. We believe that the GAN with gradient penalty converge more stable and generate images with higher quality. The experiment results in section \ref{section4} also supports this point. \subsection{Generative loss\label{section3.5}} As mentioned in Isola et al.'s work \cite{isola2016image}, location loss is really important for synthesizing images. In our network, we use the L1 loss as the location loss. So the total loss of our generator is defined as: \begin{equation} \begin{split} L = \xi_{1}L_{CGAN}(G,D) + \xi_{2}L_{Cascade} + \\ \xi_{3}L_{GP} + \xi_{4}L_{PC}+\xi_{5}\left\| y -G(x,z) \right\|_{1} \end{split} \end{equation} Here $y$ is the target image and $G(x,z)$ is the image synthesized by the generator. In our experiments, we find that the model performs well when we choose $\xi_{1}=\xi_{2}=\xi_{3}=1, \xi_{4}=10, \xi_{5}=50$. And our final loss is the sum of this three different loss. We apply this final loss to the generator with Adam optimizer of learning rate $0.0002$.In our experiments. \section{Experimental results\label{section4}} \subsection{Datasets} We test our methods on three face datasets. We use two datasets to evaluate the image generation ability on pose direction and facial expression, respectively. The first dataset we used is FEI dataset\cite{thomaz2012fei}. This dataset contains 11 different pose directions from 200 individuals. Then we evaluate our methods for facial expression generation on the Yale dataset\cite{georghiades1997yale} which consist of 15 individuals and 6 different kinds of facial expression. In these two experiments, images only contains one attribute. Thus, we only use half of the model which we mentioned in Figure.~\ref{} for comparison and evaluation. At last, we train our model on the Karolinska Directed Emotional Faces dataset (KDEF)\cite{calvo2008facial} and evaluate the image generation ability with multiple attributes. The KDEF datasets is a set of totally 4900 pictures of human facial expressions pictures with different pose. Each subject contains 5 different pose and 7 emotional expressions (totally $5*7=35$ images). So it has 35 different face images for each person. We split the datasets into training dataset and testing dataset for 4:1. \begin{figure}[t] \begin{center} \includegraphics[width=1.0\linewidth]{figure6.pdf} \end{center} \caption{Head pose generation results with different methods.} \label{} \end{figure} \begin{figure*} \begin{center} \includegraphics[width=1.0\linewidth,height=1.0\textheight,keepaspectratio]{figure7.pdf} \end{center} \caption{Head pose generation results on 10 different angles. The images from the first column on the left are the inputs and the left images are the synthesized samples.} \label{} \end{figure*} \subsection{Head pose generation\label{section4.2}} In the FEI database, each individual contains 1 standard frontal facial image and 10 kinds of images with different pose direction. We choose the standard frontal facial image as the input for our model to generate the left 10 classes. In this task, we aim to generate images with different pose. A comparison is displayed in Figure.~\ref{}. It seems that each method is able to generate an image with the correct head direction. However, a synthesized image with correct facial detail is not an easy task. The results of Pix2pix is blur and fail to generate the eyes. The results impove when we integrate more methods mentioned in section \ref{section3}. Our methods performs better when we considered parallel classification, cascade loss and gradient penalty simultaneously. Figure.~\ref{} shows more pose generation results on 10 different pose with our methods (Pix2pix + Cascade + GP + PC). In most of cases, our model can generate clear images with correct local facial information. The results also prove that the conditional vectors which we defined in \ref{section3.2} on the coded layer successfully controls the pose direction well. \subsection{Facial expression generation\label{section4.3}} In the Yale dataset, we choose 1 neutral facial image as the input of our model and the left 5 different facial expression images as the targets to evaluate the expression generation performance of our model. Figure.~\ref{} shows the comparing testing results using different methods. At the first glance, each method seems to be able to generate images with the correct expression. If we enlarge the image size, we can find that the image generated by Pix2pix has a unclear boundary surrounding the subject's face. And our model can generate a clear facial image after we integrate Pix2pix with cascade loss, gradient penalty and parallel classification. Figure.~\ref{} shows more testing results in the Yale face datasets using our methods. These plausible generations show that our model can solve these image transfer tasks with one attribute well. \begin{figure}[t] \begin{center} \includegraphics[width=1.0\linewidth]{yale-compare.pdf} \end{center} \caption{Facial expression generation results with different methods.} \label{} \end{figure} \begin{figure}[t] \begin{center} \includegraphics[width=1.0\linewidth]{yale_data.pdf} \end{center} \caption{Facial expression generation results on 5 different classes.} \label{} \end{figure} \begin{table} \begin{center} \begin{tabular}{|c|c|c|c|} \hline Method & P-SNR &MSE &R-MSE\\ \hline\hline Ours & 17.0296 & 0.01394& 0.1158\\ PE + Cascade +GP & 16.9893 & 0.01411 &0.1163 \\ PE + Cascade & 16.8210 & 0.01451 & 0.1183\\ PE + GP& 16.9636 & 0.01410 & 0.1166 \\ PE & 16.4461 & 0.01586 & 0.1237\\ EP+GP& 16.5461 & 0.01547 & 0.1222\\ EP& 16.4180 & 0.01557 & 0.1236\\ Pix2pix & 17.0004 & 0.01556 & 0.1230\\ \hline \end{tabular} \end{center} \caption{Multiple attributes generation task using different methods on KDEF dataset.} \label{} \end{table} \begin{figure*} \begin{center} \includegraphics[width=1.0\linewidth,height=1.0\textheight,keepaspectratio]{figure10.pdf} \end{center} \caption{Image generation results on different subjects with multiple attributes. (a) subject 1 with -45 degrees and a sad expression; (b) subject 2 with +45 degree and a disgust expression; (c) subject 3 with -90 degree and a happy expression; (d)subject 3 with -45 degree and a happy expression.} \label{} \end{figure*} \subsection{Face generation with multiple attributes} In this paper, our final goal is to develop a model to generate images with multiple attributes using a Pipeline generative adversarial network. We choose KDEF dataset to evaluate the performance of our model on this task. There are 4900 images including 5 different angles (-90, -45, 0, +45, +90 degrees) and 7 different expressions (afraid, angry, disgusted, happy, neutral, sad, surprised) from 70 participants (35 males and 35 females) in this dataset. Similar to the experiments mentioned in section \ref{section4.2} and \ref{section4.3}, we use one neutral straight image as the model input to generate facial images with 4 different poses using the Generator 1 (Figure.~\ref{}). In this step, we can generate 4 different poses with neutral expression. And these images are reused as the inputs for the Generator 2 to generate $4\time6=24$ facial images. The Table.~\ref{} shows the image evaluation results of our experiments. Although Pix2pix has a similar P-SNR (higher is better) comparing to our methods, it has a poor performance on MSE and R-MSE (lower is better on both criteria). Our model, which integrate PE structure, Cascade loss, gradient penalty and parallel classification, has the best performs when considering cascade loss, gradient penalty, cross entropy and Pose-Emotion (PE) pipeline structure simultaneously. It should be noted that we find that PE structure performs better than EP structure. By using this trick, our results can be clearer and can generate elaborate images with plausible parts such as teeth, eyes and so on. Because we find that pose generation task on KDEF has better performance than facial expression generation task on the Generator 1. The generation quality of the first generator will influent the performance of the next generator. That's why we choose PE structure in Figure.~\ref{}. We evaluate the generation performance with multiple attribution of each model in Figure.~\ref{}. The first column image is the input of each model, other images represent outputs. (a), (b) and (c) show the performance of each model using 3 different subjects. (c) and (d) represent the generation results using the same subject with different attribute outputs (-90 degrees + happy). In the second column, we can find that the results of a single Pix2pix model are blur with distorted facial features. In this task, we try to generate $4\time6=24$ kinds of images with a single Pix2pix model. The result show that it's possible for a Pix2pix model to generate a facial image with correct head pose. But it fail to generate the correct expression in (b), (c) and (d) cases. Images in the third and the fifth column show the results of a pipeline Pix2pix structure of PE (pose first) and EP (expression first), respectively. In these two tasks, one Pix2pix model is only used to generate images with one attribute (pose or expression). The results of PE and EP perform better than a single Pix2pix case. We guess that the expression image generation is more difficult than pose image generation on KDEF dataset. However, the image detail of these two models are still very poor. Relatively, PE performs better than EP structure. The mouth generated by PE model is more approaching to the ground truth. We get better image quality when we include gradient penalty on PE and EP structure. We further enhanced the generation quality when we integrate GP and cascade loss on PE structure. And our method achieves the best performance among these methods. \section{Conclusion} In this paper, we proposed a new architecture that can generate facial images with multiple different attributes. And the pipeline architecture with two different generators and two discriminators can reduce the difficulty of each generator and guarantee the final image translation results. This direct approach can divide the difficult task into two stages and perform better than contemporaneous work single GAN based model. Furthermore, we successfully improve the image generation performance by adding cascade loss, gradient penalty and parallel classification. Our method also provide a new way for image data augmentation under multiple attributes. Nowadays, unsupervised learning is very popular because its good performance in representation learning \cite{zhu2017unpaired}. On the other side, summarizing multiple attributes from images is not an easy work if the image is complex. How to collect attributes automatically through unsupervised learning is an interesting work. We leave this part in our future work. {\small \bibliographystyle{ieee}

\section{Introduction} Consider a surface $M$ in $3$-space. A quadratic point $p$ is a point where $M$ can be exceptionally well approximated by a quadric. Generically, the quadratic points are isolated and occur at the elliptic or the hyperbolic region of $M$, not in the parabolic curve. Examples of algebraic hyperbolic tori in projective $3$-space with many quadratic points can be found in \cite{Tabach}, where it is conjectured that any hyperbolic tori in $\P^3$ should have quadratic points. In \cite{Ronaldo} one can find hyperbolic tori in $\S^3$ with no quadratic points. In \cite{Uribe}, it is proved that any hyperbolic disk bounded by a parabolic curve should have an odd number of quadratic points. The cubic form $C$ defines a binary cubic differential equation (BCDE) outside the parabolic curve. This BCDE is totally real in the elliptic region and has one root in the hyperbolic region, and is singular exactly at the quadratic points. Thus the cubic form defines a line field ${\mathcal H}$ in the hyperbolic region and a $3$-web ${\mathcal E}$ in the elliptic region, except at quadratic points. Both fields are projectively invariant. The behavior of a totally real $3$-web close to a singularity was studied in \cite{Ballesteros}. Applying the results of this paper to the particular case of a cubic form, we verify that generically the $3$-web $\mathcal{E}$ has index $\pm\frac{1}{3}$ at quadratic elliptic points. As a corollary we have that at a generic compact convex surface, there exist at least $6$ quadratic points. For semi-homogeneous cubic forms, we show a relation between the index of the $3$-web and Loewner's conjecture to conclude that the index of a quadratic point is at most $1$. As a corollary, we obtain that, under the semi-homogeneity hypothesis, the number of quadratic points in a compact convex surface is at least $2$, which is a version of Carath\'eodory conjecture. We also give an example of a compact rotation surface with exactly $2$ quadratic points. For quadratic points in the hyperbolic region, we show that generically the line field $\mathcal{H}$ has index $\pm 1$. By considering the polar blow-up, we describe the generic phase portrait of $\mathcal{H}$ at these points. In a neighborhood of the parabolic curve, there exists a line field ${\mathcal L}$, tangent to the parabolic curve, which coincides with the line field $\mathcal{H}$ in the hyperbolic part $H$ and with a line field $\mathcal{E}_1$ of the $3$-web $\mathcal{E}$ in the elliptic part $E$. With the help of this line field, we prove a version of Poincar\'e-Hopf Theorem which says that for a generic compact surface $M$ with elliptic region $E$ and hyperbolic region $H$, $$ \sum_{p_i\in E} Ind(p_i,\mathcal{E}) +\sum_{p_i\in H} Ind(p_i,\H)=\chi(M). $$ where the sum is taken over the the quadratic points $p_i$. The paper is organized as follows: In section 2 we show the relation between quadratic points and cubic forms. In section 3 we study the behavior of the $3$-web $\mathcal{E}$ at a simple quadratic point, while in section 4 we calculate the index of a $3$-web coming from a semi-homogeneous cubic form. In section 5 we study the behavior of the line field $\mathcal{H}$ at a simple quadratic point. In section 6 we calculate the cubic form in a neighborhood of a generic parabolic curve. Finally, in section 7, we prove some Poincar\'e-Hopf type theorems. \section{Quadratic points and cubic forms} \subsection{Darboux directions and quadratic points} A general point $p\in M$ admits a $3$-dimensional space quadrics with contact of order $2$ with the surface, i.e., with the same tangent plane and the same curvatures in all directions. Among these quadrics, there are three one-parameter families with the property that the contact function is a perfect cube. If $p$ is elliptic, all three families are real, while if $p$ is hyperbolic, one of these families is real and the other two complex.The null directions of the perfect cubes are called Darboux directions (\cite[p.358]{Darboux}, \cite[p. 141-144]{Lane}). {\it Quadratic points} $p\in M$ are points that admits an osculating quadric of order $3$. Quadratic points are invariant by projective transformations of the $3$-space and admit a one-dimensional space of third order osculating quadrics. \begin{example} Let $p=(0,0,0)$ and $M$ be given by \begin{equation}\label{} z=\frac{1}{2}(x^2+y^2)+\frac{c}{6}(x^3-3xy^2)+O_4(x,y). \end{equation} Any quadric of the form $$ z=\frac{1}{2}(x^2+y^2)+\alpha xz+\beta yz+\gamma z^2 $$ has a second order contact with $M$ at $p$. For the quadrics $$ z=\frac{1}{2}(x^2+y^2)-cxz+\gamma z^2, $$ the contact function at order $3$ is a multiple of $x^3$. Thus $(0,1)$ is a Darboux direction. The other Darboux directions are $(3,\pm\sqrt{3})$. The point $p$ is quadratic if and only if $c=0$. \end{example} \begin{example} Let $p=(0,0,0)$ and $M$ be given by \begin{equation}\label{} z=xy+a^3x^3+b^3y^3+O_4(x,y). \end{equation} Any quadric of the form $$ z=xy+\alpha xz+\beta yz+\gamma z^2 $$ has a second order contact with $M$ at $p$. For the quadrics $$ z=xy-3a^2bxz-3ab^2yz+\gamma z^2, $$ the contact function at order $3$ is a multiple of $(ax+by)^3$. Thus $(-b,a)$ is the Darboux direction. The point $p$ is quadratic if and only if $a=b=0$. \end{example} \subsection{Affine differential geometry and cubic forms} In the elliptic and hyperbolic parts of $M$, consider the equation \begin{equation*} D_XY=\nabla_XY+g(X,Y)\xi \end{equation*} where $X,Y$ are vector fields on $M$, $\xi$ is a transversal vector field, $\nabla$ is an affine connection and $g$ a non-degenerate bilinear form. It turns out that there exists a unique (up to sign) vector field $\xi$ such that $D_X\xi$ is tangent to $M$, for any vector field $X$ on $M$, and the area form induced by $\xi$ coincides with the area form defined by $g$. This $\xi$ is called the {\it affine normal vector field}, the corresponding $g$ and $\nabla$ are called the {\it Blaschke metric} and {\it induced connection}, respectively. At elliptic points, the Blaschke metric is definite, while at hyperbolic points, the Blaschke metric is indefinite (see \cite{Nomizu}). The cubic form in the elliptic and hyperbolic parts of $M$ are defined by $C=\nabla g$. It turns out that the conformal class of the cubic forms is projectively invariant. At elliptic points, $C$ has three real null directions, while at hyperbolic points, $C$ has only one real null direction (see \cite{Nomizu}). The following result can be found in \cite[p.114]{Tabach1}: \begin{lemma} A point $p$ is quadratic if and only if $C(p)=0$. \end{lemma} \begin{proof} Assume first that $p$ is elliptic. Then, up to an affine change of coordinates, the surface $M$ in a neighborhood of $p$ is given by equation \eqref{}. Then $p$ is quadratic if and only if $c=0$, i.e., if and only if the cubic form at $p$ is zero. If $p$ is hyperbolic, up to an affine change of coordinates, the surface $M$ in a neighborhood of $p$ is given by equation \eqref{} The point $p$ is quadratic if and only if $a=b=0$, i.e., if and only if the cubic form at $p$ is zero. \end{proof} \section{Simple quadratic points in the elliptic region} Assume that $M$ is the graph of $f:U\subset\mathbb{R}^2\to\mathbb{R}$ given by \begin{equation}\label{} f=\frac{1}{2}(x^2+y^2)+\frac{1}{24}\left( a_{40}x^4+4a_{31}x^3y+6a_{22}x^2y^2+4a_{13}xy^3+a_{04}y^4 \right)+O_5(x,y), \end{equation} where $O_5(x,y)$ denotes the terms of order $\geq 5$ in $(x,y)$. Then the cubic form is given by \begin{equation}\label{} \omega= \omega_1+O_2(x,y), \end{equation} where \begin{equation} \omega_1= (ax+by)(dx^3-3dxdy^2)+(cx+dy)(dy^3-3dydx^2) \end{equation} and \begin{equation} a=a_{40}-3a_{22},\ b=a_{31}-3a_{13},\ c=a_{13}-3a_{31}\ d=a_{04}-3a_{22}, \end{equation} (see appendix A). We say that the singularity $(0,0)$ of $\omega$ is {\it simple} if it is also an isolated singularity of $\omega_1$. It is easy to check that this condition is equivalent to \begin{equation}\label{Hyp:11} ad-bc\neq 0. \end{equation} Observe that for a generic surface $M$, all quadratic points are simple. \subsection{Index of simple quadratic points}\label{} Using complex notation we can write $\omega=(A+iB)dz^3+(A-iB)d\bar{z}^3$, where $A=C_{111}$ and $B=C_{222}$. It is proved in \cite{Ballesteros} that the index $I=I(0,0)$ of the singularity is given by \begin{equation}\label{} I=-\frac{\deg(A+iB)}{3}. \end{equation} The complex function $A+iB$ can be approximated by $A_1+iB_1=(ax+by)+i(cx+dy)$. In fact, we have $$ A+iB=A_1+iB_1+O_2(x,y) $$ \begin{lemma}\label{} For a simple quadratic point, the degree of $A+iB$ is the same degree of $A_1+iB_1$. \end{lemma} \begin{proof} Since the image of $|z|=r$ by $A_1+iB_1$ does not pass through the origin, we can choose $r$ sufficiently small such that $$ ||A+iB-(A_1+iB_1)||\leq \frac{1}{2}||A_1+iB_1||. $$ This implies that the degrees of $A+iB$ and $A_1+iB_1$ are the same. \end{proof} Define the {\it characteristic polynomial} \begin{equation}\label{} P(x,y)=\left( ax+by \right) (x^3-3xy^2)+\left( cx+dy \right) (y^3-3yx^2), \end{equation} and let \begin{equation}\label{} Q(x,y)=-\left( ax+by \right) (y^3-3xy^2)+\left( cx+dy \right) (x^3-3xy^2). \end{equation} Observe that if the quadratic point is simple, $P$ and $Q$ have no common roots. Let $P(t)=P(\cos(t),\sin(t))$, $Q(t)=Q(\cos(t),\sin(t))$, $A_1(t)=A_1(\cos(t),\sin(t))$ and $B_1(t)=B_1(\cos(t),\sin(t))$. Then $$ P(t)+iQ(t)=(A_1(t)+iB_1(t))\exp{(3it)}. $$ Thus the degree of $P+iQ$ is $\deg(A_1+iB_1)+3$, and so we conclude that \begin{equation} I=1-\frac{\deg(P+iQ)}{3}. \end{equation} The polynomials $(P,Q)$ were considered in \cite{Ivanov} in the context of euclidean umbilic points. \subsection{Polar blow-up of the cubic form}\label{} We can factor the cubic form $\omega$ as $\omega=\lambda_1\lambda_2\lambda_3$, where $\lambda_j$ is a $1$-form given by $$ \lambda_j= \cos(\phi_j)dx -\sin(\phi_j)dy=e^{i\phi_j}dz+e^{-i\phi_j}d\bar{z}. $$ Thus $$ \omega_1=\lambda_1\lambda_2\lambda_3=\exp{(i(\phi_1+\phi_2+\phi_3))}dz^3+\exp{(-i(\phi_1+\phi_2+\phi_3))}d\bar{z}^3 $$ $$ +\left[ \exp{(i(\phi_1+\phi_2-\phi_3))}+\exp{(i(\phi_1-\phi_2+\phi_3))}+\exp{(i(-\phi_1+\phi_2+\phi_3))} \right] d\bar{z}dz^2 $$ $$ +\left[ \exp{(i(\phi_1-\phi_2-\phi_3))}+\exp{(i(-\phi_1+\phi_2-\phi_3))}+\exp{(i(-\phi_1-\phi_2+\phi_3))} \right] d\bar{z}^2dz. $$ We conclude that $\phi_1=\theta-\pi/3$, $\phi_2=\theta$, $\phi_3=\theta+\pi/3$, where \begin{equation}\label{} \tan(3\theta)=\frac{c\cos t+d\sin t}{a\cos t+b\sin t}. \end{equation} Consider the polar projection $\pi:(-\delta,\delta)\times\mathbb{R}\to\mathbb{R}^2$ given by $\pi(r,t)=(r\cos(t),r\sin(t))$. Then \begin{equation*} \pi^*\lambda_j=\cos(\phi_j+t)dr-r\sin(\phi_j+t)dt. \end{equation*} Define the $6\pi$-periodic vector fields \begin{equation} X_j=r\sin(\phi_j+t)\frac{\partial}{\partial r}+\cos(\phi_j+t)\frac{\partial}{\partial t}, \end{equation} in the kernel of $\pi^*\lambda_j$. Then $X_j$ is singular only at points $(0,t_0)$ such that $\phi_j(t_0)+t_0=k\pi+\frac{\pi}{2}$, $k\in\mathbb{Z}$. At these points $$ DX_j(0,t_0)=(-1)^k \left[ \begin{array}{cc} 1 & 0 \\ 0 & -(1+\phi_j'(t_0)) \end{array} \right]. $$ The singularity $(0,t_0)$ is hyperbolic for $X_j$ if and only if $1+\phi_j'(t_0)\neq 0$. Since $\pi^*\omega=\pi^*\lambda_1\cdot\pi^*\lambda_2\cdot\pi^*\lambda_3$, the singularities of $\pi^*\omega$ are singularities of $\pi^*\lambda_j$, for some $j=1,2,3$. The following lemma is proved in \cite{Ballesteros}: \begin{lemma} The singularities of $\pi^*\omega$ are of the form $(0,t_0)$, where $t_0$ is a root of the characteristic polynomial $P(t)$. Moreover, the singularity $(0,t_0)$ is hyperbolic for $X_j$ if and only if $t_0$ is a simple root of $P(t)$. \end{lemma} \begin{lemma} For a singular point $(0,t_0)$ of $\pi^*\lambda_j$, we have that $$ P'(t_0)=-3Q(t_0)(1+\phi_j'(t_0)). $$ \end{lemma} \begin{proof} Differentiating \eqref{} we obtain $$ \sec^2(3\theta)3\theta'=\frac{ad-bc}{(a\cos t+b\sin t)^2} $$ and so $$ 3\theta'=\frac{ad-bc}{(a\cos t+b\sin t)^2+(c\cos t+d\sin t)^2}. $$ On the other hand, $$ P'+3Q=(-a\sin t+b\cos t)\cos(3t)+(c\sin t-d\cos t)\sin(3t). $$ Assuming $P(t)=0$, we have $(a\cos(t)+b\sin(t))\cos(3t)=(c\cos(t)+d\sin(t))\sin(3t)$. Thus $$ Q=\frac{c\cos(t)+d\sin(t)}{\cos(3t)};\ P'+3Q=\frac{bc-ad}{Q};\ 3\theta'=\frac{ad-bc}{Q^2}, $$ thus proving the lemma. \end{proof} As a consequence, the singularity $(0,t_0)$ is a saddle for $X_j$ if and only if $P'(t_0)$ and $Q(t_0)$ have opposite signs. So we can determine the phase portrait of the $3$-web in a neighborhood of a simple quadratic point only from the polynomials $P$ and $Q$. \subsection{Phase portraits of simple singularities} Consider the characteristic polynomial $P$ defined by equation \eqref{}. we may write $P$ as $$ P=ax^4+(b-3c)x^3y-3(a+d)x^2y^2+(c-3b)xy^3+dy^4. $$ We can now calculate the discriminant $\Delta$ of $P$ (see \cite{Poston}). The polynomial $P$ has double roots if and only if the discriminant $\Delta$ vanishes. Straightforward calculations show that $$ \Delta= S^3-27T^2, $$ where $$ S=\frac{1}{4}\left( 10(ad-bc)+3a^2+3b^2+3c^2+3d^2 \right). $$ and $$ T=\frac{1}{8}\left( (a+d)(a-d)^2-2bc(a+d)+ac^2+db^2-3dc^2-3ab^2 \right). $$ \begin{lemma}\label{} By a projective change of coordinates, we may take $a+d=0$. \end{lemma} \begin{proof} Consider a projective transformation of the form $$ (X,Y,Z)=\frac{1}{1+\lambda z}(x,y,z) $$ with an adequate $\lambda$. \end{proof} Since we are only interested in the conformal class of the cubic form, we may take $b-c=1$. Assuming $d=-a$, $c=h-\frac{1}{2}$, $b=h+\frac{1}{2}$ we obtain $$ \Delta=27a^2h^2-(1-a^2-h^2)^3;\ \ \delta=\frac{1}{4}-a^2-h^2. $$ On the other hand, $P$ and $Q$ have common roots if and only $\delta=ad-bc=0$. The curve $\Delta=0$ is an astroid (full line in Figure \ref{}) and $\delta=0$ is a circle tangent to the astroid (dashed line in Figure \ref{}). This astroid appears in \cite{Porteous}. Note that the type of singularity is constant on the connected components of the complement of $\Delta=0$ (section \ref{}), while the index is constant on the connected components of the complement of $\delta=0$ (section \ref{}) \begin{figure}[htb] \includegraphics[width=.4 \linewidth,clip =false]{ParElliptic.png} \caption{The space of parameters $(a,h)$.} \label{} \end{figure} \begin{Proposition} We have that \begin{itemize} \item If $(a,h)$ is outside the astroid, then $P$ has $2$ roots and the index is $1/3$. \item If $(a,h)$ is inside the astroid, the $P$ has $4$ roots: \begin{itemize} \item If $(a,h)$ is outside the dashed circle, then the index is $1/3$. \item If $(a,h)$ is inside the dashed circle, then the index is $-1/3$. \end{itemize} \end{itemize} \end{Proposition} It is proved in \cite{Ballesteros} that the phase portraits of a simple singularity is of type $D_1$, $D_2$ or $D_3$. It is clear that $D_1$ corresponds to points outside the astroid, $D_2$ to points inside the astroid but outside the circle, and $D_3$ to points inside the circle. \section{Semi-homogeneous elliptic cubic forms} \subsection{Definition} Let $\omega$ be a cubic form and $p$ an isolated singularity of $\omega$. Denote by $\omega_n$ the order $n$-jet of $\omega$ at $p$. A cubic form is called {\it semi-homogeneous} (of degree $n$) if $\omega_k=0$, $k=1...,n-1$, and $\omega_{n}$ has $p$ as an isolated singularity (see \cite{Ballesteros}). Observe that for $n=1$ this definition reduces to the definition of a simple singularity. The complex function $A+iB=C_{111}+iC_{222}$ can be approximated by $A_n+iB_n=C_{111}^{n}+iC_{222}^{n}$, where $C_{ijk}^n$ denotes the $n$-jet of $C_{ijk}$. More precisely $$ A+iB=A_{n}+iB_n+O_{n+1}(z,\bar{z}). $$ The semi-homogeneity assumption means that $A_k+iB_k=0$, $k=1,...,n-1,$ and $A_n+iB_n=0$ only for $z=\bar{z}=0$. \begin{lemma}\label{} For a semi-homogeneous cubic form $\omega$ of degree $n$, the index of $\omega$ is given by $$ I=-\frac{\deg(A_n+iB_n)}{3}. $$ \end{lemma} \begin{proof} The same argument as in Lemma \ref{} shows that $\deg(A_n+iB_n)=\deg(A+iB)$. Now the result follows from formula \eqref{}. \end{proof} \subsection{Local model of a surface with semi-homogeneous cubic form} Assume that $f$ is given by equation \eqref{}. By a projective change of coordinates (see Lemma \ref{}), we may assume $a_{22}=0$ and write \begin{equation}\label{eq:1} f(x,y)=\frac{1}{2}(x^2+y^2)+\frac{1}{24}\left( a_{40}x^4+4a_{31}x^3y+4a_{13}xy^3+a_{04}y^4\right)+O_5(x,y). \end{equation} \begin{lemma}\label{lemma:2} Assume that $f$ is given by \eqref{eq:1} and the $1$-jet of the cubic form is zero. Then \begin{equation}\label{eq:fn} f(x,y)=\frac{1}{2}(x^2+y^2)+O_5(x,y). \end{equation} \end{lemma} \begin{proof} From Appendix \ref{} we have that \begin{equation*} \omega_1=(ax+by)(dx^3-3dxdy^2)+(cx+dy)(dy^3-3dydx^2), \end{equation*} where \begin{equation*} a=a_{40},\ b=a_{31}-3a_{13},\ c=a_{13}-3a_{31}\ d=a_{04}. \end{equation*} Since, by hypothesis, $a=b=c=d=0$, we conclude that $a_{40}=a_{04}=a_{31}=a_{13}=0$. \end{proof} \begin{lemma}\label{lemma:h} Assume that $f$ is given by \eqref{eq:1} and the $(n-1)$-jet of the cubic form is zero. Then necessarily \begin{equation}\label{eq:fn} f(x,y)=\frac{1}{2}(x^2+y^2)+h(x,y)+O_{n+4}(x,y), \end{equation} where $h=h_{n+3}$ is a homogeneous polynomial of degree $n+3$. \end{lemma} \begin{proof} Induction. The case $n=2$ was proved in Lemma \ref{lemma:2}. For $n\geq 3$, observe that the $(n-1)$-jet of the cubic form has $2n$ coefficients which are linear combinations of the $n+3$ coefficients of $h_{n+2}$, as proved in Appendix \ref{}. We have only to verify that this linear map is injective, which is straightforward. \end{proof} Lemma \ref{lemma:h} says that if the cubic form is homogeneous of degree $n$, the surface is the graph of a function $f$ given by formula \eqref{eq:fn}. Then Appendix \ref{} says that $$ \omega_n=A_n(dx^3-3dxdy^2)+B_n(dy^3-3dydx^2), $$ where \begin{equation}\label{} A_n+iB_n=h_{xxx}-3h_{xyy}+i(h_{yyy}-3h_{yxx}). \end{equation} The semi-homogeneity hypothesis states that this function has $(0,0)$ as an isolated singularity. \subsection{Index of a semi-homogeneous cubic form} \begin{thm} \label{} Consider an isolated singularity $p$ of the cubic form in the elliptic region of a surface. Assume that the cubic form at $p$ is semi-homogeneous. Then the index of the $3$-web is at most $1$. \end{thm} \begin{proof} The index of $p$ coincides with the degree of $A_n+iB_n$ which is given by equation \eqref{}. Since $$ \frac{\partial^3 h}{\partial{\bar{z}^3}}=h_{xxx}-3h_{xyy}-i(h_{yyy}-3h_{yxx}) $$ Loewner's conjecture (\cite{Titus}) implies that the degree of $A_n+iB_n$ is $\geq -3$. Then Lemma \ref{} implies that the index of the $3$-web is at most $1$. \end{proof} A natural conjecture here is that the index of the cubic form is always $\leq 1$, even without the semi-homogeneity hypothesis. \section{Quadratic points in the hyperbolic region} Assume that $M$ is the graph of $f:U\subset\mathbb{R}^2\to\mathbb{R}$ given by \begin{equation}\label{} f(x,y)=xy+\frac{1}{24}\left( ax^4+4bx^3y+6ex^2y^2+4cxy^3+dy^4 \right)+O_5(x,y). \end{equation} Then the cubic form is given by \begin{equation}\label{} \omega= \omega_1+O_2(x,y), \end{equation} where \begin{equation}\label{} \omega_1= (ax+by)dx^3+(cx+dy)dy^3, \end{equation} (see Appendix \ref{}). We say that the isolated singularity $(0,0)$ of $\omega$ is {\it simple} if it is also an isolated singularity of $\omega_1$, which is equivalent to \begin{equation}\label{Hyp:1} \delta=ad-bc\neq 0. \end{equation} We observe that for a generic surface $M$, all quadratic points are simple. \subsection{Index of simple singularities}\label{} Denote by $X$ a vector field in the null direction of $\omega$. It follows from equation \eqref{} that we can find a vector field $X_1$ in the null direction of $\omega_1$ such that $$ X-X_1=O_2(x,y). $$ Then the same argument as in the proof of Lemma \ref{} shows that the index of $X$ is equal to the index of $X_1$, which is equivalent to say that the indices of $\omega$ and $\omega_1$ coincide. Observe that the index of $\omega_1$ varies continuously with the parameters $(a,b,c,d)$ in the set $\delta\neq 0$, and thus it is constant in each connected component of this set. Define the {\it characteristic polynomial} \begin{equation}\label{DefineCharPol} P(x,y)=(ax+by)x^3+(cx+dy)y^3 \end{equation} and let \begin{equation}\label{} Q(x,y)=xy\left((ax+by)x-(cx+dy)y\right). \end{equation} \begin{lemma}\label{} For simple singularities satisfying $ad\neq 0$, $P$ and $Q$ have no common roots. \end{lemma} \begin{proof} If $(x,y)$ is a root of $Q$ with $x\neq 0$ and $y\neq 0$, $(ax+by)x=(cx+dy)y$, and so $$ P(x,y)=(ax+by)x(x^2+y^2). $$ If $(x,y)$ is also a root of $P$, $ax+by=0$, which implies $cx+dy=0$, thus contradicting hypothesis \eqref{Hyp:1}. \end{proof} \begin{lemma} Denote $P(t)=P(\cos(t),\sin(t))$ and $Q(t)=Q(\cos(t),\sin(t))$. Up to a factor, we can write $\omega_1$ in polar coordinates as \begin{equation}\label{eq:l3} \omega_1= P(t)dr^3-3rQ(t)dr^2dt+ O_2(r)drdt^2+O_3(r)dt^3, \end{equation} where $O_n(r)$ denotes terms of order $\geq n$ in $r$. \end{lemma} \begin{proof} Straightforward calculations using $x=r\cos(t)$ and $y=r\sin(t)$. \end{proof} \begin{lemma} Assume $ad\neq 0$ and denote $\bar{X}_1$ the vector field $$ \bar{X}_1=(3xQ-yP) \frac{\partial}{\partial x} +(3yQ+xP)\frac{\partial}{\partial y}. $$ Then $$ ||\bar{X}_1-X_1||=O_2(r). $$ \end{lemma} \begin{proof} In polar coordinates, $$ \bar{X}_1=3rQ(t)\frac{\partial}{\partial r}+P(t)\frac{\partial}{\partial t}. $$ Using formula \eqref{eq:l3}, we conclude the proof of the lemma. \end{proof} \begin{Proposition}\label{} The index of the vector fields $X_1$ and $\bar{X}_1$ coincide. \end{Proposition} \begin{proof} Similar to the proof of Lemma \ref{}. \end{proof} \begin{corollary} Assume $ad\neq 0$. Then the index of a simple singularity is equal to the degree of the map $(x,y)\to(Q(x,y),P(x,y))$ plus $1$. \end{corollary} \begin{proof} By Proposition \ref{}, the index of the simple singularity is given by the degree of the map $(x,y)\to(3xQ-yP,3yQ+xP)$. But the degree of this map is equal to the degree of the map $(x,y)\to (3Q,P)$ plus $1$. Finally the degree of $(3Q,P)$ is equal to the degree of $(Q,P)$. \end{proof} \subsection{Polar blow-up of the linear form}\label{} We can write $$ \lambda_1= \cos(\phi)dx -\sin(\phi)dy, $$ where \begin{equation}\label{} \tan^3(\phi)=-\frac{cx+dy}{ax+by}. \end{equation} Denote by $X_1$ a vector field in the kernel of $\lambda_1$. Consider the polar projection $\pi:(-\delta,\delta)\times\mathbb{R}\to\mathbb{R}^2$ given by $\pi(r,t)=(r\cos(t),r\sin(t))$. Then straightforward calculations show that \begin{equation}\label{eq:l1} \pi^*\lambda_1=\cos(\phi+t)dr-r\sin(\phi+t)dt. \end{equation} \begin{lemma} Assume $ad\neq 0$. The singularities of $\pi^*\omega_1$ are of the form $(0,t_0)$, where $t_0$ is a root of the characteristic polynomial $P$. \end{lemma} \begin{proof} We use Formula \eqref{eq:l3} with $\pi^*\omega_1$ in place of $\omega_1$. Thus, $\pi^*\omega_1(r,t)=0$ implies that $P(t)=0$. And since, by Lemma \ref{}, $P$ and $Q$ have no common roots, we conclude that $r=0$. Conversely if $P(t)=0$ and $r=0$ is is obvious that $\pi^*(r,t)=0$. \end{proof} The vector field $$ \pi^*X_1=r\sin(\phi+t)\frac{\partial}{\partial r}+\cos(\phi+t)\frac{\partial}{\partial t} $$ is in the kernel of $\pi^*\lambda_1$. $\pi^*X_1$ is singular only at points $(0,t_0)$ such that $\phi(t_0)+t_0=k\pi+\frac{\pi}{2}$, $k\in\mathbb{Z}$. At these points $$ D(\pi^*X_1)(0,t_0)=(-1)^k \left[ \begin{array}{cc} 1 & 0 \\ 0 & -(1+\phi'(t_0)) \end{array} \right]. $$ Observe that $(0,t_0)$ is hyperbolic for $\pi^*X_1$ if and only if $1+\phi'(t_0)\neq 0$. \begin{lemma} Assume $ad\neq 0$. Then, if $P(t_0)=0$, we have that $$ P'(t_0)=-3Q(t_0)(1+\phi'(t_0)). $$ \end{lemma} \begin{proof} Differentiating equation \eqref{} we obtain $$ -3\tan^2(\phi)\sec^2(\phi)\phi'=\frac{ad-bc}{(a\cos t+b\sin t)^2}. $$ If $P(t)=0$, we have $\phi+t=\frac{\pi}{2}+k\pi$, and so $\sin(\phi)=\pm \cos(t)$, $\cos(\phi)=\pm\sin(t)$. Thus $$ 3\phi'= \frac{bc-ad}{(a\cos t+b\sin t)^2}\frac{\sin^4t}{\cos^2t}. $$ Moreover, since $P(t)=0$, $Q(t)=\frac{\cos^2t}{\sin t}(a\cos t+b\sin t)$. Thus $$ 3Q\phi'= \frac{bc-ad}{a\cos t+b\sin t}\sin^3t. $$ On the other hand, $$ P'+3Q=(-a\sin t+b\cos t)\cos^3t+(-c\sin t+d\cos t)\sin^3t. $$ From $P(t)=0$ we obtain $$ P'+3Q=\sin^3t \left( (a\sin t-b\cos t) \frac{(c\cos t+d\sin t)}{a\cos t+b\sin t} +(-c\sin t+d\cos t) \right) $$ Thus $P'+3Q=-3Q\phi'$. \end{proof} \begin{Proposition} Assume $ad\neq 0$. The singularity $(0,t_0)$ of $\pi^*X_1$ is hyperbolic if and only if the root $t_0$ of $P$ is simple. Moreover, it is a saddle if and only if $P'(t_0)$ and $Q(t_0)$ have opposite signs. \end{Proposition} \subsection{Parameter space in the hyperbolic region} We may write the characteristic polynomial as $$ P(x,y)=ax^4+bx^3y+cxy^3+dy^4. $$ This polynomial has multiple roots if and only if its discriminant $\Delta$ vanishes. In this case, we have that $\Delta=S^3-27T^2$, where $$ S=ad-\frac{bc}{4}; \ \ T=-\frac{1}{16}(ac^2+db^2), $$ (see\cite{Poston}). By section \ref{}, the type of singularity is constant in the connected components of $\Delta=0$ and $ad=0$. On the other hand, by section \ref{}, the index is constant in the connected components of $\delta=0$. If $bc\neq 0$, by a projective change of coordinates, we may assume $b=\pm 4$, $c=\pm 4$ (see \cite{Tabach}), and we consider two different cases, $bc=16$ and $bc=-16$. The third case to consider is $bc=0$. \subsubsection{Case $bc=16$} For $bc=16$, we obtain $$ \Delta=(ad-4)^3-27(a+d)^2;\ \ \delta=ad-16. $$ \begin{Proposition} We have that \begin{itemize} \item If $\Delta>0$ (outside the full line in Figure \ref{}), then $P$ has no roots and the index is $1$. \item If $\Delta<0$ (inside the full line in Figure \ref{}), then $P$ has $2$ roots: \begin{itemize} \item If $\delta>0$ (outside the dashed line in Figure \ref{}), then the index is $1$. \item If $\delta<0$ (inside the dashed line in Figure \ref{}), then the singularity is a saddle and the index is $-1$. \end{itemize} \end{itemize} \end{Proposition} \begin{figure}[htb] \includegraphics[width=.4 \linewidth,clip =false]{ParHyp1.png} \caption{The space of parameters $(a,d)$ in case $bc=16$.} \label{} \end{figure} \subsubsection{Case $bc=-16$} For $bc=-16$, we obtain $$ \Delta=(ad+4)^3-27(a+d)^2;\ \ \delta=ad+16. $$ \begin{Proposition} The full line curve in Figure \ref{} is $\Delta=0$. The region $\Delta>0$ is divided in a bounded region, that we shall call region $I$, and two unbounded regions, that we shall call region $II$. The region $\Delta<0$ is divided in two unbounded regions, that we shall call region $III$. \begin{itemize} \item If $(a,d)$ belongs to region $I$, then $P$ has $4$ roots and the index is $1$. \item If $(a,d)$ belongs to region $II$, then $P$ has no roots and the index is $1$. \item If $(a,d)$ belongs to region $III$, then $P$ has $2$ roots: \begin{itemize} \item If $\delta>0$ (inside the dashed line in Figure \ref{}), then the index is $1$. \item If $\delta<0$ (outside the dashed line in Figure \ref{}), then the singularity is a saddle and the index is $-1$. \end{itemize} \end{itemize} \end{Proposition} \begin{figure}[htb] \includegraphics[width=.4 \linewidth,clip =false]{ParHyp2.png} \caption{The space of parameters $(a,d)$ in case $bc=-16$.} \label{} \end{figure} \subsubsection{Case $bc=0$} We shall assume without loss of generality that $b=0$ and $c=4$. In this case, $$ \Delta=a^2(ad^3-27);\ \delta=ad. $$ \begin{figure}[htb] \includegraphics[width=.3 \linewidth,clip =false]{ParHyp3.png} \caption{The space of parameters $(a,d)$ in case $bc=0$.} \label{} \end{figure} The full line in Figure \ref{} is $\Delta=0$, while the dashed lines are $\delta=0$. \begin{Proposition} We have that: \begin{itemize} \item If $(a,d)$ is in the second or fourth quadrants, then the index is $-1$. \item If $(a,d)$ is in the first or third quadrants, then the index is $+1$. \begin{itemize} \item If $(a,d)$ is outside the full lines, then $P$ has no roots. \item If $(a,d)$ is inside the full lines, then $P$ has $2$ roots. \end{itemize} \end{itemize} \end{Proposition} \subsection{Phase portraits of simple singularities} In this section we assume that the singularity is simple, $ad\neq 0$ and the characteristic polynomial $P$ has only simple roots. Under these hypothesis, we shall show that there are only four possible types of phase portraits, and they correspond to index $1$, $P$ with $0$, $2$ or $4$ roots and index $-1$, $P$ with $2$ roots. We begin with some general facts: At consecutive roots of $P$, $\frac{dP}{dt}$ changes its sign. To find the sign of $Q$, observe that if $P=0$ then $$ (a\cos(t)+b\sin(t))\cos^3(t)=-(c\cos(t)+d\sin(t))\sin^3(t). $$ At such points $$ Q(t)=-\sin^2(t) \left(c+d\tan(t) \right). $$ So $Q(t)>0$, for $\tan(t)<-\frac{c}{d}$ and $Q(t)<0$, for $\tan(t)>-\frac{c}{d}$. \paragraph{Case 1: $P$ has no real roots} In this case the index is $+1$. \begin{example}\label{} Consider $a=1$, $b=2$, $c=-4$, $d=12$. $$ P=x^4+2x^3y-4xy^3+12y^4=(x^2-2xy+2y^2)(x^2+4xy+6y^2). $$ The phase portrait of the line field can be seen in Figure \ref{}. \begin{figure}[htb] \centering \includegraphics[width=0.3\linewidth]{foco1a.pdf} \caption{ Line field of Example \ref{}. } \label{} \end{figure} \end{example} \paragraph{Case 2: $P$ has $4$ real roots} From the previous section, we may assume that $bc=-16$ and $(a,d)$ belongs to region $I$ (see Figure \ref{}). It is also clear from our analysis that the type of the hyperbolic singularities of $\pi^*X_1$ are constant in each of the four regions $I1=\{(a,d)\in I| a>0,d>0\}$, $I2=\{(a,d)\in I| a<0,d>0\}$, $I3=\{(a,d)\in I| a<0,d<0\}$ and $I4=\{(a,d)\in I| a>0,d<0\}$. Thus to see the type of the hyperbolic singularities we must choose a pair $(a,d)$ in each of these four regions and compare $-\frac{c}{d}$ with the roots $s_1<s_2<s_3<s_4$ of $P$. We have chosen the following pairs: For $(a,b,c,d)=(1,-4,4,1)\in I1$, we obtain $-\frac{c}{d}=-4<-1<s_1$. For $(a,b,c,d)=(-1,-4,4,-1)\in I3$, we obtain $-\frac{c}{d}=4>1>s_4$. For $(a,b,c,d)=(-\frac{8}{5},-4,4,\frac{8}{5})\in I2$, we obtain $-\frac{c}{d}=-\frac{5}{2}<s_1=-2$. For $(a,b,c,d)=(\frac{8}{5},-4,4,-\frac{8}{5})\in I4$, we obtain $-\frac{c}{d}=\frac{5}{2}>s_4=2$. We conclude that in any case, the sign of $Q$ at the roots of $P$ keep constant. Thus we have only parabolic sectors. \begin{example}\label{} Consider $$ P(t)=\cos^4(t)(\tan(t)-s_1)(\tan(t)-s_2)(\tan(t)-s_3)(\tan(t)-s_4), $$ where $s_1=-2$, $s_2=-1$, $s_3=1$ and $s_4=-\frac{1}{2}$. Then $$ a=-1,\ b=-\frac{5}{2},\ c=\frac{5}{2},\ d=1. $$ The phase portrait of the polar line field can be seen in Figure \ref{}. \begin{figure}[htb] \centering \includegraphics[width=0.3\linewidth]{index1No.pdf} \caption{ Line field of Example \ref{}. Polar blowing-up resolution with eight hyperbolic singular points, four saddles and four nodes.} \label{} \end{figure} \end{example} \paragraph{Case 3: $P$ has $2$ real roots} Assume that the roots of $P(t)$ are $t_1<t_2$. We write $$ P(t)=\cos^2(t)P_2(\cos t,\sin t)(\tan(t)-\tan(t_1))(\tan(t)-\tan(t_2)), $$ for some quadratic polynomial $P_2(x,y)>0$. Then $$ P'(t_1)=P_2(\cos t_1, \sin t_1)(\tan(t_1)-\tan(t_2))<0, $$ and similarly $P'(t_2)>0$. If the signs of $Q(t_1)$ and $Q(t_2)$ are the same, then there are only parabolic sectors, the singular point is a node and the index is $+1$. If $Q(t_1)$ and $Q(t_2)$ have different signs, then there are $4$ hyperbolic sectors, the singular point is a saddle and the index is $-1$. \begin{example}\label{} Consider $P_2(x,y)= (y-x)^2+x^2$, $\tan(t_1)=1$, $\tan(t_2)=-\frac{4}{3}$. In other words, $a=-8/3$, $b=10/3$, $c=-5/3$, $d=1$. Then there are only parabolic sectors. The phase portrait of the line field can be seen in Figure \ref{}. \begin{figure}[htb] \centering \includegraphics[width=0.3\linewidth]{index1No2.pdf} \caption{ Line field of Example \ref{}. Polar blowing-up resolution with four hyperbolic singular points, two saddles and two nodes. } \label{} \end{figure} \end{example} \begin{example}\label{} Consider $a=-1$, $d=1$, $b=-1/2$, $c=1/2$. Then there are $2$ singular points of the blow-up, $\tan(t)=\pm 1$. There are $4$ hyperbolic sectors, thus the index is $-1$. The phase portrait of the line field can be seen in Figure \ref{}. \begin{figure}[htb] \centering \includegraphics[width=0.3\linewidth]{indexNeg.pdf} \caption{ Line field of Example \ref{}. Polar blowing up resolution with four hyperbolic singular points, all saddles.} \label{} \end{figure} \end{example} \section{Parabolic curve} Let $C$ be a parabolic curve of a surface $M$. Close to $C$ there are hyperbolic and elliptic regions $H$ and $E$. In this section we prove that generically, we can extend the line field $\H$ continuously to $C$ by considering the tangent lines $\L$ to $C$. We prove also that we can select a line field $\mathcal{E}_1$ of the $3$-web $\mathcal{E}$ of $E$ such that $\mathcal{E}_1$ can be continuously extended to $\L$. Generically, a parabolic curve has fold points and a finite number of Gauss points, also called godrons (\cite{Izumiya}). \subsection{The cubic form in a neighborhood of a fold point} A generic surface $M$ at a fold point is given by \begin{equation}\label{} z=\frac{1}{2}x^2+\frac{1}{6}y^3. \end{equation} (see \cite{Izumiya}). \begin{Proposition}\label{} Assume that $M$ is given by equation \eqref{}. Then the cubic form is conformally equivalent to $-3dx^2dy+ydy^3$. \end{Proposition} \begin{proof} See Appendix \ref{}. \end{proof} From the above proposition we conclude that, at a fold point, there exists a line field $\L$ that coincides with $\H$ in the hyperbolic part, with a line field $\mathcal{E}_1$ of the $3$-web $\mathcal{E}$ in the elliptic part at is tangent to the parabolic curve. \subsection{The cubic form in a neighborhood of a Gauss cusp} A generic surface in a neighborhood of a Gauss cusp is given by \begin{equation}\label{} z=\frac{1}{2}x^2+\frac{1}{2}xy^2+\frac{\lambda}{24}y^4, \end{equation} for some $\lambda\neq 0,3$ (see \cite{Platonova} and also \cite{Uribe}). \begin{Proposition}\label{} Assume that $M$ is given by equation \eqref{}. Then the cubic form is conformally equivalent to $$ C_{111}=-3, \ C_{112}=-\lambda y, \ C_{122}= 3x-\lambda\frac{y^2}{2}, $$ $$ C_{222}=y\left(x(6+\lambda)+\lambda(\lambda-2) \frac{y^2}{2}\right). $$ \end{Proposition} \begin{proof} See Appendix \ref{}. \end{proof} From the above proposition we conclude that, at a Gauss cusp, there exists a line field $\L$ that coincides with $\H$ in the hyperbolic part, with a line field $\mathcal{E}_1$ of the $3$-web $\mathcal{E}$ in the elliptic part at is tangent to the parabolic curve. \section{Poincar\'e-Hopf type theorems} \subsection{Elliptic region} Consider now an elliptic region $E$ bounded by a parabolic curve $C$. Let $\mathcal{E}$ denote the $3$-web of the cubic form in the region $E$. As we have seen above, in a neighborhood of $C$ one of the line field $\mathcal{E}_1$ of the $3$-web $\mathcal{E}$ can be continuously extended to $C$ by considering $\mathcal{E}_1$ tangent to $C$. \begin{Proposition}\label{} We have that $$ \sum_{p_i\in E} Ind(p_i,\mathcal{E}) =\chi(E). $$ where the sum is taken over the quadratic points $p_i$. \end{Proposition} \begin{proof} In this proof we follow \cite{Manfredo} and \cite{Ballesteros}. Assume first that $M$ is orientable, for non-orientable surfaces consider a double covering. Fix an orientation and a Riemannian metric on $M$. Take a triangulation $\mathcal{T}$ such that each triangle $T\in\mathcal{T}$ contains at most one singular point $p_T$ in its interior. Take $X$ any vector field in the kernel of $\omega$. Then $$ \int_{T}Kd\sigma-2\pi Ind(\omega,p_T)=\frac{1}{6}\int_{6\partial T} \left[ \frac{DX}{ds} \right] ds, $$ where $\left[ \frac{DX}{ds} \right]$ denotes the algebraic value of the covariant derivative and the notation $6\partial{T}$ means $6$ times the boundary of $T$. Summing we obtain $$ \int_{E}Kd\sigma-2\pi \sum_{p\in E} Ind(\omega,p)=\int_C \left[ \frac{DX}{ds} \right] ds=-\int_C k_gds, $$ where the last equality comes from the fact that we can choose $X$ tangent to $C$. Now the Gauss-Bonnet theorem implies that $$ \sum_{p\in E} Ind(\omega,p)=\chi(E), $$ thus proving the proposition. \end{proof} \begin{corollary} A generic compact convex surface $M$ in $\P^3$ has at least $6$ quadratic points. \end{corollary} \begin{corollary}\label{} Let $M$ be a compact convex surface in $\P^3$ and assume that the cubic forms at the quadratic points are semi-homogeneous. Then $M$ has at least $2$ quadratic points. \end{corollary} Corollary \ref{} is a version of Carath\'edory conjecture for quadratic points, assuming that the cubic forms are semi-homogeneous. One can conjecture that Corollary \ref{} remains valid without this assumption. \subsection{A compact convex surface with two quadratic points} Consider a smooth rotation surface $S$ with axis $L$. The points of $L\cap S$ are quadratic of index $1$. We describe below a compact rotation surface $S$ with $2$ points of intersection with the axis $L$ and no other quadratic point. In rectangular coordinates $(x_1,x_2,y)$, the equation of $S$ is \begin{equation*} x_1^2+x_2^2+y^2(1+\lambda y)^2=(1+\lambda y)^2, \end{equation*} for some constant $\lambda>0$ small. Let $(x(s),y(s))$ be a curve satisfying \begin{equation}\label{} x'y''-y'x''=1 \end{equation} and consider the rotation surface \begin{equation}\label{} \psi(s,\theta)=\left( x(s)\cos\theta, x(s)\sin\theta, y(s) \right). \end{equation} \begin{Proposition}\label{} Assume $M$ is given by equation \eqref{}. Then the quadratic points are given by the equation $xy''=x'y'$. \end{Proposition} \begin{proof} See Appendix \ref{}. \end{proof} \begin{example} Take $$ (x(t),y(t))=\left( \cos(t) (1+\lambda\sin(t)), \sin(t) \right), \ \ -\frac{\pi}{2}\leq t\leq \frac{\pi}{2}, $$ for some constant $\lambda>0$ small (see Figure \ref{}). \begin{figure}[htb] \centering \includegraphics[width=0.3\linewidth]{Rotation.png} \caption{ A section of the rotation surface ($\lambda=0.2$). } \label{} \end{figure} Then $$ \Delta=x_ty_{tt}-x_{tt}y_t=1+3\lambda\sin t-2\lambda\sin^3t. $$ Observe that $y'=y_t\frac{dt}{ds}=\cos(t)\Delta^{-1/3}$ and so $$ \frac{y'}{x}=\frac{\Delta^{-1/3}}{1+\lambda\sin t}. $$ Then $$ \left( \frac{y'}{x} \right)_t= -\frac{\Delta^{-4/3}}{3(1+\lambda\sin t)^2} \left( \Delta_t(1+\lambda\sin t)+3\Delta \lambda\cos t \right) $$ Since $$ \Delta_t(1+\lambda\sin t)+3\lambda\cos t\Delta =6\lambda\cos^3t(1+2\lambda\sin t), $$ we conclude that $\left( \frac{y'}{x} \right)_t<0$, or equivalently $\left( \frac{y'}{x} \right)'<0$. This implies that there no quadratic points other than $(0,0,\pm 1)$. \end{example} \subsection{Hyperbolic region} Consider an hyperbolic region $H$ bounded by a parabolic curve $C$. Let $\H$ be the line field of the cubic form in the region $H$. As we have seen above, generically $\H$ can be extended continuously to $C$ by considering $\H$ tangent to $C$. Thus the classical Poincar\'e-Hopf Theorem implies the following: \begin{Proposition}\label{} Let $H$ be the hyperbolic region of a generic surface $M$. Then $$ \sum_{p_i\in H} Ind(p_i,\H) =\chi(H), $$ where the sum is taken over the quadratic points $p_i$. \end{Proposition} The following corollary is proved in \cite{Uribe}: \begin{corollary} A generic hyperbolic disc has an odd number of quadratic points. \end{corollary} \subsection{A general result} The following general result is a consequence of Propositions \ref{} and \ref{}: \begin{thm} Denote by $E$ and $H$ the elliptic and hyperbolic regions of a generic compact surface $M$ in the projective space $\P^3$. Then $$ \sum_{p_i\in E} Ind(p_i,\mathcal{E}) +\sum_{p_i\in H} Ind(p_i,\H)=\chi(M), $$ where the sum is taken over the quadratic points $p_i$. \end{thm}

\section{Introduction} \label{introduction} Since their inception, high throughput materials science frameworks such as {\small AFLOW}\ \cite{aflowPAPER,, have been amassing large databases of materials properties. For instance, the {\small AFLOW}\ database \cite{aflowlibPAPER, alone contains over 1.7 million material compounds with over 170 million calculated properties, generated from the \underline{I}norganic \underline{C}rystal \underline{S}tructure \underline{D}atabase ({\small ICSD}) \cite{ICSD,ICSD0,Belsky_ActaCristB_2002}, as well as by decorating crystal structure prototypes \cite{}. Combined with other online databases such as the Materials Project \cite{materialsproject.org}, NoMaD \cite{nomad} and {\small OQMD} \cite{oqmd.org}, materials data is abundant and available. As a result, \underline{m}achine \underline{l}earning ({\small ML}) methods have emerged as the ideal tool for data analysis \cite{Zeng_ChemMat_2002,Gu_SSS_2006,Ghiringhelli_PRL_2015,PyzerKnapp_AdFM_2015,Guzik_NMat_2016,Ziatdinov_NPJCM_MachineVision_2017}, by identifying the key features in a data set \cite{} to construct a model for predicting the properties of a material. Recently, several models were developed to predict the properties of various material classes such as perovskites \cite{}, oxides \cite{Ceder_Chem_Materials}, elpasolites \cite{Faber_PRL_Elpasolite_2016,Pilania_CMS_Elpasolite_2017}, thermoelectrics \cite{}, and metallic glasses \cite{acs.jpclett.7b01046}. Additionally, generalized approaches have been devised for inorganic materials \cite{} and for systematically identifying efficient physical models of materials properties \cite{}. While predictions are powerful tools for rational materials design, the discipline is still reasonably new within the realm of materials science. As a result, a working understanding of machine learning principles, along with a high level of technical expertise, is required for using code bases effectively. This inhibits accessibility, where an average end user aims to utilize the codes to retrieve predictions with as little complication as possible. With the ever increasing number of predictive models, a unique challenge emerges: how does one create an accessible means to integrate machine learning frameworks into a materials discovery workflow? {\small AFLOW-ML}\ alleviates this issue by providing a simplified abstraction on top of sophisticated predictive models. Predictions are exposed through a web accessible \underline{A}pplication \underline{P}rogramming \underline{I}nterface ({\small API}) where functionality is distilled down to its essence: from the user input, return a prediction. {\small AFLOW-ML}\ can be added into any code base through use of an {\small HTTP} request library, native to most languages. Alternatively, {\small AFLOW-ML}\ can be utilized though use of the included Python client and command line interface. Through such abstractions, {\small AFLOW-ML}\ is accessible to a wide audience: it unburdens the user from having to understand the intricacies of machine learning and eliminates the technical expertise required to set up such codes. \section{REST API} \label{Features} The {\small AFLOW-ML}\ {\small API}\ is structured around a \underline{RE}presentational \underline{S}tate \underline{T}ransfer ({\small REST}) architecture, which allows resources to be accessed using {\small HTTP} request methods. Each resource is located at an endpoint, which is identified by a {\small URL}\ comprised of descriptive nouns. A {\small URL}\ may also include special variables in the form of path parameters and a query string. A path parameter is a segment of the {\small URL}\ that specifies a particular resource and is denoted by a noun within braces (\{...\}) inside the endpoint. The query string is a list of key-value pairs placed at the end of a {\small URL}, which control the representation (e.g. format) of the resource. Typically, it is used to apply filters or define the structure of the returned representation. Resources within the {\small API}\ are represented in \underline{J}ava\underline{S}cript \underline{O}bject \underline{N}otation ({\small JSON}), and are referred to as objects. Once at an endpoint, the user must specify how to interact with the object. This is referred to as an action, and is an {\small HTTP} request method. The {\small API}\ supports the two most common {\small HTTP} request methods, \verb|GET| and \verb|POST|, where \verb|GET| fetches an object and \verb|POST| sends user defined data to the server. Therefore, users will interact with the {\small API}\ by performing actions (\verb|GET|, \verb|POST|) on endpoints (URLs) to retrieve objects (resources). \begin{table} \begin{tabular}{|c|l|p{1.6cm}|} \hline \textbf{Action} & \textbf{Endpoint} & \textbf{Object} \\ \hline \verb|POST| & \verb /plmf/prediction & task \\ \verb|POST| & \verb /mfd/prediction & task \\ \verb|GET| & \verb /prediction/result/{id} & status or prediction \\ \hline \end{tabular} \caption{A list of all endpoints in the {\small API}. Actions specify the supported {\small HTTP} request method, endpoints define the {\small URL}\ and objects are the returned resource. An endpoint with \{...\} denotes a path parameter.} \label{} \end{table} \section{API Structure} \begin{figure*} \centering \includegraphics[width=0.8\textwidth]{fig01.pdf} \caption{A flowchart illustrating the typical use-case of the API. First a {\small POSCAR}\ file is posted to the \texttt{<model>/prediction} endpoint to retrieve a task object. The \texttt{id} field of the task object is used to poll the \texttt{<model>/prediction/\{id\}} for the status of the prediction. If the status field of the response equals \texttt{SUCCESS} then the prediction object is returned.} \label{} \end{figure*} The overall structure of the {\small API}\ can be seen in Table \ref{}. All endpoints are located at the base {\small URL}\ \textbf{} and are organized by the model and the returned object. {\small AFLOW-ML}\ currently supports two models: \underline{m}olar \underline{f}raction \underline{d}escriptor (\verb|mfd|) \cite{} and \underline{p}roperty-\underline{l}abeled \underline{m}aterials \underline{f}ragments (\verb|plmf|) \cite{}. It is designed to be extensible, and additional models will be added in the future as they are developed. The \verb|mfd| model \cite{} predicts the material properties based on the chemical formula only: the vector of descriptors has 87 components, each component $b_{i}$ being the mole fraction of element $Z_{i}$ in the compound ($Z_{1}$ is H, $Z_{2}$ is He, etc.). The model is built with nonlinear support vector machines and a radial basis function kernel. The model is trained using a data set of 292 randomly selected compounds of the {\small ICSD}\ for which the vibrational properties are computed with {\small DFT} calculations. The \verb|plmf| model \cite{} represents each crystal structure as a colored graph, where the atomic vertices are decorated by the reference properties of the corresponding elemental species. Topological neighbors are determined using a Voronoi tessellation, and these nodes are connected to form the graph. The final feature vector for the ML model is obtained by partitioning the full graph into smaller subgraphs, termed fragments in analogy with the fragment-based descriptors used in cheminformatics \cite{Ruggiu_MI_2010}. All \verb|plmf| models are built with the \underline{G}radient \underline{B}oosting \underline{D}ecision \underline{T}ree ({\small GBDT}) method \cite{Friedman_AnnStat_2001}. Models for electronic and thermo-mechanical properties were trained on 26,674 and 2,829 materials entries from the {\small AFLOW}\ repository, respectively. All models are validated through $Y$-randomization (label scrambling) and five-fold cross validation. In general, {\small API}\ usage involves uploading a material structure to a \verb|POST| endpoint, \verb|<model>/prediction|, and retrieving a prediction object from a \verb|GET| endpoint, \verb|/prediction/result/{id}|, as shown in the flowchart in Figure \ref{}. \verb|POST| endpoints are responsible for the submission of a material structure for a prediction. In their request body, the file keyword is required. It must contain a string representation of the material's crystal structure, in {\small POSCAR 5}\ format (the lattice geometry and atomic position input file for version 5 of the {\small VASP}\ {\small DFT} package \cite{kresse_vasp, vasp}). Upon receiving a request, the response body returns a task object containing information about the submitted structure, which has the following format: \begin{verbatim} { "id": String, "model": String, "results_endpoint": String } \end{verbatim} When a material is posted to the {\small API}, a prediction task is created and added to a queue. Each task is assigned an identifier, the \verb|id| keyword, used to fetch the the prediction object at the endpoint referenced in the \verb|results_endpoint| keyword. This endpoint, \verb|/prediction/result/{id}|, supports the \verb|GET| method and requires the id as a path parameter. Depending on the status of the prediction task, the response body returns a status object or prediction object. When the task is still in the queue, the status object is returned: \begin{verbatim} { "status": String, "description": String } \end{verbatim} The status object details the state of the prediction task. The state is identified by the keyword \verb|status| which takes one of the following values: \verb|STARTED|, \verb|PENDING|, \verb|SUCCESS| and \verb|FAILURE|, while a description of each status type is given by the \verb|description| keyword. When attempting to retrieve the prediction object, it is best to poll the endpoint periodically to check the status. Tasks that are still within the queue are given the \verb|STARTED| or \verb|PENDING| status, while a completed task status will read \verb|SUCCESS|. In instances where the uploaded file is incorrectly formatted, a failed task will occur, status: \verb|FAILURE|. When the task is completed, the response contains the prediction object. The prediction object is an extension of the status object and contains different keywords depending on the model used. For \verb|plmf|, the prediction object, known as a \verb|plmf| prediction, takes the following form: \begin{verbatim} { "status": String, "description": String, "model": String, "citation": String, "ml_egap_type": String, "ml_egap": Number, "ml_energy_per_atom": Number, "ml_ael_bulk_modulus_vrh": Number, "ml_ael_shear_modulus_vrh": Number, "ml_agl_debye": Number, "ml_agl_heat_capacity_Cp_300K": Number, "ml_agl_heat_capacity_Cv_300K": Number, "ml_agl_heat_capacity_Cp_300K_per_atom": Number, "ml_agl_heat_capacity_Cv_300K_per_atom": Number, "ml_agl_thermal_conductivity_300K": Number, "ml_agl_thermal_expansion_300K": Number } \end{verbatim} \noindent The \verb|mfd| model returns an \verb|mfd| prediction object: \begin{verbatim} { "status": String, "description": String, "model": String, "citation": String, "ml_Cv": Number, "ml_Fvib": Number, "ml_Svib": Number } \end{verbatim} \noindent The details of each object and their keywords are found in the List of {\small API}\ Endpoints and Objects section. \begin{widetext} \section{Using the API} The process to retrieve a prediction is as follows: First, the contents of a {\small POSCAR 5}\ file, titled test.poscar, are uploaded to the submission endpoint. This can be achieved by using an {\small HTTP} library such as requests (Python) \cite{pythonRequests}, URLSession (iOS {\small SDK}), HttpURLConnection (Android {\small SDK}), Fetch (JavaScript) or using a command line tool such as wGET or cURL. For this example, cURL will be used. The contents of the {\small POSCAR}\ are posted to the submission endpoint as follows: \begin{verbatim} curl --data-urlencode \end{verbatim} \noindent where the \verb|--data-urlencode| flag handles encoding the contents of the {\small POSCAR}, located in the current directory, and associating it to the file keyword. Note that as mentioned previously, the \verb|POST| will return a {\small JSON}\ response including the task id. The task id is then used to poll the results endpoint: \begin{verbatim} curl } \end{verbatim} The status keyword is used as an indicator to determine if any additional polling is required. Depending on the status of the job, the endpoint will return either a task status object or a prediction object. If the status keyword's value is \verb|SUCCESS| then no additional polling is required, since the endpoint will return a prediction object, which is an extension of the task status object. \end{widetext} \section{List of API endpoints and objects} This section includes the details of each endpoint and object accessible in the {\small API}. Endpoint information contains the associated {\small HTTP} method, request parameters, request body data and response object for each endpoint. Object properties are listed along with their type and description. \subsection{Endpoints} \begin{itemize} \item \verb|POST plmf/prediction| \begin{itemize} \item \textit{Description.} Uploads the contents of a {\small POSCAR 5}\ to retrieve a prediction using the \verb|plmf| model. \item \textit{Query parameters.} \begin{itemize} \item \verb|file| (required) - The contents of the {\small POSCAR 5}. \end{itemize} \item \textit{Response format.} On success, the response header contains the {\small HTTP} status code 200 OK and the response body contains a task object, in {\small JSON}\ format. \item \textit{Example.} \begin{verbatim} curl v1.0/plmf/prediction --data-urlencode \end{verbatim} \end{itemize} \item \verb|POST mfd/prediction| \begin{itemize} \item \textit{Description.} Uploads the contents of a {\small POSCAR 5}\ to retrieve a prediction using the \verb|mfd| model. \item \textit{Query parameters.} \begin{itemize} \item \verb|file| (required) - The contents of the {\small POSCAR 5}. \end{itemize} \item \textit{Response format.} On success, the response header contains the {\small HTTP} status code 200 OK and the response body contains a task object, in {\small JSON}\ format. \item \textit{Example.} \begin{verbatim} curl v1.0/mfd/prediction --data-urlencode \end{verbatim} \end{itemize} \item \verb|GET /prediction/result/{id}| \begin{itemize} \item \textit{Description.} Fetches the status object or returns the prediction object when the task is completed. \item \textit{Path parameters.} \begin{itemize} \item \verb|id| (required) - The unique identifier retrieved from the task object on submission. \end{itemize} \item \textit{Query string arguments.} \begin{itemize} \item \verb|fields| - A comma separated list of the fields to include in the {\small JSON}\ response object. Note that specified fields will only affect the prediction object. \end{itemize} \item \textit{Response format.} On success, the response header contains the {\small HTTP} status code 200 OK. If the task is still pending, the response body contains a task status object in {\small JSON}\ format. Upon completion the response body contains a prediction object in {\small JSON}\ format. \end{itemize} \item \textit{Example.} \begin{verbatim} curl v1.0/prediction/result /59ea0f78-4868-4a1e-9a20-e7343f00907d \end{verbatim} \end{itemize} \subsection{Objects} \begin{itemize} \item Task \begin{itemize} \item \textit{Description.} Describes the task for a submitted prediction. Includes the unique identifier for the prediction and results endpoint. \item \textit{Keys.} \begin{itemize} \item \verb|id| \begin{itemize} \item \textit{Description.} The unique identifier of the prediction task. Used at the fetch prediction endpoint to retrieve the status of a prediction and return the results on completion. \item \textit{Type.} String. \end{itemize} \item \verb|results_endpoint| \begin{itemize} \item \textit{Description.} The path of the endpoint where the prediction task status and results are retrieved. \item \textit{Type.} String. \end{itemize} \item \verb|model| \begin{itemize} \item \textit{Description.} The name of the machine learning model used to generate the prediction. \item \textit{Type.} String. \end{itemize} \end{itemize} \item \textit{Example.} \begin{verbatim} { "id": "d29af704-06bf -4dc8-8928 -cd2c41aea454", "model": "plmf", "results_endpoint": "/prediction/result/ d29af704-06bf -4dc8-8928- cd2c41aea454" } \end{verbatim} \end{itemize} \item Status \begin{itemize} \item \textit{Description.} Provides the status of a (prediction) task. \item \textit{Keys.} \begin{itemize} \item \verb|status| \begin{itemize} \item \textit{Description.} The status of the task. Takes the following values: \verb|STARTED|, \verb|PENDING|, \verb|SUCCESS| and \verb|FAILURE|. When a task is added to the queue its status will read \verb|PENDING|. Once it reaches the top of the queue the status will read \verb|STARTED| and the prediction will run. If the prediction is successful the status will read \verb|SUCCESS|. \item \textit{Type.} String. \end{itemize} \item \verb|Description| \begin{itemize} \item \textit{Description.} Describes the status of the task. \item \textit{Type.} String. \end{itemize} \end{itemize} \item \textit{Example.} \begin{verbatim} { "status": "PENDING" "description": "The calculation is running" } \end{verbatim} \end{itemize} \item \verb|plmf| prediction \begin{itemize} \item \textit{Description.} The results of the prediction using the \verb|plmf| model. This is an extension of the task status object. \item \textit{Keys.} \begin{itemize} \item \verb|status| \begin{itemize} \item \textit{Description.} The status of the task. Takes the following values: \verb|STARTED|, \verb|PENDING|, \verb|SUCCESS| and \verb|FAILURE|. When a task is added to the queue its status will read \verb|PENDING|. Once it reaches the top of the queue the status will read \verb|STARTED| and the prediction will run. If the prediction is successful the status will read \verb|SUCCESS|. \item \textit{Type.} String. \end{itemize} \item \verb|description| \begin{itemize} \item \textit{Description.} Describes the status of the task. \item \textit{Type.} String. \end{itemize} \item \verb|model| \begin{itemize} \item \textit{Description.} The model used in the prediction. \item \textit{Type.} String. \end{itemize} \item \verb|citation| \begin{itemize} \item \textit{Description.} The DOI for the model's publication. \item \textit{Type.} String. \end{itemize} \item \verb|ml_egap_type| \begin{itemize} \item \textit{Description.} Specifies if the material is a metal or an insulator. Takes the following values: \verb|Metal|, \verb|Insulator|. \item \textit{Type.} String. \end{itemize} \item \verb|ml_egap| \begin{itemize} \item \textit{Description.} The electronic band gap. \item \textit{Units.} eV. \item \textit{Type.} Number. \end{itemize} \item \verb|ml_energy_per_atom| \begin{itemize} \item \textit{Description.} The energy per atom. \item \textit{Units.} eV/atom. \item \textit{Type.} Number. \end{itemize} \item \verb|ml_ael_bulk_modulus_vrh| \begin{itemize} \item \textit{Description.} The bulk modulus, trained using the Automatic Elasticity Library ({\small AEL}). \cite{} \item \textit{Units.} GPa. \item \textit{Type.} Number. \end{itemize} \item \verb|ml_ael_shear_modulus_vrh| \begin{itemize} \item \textit{Description.} The shear modulus, trained using {\small AEL}. \item \textit{Units.} GPa. \item \textit{Type.} Number. \end{itemize} \item \verb|ml_agl_debye| \begin{itemize} \item \textit{Description.} The Debye temperature, trained using the Automatic GIBBS Library ({\small AGL}) \cite{}. \item \textit{Units.} K. \item \textit{Type.} Number. \end{itemize} \item \verb|ml_agl_heat_capacity_Cp_300K| \begin{itemize} \item \textit{Description.} The heat capacity at 300K and constant pressure, trained using {\small AGL}. \item \textit{Units.} $k_\mathrm{B}$/cell. \item \textit{Type.} Number. \end{itemize} \item \verb|ml_agl_heat_capacity_Cp_300K_per_atom| \begin{itemize} \item \textit{Description.} The heat capacity per atom at 300K and constant pressure, trained using {\small AGL}. \item \textit{Units.} $k_\mathrm{B}$/atom. \item \textit{Type.} Number. \end{itemize} \item \verb|ml_agl_heat_capacity_Cv_300K| \begin{itemize} \item \textit{Description.} The heat capacity at 300K and constant volume, trained using {\small AGL}. \item \textit{Units.} $k_\mathrm{B}$/cell. \item \textit{Type.} Number. \end{itemize} \item \verb|ml_agl_heat_capacity_Cv_300K_per_atom| \begin{itemize} \item \textit{Description.} The heat capacity per atom at 300K and constant volume, trained using {\small AGL}. \item \textit{Units.} $k_\mathrm{B}$/atom. \item \textit{Type.} Number. \end{itemize} \item \verb|ml_agl_thermal_conductivity_300K| \begin{itemize} \item \textit{Description.} The lattice thermal conductivity at 300K, trained using {\small AGL}. \item \textit{Units.} W/(m K). \item \textit{Type.} Number. \end{itemize} \item \verb|ml_agl_thermal_expansion_300K| \begin{itemize} \item \textit{Description.} The thermal expansion coefficient at 300K, trained using {\small AGL}. \item \textit{Units.} K$^{-1}$. \item \textit{Type.} Number. \end{itemize} \end{itemize} \item \textit{Example.} \begin{verbatim} { "status": "SUCCESS" "description": "The job has completed.", "model": "plmf", "citation": "10.1038/ncomms15679", "ml_egap_type": "Insulator", "ml_egap": 0.923, "ml_energy_per_atom": -5.760, "ml_ael_bulk_modulus_vrh": 178.538, "ml_ael_shear_modulus_vrh": 140.121, "ml_agl_debye": 713.892, "ml_agl_heat_capacity_Cp_300K": 23.362, "ml_agl_heat_capacity_Cp_300K _per_atom": 2.333, "ml_agl_heat_capacity_Cv_300K": 22.625, "ml_agl_heat_capacity_Cv_300K _per_atom": 2.311, "ml_agl_thermal_conductivity _300K": 2.792, "ml_agl_thermal_expansion_300K": 6.093e-05 } \end{verbatim} \end{itemize} \item \verb|mfd| prediction \begin{itemize} \item \textit{Description.} The results of a prediction using the \verb|mfd| model. This is an extension of the task status object. \item \textit{Keys.} \begin{itemize} \item \verb|status| \begin{itemize} \item \textit{Description.} The status of the task. Takes the following values: \verb|STARTED|, \verb|PENDING|, \verb|SUCCESS| and \verb|FAILURE|. When a task is added to the queue its status will read \verb|PENDING|. Once it reaches the top of the queue the status will read \verb|STARTED| and the prediction will run. If the prediction is successful the status will read \verb|SUCCESS|. \item \textit{Type.} String. \end{itemize} \item \verb|description| \begin{itemize} \item \textit{Description.} Describes the status of the task. \item \textit{Type.} String. \end{itemize} \item \verb|model| \begin{itemize} \item \textit{Description.} The model used in the prediction. \item \textit{Type.} String. \end{itemize} \item \verb|citation| \begin{itemize} \item \textit{Description.} The DOI for the model's publication. \item \textit{Type.} String. \end{itemize} \item \verb|ml_Cv| \begin{itemize} \item \textit{Description.} The heat capacity at constant volume. \item \textit{Units.} meV/(atom K). \item \textit{Type.} Number. \end{itemize} \item \verb|ml_Fvib| \begin{itemize} \item \textit{Description.} The vibrational free energy per atom. \item \textit{Units.} meV/atom. \item \textit{Type.} Number. \end{itemize} \item \verb|ml_Svib| \begin{itemize} \item \textit{Description.} The vibrational entropy per atom. \item \textit{Units.} meV/(atom K). \item \textit{Type.} Number. \end{itemize} \end{itemize} \item \textit{Example.} \begin{verbatim} { "description": "The job has completed.", "model": "mfd", "citation": "10.1021/acs.chemmater.7b00789", "status": "SUCCESS", "ml_Cv": 0.221, "ml_Fvib": 21.188, "ml_Svib": 0.211 } \end{verbatim} \end{itemize} \end{itemize} \section{Python Client} \label{python-client} A Python client is available for the {\small AFLOW-ML}\ {\small REST}\ {\small API}\ that provides researchers and developers a means to integrate {\small AFLOW-ML}\ into their applications or workflows, such as {\small AFLOW$\pi$}\ \cite{}. The client includes the \verb|AFLOWmlAPI| class which provides all the functionality needed to interface with the {\small AFLOW-ML}\ {\small API}, and can be downloaded at \verb|. From the client, a prediction is retrieved by passing the contents of a {\small POSCAR}\ to the \verb|get_prediction| method. The \verb|AFLOWmlAPI| class can be incorporated into a Python framework using code similar to the example illustrated in Figure \ref{aflowmlclientcall}. \begin{figure}[h!] \centering \includegraphics[width=0.5\textwidth]{fig02.pdf} \caption{Example showing how to retrieve a prediction using the {\small AFLOW-ML}\ Python client.} \label{aflowmlclientcall} \end{figure} This method takes two arguments: \verb|poscar| and \verb|model|, where \verb|poscar| is a file object (reading from the file \verb|test.poscar| in the example in Figure \ref{aflowmlclientcall}) and \verb|model| is a string specifying the model to use (\verb|plmf| or \verb|mfd|). This method returns a Python dictionary, in which the keys and respective predicted values are model dependent. For a list of each prediction object's key and value pair, please refer to the previous section. The client's \verb|AFLOWmlAPI| class includes two additional methods, \verb|submit_job| and \verb|poll_job|, that provide more control when submitting a prediction, and which can be used in place of the \verb|get_prediction| method in the example shown in Figure \ref{aflowmlclientcall}. The \verb|submit_job| method targets the \verb|<model>/prediction| endpoint and returns the jobs task id. From this id, the job can be polled using the \verb|poll_job| method which will return a prediction object upon completion. These methods are ideal for cases where the user would prefer to postpone polling to a later time. \section{Command Line Interface} Upon installation, the Python {\small AFLOW-ML}\ client provides a \underline{c}ommand \underline{l}ine \underline{i}nterface ({\small CLI}) titled \verb|aflow-ml|. The {\small CLI}\ exposes all the functionality of the Python client and is targeted at users who are not familiar with Python or using {\small REST}\ APIs. To receive a prediction the path of the {\small POSCAR 5}\ file is passed to the {\small CLI}\ as a positional argument. Additionally, the model type is specified via the \verb|--model| flag: \begin{verbatim} aflow-ml test.poscar --model=plmf \end{verbatim} By default, the {\small CLI}\ will output the results to the terminal. The default functionality is modified by the use of additional flags. For instance, results can be saved to an out file by use of the \verb|-s| flag: \begin{verbatim} aflow-ml test.poscar --model=plmf -s \end{verbatim} \noindent where the predicted results are saved to a file titled \verb|prediction.txt|. Additional flags exist which provide various levels of customization such as specifying the predicted values to return or the format of the output. A list of each flag is found below. This list is also viewable from the {\small CLI}\ using the \verb|-h| or \verb|--help| flags. \subsection{{\small CLI}\ flags} \label{CLI-flags} \begin{itemize} \item Model \begin{itemize} \item \textit{Flag.} \verb|-m| or \verb|--model| \item \textit{Description.} (Required) Specifies the model to use in the prediction. \item \textit{Example.} \\ \verb|aflow-ml test.poscar --model=plmf| \end{itemize} \item Save \begin{itemize} \item \textit{Flag.} \verb|-s| or \verb|--save| \item \textit{Description.} Saves the prediction to a file. If the out file is not specified contents are saved to a file named \verb|prediction.txt|. \item \textit{Example.} \\ \verb|aflow-ml test.poscar -m plmf -s| \end{itemize} \item Outfile \begin{itemize} \item \textit{Flag.} \verb|--format| \item \textit{Description.} Specifies the path and name of the out file. \item \textit{Example.} \\ \verb|aflow-ml test.poscar -m plmf -s| \verb|--outfile=prediction.txt| \end{itemize} \item Format \begin{itemize} \item \textit{Flag.} \verb|--format| \item \textit{Description.} Specifies the format of the out file. Currently, text and {\small JSON}\ are supported. \item \textit{Example.} \\ \verb|aflow-ml test.poscar -m plmf -s| \verb| --format=JSON| \end{itemize} \item Fields \begin{itemize} \item \textit{Flag.} \verb|--fields| \item \textit{Description.} State the predicted fields to show in the output. Expects fields as a comma separated list. If the flag is not present, all fields will be shown. \item \textit{Example.} \\ \verb|aflow-ml test.poscar -m plmf -s| \verb| --fields=egap_ml,egap_type_ml| \end{itemize} \item Verbose \begin{itemize} \item \textit{Flag.} \verb|-v| or \verb|--verbose| \item \textit{Description.} Toggle verbose mode. When enabled the {\small CLI}\ will log the progress of the prediction. \item \textit{Example.} \\ \verb|aflow-ml test.poscar -m plmf -v| \end{itemize} \end{itemize} \section{Conclusion} {\small AFLOW-ML}\ enhances materials discovery by providing streamlined open access to predictive models. The {\small REST}\ {\small API}\ promotes resource sharing, where any application, workflow or website may leverage our models. Additionally, the Python client provides a closed solution, which requires little programming knowledge to get started. With this flexibility, {\small AFLOW-ML}\ presents the accessible option for machine learning in the materials design community. \section*{Acknowledgments} This work is partially supported by DOD-ONR (N00014-14-1-0526, N00014-16-1-2326, N00014-16-1-2583, N00014-17-1-2090), and by the Duke University---Center for Materials Genomics. C.O. acknowledges support from the National Science Foundation Graduate Research Fellowship under Grant No. DGF-1106401. A.T. and O.I. acknowledge support from DOD-ONR N00014-16-1-2311 and Eshelman Institute for Innovation Award. O.I. acknowledges Extreme Science and Engineering Discovery Environment (XSEDE) award DMR-110088, which is supported by National Science Foundation grant number ACI-1053575. S.C. acknowledges the Alexander von Humboldt Foundation for financial support. \section*{References} \small

\section{INTRODUCTION} The overwhelming majority of structural phase transitions in crystalline materials is associated with changes of the symmetry or modifications of atomic positions in the unit cell. {\color{black}Only very few systems are known so far to exhibit a symmetry-conserving or `isostructural' phase transition involving pronounced variations in the metric of the unit cell.} These so-called `type 0' transformations are first-order transitions for fundamental reasons \cite{Cowley_1976,Christy_1995}. The examples are driven by a diversity of mechanisms that change the internal state of the material without breaking crystalline symmetry, {\color{black}\textit{e.g.}, changes of coordination in complex framework lattices {\color{black}\cite{Saha_2013}} or electronic transitions, including valence transitions \cite{Larson_1948,Lanata13,Sarrao96}, metal-insulator transitions \cite{Akahama_1995}, or other changes of the electronic structure \cite{Rosner_2006}.} Transformations involving spin-order with magneto-elastic couplings can appear as isostructural transitions, although time-reversal symmetry is broken. {\color{black} If a transition is driven by magnetic ordering, then marked jumps of the lattice anisotropy are not expected in $3d$-electron systems, such as Fe$_{1+y}$Te. However, another type of transition with magneto-elastic coupling can be envisaged which is {\em driven by a symmetry-conserving instability of the lattice}. Such a `type 0' transition mode gives rise to magnetic ordering as secondary effect.} The resulting phase, an anisotropically deformed isostructural lattice with magnetic order as by-product, is reached through a strongly discontinuous first-order transformation process. In metallic systems, abrupt changes in the unit-cell dimensions can be associated with modifications of the Fermi-surface (FS) topology, resulting in an electronic topological transition (ETT) \cite{Novikov_1999,Liu_2010,Glazyrin_2013,Manjon_2013,Dubrovinsky_2015}. {\color{black} We note, the changes in the FS topology discussed here are not related with topologically protected surface states.} In this work, a thermal transition in Fe$_{1.08}$Te under pressure is identified as an isostructural transition that is marked by a pronounced change of the axis ratio $c/a$ in its tetragonal phase. {\color{black}This type-0 transformation is driven by lowering temperature, in contrast to the vast majority of type-0 transitions known in other materials which are normaly driven by a non-thermal control parameter, such as hydrostatic pressure. Furthermore, in Fe$_{1.08}$Te the type-0 transformation triggers a ferromagnetic (FM) ordering as a secondary effect. } Electronic structure calculations explain the transition by a drastic change of the FS topology, as suggested originally by Lifshitz \cite{Lifshitz_1960}. {\color{black} The non-superconducting parent compound of the Fe-based superconductors, Fe$_{1+y}$Te, has prompted the interests of the condensed matter community at large \cite{Deguchi_2012}.} Bulk Fe$_{1+y}$Te exhibits a plethora of structural and magnetic phase transitions as a function of excess Fe content ($y$), temperature ($T$) \cite{Rodriguez_2011,Zaliznyak_2012,Koz_2013,Rodriguez_2013,Machida_2013,Cherian_2014}, and pressure ($p$) \cite{Okada_2009,Takahashi_2010,Koz_2012,Bendele_2013,Jorgensen_2013,Jorgensen_2015}. At low concentrations of $y\leq 0.11$, Fe$_{1+y}$Te undergoes a structural and magnetic phase transition from a high-$T$ tetragonal and paramagnetic (PM) semimetal to a low-$T$ monoclinic antiferromagnet (AFM). For $y \geq 0.12$, an orthorhombic incommensurate AFM state is realized at low temperatures \cite{Bao_2009,Rodriguez_2011,Koz_2013,Materne_2015}. Similarly, applying pressure on a sample drives the system through a series of phase transitions that closely resemble those induced by excess Fe content \cite{Koz_2012}. One such transition, being completely unanticipated, is of particular interest: a pressure-induced FM phase transition at low temperatures {\color{black} \cite{Bendele_2013,Jorgensen_2016,Fedorchenko_2011,Monni_2013,Ciechan_2014}.} This phase has not been observed at ambient pressure for samples of any level of Fe excess. {\color{black} Details on the preparation of polycrystalline Fe$_{1.08}$Te samples by a solid-state reaction and on the electrical resistivity, magnetization, and x-ray diffraction (XRD) experiments under pressure as well as on the density-functional theory based calculations can be found in the Supplemental Material \cite{suppl}.} The temperature dependence of the electrical resistivity $\rho$ of Fe$_{1.08}$Te taken in cooling and heating cycles for selected pressures $1.53 {\rm ~GPa} \leq p \leq 2.72$ GPa is displayed in Fig.\,\ref{data}a. At high temperatures, above the magnetic ordering, we observe a characteristic change in $\rho$ around 2.2~GPa. While $\rho(T)$ increases slightly upon lowering temperature at $p\leq 2.20$~GPa, indicating a semi-conducting-like behavior, $\rho(T)$ decreases at higher pressures as expected in a metal. \begin{figure}[b!] \centering \includegraphics[width=.9\linewidth]{Fig1} \caption{(a) $\rho(T)$ data of Fe$_{1.08}$Te for selected pressures collected upon cooling and heating as marked by arrows. $T_N$, $T_{C1}$, and $T_{C2}$ are indicated. (b) $M(H)$ hysteresis loops of Fe$_{1.08}$Te at 5~K and at different pressures. Inset: field-cooled $M(H)$ data for selected pressures.} \label{data} \end{figure} At 1.53 GPa, a small step-like feature in $\rho(T)$ around $T_N=65$~K indicates the phase transition from a tetragonal PM to an orthorhombic incommensurate AFM phase \cite{Jorgensen_2013}. This step-like feature is in contrast to the continuous anomaly observed at lower pressures. It does not show any thermal hysteresis, pointing at a second-order character of the phase transition. Upon further cooling, metallic behavior is observed down to 7~K, below which $\rho(T)$ exhibits a small upturn. The latter feature is less pronounced at higher pressures, but remains visible up to 2.20~GPa. The height of the step-like feature marking $T_{N}$ is suppressed upon increasing pressures. It is still clearly visible at 1.95~GPa, where a thermal hysteresis in $\rho(T)$, centered around 30~K and extending over a wide temperature range, is detected well below $T_{N} = 63$~K. The hysteresis is related to a first-order phase transition at $T_{C1}\approx30$~K. Measurements of $\rho(T)$ in different magnetic fields (see Supplemental Material \cite{suppl}) give a first hint that $T_{C1}$ is associated with a phase transition from the incommensurate AFM phase to a ferromagnetically ordered state at lower temperatures. A very weak anomaly in $\rho(T)$ at $T_{N}$ still exists at 2.20~GPa. Increasing pressure drives $T_{C1}(p)$ rapidly toward higher temperatures. Concurrently, the temperature interval of the thermal hysteresis decreases, but extends almost up to $T_{N}$. This observation suggests a competition between the FM and incommensurate AFM ordering. For $p \gtrsim 2.4$~GPa, only a broad feature with a large thermal hysteresis remains in $\rho(T)$ signaling a first-order phase transition from the PM to the FM state at $T_{C2}$ [$T_{C2}$ is defined by the kink in the warming curve]. Furthermore, the resistivity decreases monotonously toward the lowest temperatures. $T_{C2}(p)$ increases strongly with increasing pressure accompanied by an increase in the width of the thermal hysteresis. The positive field dependence of the anomaly at $T_{C2}$ is consistent with a FM nature of the ordered state (see Supplemental Material \cite{suppl}). \begin{figure}[t!] \centering \includegraphics[width=.9\linewidth]{Fig2} \caption{(a) $T-p$ phase diagram for Fe$_{1.08}$Te. $\blacktriangle$, $\bigstar$, and $\bullet$ mark the transition temperatures extracted from $\rho(T)$ and $\blacksquare$ that from $M(T)$. The letters T and O represent tetragonal and orthorhombic phases according to XRD measurements. The single line and the double lines indicate second-order and first-order phase transitions, respectively. The inset displays the $T-p$ phase diagram at zero-field (solid symbols, dashed lines) and at 9~T (open symbols, dotted lines). (b) and (c) Unit-cell volume $V$ (left axis) and $c/a^*$ ratio (right axis) at 40~K as function of pressure, and at $\sim2.9$~GPa as function of temperature, respectively.} \label{phasediagram} \end{figure} At 5~K, the hysteresis in the magnetization $M(H)$ loops confirms the FM nature of the high-pressure phase (see Fig.\,\ref{data}b). No signal can be resolved in the AFM phases at low pressures. \textcolor{black}{Due to the large uncertainty in the determination of the sample mass, we cannot provide absolute values of the magnetization. However, this does not affect the relative changes between different pressures. Our finding is further supported by neutron diffraction data on Fe$_{1.141}$Te, which indicate that all Fe moments are oriented along the $c$-axis in the pressure-induced FM phase \cite{Jorgensen_2016}.} The saturated high-field magnetization increases with increasing pressure, signifying a stabilization of ferromagnetism under pressure. $M(T)$ taken upon cooling in 10~mT shows a strong increase toward low temperatures upon entering the FM phase (inset of Fig.\,\ref{data}b). The extracted transition temperatures $T_{C2}(p)$ are in good agreement with the ones obtained from $\rho(T)$. The $T-p$ phase diagram in Fig.\,\ref{phasediagram}a summarizes the results of our $\rho(T)$ and $M(T)$ investigations and, additionally, includes the XRD data taken on Fe$_{1.08}$Te. One second-order ($T_N$) and two first-order ($T_{C1}$ and $T_{C2}$) phase-transition lines meet at a multicritical point at $p_{\rm MP} \approx 2.4$~GPa and $T_{\rm MP} \approx 58$ K. A field of 9~T suppresses $T_{N}$ only weakly, but strongly enhances $T_{C1}$ and $T_{C2}$, except for the very vicinity of the multicritical point (Fig.\,\ref{phasediagram}a, inset). The strong increase in $T_{C1}$ and $T_{C2}$ is expected for a ferromagnet. We note that while at 1.75~GPa in zero field no $T_{C1}$ anomaly in $\rho(T)$ is resolved down to 1.8~K, a field of 9~T induces FM ordering below $T_{C1}^{9{\rm T}}\approx30$~K. By entering the FM phase from the incommensurate AFM upon increasing pressure the lattice symmetry changes from orthorhombic to tetragonal. The XRD data taken at 40~K upon increasing pressure indicate a broad two-phase region dominated by the tetragonal phase in a pressure window between 2.1 and 2.5~GPa consistent with the large hysteresis observed in the electrical resistivity. The transition from the orthorhombic-incommensurate AFM to the tetragonal-FM phase at $T_{C1}$ is accompanied by a reduction of the unit-cell volume $V$ and an increase in the ratio of the $c$ and $a^*$-axis lattice parameters [for orthorhombic and monoclinic symmetries $a^*$ = $\frac{1}{2}(a+b)$]. At $T_{C1} = 40$~K we find a marked change in both $V$ and $c/a^*$ by $-0.6\%$ and $2.1\%$, respectively, upon the transformation from the orthorhombic AFM to the tetragonal FM state (see Fig.\,\ref{phasediagram}b). This transformation connects two phases with different and unrelated magnetic symmetries, which explains why a classical first-order-type behavior with hysteresis is observed. At $T_{C2}$, in contrast, the {\color{black} transformation} from the high-temperature PM to the low-temperature FM state also takes place as a first-order process, evidenced by the large hysteresis observed in the $\rho(T)$ data, but without change of the tetragonal lattice symmetry. At this symmetry-conserving phase transition at $T_{C2}$, again marked variations in $V$ and $c/a$ are seen, \textit{e.g.}\ $-0.8\%$ and $1.5\%$ at $\sim2.9$~GPa, respectively (see Fig.\,\ref{phasediagram}c). This testifies that {\color{black} this phase transition is not driven by magnetic ordering}. Rather, the marked change of the lattice aspect ratio and its volume suggests that the internal electronic structure of the material is fundamentally transformed when entering the FM state. {\color{black} The reason for the transformation, therefore, is not a symmetry-breaking in the spin-structure, but an isostructural instability of the tetragonal lattice.} We note that the changes in $V$ and $c/a$ upon entering the tetragonal-FM phase are similar, independent of the path, \textit{i.e.} upon entering from the orthorhombic-incommensurate AFM or from the tetragonal-PM phase. In the following we turn to electronic-structure calculations to track the FS evolution across the symmetry-conserving phase transition. \begin{figure}[t!] \centering \includegraphics[width=.9\linewidth]{Fig3} \caption{Energy and Fe moment (calculated and experimental) as a function of the $c/a$ ratio for stoichiometric FeTe (Fe$_{1.08}$Te). In the experimental data, $p$ is an implicit parameter} \label{E_mom} \end{figure} \begin{figure*}[t!] \centering \includegraphics[width=1\linewidth]{Fig4} \caption{Topological changes of the FS sheets for two closely spaced $c/a$ ratios in FeTe. The left four FS sheets correspond to the spin-up channel (green) and the right five ones to the spin-down channel (blue). Upon increasing the $c/a$ ratio by only 2.5\%, four FS sheets (two in each spin channel) disappear, illustrative of an ETT.} \label{fermi} \end{figure*} For simplicity, we have limited our band-structure analysis to the tetragonal phase and stoichiometric composition. The energetics and magnetic moments are analyzed as a function of the $c/a$ ratio, whilst constraining the volume to that measured for 2.9~GPa (see previous paragraph). Firstly, we observe a salient feature in the total energy with increasing $c/a$ (see Fig.\,\ref{E_mom}): a non-monotonic flattening of the curve around $c/a \approx 1.6$, resulting in a shallow minimum. Upon further increase in $c/a$, the total energy then continues to decrease monotonically. Concomitantly, around $c/a \approx 1.6$, the Fe moment shows a sharp increase, strengthening from $0.4~\mu_{\mathrm{B}}$ to more than $1.5~\mu_{\mathrm{B}}$. Although the change in the total energy is subtle, the jump in the Fe moment is quite abrupt, and can be construed as being emblematic of the first-order character of the symmetry-conserving phase transition, consistent with our experiments. Secondly, we monitored the change in the electronic structure of FeTe. Collected in Fig.\,\ref{fermi} are the FS's for two closely spaced $c/a$ ratios adjoining the jump in Fig.\,\ref{E_mom}. Upon increasing $c/a$ from 1.59 by 2.5\%, which is comparable to the experimentally observed increase at the symmetry-conserving transition, four FS sheets (two in each spin channel) rapidly reduce in volume and subsequently disappear. This is illustrative of the anticipated ETT, in the same vein as a Lifshitz transition \cite{Lifshitz_1960}. We note that we find a similar change in $c/a^*$ at the orthorhombic-incommensurate AFM to tetragonal-FM transition upon varying pressure. Corresponding to the $c/a$ variation, a pronounced narrow peak in the density of states (see Supplemental Material \cite{suppl}) jumps from below the Fermi level to above, thereby shifting the Fermi energy away from the van Hove-like singularity. The loss of charge carriers due to the disappearing FS sheets is more than compensated by the remaining sheets which increase significantly in volume, consistent with the electrical resistivity becoming more metallic-like at higher pressures. Hence, the combined changes in the topology of the FS, concomitant with the depopulation of the narrow density of states peak, and the sharp increase of the Fe moment, demonstrate a Lifshitz-ETT in the spin-split bandstructure of Fe$_{1.08}$Te \cite{Lifshitz_1960,Rosner_2006}{\color{black}. Lifshitz already demonstrated that ETTs can drive type-0 transitions \cite{Lifshitz_1960}, and it is known that electron-phonon coupling can exacerbate anisotropic inharmonic response and instability of the lattice \cite{dagens79}. However, the remarkable change of the $c/a$ ratio in Fe$_{1.08}$Te is unconventional and suggests that particular electron-phonon interactions exist in this system, so that the ETT results in an anisotropic lattice distortion. As we are observing a transition at finite temperature, its microscopic understanding should rely on the coupling of thermally smeared electronic states near singular Fermi-surface pieces to thermally excited acoustic phonons. Theoretically, such a strongly anisotropic effect in a $3d$-electron metallic system has not been anticipated, see review Ref.~\cite{blanter94}.} {\color{black} Phenomenologically, the first-order thermal} PM-FM transition can be understood by considering the lattice strain, i.e. change of the $c/a$-ratio, as order-parameter, $\eta=e_{zz}$. It transforms as identity representation of the lattice space-group. Coupled to the magnetization $m$, the Landau potential for this transition reads \(f=\alpha\,m^2+\beta\,m^4 + A\,\eta^2+B\,\eta^3+C\,\eta^4+d\,\eta\,m^2\), where the last (attractive) term, $d<0$, encodes the unconventionally strong magneto-elastic coupling that relies on the ETT in the spin-split bandstructure. A first-order transition in the primary order parameter $\eta$ is prompted by the cubic term $B$, and can lead to a jump-like onset of magnetic order, if $|d|$ is comparable to $\alpha > 0$, as in similar systems with coupled ordering modes \cite{holakowsky1973,Salje_2011}. {\color{black} This fundamental thermodynamic model describes all qualitative features and the strengths of the observed thermal transition in Fe$_{1.08}$Te, as a `type 0' process driven by the lattice instability $\eta$. In contrast, a magnetically driven transition would require to reach $\alpha < 0$ which entails only a weak response of the lattice (details in Supplemental Material \cite{suppl}).} Another important structural aspect aiding the ETT induced FM state in Fe$_{1.08}$Te is the anisotropic {\it enhancement} of $c/a^{(*)}$ under pressure. An anomalous expansion of the lattice parameters under pressure has been observed in various members of the Fe-based superconductors, but they tend to reduce relatively quickly afterwards. They include the realization of a collapsed state, wherein the $c/a$ ratio is reduced considerably \cite{Kasinathan_2009,Kasinathan_2011,Chu_2009,Gati_2012,Guterding_2015}. The formation of interlayer bonds is seen as a driving force for the collapsed state, which in turn becomes nonmagnetic \cite{Kasinathan_2011}. In contrast, in Fe$_{1.08}$Te $c/a$ is {\it enhanced} in the FM state upon entering from the orthorhombic-incommensurate AFM phase or upon cooling from the tetragonal-PM one (Fig.\,\ref{phasediagram}). As seen in Fig.\,\ref{E_mom}, such an increase in $c/a$ is decisive in stabilizing the FM moment in Fe$_{1.08}$Te. The formation of different magnetic structures due to an ETT is commonly observed in rare-earth containing metals (heavy-fermion systems) \cite{Chu_1970,Andrianov_2014}, wherein the localized moments of the rare-earth ions are interacting via the Ruderman-Kittel-Kasuya-Yosida (RKKY) mechanism \cite{Jensen_1991}. Small changes in the FS topology of the conduction electrons which regulate the exchange mechanism can result in drastic changes of the magnetic properties. Dynamical mean-field theory studies on FeTe reported a mass enhancement of more than seven for the $3d_{xy}$ orbital, which is typical for heavy-fermion systems, and surmised that FeTe can promote an orbital-selective localization \cite{Yin_2010}. Another study based on semi-phenomenological models with coupled localized-itinerant moments in Fe$_{1+y}$Te, discussed the appearance of new magnetic phases by the $y$-dependent RKKY part of the interaction \cite{Ducatman_2014}. In our work, projecting out the $3d$ orbital character, we notice the FS sheets to be comprised predominantly by the itinerant $3d_{xz/yz}$ and $3d_{x^{2}-y^{2}}$ orbitals (see Supplemental Material \cite{suppl}) while the $3d_{xy}$ orbital remains localized and away from the Fermi level. Consequently, the calculated rapid change in the FS topology, the sudden increase of Fe moment and the stabilization of the FM phase in Fe$_{1.08}$Te bear some similarity to heavy-fermion materials in that the moments on the localized $3d_{xy}$ orbital are mediated by the other itinerant electrons. Thus, Fe$_{1.08}$Te presents a rare example of an ETT-induced FM transition under pressure in a $3d$ system, wherein only a handful of examples (mostly containing rare-earth ions) exist in the literature \cite{Kadomatsu_1986,Hauser_1994,Ishizuka_1997,Nakamura_2003,Tateiwa_2014}. Our results identify a topological transition of the Fermi surface in the (spin-split) electronic band-structure of Fe$_{1+y}$Te that leads to a coupled instability of the lattice geometry with magnetic ordering and explains its unconventional magneto-structural 'type-0' transformation in the thermal energy range at a pressure of a few GPa. The results demonstrate the subtle competition between various electronic degrees of freedom and the coupling to (the geometry of) the lattice as the relevant parameter to explain the existence of the many different phases in this Fe-chalcogenide. D.K. acknowledges funding by the DFG within FOR 1346. U.S., S.W., and S.R. acknowledge support by the DFG within SPP 1458.

\section{Introduction} \begin{figure} \centering \includegraphics[width=1.0\columnwidth]{computer-assisted-collaboration.pdf} \caption{RoboJam takes the role of a remote collaborator, providing musical responses to a user's touchscreen improvisations on demand. In this way, the user can gain feedback from a simulated collaboration, while waiting for other users of to respond.} \label{} \end{figure} New interfaces for musical expression (NIMEs) are often aimed at casual or novice musicians, this is especially the case for touchscreen instruments that can be deployed on popular mobile devices. While these interfaces often emphasise a solo musical production paradigm, the concept of ensemble, or collaborative, performance has often been under-explored. That mobile devices are typically used while alone, in transit, or in public, significantly limits opportunities for touchscreen musicians to jam with others. This limitation restricts the user's ability to gain feedback, respond to others' ideas, and to refine their own music-making through collaboration. In this research, we propose a machine-learning system for generating music that can be used to assist the user by composing responses in near real-time, thus emulating the experience of collaborating with another user. Our system, called \emph{RoboJam}, has been integrated into an existing touchscreen music app. In this app, the user is able to compose short musical pieces using the touchscreen, share them with friends, or collaborate by ``replying'' to other performances. The user taps, swipes and swirls in a free-form manner on the touchscreen and these interactions are translated into musical sounds. Several sound schemes are possible, with the touches mapped to different kinds of synth and instrumental sounds. Recordings, which are limited to 5 seconds in length, are uploaded automatically to a cloud server where they can be played back by other users. Users can collaborate by replying to each others' performances, forming more complex pieces where replies are composed as sonic layers with the original performance. In this way, the collaborative interaction mirrors a call-and-response style of performance; the first performer plays the call, and subsequent performers reply with responses. This process of musical collaboration can potentially offer important feedback to the original performer. Hearing their own performances layered with other responses might change how they approach further improvisations and they might gain new ideas from others' responses. In a co-located situation, as shown in the left hand side of Figure \ref{}, this feedback loop could be tight and allow rapid improvement and increased engagement with the touchscreen interface. RoboJam is an agent designed to emulate this call-and-response interaction in situation when other users aren't present or available, this process is shown in the right side of Figure \ref{}. The user can call on RoboJam to provide a response to their performances whenever--and however many times--it is required. RoboJam is designed to predict what a touchscreen performer might do next, having heard the first performance. The advantage of this design is that the user can hear their own performance with multiple accompanying replies, thus the user can practice and refine their own performance ideas in context. The machine-learning system behind RoboJam uses an artificial neural network (ANN) to generate responses on demand. Similar to other music generation systems, RoboJam uses a recurrent neural network (RNN) with long short-term memory (LSTM) cells to predict a temporal sequence of discrete events. In this sequence-predicting configuration, the input to the network is the current event, and the output is the next predicted event; thus, the network can predict a sequence one event at a time. When used to predict responses, our RoboJam network first ``listens'' to the user's performance, i.e., the performance is propagated through the network (ignoring predictions) to condition the LSTM memory state. A new performance, 5 seconds in length, is then predicted by the network to form the response. A key point of difference for RoboJam is that it models the input data, a stream of touch-screen events, rather than the musical data, frequently MIDI-note pitches, modelled by other ANNs for generating music. This is made possible by the novel application of a mixture density network (MDN) to creative touchscreen interactions in RoboJam. In this research we show how this system models musical control data, and discuss evaluations from the perspectives of model validation, performance generation and user experience. The results support RoboJam's ability to generate responses that are related to the call performance and improve their sound as well as to enhance the human performer's experience. The structure of this paper is as follows: In Section \ref{} we will discuss related work in the use of ANNs to generate music and in co-creative musical interfaces. In Section \ref{} we will discuss the neural network design and training for RoboJam. Section \ref{} will describe how this network is integrated into the touchscreen app and its interaction design. Evaluations will be discussed in Section \ref{}. \section{Related Work}\label{} \subsection{RNN Music Generation}\label{} ANNs have long been used to model and generate music. RNNs, able to learn temporal information in between computations, are particularly applicable to modelling musical sequences, where upcoming events strongly depend on those that have previously occurred. In general, these networks generate musical event sequences in a one-by-one manner; the input is the present note, and the output predicts the next note in the sequence. Mozer's CONCERT system~\cite{} was an early attempt to compose music using an RNN; this work emphasised the advantage in learning long range dependencies in music without handling extremely large transition tables in, for example, a Markov model. Eck and Schmidhuber contributed further work in this area, using an RNN to generate (potentially) endless blues music~\cite{}. This project notably used long short-term memory (LSTM) units \cite{} to help alleviate the problem of vanishing or exploding gradients when training RNNs. In recent times, the use of GPU computation and large datasets have enabled a variety of creative applications for RNNs including the popular CharRNN model~\cite{} as well as in music generation. Sturm et al. focused on a textual representation of music in their FolkRNN project \cite{}. This system was trained on a large dataset of folk melodies available online in the plain-text ``ABC'' format. Hadjeres et al. created an RNN generator of J.S. Bach-styled chorales that can be steered towards particular melodies~\cite{}. Colombo et al. used an RNN to predict pitch and rhythm separately~\cite{}. Malik and Ek used an RNN trained on MIDI-recordings of piano performances to augment existing MIDI compositions with dynamic (volume) instructions~\cite{}. Hutchings and McCormack used an RNN to generate harmonic sequences in an agent-based jazz improvisation system~\cite{}. Google's Magenta project\footnote{\url{}} have released various trained RNN models for music generation and made steps towards integrating them with popular music production software. \subsection{Mixture Density Networks}\label{} A commonality of the above models is that they predict sequences of events from a finite set of symbols (e.g., one of the 128 integer MIDI pitches). Mapping the output of an RNN to a prediction from a finite number of classes can be effectively managed using the softmax function. In our application, we wish to predict locations and timings of interactions with a touchscreen, which have real---not finite--values, so a softmax output cannot be used without a significant loss of precision. A similar task, generating simple line drawings, was recently tackled by the Magenta group resulting in SketchRNN~\cite{}. This system is able to generate drawings which, unlike pixel representations of images, are constructed from sequences of pen movements in a 2D plane. The approach used here, and in previous experiments in generating handwriting~\cite{} was to replace the categorical softmax model at the outputs of the network with a mixture model, which provide more flexible predictions. Mixture density networks (MDNs), where the output of a neural network is used as the parameters of a mixture model, was first explored by Bishop as a way of learning to predict multimodal problems that do not fit a normal distribution and are poorly modelled by a least-squares approach to training the output of an ANN~\cite{}. The probability density function (PDF) of a mixture of normal distributions can be calculated by taking the linear combination of a number of normal distributions with weighting coefficients. Bishop observed that an ANN could be trained to produce these coefficients ($\pi_i$), and the centres ($\mu_i$) and variances ($\sigma_i$) for each component of the mixture. Thus a loss function for the network could be given by the negative logarithm of the following likelihood function, given $m$ mixtures, of target $\mathbf{t}$ occurring in the distribution generated by input data $\mathbf{x}$: \begin{equation} \mathcal{L} = \sum_{i=1}^m\pi_i(\mathbf{x})\mathcal{N}\bigl(\mu_i(\mathbf{x}), \sigma_i^2(\mathbf{x}); \mathbf{t} \bigr) \end{equation} After training, the mixture model can then be sampled to generate output data. While Bishop used an MDN with a non-recurrent network, it has also been applied with RNNs as in Graves' experiments~\cite{}, and SketchRNN~\cite{} mentioned above. Such a configuration, where mixture density layers follow an RNN, could be termed an MDRNN. While MDNs have advantages, they also suffer from difficulties in training~\cite{}, and are not widely used. To our knowledge, an MDN design has not previously been applied to a musical task. \subsection{Predictive Musical Interfaces}\label{} There are many examples of generative music systems that are designed to collaborate with human performers. These are often used in a soloist-ensemble configuration where a human performer is ``accompanied'' by artificially generated musicians. GenJam~\cite{} is a generative backing band that knows how to play jazz standards. The Reflexive Looper~\cite{} can arrange short recordings of the performer's own sounds into parts fitting a given song structure. Systems that \emph{learn} some aspect of the user's style and can interact with them in performance are more relevant to this work. The \emph{Continuator}~\cite{} did this very effectively using Markov models, and defined an interaction model of continuing to play when the user stops. More recently, it has become possible to apply RNN music models in a real-time system as shown in Magenta AI Duet~\cite{}. In this instance, the model does not learn from the user in the sense of updating ANN weights; rather, the user's performance can be used to condition the LSTM cells' memory state which then governs the style of generated notes. Due to this conditioning and generation process, AI Duet can be directed towards different styles after having learned a more general model of music. \section{Neural Network Design}\label{} RoboJam's artificial neural network is a model of musical touchscreen interactions. These interactions consist of touch points in a 2D plane as well as the location of these interactions in time. In our touchscreen music app, the user interacts with an area of the touchscreen in a free-form manner to perform music, and is not constrained by UI elements such as virtual faders or buttons. This means that touchscreen interaction data can be considered as a record of the user's control gestures: interactions with the touchscreen that produce a musical result~\cite{}. Thus, by modelling and generating this control data, RoboJam can produce new performances. The focus on control gesture data and temporal locations means that this model differs from most music generation RNNs. FolkRNN~\cite{} (among others) predicts symbols from a finite dictionary defined before training. This means that the output of the network can be generated by sampling from a categorical distribution given by a softmax layer. As our network must predict real-valued locations in 2D space, a softmax layer cannot be used without quantising the output to an unacceptably low resolution. Other musical generation systems such as DeepBach~\cite{} generate notes on a predefined semiquaver pulse. Our network predicts temporal locations for each touch interaction, again as a positive, real-valued number of seconds. Generating real-valued control data provides precision, but has a cost in terms of the amount of high-level information (e.g., harmony) that can be learned. This approach would not be appropriate in all musical systems, but allows RoboJam to focus on free-form touch expression and means that it can be used with many different musical mappings, and even in different creative apps. \subsection{Dataset}\label{} Training data for the network consisted of musical touchscreen improvisations. This data was derived from a corpus of performances by touchscreen ensembles~\cite{metatone_analysis_data}. The dataset included 163 collaborative sessions corresponding to 20 hours of performance and 4.3M individual touch interaction events--the musical control data for these performances. It should be noted that this dataset was collected in a variety of apps. Each of these apps mapped free-form touchscreen interactions directly to sound so this corpus is an appropriate analogue for the host app for RoboJam. The dataset was prepared first by extracting each performer's contribution from each session. The touch locations of each record were transformed from the original recording resolution to be in $[0,1] \times [0,1]$. Absolute times were changed to time deltas with a maximum value of 5 seconds. This means that each touch event became a vector $(x, y, dt) \in [0,1] \times [0,1] \times [0,5]$. For training, the performances were placed end-to-end to form one long sequence of touch events. Overlapping examples of 256 events in length were extracted from this sequence resulting in almost 4.3M training examples. \subsection{Implementation} \begin{figure} \centering \includegraphics[width=1.0\columnwidth]{robojam-mdrnn-design.pdf} \caption{RoboJam uses the output of an RNN to parameterise mixture models that handle touchscreen location and the timing of events. Two models are used: a mixture of 2D normal distributions to predict location of touches, and a mixture of 1D normals to predict time deltas in between events.} \label{} \end{figure} To model this data, we use an MDRNN design, inspired by SketchRNN~\cite{} and Graves' handwriting generation network~\cite{} (discussed in Section \ref{}). Both of these systems are intended to generate sequences of 2D surface interactions, very similar to our touchscreen data; however, sketches and handwriting do not have a significant temporal component except for the ordering of strokes. Both of these previous ANNs used the outputs of LSTM cells as the parameters of a mixture of probability models. In both cases, a mixture of bi-dimensional normal distributions were used to predict the location of the next pen movement in 2D space, the number of normal distributions in a mixture was a hyperparameter in training these networks. In our case, location of touches is accompanied by a positive time value, indicating the number of seconds in the future that the interaction should occur. The same 2D mixture distribution as in SketchRNN is used to predict spatial location of touches while a separate mixture of 1D normal distributions is used to predict the temporal location of the touch. Our MDRNN is implemented in TensorFlow, and the source code is available online\footnote{RoboJam's source code is available at \url{}}. An overview of our MDRNN design is shown in Figure \ref{}. Input data are vectors $[x,y,dt] \in \mathbb{R}^3$ corresponding to absolute location in the plane $[0,1] \times [0,1]$ and value in $[0,5]$ indicating the number of seconds since the last interaction. The input vector flows through three fully-connected layers of LSTM cells. The output of the last LSTM layer is projected using a fully connected layer to a vector $p$ that is used to generate parameters for our two mixture models that predict time and space values respectively. The time model is a mixture of $M$ 1D normal distributions, parameterised as given in Section \ref{} by a mean ($\mu_t$), standard deviation ($\sigma_t$), and a coefficient $\pi_t$ for each mixture component. The space model is a mixture of $M$ 2D normal distributions, each having two means ($\mu_x, \mu_y$), standard deviations ($\sigma_x, \sigma_y$), one correlation ($\rho$) and coefficient ($\pi_{xy}$). So, $p$ has size $3M + 6M$. The components of $p$ are transformed as recommended by Brando~\cite{}, Graves~\cite{} and Bishop~\cite{}. This ensures that the standard deviations are greater than zero, correlations are in $(-1,1)$, and that the coefficients for each mixture sum to one. In operation, a sample is drawn from these two mixture models to generate the next vector: $[x_{t+1}, y_{t+1}, dt_{t+1}]$. In training, we measure the likelihood of the next (known) vector occurring in the generated mixture models. The loss function has two components that are added --- the average negative log likelihood over the batch for the time model and for the space model: \begin{equation} \mathcal{L}_{space} = -\frac{1}{N} \sum_{i=1}^N \log \sum_{j=1}^{M} \bigl(\pi_{xy,j}\mathcal{N}_{2D}(\mu_{x,j},\mu_{y,j},\sigma_{x,j},\sigma_{y,j},\rho_{j}; x_i,y_i)\bigr) \end{equation} \begin{equation} \mathcal{L}_{time} = -\frac{1}{N} \sum_{i=1}^N \log \sum_{j=1}^M\pi_{dt,j}\mathcal{N}_{1D}(\mu_{dt,j}, \sigma_{dt,j}; dt_i) \end{equation} \begin{equation} \mathcal{L}_{tot} = \mathcal{L}_{space} + \mathcal{L}_{time} \end{equation} Where $\mathcal{N}_{2D}$ is as given in equations 24 and 25 of \cite{}, and $\mathcal{N}_{1D}$ is as given in equation 23 of \cite{}. Code for $\mathcal{N}_{2D}$ follows Ha and Eck's work~\cite{} and a similar TensorFlow implementation was developed for $\mathcal{N}_{1D}$. \subsection{Training}\label{} As discussed above in Section \ref{}, the training dataset consisted of 4.3M overlapping examples of 256-event performance excerpts. These examples were shuffled before training and divided into 33579 batches of 128 examples. The Adam optimiser was used for gradient descent with an initial learning rate of $1 \times 10^{-4}$. RoboJam's network was configured with 3 layers of LSTM units and $M = 16$ mixtures. Models were trained up to five epochs with LSTM layer sizes of 64, 128, 256, and 512 units. The results of this training is discussed below in Section \ref{}. Numerical stability in training is a significant challenge for MDNs. We follow advice from Bishop~\cite{}, Brando~\cite{} and others to transform and clip the mixture parameters, and to apply gradient clipping. Nevertheless, avoiding division by zero when calculating $\mathcal{L}_{tot}$ continued to be a challenge in this research and is a topic for future investigation. \subsection{Model Limitations}\label{} Our implementation separates space and time predictions into two mixture models, both connected to the output of one RNN. It may be possible to represent these in a single mixture of 3D normal distributions, but the PDF for a 3D normal is more complex. Our decision here was pragmatic in that we were able to extend previously effective MDN designs with parameters for space and time predictions linked via the fully-connected LSTM and projection layers. Future work could determine whether a 3D model, or indeed three 1D models would be more suitable. Our model does not explicitly consider the state of a touch event (whether it is a new, or moving touch), even though this data is available in the corpus. Given the temporal dimension, this aspect of the data is quite predictable so we used a heuristic threshold of $dt > 0.1$ to determine a touch as new. Potentially this data could be predicted by the MDRNN as in earlier handwriting~\cite{} and sketch~\cite{} models. \section{Interaction Design}\label{} \begin{figure} \centering \includegraphics[width=1.0\columnwidth]{robojam-interaction.pdf} \caption{RoboJam is a web-application that generates responses to touchscreen performances. User performances are sent to the server, used to condition the memory state of an MDRNN, this can then generate a new performance that is returned to the user as layer of accompaniment.} \label{} \end{figure} RoboJam is a web-application that uses the MDRNN described above to predict appropriate responses to user-created touchscreen performances. The interaction paradigm mirrors a call-and-response improvisation; the performer creates a short improvisation, RoboJam ``listens'' to this performance and creates a response, then both contributions can be played back at the same time\footnote{This process is illustrated in the video figure: \url{}}. The system architecture is illustrated in Figure \ref{}: First, the user performs a short piece of music with the touchscreen interface; the app limits these performances to 5 seconds in length. The user can then request a RoboJam response by tapping a button in the app's graphical interface. The performance, consisting of a sequence of touch interaction events, is encoded in JSON format and sent to the server via a REST API. When the server receives a performance, it feeds each touch event in sequence through the MDRNN in order to condition the neural net's memory state. When this is complete, it generates a new performance from the MDRNN by sampling a new sequence of touch-events until 5 seconds of interactions have been saved. This sampled performance is sent back to the user's device and displayed as a second layer of performance on the screen. When the user hits ``play'', both performances are played back simultaneously as separate parts. In this research, RoboJam has been used with a prototype touchscreen music app for iOS devices. This app allows touch interactions to be interpreted as different instrumental sounds such as drums, strings, bass, or different synthesiser sounds. In general, these mappings allow the user continuous control of pitch on the x-axis and tone or effects on the y-axis. The mappings are available to the user as selectable presets for their performances. Response performances from RoboJam's MDRNN can be performed by any one of these mappings, but our implementation assigns one randomly to each response such that the mapping is different than that currently selected by the user. RoboJam is implemented in Python using the Flask web development framework. This design allowed RoboJam's TensorFlow-based MDRNN (also defined in Python) to be accessed easily. Future implementations of RoboJam could be integrated directly into a touchscreen app using TensorFlow, or manufacturer-specific deep learning frameworks such as Apple's CoreML. In the short-term, however, the simplicity of using the MDRNN in a client/server architecture outweighed the potential benefits of on-device predictions. \section{Evaluation}\label{} Evaluation of RoboJam, as with many co-creative interfaces, presents challenges. We have chosen to split evaluation into three stages: validation, generation, and interaction. These stages respectively address the following questions: Has the networked learned anything? Does the network generate realistic performances? Is the interactive system useful for its purpose of enhancing the users' musical experience? \subsection{Validation}\label{} \begin{figure} \centering \includegraphics[width=0.9\columnwidth]{mdn_training_experiment.pdf} \caption{Training loss for RoboJam networks with 64, 128, 256, and 512-unit LSTM layers. Training and validation loss for the 512 unit network was lowest. The 128-unit network failed to train due to numerical errors.} \label{} \end{figure} \begin{table}[] \centering \begin{tabular}{@{}p{1.5cm}llllll@{}} \toprule \multirow{2}{2.0cm}{Training Step (1000s)} & \multicolumn{2}{c}{64 units} & \multicolumn{2}{c}{256 units} & \multicolumn{2}{c}{512 units} \\ & Training& Valid'n & Training& Valid'n& Training & Valid'n \\ \midrule 2 & -3.72 & -4.21 & -3.98 & -4.39 & -3.56 & -3.81 \\ 34 & -6.73 & -7.58 & -7.30 & -8.58 & -6.96 & -7.92 \\ 68 & -7.61 & -8.21 & -8.42 & -8.98 & -8.86 & -9.50 \\ 100 & -7.67 & -8.39 & -8.14 & -8.85 & -9.14 & -9.88 \\ 132 & -7.93 & -8.61 & -8.19 & -9.09 & -9.47 & -10.00 \\ 166 & -8.07 & -8.76 & -8.51 & -8.84 & -9.42 & -10.31 \\ \bottomrule\\ \end{tabular} \caption{Training and validation loss for RoboJam's MDRNN at throughout training and with different RNN-layer sizes. (Lower is better.)} \label{} \end{table} Four versions of the model were trained up to five epochs (166000 batches) using 64, 128, 256, and 512-unit LSTM layers respectively. Training loss for these models can be seen in \ref{}; notably, the 128-unit network failed to train due to numerical errors. Validation was performed by calculating loss values when the network predicts excerpts of touchscreen performances from RoboJam's host touchscreen app. This validation dataset was not used in training and consisted of 1047 performances with a total of 168039 touch interactions; these data were processed similarly to the training set. A validation experiment was performed on the three successfully trained models with training and validation loss recorded at model checkpoints near the end of each epoch. The results in Table \ref{} show that both training loss and validation loss decrease during training with the lowest scores occurring with the 512-unit network after five epochs. This model was used for the two evaluations below. It is notable that the loss on the validation set is consistently lower than on the training set -- an unusual situation. This suggests that the validation data may actually be more predictable than the training data. Given that the training data was collected from a variety of different apps, it may contain more unpredictable or arbitrary interactions than we thought. While validation shows that the network's predictive ability improved during training, it may be better in future to focus training on a dataset that has been restricted to the most relevant interactions. \subsection{Performance Generation}\label{} The output of RoboJam's MDRNN are performances that should be experienced in time; however, as a quick overview, we can view plots of the touches from these performance over their whole duration. Figure \ref{} shows five-second performances generated without conditioning the network's memory state. More relevant to the application, Figure \ref{} shows examples conditioned performances, the input performance is shown in blue and RoboJam's response is shown in red. In both figures, moving touches are connected while individual touches are not. Examples of call-and-response RoboJam performances can be found in a video figure for this paper which is available online\footnote{Video figure: \url{}}. \begin{figure} \centering \includegraphics[width=1.0\columnwidth]{uncond-montage.jpg} \caption{RoboJam's network generating 5-second performances in ``unconditional'' mode, i.e., with an empty memory state and starting from the centre of the performance area.} \label{} \end{figure} \begin{figure} \centering \includegraphics[width=1.0\columnwidth]{cond-montage.jpg} \caption{Examples of 5-second responses to performances from the validation set. The input performance is shown in blue and RoboJam's response is shown in red.} \label{} \end{figure} Figures \ref{} and \ref{} show a variety of touchscreen interaction behaviours, but reveal limitations in terms of generating realistic performances. Compared with example touchscreen performances (shown in blue in Figure \ref{}), the network generated performances do not have as many smooth paths, and have many more large jumps across the screen with very short time delays (resulting in long connected paths). While the unconditional performance (Figure \ref{}) show uncontrolled behaviour, many of the conditioned performances \ref{} bear some relation to the ``style'' of the input performance, in terms of location of touch points, direction and shape of motion. Some rhythmic relationships can also be seen, input performances with sparse disconnected paths indicate a rhythmic performance, these seem to lead to more rhythmic variation in the generated response. Frequently in the generated performances, the touch point tends to move to the upper left corner of the touch area, corresponding to location $(0,0)$ in the 2D plane. This tendency is as confusing as it is worrying, given that the training data is much more evenly spread across the touch field. Exploring the reasons and possible solutions for this issue is a topic to be addressed in future research. \subsection{Interaction}\label{} \begin{figure} \centering \includegraphics[width=1.0\columnwidth]{robojam_user_study_results.pdf} \caption{Distribution of responses to each question in the preliminary user study of RoboJam. Most users agreed that the responses were related to their performances and that interacting with RoboJam enhanced their experience.} \label{} \end{figure} A preliminary evaluation of interaction was performed with a small group of users with RoboJam connected to a touchscreen music application. The aim of the study was to gain a rough picture of how users viewed the quality of RoboJam's responses and the experience of interacting with this call-and-response agent. Twelve participants were included in the study: students, researchers, and teachers of computer science and music with a mix of musical experience. The study procedure was as follows: Each participant was given a short tutorial and practice session (around 5 minutes) in creating performances in the app. They were then asked to create several performances and generate as many responses from RoboJam as they liked, listening to each resulting layered performance. After the test session, the participants responded to five Likert-style statements on 5-point agreement scales (\emph{strongly disagree, disagree, neutral, agree, strongly agree}). The questions were: \begin{enumerate} \item The responses were related to my performance. (\emph{related}) \item The responses had a high musical quality. (\emph{quality}) \item The responses showed musical creativity. (\emph{creativity}) \item The response layer made my performance sound better. (\emph{improvement}) \item Interacting with RoboJam enhanced my experience. (\emph{experience}) \end{enumerate} The results of this study are shown in Figure \ref{}. Almost all the users felt that the responses were related to their performance and that interacting with RoboJam enhanced their experience. Most participants agreed that RoboJam's responses showed creativity and made their performances sound better; however, they were more uncertain about the overall quality of the responses. These results tell us that interacting with a call-and-response agent is a potentially useful addition to this app. In some ways, it would appear that this feature is useful even when the agent produces responses of uneven quality. The participants appear to have appreciated hearing their performances in context with a response and felt that this improved the sound of their own contributions. They were able to naturally cherry-pick responses until they found one that appealed to them, this involved critically engaging with their own performances as well as RoboJam's. These results encourage us that RoboJam could be a useful and significant addition to our touchscreen music app. A more in depth user study could compare RoboJam responses with other agents that generate responses via alternative generative music processes. \section{Conclusion}\label{} This work has described RoboJam, an ANN-based musical sequence generator connected to an interactive touchscreen music app as a agent for call-and-response style improvisation. This system has the potential to enhance the user's experience, particularly when they are not able to collaborate with other users. Our agent, RoboJam, uses a novel application of MDN and RNN to model musical control data. Not only does this network model the location of interactions on a touchscreen, but it also the \emph{rhythm} of these interactions in absolute time. This configuration distinguishes RoboJam from other typical approaches in ANN music generation; our network learns from data at the control gesture level, rather than at the note level, it also learns to perform in absolute time, rather than at preset rhythmic subdivisions. In the context of an interactive digital musical instrument, this configuration allows RoboJam to perform music in exactly the same way as users. Rather than learning to compose music, RoboJam is actually learning to perform on a touchscreen instrument. This design choice has several implications in our results. Learning from low-level interaction data seems to be a harder task than learning from higher-level musical notes. From a musical perspective, RoboJam's performances seem underwhelming compared with high-level ANN music generators. However, there are advantages from learning to \emph{perform} rather than to \emph{compose}. Since a touchscreen interface can be used to control different kinds of instruments, RoboJam's output can be mapped to different kinds of synthesis processes as in our touchscreen app. Furthermore, RoboJam's responses are related to the body movements of the app user, an embodied approach that is appropriate in our application where free-form gestural exploration is emphasised. In this research, we examined the results of evaluations of RoboJam in terms of model validation, generative power, and user experience. These studies have demonstrated that this system can generate responses that are related in movements and rhythm to the call performance. Our preliminary human-centred evaluation has shown that users felt the responses improved their performances and enhanced their experiences. This was the case even when response quality was not always highly rated. Our results are encouraging, but they are tempered by the difficulty of training an MDRNN. Our musical data contains a wider and more abstract variety of interactions than, for example, handwritten letters of the alphabet. This may explain some of our difficulties with training. Future improvements to RoboJam could explore more curated training data and alternative mixture model designs. The generation of music and other creative data is an exciting topic in deep learning. While it is clear that computer-generated art spurs the imagination, it is less clear how music generators can be integrated into human-centred creative processes. We have focused on modelling musical touchscreen control data and on developing a call-and-response interaction between user and agent. As a result, our system can easily be integrated into mobile musical interfaces and appears to enhance the musical experience of users in our study. Future research could expand on how models of musical control data and call-and-response agents could be used to accompany, modulate, or assess human musical performances. \subsection*{Acknowledgements} This work was supported by The Research Council of Norway as a part of the Engineering Predictability with Embodied Cognition (EPEC) project, under grant agreement 240862. \bibliographystyle{plainurl}

\section{Introduction} The \emph{comprehensive factorization} of a functor was introduced by Street and Walters in \cite{StW}, where they showed that any functor $F \colon \bC \to \bD$ between arbitrary categories is the composite of an initial functor followed by a discrete opfibration, and these two classes of morphisms form a factorization system for \Cat. By duality, a similar factorization can be obtained by means of a final functor followed by a discrete fibration, and the latter too is known under the name of comprehensive factorization. Let us recall that a functor $F \colon \bC \to \bD$ is called \emph{final} if, for any object $y$ in \bD, the comma category $(y \downarrow F)$ is non-empty and connected. On the other hand, a \emph{discrete fibration} is a functor $F \colon \bC \to \bD$ such that, for any object $x$ in \bC\ and any arrow $g \colon y \to F(x)$ in \bD, there exists a unique arrow $f$ in \bC\ with codomain $x$ and such that $F(f)=g$. Initial functors and discrete opfibrations are the obvious dualizations of the previous concepts. It is an easy observation that if we restrict our attention to groupoids, then initial and final functors coincide, and the same is true for (discrete) fibrations and opfibrations. Moreover the following elementary result holds. \begin{Proposition} \label{} A functor between groupoids is final if and only if it is full and essentially surjective. \end{Proposition} \begin{proof} Let $F \colon \bC \to \bD$ be a functor between groupoids. Asking that, for any object $y$ in \bD, the comma category $(y \downarrow F)$ is non-empty is equivalent to asking that $F$ is essentially surjective. Suppose now that $F$ is full and consider two objects $(g\colon y \to F(x),x)$ and $(g'\colon y \to F(x'),x')$ in the comma category $(y \downarrow F)$. Then the arrow $g'g^{-1}\colon F(x) \to F(x')$ has an inverse image $f\colon x \to x'$ in \bC, and $(g'g^{-1},f)$ is an arrow in $(y \downarrow F)$ connecting $(g,x)$ and $(g',x')$. Conversely, if $F$ is final, then given an arrow $g\colon F(x) \to F(x')$, an inverse image of $g$ is the second projection of a connecting isomorphism between $(1_{F(x)},x)$ and $(g,x')$ in the comma category $(F(x) \downarrow F)$ \end{proof} An internal version of the comprehensive factorization system for functors between groupoids in a Barr-exact category has been developed by Bourn in \cite{Bourn87}, where he provided an explicit construction of the above factorization by means of the \emph{d\'ecalage} functor. In that paper, Bourn takes discrete fibrations as basic notion and defines final functors as their orthogonal class. As far as we know, the characterization of Proposition \ref{} has no internal counterpart in the literature. It is the aim of the present paper to fill this gap. Throughout the paper \cC\ will be a Barr-exact category. \section{Internal groupoids} We fix here some notation and recall some basic facts about internal groupoids in Barr-exact categories. An internal category in \cC\ is given by a diagram \[ \xymatrix{ H_1 \times_{(c,d)} H_1 \ar@<1ex>[r]^-{p_2} \ar@<-1ex>[r]_-{p_1} \ar[r]|-{m_H} & H_1 \ar@<1ex>[r]^-{c_H} \ar@<-1ex>[r]_-{d_H} & H_0 \ar[l]|-{e_H} } \] satisfying the usual axioms for a category. It is a groupoid when, in addition, it comes equipped with an inversion morphism for arrows \[ i_H\colon H_1 \to H_1 \] satisfying the well known axioms. We will avoid the use of subscripts as far as no confusion arises. An internal functor, denoted by $F\colon \bH \to \bG$, is given by a pair of morphisms $(f_0,f_1)$ in \cC\ such that $f_0d=df_1$, $f_0c=cf_1$, $ef_0=f_1e$ and $m_G(f_1\times_{f_0}f_1)=f_1m_H$: \[ \xymatrix{ H_1 \ar@<1ex>[d]^c \ar@<-1ex>[d]_d \ar[r]^{f_1} & G_1 \ar@<1ex>[d]^c \ar@<-1ex>[d]_d \\ H_0 \ar[u]|e \ar[r]_{f_0} & G_0 \ar[u]|e } \] We denote by $\Gpd(\cC)$ the category of internal groupoids in \cC\ and functors between them. \subsection{Some relevant classes of functors} \label{} An internal \emph{discrete fibration} is a functor $F\colon \bH \to \bG$ of internal categories such that the following square of solid arrows is a pullback: \[ \xymatrix{ H_1 \ar@<.7ex>[d]^c \ar@<-.7ex>@{-->}[d]_d \ar[r]^{f_1} & G_1 \ar@<.7ex>[d]^c \ar@<-.7ex>@{-->}[d]_d \\ H_0 \ar[r]_{f_0} & G_1 } \] In the case of groupoids, thanks to the inversion morphisms, the commutative square with dashed downwarded arrows is a pullback too, hence $F$ is also a discrete opfibration. It is easy to prove that discrete (op)fibrations are pullback stable. Given a groupoid \bG\ and a morphism $f\colon X \to G_0$ in \cC, the following procedure yields an internal functor. Consider the pullback \[ \xymatrix{ P \ar[d]_{p_1} \ar[r]^{p_2} & G_1 \ar[d]^{(d,c)} \\ X \times X \ar[r]_-{f \times f} & G_0 \times G_0 } \] Then the following is an internal functor between groupoids \[ \xymatrix{ P \ar@<.7ex>[d]^{c} \ar@<-.7ex>[d]_{d} \ar[r]^{p_2} & G_1 \ar@<.7ex>[d]^c \ar@<-.7ex>[d]_d \\ X \ar[r]_{f} & G_0 } \] (where the $d$ and $c$ on the left are the composites of $p_1$ with the product projections) which is indeed a cartesian lifting of $f$ at \bG, with respect to the fibration $()_0\colon \Gpd(\cC)\to \cC$ associating with any groupoid in \cC\ its object of objects. So we are allowed to denote by $f^*\bG$ the domain of the functor $(f,p_2)$. The square above is also called the joint pullback of $(d,c)$ along $f$. In particular, given a functor $F\colon \bH\to\bG$ in $\Gpd(\cC)$, one can factor $F$ through $f_0^*\bG$, as in the following diagram: \[ \xymatrix{ H_1 \ar@/^3ex/[rr]^{f_1} \ar[r]_{\phi_F} \ar@<.7ex>[d]^c \ar@<-.7ex>[d]_d & P \ar@<.7ex>[d]^{c} \ar@<-.7ex>[d]_{d} \ar[r]_{p_2} & G_1 \ar@<.7ex>[d]^c \ar@<-.7ex>[d]_d \\ H_0 \ar@{=}[r] & H_0 \ar[r]_{f_0} & G_0 } \] $F$ is said to be \emph{full} if $\phi_F$ is a regular epimorphism and \emph{faithful} if $\phi_F$ is a monomorphism. In fact, the above procedure yields a factorization system for $\Gpd(\cC)$ given by bijective on objects and fully faithful functors. \begin{Proposition} \label{} Let the diagram below be a pullback in $\Gpd(\cC)$: \[ \xymatrix{ \bH \times_{\bG} \bH' \ar[r]^-{\overline{F}} \ar[d]_{\overline{F}'} & \bH' \ar[d]^{F'} \\ \bH \ar[r]_{F} & \bG } \] The following properties hold: \begin{enumerate} \item If $F$ is full, $\overline{F}$ is full, and the converse is true if the arrow component $f_1'$ of $F'$ is a regular epimorphism; \item If $F$ is faithful, $\overline{F}$ is also faithful. \end{enumerate} \end{Proposition} \begin{proof} Let us consider the following commutative diagram, where the front and back faces of the cube are the pullbacks yielding the fully faithful liftings of $f_0$ and $\overline{f}_0$ at \bG\ and $\bH'$ respectively: \[ \xymatrix@=4ex{ P' \ar[rr]^{u'} \ar[dd] \ar[dr]_{v} & & H_1' \ar[dd]_(.3){(d,c)} \ar[dr]^{f_1'} \\ & P \ar[rr]_(.3){u} \ar[dd] & & G_1 \ar[dd]^{(d,c)} \\ (H_0 \times_{G_0} H_0') \times (H_0 \times_{G_0} H_0') \ar[dr]_{\overline{f}_0'\times \overline{f}_0'} \ar[rr]_(.75){\overline{f}_0\times \overline{f}_0} & & H_0' \times H_0' \ar[dr]^{f_0'\times f_0'} \\ & H_0 \times H_0 \ar[rr]_{f_0 \times f_0} & & G_0 \times G_0 } \] Since the bottom face is a pullback, then so is the top face. As a consequence, in the diagram below, since the whole rectangle and the right square are pullbacks, so is the square on the left hand side. \[ \xymatrix{ H_1 \times_{G_1} H_1' \ar[d]_{\overline{f}_1'} \ar[r]^-{\phi_{\overline{F}}} & P' \ar[d]_v \ar[r]^{u'} & H_1' \ar[d]^{f_1'} \\ H_1 \ar[r]_{\phi_{F}} & P \ar[r]_{u} & G_1 } \] Now the theses follow by definition of full and faithful functor and from the fact that monomorphisms and regular epimorphisms are pullback stable in \cC. \end{proof} Finally, given a functor $F\colon\bH\to\bG$, consider the following pullback in \cC: \[ \xymatrix{ E_0 \ar[r]^-{\overline{f_0}} \ar[d]_{p_1} & G_1 \ar[d]^{d} \\ H_0 \ar[r]_{f_0} & G_0 } \] $F$ is said to be \emph{essentially surjective} if the composite $c\cdot\overline{f_0}\colon E_0\to G_0$ is a regular epimorphism. \subsection{Support and connected components} For any groupoid \bH, the pair $(d,c)$ factors through an equivalence relation $\Sigma\bH$ on $H_0$ \[ \xymatrix{ \Sigma H_1 \ar@<1ex>[r]^{r_2} \ar@<-1ex>[r]_{r_1} & H_0 \ar[l]|{s} } \,, \] where $\Sigma H_1$ is the regular image of $H_1$ in $H_0 \times H_0$ \[ \xymatrix{ H_1 \ar@{->>}[r]^-{\sigma_H} \ar@/_3ex/[rr]_{(d,c)} & \Sigma H_1 \ar@{ >->}[r]^-{(r_1,r_2)} & H_0 \times H_0 } \,. \] \cC\ being Barr-exact, $\Sigma\bH$ is effective and we will denote by $q_H \colon H_0 \to \pi_0(\bH)$ its quotient, which is also the coequalizer of $(d,c)$. In fact, the above procedure defines three functors: \[ \Sigma \colon \Gpd(\cC) \to \Eq(\cC) \,, \qquad \bH \mapsto \Sigma\bH \] the \emph{support} functor, sending each groupoid \bH\ to its associated equivalence relation, also called the support of \bH. \[ Q \colon \Eq(\cC) \to \cC \] sending each equivalence relation to its quotient object. \[ \pi_0 \colon \Gpd(\cC) \to \cC \] the \emph{connected components} functor, which is nothing but the composite $\pi_0=Q\cdot\Sigma$, sending each groupoid \bH\ to its object of connected components $\pi_0(\bH)$. From now on, we will use the notation $\pi_0$ also for the quotient of an equivalence relation, identifying it with the associated internal groupoid. \medskip The next two results are based on Proposition 1.1 in \cite{Bourn03} and will be useful afterwards. \begin{Proposition} \label{prop:ff<->mono} A functor $F\colon\bR\to\bS$ between internal equivalence relations in \cC\ is fully faithful if and only if $\pi_0(F)$ is monomorphic. \end{Proposition} \begin{proof} Let us draw the components of $F$ vertically, and compute $\pi_0(F)$ as the induced arrow between the quotient objects of the domain and codomain: \[ \xymatrix{ R_1 \ar@<.7ex>[r]^{r_2} \ar@<-.7ex>[r]_{r_1} \ar[d]_{f_1} & R_0 \ar[d]^{f_0} \ar@{->>}[r] & \pi_0(\bR) \ar[d]^{\pi_0(F)} \\ S_1 \ar@<.7ex>[r]^{r_2} \ar@<-.7ex>[r]_{r_1} & S_0 \ar@{->>}[r] & \pi_0(\bS) } \] Since the two rows in the above diagram are exact forks (a regular epimorphism with its kernel pair), the thesis follows by Proposition 1.1 in \cite{Bourn03}. \end{proof} \begin{Corollary} \label{>mono} If a functor $F\colon\bH\to\bG$ between internal groupoids in \cC\ is full, then $\pi_0(F)$ is monomorphic. The converse is true if\, \bG\ is an internal equivalence relation. \end{Corollary} \begin{proof} Let us consider the following commutative diagram, where the left hand side squares are pullbacks (i.e.\ $P_1$ and $P_2$ yield the full and faithful liftings of $f_0$ at $\Sigma G_1$ and $G_1$ respectively): \[ \xymatrix{ H_1 \ar@{->>}[d]_{\sigma_H} \ar@/^3ex/[rr]^{f_1} \ar[r]_{\phi_F} & P_2 \ar[r] \ar@{->>}[d] & G_1 \ar@{->>}[d]^{\sigma_G} \\ \Sigma H_1 \ar@{ >->}[d] \ar@/^4ex/[rr]^(.7){\Sigma f_1} \ar[r]_{\phi_{\Sigma F}} & P_1 \ar[r] \ar@{ >->}[d] & \Sigma G_1 \ar@{ >->}[d] \\ H_0 \times H_0 \ar@{=}[r] & H_0 \times H_0 \ar[r]_{f_0 \times f_0} & G_0 \times G_0 } \] If $F$ is full, $\phi_F$ is a regular epimorphism by definition, hence $\phi_{\Sigma F}$ is a regular epimorphism and a monomorphism at the same time, so it is an isomorphism, i.e.\ $\Sigma F$ is full and faithful. Hence $\pi_0(F)$ is monomorphic by Proposition \ref{prop:ff<->mono}. Conversely, if $\pi_0(F)$ is a monomorphism then $\Sigma F$ is fully faithful, i.e.\ $\phi_{\Sigma F}$ is an isomorphism. If in addition \bG\ is an equivalence relation, then $\sigma_G$ is an isomorphism, hence $P_2\cong P_1$ and $\phi_F$ is a regular epimorphism, being isomorphic to $\sigma_H$. By definition $F$ is then full. \end{proof} \begin{Proposition} \label{prop:es} A functor $F\colon\bH\to\bG$ between internal groupoids in \cC\ is essentially surjective if and only if $\pi_0(F)$ is a regular epimorphism. \end{Proposition} \begin{proof} It suffices to focus on the following commutative diagram: \[ \xymatrix@=7ex{ H_0 \times_{(f_0,d)} G_1 \ar@{}[dr]|{(a)} \ar@/_10ex/[dd]_{p_1} \ar[r]^-{\overline{f_0}} \ar[d]^{1\times\sigma_G} & G_1 \ar@/_5ex/[dd]^(.7){d} \ar[r]^{c} \ar[d]^{\sigma_G} & G_0 \ar@{=}[d] \\ H_0 \times_{(f_0,r_1)} \Sigma G_1 \ar@{}[dr]|{(b)} \ar[d] \ar[r]^(.6){p_2} & \Sigma G_1 \ar@{}[dr]|{(c)} \ar[d]^{r_1} \ar[r]_{r_2} & G_0 \ar[d]^{q_G} \\ H_0 \ar[r]_{f_0} & G_0 \ar[r]_{q_G} & \pi_0(\bG) } \] If $F$ is essentially surjective, then $c\overline{f_0}$ is a regular epimorphism, and so is $q_Gc\overline{f_0}=q_Gf_0p_1=\pi_0(F)q_Hp_1$, hence $\pi_0(F)$ is also a regular epimorphism. Conversely, if $\pi_0(F)$ is a regular epimorphism, then so is $q_Gf_0=\pi_0(F)q_H$. Now let us observe that $(c)$, $(b)$, $(a)+(b)$, and hence $(a)$ are all pullbacks. So $r_2p_2$ is a regular epimorphism as a pullback of $q_Gf_0$, and $1\times \sigma_G$ is a regular epimorphism as a pullback of $\sigma_G$. Then their composite $c\overline{f_0}$ is a regular epimorphism and $F$ is essentially surjective. \end{proof} \section{The comprehensive factorization} We borrow from \cite{Bourn87} the definition and some needed results about the shift functor $\Dec\colon \Gpd(\cC)\to\Gpd(\cC)$. Let us recall that, for any groupoid \bH, the pullback $H_1 \times_{(c,d)} H_1$ is also isomorphic to the kernel pair $R_d$ of $d$ (or the kernel pair $R_c$ of $c$): \[ \xymatrix{ R_d \ar[r]_-\sim^-{(ir_1,r_2)} & H_1 \times_{(c,d)} H_1 & R_c \ar[l]^-\sim_-{(r_1,ir_2)} } \] For the sake of convenience, we denote \[ \overline{d}=m(r_1,ir_2)\colon R_c \to H_1 \qquad \mbox{and} \qquad \overline{c}=m(ir_1,r_2)\colon R_d \to H_1 \] We define here $\Dec$ by means of $R_c$ as the functor associating with any groupoid \bH\ in \cC\ the following internal groupoid: \[ \xymatrix{ R_c \ar@<1ex>[r]^-{r_2} \ar@<-1ex>[r]_-{r_1} & H_1 \ar[l]|-{s} } \,, \] and we denote by $\eps\bH\colon\Dec\bH\to\bH$ the internal discrete fibration \[ \xymatrix@C=7ex{ R_c \ar@<1ex>[d]^{r_2} \ar@<-1ex>[d]_{r_1} \ar[r]^{\overline{d}} & H_1 \ar@<1ex>[d]^c \ar@<-1ex>[d]_d \\ H_1 \ar[u]|s \ar[r]_{d} & H_0 \ar[u]|e } \] The following is an exact fork in $\Gpd(\cC)$: \[ \xymatrix{ \Dec^2\bH \ar@<.7ex>[r]^{\eps{\Dec\bH}} \ar@<-.7ex>[r]_{\Dec\eps\bH} & \Dec\bH \ar[r]^-{\eps\bH} & \bH } \,. \] Following \cite{BR}, we give here a description of the comprehensive factorization in $\Gpd(\cC)$. Further details can be found also in \cite{Bourn87}. Let $F\colon\bH\to\bG$ be an internal functor, then the pair $(\Dec F,\Dec^2 F)$ gives rise to a morphism between equivalece relations in $\Gpd(\cC)$: \[ \xymatrix@C=9ex{ \Dec^2\bH \ar@<.7ex>[d]^{\eps{\Dec\bH}} \ar@<-.7ex>[d]_{\Dec\eps\bH} \ar[r]^{\Dec^2 F} & \Dec^2\bG \ar@<.7ex>[d]^{\eps{\Dec\bG}} \ar@<-.7ex>[d]_{\Dec\eps\bG} \\ \Dec\bH \ar[r]_{\Dec F} & \Dec\bG } \] We consider the following factorization of the above functor, where all the right hand side squares are pullbacks: \begin{equation} \label{} \begin{aligned} \xymatrix{ \Dec^2\bH \ar@<.7ex>[d]^{\eps{\Dec\bH}} \ar@<-.7ex>[d]_{\Dec\eps\bH} \ar[r] & \bR_{\Delta} \ar@<.7ex>[d] \ar@<-.7ex>[d] \ar[r] & \Dec^2\bG \ar@<.7ex>[d]^{\eps{\Dec\bG}} \ar@<-.7ex>[d]_{\Dec\eps\bG} \\ \Dec\bH \ar[d]_{\eps\bH} \ar[r] & \bE \ar[r]^-{\overline{F}} \ar[d]_{\Delta} \ar@{}[dr]|{(\ast)} & \Dec\bG \ar[d]^{\eps\bG} \\ \bH \ar@{=}[r] & \bH \ar[r]_{F} & \bG } \end{aligned} \end{equation} Finally, applying the functor $\pi_0$ to the upper rectangle, we get the factorization of $F$ into a final functor $J=(j_0,j_1)$ followed by a discrete fibration $K=(k_0,k_1)$: \begin{equation} \label{diag:cf} \begin{aligned} \xymatrix{ H_1 \ar@<.7ex>[d]^{c} \ar@<-.7ex>[d]_{d} \ar[r]^{j_1} & T_1 \ar@<.7ex>[d]^{c} \ar@<-.7ex>[d]_{d} \ar[r]^{k_1} & G_1 \ar@<.7ex>[d]^{c} \ar@<-.7ex>[d]_{d} \\ H_0 \ar[r]_{j_0} & T_0 \ar[r]_{k_0} & G_0 } \end{aligned} \end{equation} \section{Final functors} The idea behind the characterization of final functors lies in the following lemma. \begin{Lemma} \label{} A functor $F\colon\bH\to\bG$ between internal groupoids in \cC\ is final if and only if its pullback $\overline{F}\colon\bE\to\Dec\bG$ along $\eps\bG$ is inverted by $\pi_0$. \end{Lemma} \begin{proof} First of all, notice that $\pi_0(\overline{F})$ is nothing but the arrow $k_0$ in diagram (\ref{diag:cf}), so saying that $\overline{F}$ is inverted by $\pi_0$ means that the arrow $k_0$ is an isomorphism. Now, if $F$ is final, $K$ is an isomorphism, and so is $k_0$. Conversely, since $K$ is a dicrete fibration, the right hand side commutative squares in (\ref{diag:cf}) are pullbacks, hence if $k_0$ is an isomorphism, so is $k_1$, and $F$ is final. \end{proof} We are now ready for an internal version of Proposition \ref{}. \begin{Theorem} \label{} A functor between internal groupoids in a Barr-exact category is final if and only if it is full and essentially surjective. \end{Theorem} \begin{proof} Let us consider the vertical expansion of the pullback $(\ast)$ in diagram (\ref{}) and take the (bijective on objects, fully faithful) factorizations of the internal functors $F$ and $\overline{F}$: \[ \xymatrix{ E_1 \ar[dr] \ar[rr]^{\phi_{\overline{F}}} \ar@<.7ex>[dd]^{c} \ar@<-.7ex>[dd]_{d} & & \overline{P} \ar[rr] \ar[dr]^{v} \ar@<.7ex>[dd] \ar@<-.7ex>[dd] & & R_c \ar[dr]^{\overline{d}} \ar@<.7ex>[dd]^(.35){r_2} \ar@<-.7ex>[dd]_(.35){r_1} \\ & H_1 \ar@<.7ex>[dd]^(.35){c} \ar@<-.7ex>[dd]_(.35){d} \ar[rr]^(.3){\phi_{F}} & & P \ar[rr] \ar@<.7ex>[dd] \ar@<-.7ex>[dd] & & G_1 \ar@<.7ex>[dd]^{c} \ar@<-.7ex>[dd]_{d} \\ E_0 \ar@{=}[rr] \ar[dr] & & E_0 \ar[rr]_(.75){\overline{f_0}} \ar[dr] & & G_1 \ar[dr]_{d} \\ & H_0 \ar@{=}[rr] & & H_0 \ar[rr]_{f_0} & & G_0 } \] Thanks to Lemma \ref{}, we only have to prove that $\overline{F}$ is inverted by $\pi_0$ if and only if $F$ is full and essentially surjective. Suppose $\pi_0(\overline{F})$ is an isomorphism. Then, by Corollary \ref{>mono}, since $\Dec\bG$ is an equivalence relation, $\overline{F}$ is full, i.e.\ $\phi_{\overline{F}}$ is a regular epimorphism. But $v$ is a split epimorphism, being a pullback of $\overline{d}$, so $\phi_{F}$ is a regular epimorphism and $F$ is full. Moreover, $\pi_0(\eps\bG)$ is a regular epimorphism since $\eps\bG$ is, so $\pi_0(F)\cdot\pi_0(U)=\pi_0(\eps\bG)\cdot\pi_0(\overline{F})$ is a regular epimorphism. This implies that $\pi_0(F)$ is a regular epimorphism, hence, by Proposition \ref{prop:es}, $F$ is essentially surjective. Conversely, if $F$ is full, then $\overline{F}$ is full by Proposition \ref{}, hence $\pi_0(\overline{f})$ is a monomorphism by Corollary \ref{>mono}. If in addition $F$ is essentially surjective, $c\overline{f_0}=\pi_0(\overline{F})q_{\bE}$ is a regular epimorphism. So $\pi_0(\overline{F})$ is a regular epimorphism, hence an isomorphism. \end{proof} \begin{Corollary} \label{} A functor $F\colon\bH\to\bG$ between internal groupoids in \cC\ is final if and only if \begin{enumerate} \item[(i)] $\pi_0(F)$ is an isomorphism; \item[(ii)] the arrow $\psi_F$ in the following commutative diagram is a regular epimorphism: \begin{equation} \label{} \begin{aligned} \xymatrix@=2.5ex{ H_1 \ar[ddd]_{\sigma_H} \ar[rrr]^{f_1} \ar[dr]^{\psi_F} & & & G_1 \ar[ddd]^{\sigma_G} \\ & \Sigma H_1 \times_{\Sigma G_1} G_1 \ar[ddl]^\sigma \ar[urr] \\ \\ \Sigma H_1 \ar[rrr]_{\Sigma f_1} & & & \Sigma G_1 } \end{aligned} \end{equation} \end{enumerate} \end{Corollary} \begin{proof} By Proposition \ref{prop:ff<->mono}, $\pi_0(F)$ is a monomorphism if and only if $\Sigma F$ is fully faithful. In this case, the arrow $\psi_F$ coincides with the arrow $\phi_F$ defined in Section \ref{}. Suppose $F$ is final, then by Theorem \ref{} it is full and essentially surjective, hence, by Proposition \ref{prop:es} and Corollary \ref{>mono}, $\pi_0(F)$ is an isomorphism and $\phi_F=\psi_F$ is a regular epimorphism. Conversely, if $(i)$ and $(ii)$ hold, then $\phi_F=\psi_F$ is a regular epimorphism, so $F$ is full, and moreover $\pi_0(F)$ is a regular epimorphism, i.e.\ $F$ is essentially surjective. \end{proof}

\section{Introduction} To protect sensitive data, it is required to associate some access policies to the data so that receivers who have the necessary permissions can access it. Receiver should possess attribute sets in order to retrieve the data. Upon controlling access to the sensitive data, the associated access policy leaks too much information. For instance, consider a scenario of a medical database consisting of patients records. The access policy associated with each patient's record includes patient's identity, names of treating doctors, names of diseases, medicines recommended, treatments done etc. If access policies are public, then they reveal about the patient's personal information. Many patients do not want to disclose their identities when they undergo a plastic surgical operation, a psychiatric treatment etc. Sometimes, even doctors want to remain anonymous while treating a famous person. The solution to such privacy problems is \emph{adaptive oblivious transfer realizing hidden access policy (${\sf AOT\mhyphen HAP}$)}. \noindent The ${\sf AOT\mhyphen HAP}$ involves a sender, multiple receivers and an issuer. The sender carries a database with $N$ messages and generates a ciphertext database. The access policies are embedded in ciphertexts and are kept veiled. Only those receivers who have the necessary attributes can decipher the ciphertext using secret key issued by the issuer corresponding to its attribute set. A receiver decode the message correctly if his attribute set satisfies the access structure embedded in the ciphertext. Otherwise, a wrong message is retrieved by the receiver. In an adaptive $k\mhyphen {\rm out}\mhyphen {\rm of}\mhyphen N$ oblivious transfer, each receiver can recover only $k$ messages and remains oblivious to other $N-k$ messages. The term adaptive means, a receiver recovers one message in each of $k$ transactions. The ${\sf AOT\mhyphen HAP}$ consists of three phases -- initialization, issue and transfer. The ciphertext database is generated at initialization phase. In issue phase, the secret key is generated corresponding to a receiver's attribute set. Using secret key, a receiver contacts with the sender to decode the ciphertext in transfer phase. Initialization phase is executed once at the outset of the protocol. Issue phase is also executed once between the issuer and a receiver. On the other hand, transfer phase is run $k$ times between a receiver and the sender. \noindent Rabin \cite{rabin1981exchange} introduced the first oblivious transfer which was subsequently followed by Brassard {\em et al.} \cite{brassard1987all} and Even {\em et al.} \cite{even1985randomized} to construct secure protocols for multiparty computation. Tauman \cite{kalai2005smooth} proposed an efficient contruction of oblivious transfer protocol using projective hash framework of Cramer and Shoup \cite{cramer2002universal}. Naor and Pinkas \cite{naor2001efficient} generalized the concept of Tauman. A wide variety of oblivious transfer protocols are available in the literature \cite{aiello2001priced, kushilevitz1997replication, tassa2011generalized, lindell2008efficient, peikert2008framework, damgaard2009essentially, bresson2003simple, abdallasphf}. \noindent The first ${\sf AOT}$ was introduced by Naor and Pinkas \cite{naor1999oblivious} consequently followed by \cite{camenisch2007simulatable}, \cite{green2007blind}. Aforementioned ${\sf AOT}$ do not restrict the receivers to access the content of the database. Coull {\em et al.} \cite{coull2009controlling} presented the first ${\sf AOT}$ with access policy which was subsequently generalized by Camenisch {\em et al.} \cite{camenisch2009oblivious} and Zhang {\em et al.} \cite{zhang2010oblivious}. Recently, Guleria and Dutta \cite{guleria2015universally} presented the first issuer-free ${\sf AOT}$ with access policy in universal composable (UC) framework. However, the access policies are public in \cite{camenisch2009oblivious}, \cite{coull2009controlling}, \cite{guleria2015universally} and \cite{zhang2010oblivious}. To overcome this, Camenisch {\em et al,} \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious} and Guleria and Dutta \cite{guleria2014adaptive} introduced {\sf AOT-HAP} which are the only oblivious transfer protocols with hidden access policies. The security framework adapted in \cite{naor1999oblivious} is half-simulation model while that in \cite{camenisch2007simulatable}, \cite{green2007blind}, \cite{coull2009controlling}, \cite{camenisch2009oblivious}, \cite{zhang2010oblivious}, \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious} and \cite{guleria2014adaptive} is full-simulation model. In the half-simulation model, the security of one party is examined by comparing the output of the real world against the turnout of the ideal world, while the security of the other party is examined by an argument. The full-simulation model addresses the security of the sender and a receiver in real/ideal world paradigm. None of these models support concurrent execution. Abe {\em et al.} \cite{abe2013universally} proposed the first UC secure ${\sf AOT}$ with access policy. The UC secure framework follows real/ideal world paradigm and also remains secure even when composed with each other or with the instances of other protocols. \noindent \textbf{Our Contribution.} The aforesaid ${\sf AOT\mhyphen HAP}$ protocols do not consider collusion of an issuer with a collection of receivers. If the issuer colludes with any receiver, the receiver can decipher the ciphertext database by getting private keys for all the attributes in the universe of attributes, hence, making these protocols insecure. Besides, the existing ${\sf AOT\mhyphen HAP}$ protocols \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious}, \cite{guleria2014adaptive} are secure in full-simulation model which are vulnerable to practical attacks against concurrent execution. Designs without such constraints are more desirable in practice from both security and efficiency point of view. In this paper, we concentrate on designing ${\sf AOT\mhyphen HAP}$ by eliminating these limitations. Our contribution in this paper is-- \begin{itemize} \vspace{-.2cm} \item firstly to contruct an issuer-free ${\sf AOT\mhyphen HAP}$,\vspace{-.2cm} \item secondly to prove security in UC framework so that it does not loose security when composed with each other or with other protocols. \vspace{-.2cm} \end{itemize} \noindent Our issuer-free $k\mhyphen {\rm out}\mhyphen {\rm of}\mhyphen N$ ${\sf AOT\mhyphen HAP}$ is executed between a sender and multiple receivers. We combine Boneh-Boyen (BB) \cite{boneh2004shortsignature} signature and Nishide {\em et al.} \cite{nishide2008attribute} ciphertext policy attribute based encryption (CP-ABE). Groth-Sahai proofs \cite{groth2008efficient} are employed for non-interactive verification of pairing product equations. The sender holds a database of $N$ messages, and each receiver carries some attributes. Each message is associated with an access policy. The sender generates ciphertext database in initialization phase by encrypting each message using CP-ABE under its access policy. The ciphertext database is made public, but the access policies are kept hidden. Subsequently, the protocol makes a motion to issue phase, in which a receiver interacts with the sender to get private key corresponding to its attribute set without leaking any information about its attributes to the sender. Finally, the receiver invokes transfer phase to recover the message. The receiver can retrieve at most $k$ messages, one in each transfer phase in our $k\mhyphen {\rm out}\mhyphen {\rm of}\mhyphen N$ ${\sf AOT\mhyphen HAP}$. \noindent The security of our ${\sf AOT\mhyphen HAP}$ is analyzed in UC framework in static corruption model in which, the adversary pre-decides which party to corrupt at the beginning of the protocol. This allows security against individual malicious receiver as well as malicious sender. The malicious adversary do not follow the protocol specifications and its malicious behaviour is controlled by providing non-interactive zero-knowledge \cite{groth2008efficient} proofs. We use Groth-Sahai non-interactive witness indistinguishability (${\sf NIWI}$) and non-interactive zero-knowledge (${\sf NIZK}$) proofs to detect the malicious activities of a receiver and the sender, respectively. The proposed ${\sf AOT\mhyphen HAP}$ provides the following security guarantees -- \begin{itemize} \vspace{-.2cm} \item the sender does not know who queries a message, which message is being queried and what is the access policy of the message,\vspace{-.2cm} \item the receiver learns one message per query,\vspace{-.2cm} \item the receiver recovers the message correctly if its attribute set qualifies the access policy associated with the message,\vspace{-.2cm} \item the receiver does not learn anything about the access policy of the message.\vspace{-.2cm} \end{itemize} \noindent The proposed ${\sf AOT\mhyphen HAP}$ is UC secure assuming $q$-strong Diffie-Hellman (SDH), decision Linear (DLIN) and decision bilinear Diffie-Hellman (DBDH) assumptions in the presence of malicious adversary. \noindent We compare the proposed ${\sf AOT\mhyphen HAP}$ with the existing similar schemes, \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious} and \cite{guleria2014adaptive}, which are the only schemes available in literature. The ${\sf AOT\mhyphen HAP}$ \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious} and \cite{guleria2014adaptive} assume an issuer and are secure in full-simulation model. The ${\sf AOT}$ scheme of \cite{abe2013universally} is secure in UC framework. However, it invokes an issuer and access policies are public. In contrast, our proposed ${\sf AOT\mhyphen HAP}$ relaxing the need of an issuer and is secure in UC framework. In \cite{abe2013universally}, \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious} and \cite{guleria2014adaptive}, an issuer is included to certify whether each of the receiver possesses the claimed attributes. In such protocols, when a receiver sends an attribute to the ideal functionality, the functionality passes the attribute and receiver identity to the issuer. The issuer informs the functionality whether the attribute should be rejected or accepted. In our protocol, when a receiver sends an attribute to the ideal functionality, the functionality passes the randomized attribute to the sender. The sender checks whether the randomized attribute satisfies the associated pairing product equation. If yes, the sender tells the functionality that the attribute should be accepted, otherwise, rejected. The access policies associated with the messages are completely hidden. The encryption of the messages is such that only those receivers whose attribute set satisfies the access policy attached with messages could recover a correct message otherwise a random message is recovered. \section{Preliminaries}\label{preliminaries} \noindent \textbf{Notations:} Throughout, $\rho$ is taken as the security parameter, $x \xleftarrow{\$} A$ is a random element sampled from the set $A$, $y \leftarrow B$ indicates algorithm $B$ outputs y, $X \stackrel{c}{\approx} Y$ denotes $X$ and $Y$ are indistinguishable computationally and $\mathbb{N}$ denotes natural numbers. \begin{definition}(Bilinear Pairing) The map $e : \mathbb{G}_{1} \times \mathbb{G}_{2} \rightarrow \mathbb{G}_T$ is \emph{bilinear} if the following conditions are stisfied. (i) $e(x^{a}, y^{b}) = e(x, y)^{ab} ~\forall~ x \in \mathbb{G}_1, y \in \mathbb{G}_2, a, b \in \mathbb{Z}_{p}$. (ii) $e(x, y)$ generates $\mathbb{G}_T$, $~\forall~ x \in \mathbb{G}_1, y \in \mathbb{G}_2, x \neq 1, y \neq 1$. (iii) $e(x, y)$ is efficiently computable $~\forall~ x \in \mathbb{G}_1, y \in \mathbb{G}_2$, where $\mathbb{G}_1, \mathbb{G}_2$ and $\mathbb{G}_T$ are three multiplicative cyclic groups of prime order $p$ and $g_1$ and $g_2$ are generators of $\mathbb{G}_1$ and $\mathbb{G}_2$ respectively. \end{definition} \noindent If $\mathbb{G}_1 \neq \mathbb{G}_2$, then $e$ is {\em asymmetric} bilinear pairing. Otherwise, $e$ is {\em symmetric} bilinear pairing. \begin{definition}($q$-SDH \cite{boneh2004shortsignature}) The advantage ${\sf Adv}^{q-SDH}_{\mathbb{G}}(\mathcal{A}) = {\sf Pr}[\mathcal{A}(g, g^x$, $g^{x^2}, \ldots, g^{x^q}) = (c$, $g^{\frac{1}{x + c}})]$ of $q$-Strong Diffie-Hellman (SDH) assumption is negligible in $\rho$, where $g \xleftarrow{\$} \mathbb{G}, x \xleftarrow{\$} \mathbb{Z}_{p}, c \in \mathbb{Z}_{p}$ and $\mathcal{A}$ is a PPT algorithm with running time in $\rho$. \end{definition} \begin{definition}(DLIN \cite{boneh2004short}) The advantage ${\sf Adv}^{DLIN}_{\mathbb{G}}(\mathcal{A}) = {\sf Pr}[\mathcal{A}(g, g^a, g^b$, $g^{ra}, g^{sb}, g^{r+s})] - {\sf Pr}[\mathcal{A}(g, g^a, g^b, g^{ra}, g^{sb}, t)]$ of Decision Linear (DLIN) assumption in is negligible in $\rho$, where $g \xleftarrow{\$} \mathbb{G}, t \xleftarrow{\$} \mathbb{G}, a, b, r, s \in \mathbb{Z}_{p}$ and $\mathcal{A}$ is a PPT algorithm with running time in $\rho$. \end{definition} \begin{definition}(DBDH \cite{waters2005efficient}) The advantage ${\sf Adv}^{DBDH}_{\mathbb{G}, \mathbb{G}_T}(\mathcal{A}) = {\sf Pr}[\mathcal{A}(g, g^a, g^b, g^c, e(g, g)^{abc})] - {\sf Pr}[\mathcal{A}(g, g^a$, $g^b, g^c, Z)]$ of Decision Bilinear Diffie-Hellman (DBDH) assumption is negligible in $\rho$, where $g \xleftarrow{\$} \mathbb{G}, Z \xleftarrow{\$} \mathbb{G}_T, a, b, c \in \mathbb{Z}_{p}$ and $\mathcal{A}$ is a PPT algorithm with running time in $\rho$. \end{definition} \noindent\begin{boxit} \noindent \textbf{${\sf BilinearSetup}$: } The ${\sf BilinearSetup}$ on input $\rho$ outputs ${\sf params} = (p, \mathbb{G}_1, \mathbb{G}_2, \mathbb{G}_{T}, e, g_1, g_2)$, where $e : \mathbb{G}_1 \times \mathbb{G}_2 \rightarrow \mathbb{G}_T$ is an asymmetric bilinear pairing, $g_1$ generates a group $\mathbb{G}_1$, $g_2$ generates a group $\mathbb{G}_2$ and $p$, the order of the groups $\mathbb{G}_1$, $\mathbb{G}_2$ and $\mathbb{G}_T$, is prime, i.e ${\sf params} \leftarrow {\sf BilinearSetup}(1^{\rho})$. \end{boxit} \subsection{Non-Interactive Verification of Pairing Product Equation \cite{groth2008efficient}}\label{noninteractive} A prover and a verifier runs the Groth-Sahai proofs for non-interactive verification of a pairing product equation \begin{eqnarray}\label{eq3} \prod_{q=1}^{Q}e(a_q\prod_{i=1}^{n}x_{i}^{\alpha_{q, i}}, b_q\prod_{i=1}^{n}y_{i}^{\beta_{q, i}}) &=&t_{T}, \end{eqnarray} where $a_{q} \in \mathbb{G}_1, b_{q} \in \mathbb{G}_2$, $\alpha_{q, i}, \beta_{q, i} \in \mathbb{Z}_{p}$ and $t_{T} \in \mathbb{G}_{T}$, the coefficients of the pairing product equation \ref{eq3} are given to the verifier, $q =1, 2, \ldots, Q, i =1, 2, \ldots, n$. The prover knows the secret values (also called witnesses) $x_{i =1, 2, \ldots, n} \in \mathbb{G}_1, y_{i =1, 2, \ldots, n} \in \mathbb{G}_2$ that satisfy the equation \ref{eq3}. The prover proves in a non-interactive way that he knows $x_{i}$ and $y_{i}$ without giving any knowledge about secret values to the verifier. Let $\mathcal{W} = \{x_{i =1, 2, \ldots, n}, y_{i =1, 2, \ldots, n}\}$ denotes the set of all secret values equation \ref{eq3}. The set $\mathcal{W}$ is called witnesses. \noindent For non-interactive verification of the pairing product equation \ref{eq3}, a common reference string ${\sf GS}$ is generated upon input a security parameter $\rho$ as follows. Let ${\sf params} = (p, \mathbb{G}_1, \mathbb{G}_2, \mathbb{G}_{T}, e, g_1$, $g_2)$ ~$\leftarrow {\sf BilinearSetup}(1^\rho)$, set $u_1 = (g_{1}^a$, $1, g_1) \in \mathbb{G}_{1}^{3}, u_2 = (1, g_{1}^b, g_1) \in \mathbb{G}_{1}^{3}$, $u_3 = u_{1}^{\xi_{1}}u_{2}^{\xi_{2}} = (g_{1}^{a\xi_{1}}, g_{1}^{b\xi_{2}}$, $g_{1}^{\xi_{1} + \xi_{2}}) \in \mathbb{G}_{1}^{3}$, $v_1 = (g_{2}^a, 1, g_2) \in \mathbb{G}_{2}^{3}, v_2 = (1, g_{2}^b, g_2) \in \mathbb{G}_{2}^{3}$, $v_3 = v_{1}^{\xi_{1}}v_{2}^{\xi_{2}} = (g_{2}^{a\xi_{1}}, g_{2}^{b\xi_{2}}$, $g_{2}^{\xi_{1} + \xi_{2}}) \in \mathbb{G}_{2}^{3}$, where $\xi_{1}, \xi_{2} \xleftarrow{\$} \mathbb{Z}_{p}$, $a, b \xleftarrow{\$} \mathbb{Z}_{p}$ and $\mu_1 : \mathbb{G}_1 \rightarrow \mathbb{G}_{1}^3$, $\mu_2 : \mathbb{G}_2 \rightarrow \mathbb{G}_{2}^3$, $\mu_T : \mathbb{G}_T \rightarrow \mathbb{G}_{T}^9$ are three efficiently computable embeddings such that $\mu_1(g_1) = (1, 1, g_1), \mu_2(g_2) = (1, 1, g_2), {\rm ~and~} \mu_T(t_T) = \left( \begin{array}{ccc} 1 & 1 & 1 \\ 1 & 1 & 1 \\ 1 & 1 & t_T \\ \end{array} \right) \forall~g_1 \in \mathbb{G}_1, g_2 \in \mathbb{G}_2, t_T \in \mathbb{G}_T. $ Note that $\mu_T(t_T) \in\mathbb{G}_{T}^9$ written in matrix form for convenience. Two elements of $\mathbb{G}_{T}^9$ are multiplied component wise. The common reference string ${\sf GS} = (u_1, u_2, u_3, v_1, v_2, v_3)$ is made public to the prover and the verifier. The commitments of $x_{i =1, 2, \ldots, n}$ and $y_{i =1, 2, \ldots, n}$ are generated using ${\sf GS}$ by the prover. To commit $x_i \in \mathbb{G}_1$ and $y_i \in \mathbb{G}_2$, the prover takes $r_{1i}, r_{2i}, r_{3i} \xleftarrow{\$} \mathbb{Z}_{p}$ and $s_{1i}, s_{2i}, s_{3i} \xleftarrow{\$} \mathbb{Z}_{p}$, sets $$ c_i = {\sf Com}(x_i) = \mu_1(x_i)u_{1}^{r_{1i}}u_{2}^{r_{2i}}u_{3}^{r_{3i}},~~ d_i = {\sf Com'}(y_i) = \mu_2(y_i)v_{1}^{s_{1i}}v_{2}^{s_{2i}}v_{3}^{s_{3i}}, $$ \noindent where ${\sf Com}(x_i) \in \mathbb{G}_{1}^{3}, {\sf Com'}(y_i) \in \mathbb{G}_{2}^{3}$ for each $i=1, 2, \ldots, n$. The following proof components are generated by the prover $$ P_j = \prod_{q=1}^{Q}\left(\widehat{d_{q}}\right)^{\sum_{i=1}^{n}\alpha_{q, i}r_{ji}},~~ P'_j = \prod_{q=1}^{Q}\left(\mu_1(a_q)\prod_{i=1}^{n}\mu_1(x_i)^{\alpha_{q, i}}\right)^{\sum_{i=1}^{n}\beta_{q, i}s_{ji}}$$ using random values $r_{ji}, s_{ji}$, which were used for generating commitments to $x_{i}, y_{i}$, and gives proof $\pi = (c_1, c_2, \ldots, c_n, d_1$, $d_2, \ldots, d_n, P_1, P_2, P_3, P'_1, P'_2, P'_3)$ to the verifier, where $$\widehat{d_q} = \mu_2(b_q)\prod_{i=1}^{n}d_{i}^{\beta_{q, i}}, ~~1\leq i \leq n, ~~j = 1, 2, 3.$$ The verifier computes $$ \widehat{c_q} = \mu_1(a_q)\prod_{i=1}^{n}c_{i}^{\alpha_{q, i}}, \widehat{d_q} = \mu_2(b_q)\prod_{i=1}^{n}d_{i}^{\beta_{q, i}}, $$ using $c_i, d_i$, coefficients $\alpha_{q, i}, \beta_{q, i}$ and outputs {\sf VALID} if the following equation holds \begin{eqnarray}\label{eq1} \prod_{q=1}^{Q}F(\widehat{c_q}, \widehat{d_q}) = \mu_{T}(t_T)\prod_{j=1}^{3}F(u_j, P_j)F(P'_j, v_j), \end{eqnarray} where $F : \mathbb{G}_{1}^3 \times \mathbb{G}_{2}^3 \rightarrow \mathbb{G}_{T}^9$ is defined as $$F((x_1, x_2, x_3), (y_1, y_2, y_3)) = \left( \begin{array}{ccc} e(x_1, y_1) & e(x_1, y_2) & e(x_1, y_3) \\ e(x_2, y_1) & e(x_2, y_2) & e(x_2, y_3) \\ e(x_3, y_1) & e(x_3, y_2) & e(x_3, y_3) \\ \end{array} \right).$$ \noindent The equations \ref{eq1} and \ref{eq3} holds simultaneously. For a linear equation in which $y_{i}$ are secrets, the verifier verifies the following equation \begin{align} &\prod_{q=1}^{Q}F\left(\mu_1(a_q)\prod_{i=1}^{n}\mu_1(x_i)^{\alpha_{q, i}}, \widehat{d_q}\right) = \mu_{T}(t_T)\prod_{j=1}^{3}F(P'_j, v_j),\\ &{\rm where~~} P'_j = \prod_{q=1}^{Q}\left(\mu_1(a_q)\prod_{i=1}^{n}\mu_1(x_i)^{\alpha_{q, i}}\right)^{\sum_{i=1}^{n}\beta_{q, i}s_{ji}}, j = 1, 2, 3.~~ \end{align} If $x_{i}$ are secrets, the verifier has to verify the following equation \begin{align} &\prod_{q=1}^{Q}F\left(\widehat{c_q}, \mu_2(b_q)\prod_{i=1}^{n}\mu_2(y_i)^{\beta_{q, i}}\right) = \mu_{T}(t_T)\prod_{j=1}^{3}F(u_j, P_j),\\ &{\rm where~~} P_j = \prod_{q=1}^{Q}\left(\mu_2(b_q)\prod_{i=1}^{n}\mu_2(y_i)^{\beta_{q, i}}\right)^{\sum_{i=1}^{n}\alpha_{q, i}r_{ji}}, j = 1, 2, 3.~~ \end{align} \begin{theorem}\label{crs}\cite{groth2008efficient} The common reference strings in perfectly sound setting and witness indistinguishability setting under DLIN assumption are computationally indistinguishable . \end{theorem} \begin{definition}(${\sf NIWI}$) The advantage $${\sf Adv}^{{\sf NIWI}}_{\mathbb{G}, \mathbb{G}_T}(\mathcal{A}) = {\sf Pr}\left[\mathcal{A}({\sf GS}, \mathcal{S}, \mathcal{W}_0) = \pi\right] - {\sf Pr}\left[\mathcal{A}({\sf GS}, \mathcal{S}, \mathcal{W}_1) = \pi\right]$$ of non-interactive witness- indistinguishable (${\sf NIWI}$) proof is negligible in $\rho$ under DLIN assumption, where ${\sf GS}$ denotes perfectly sound setting's common reference string, $\mathcal{S}$ is a pairing product equation, $\mathcal{W}_0, \mathcal{W}_1$ denote two different set of witnesses satisfying equation $\mathcal{S}$, $\pi$ denotes proof for equation $\mathcal{S}$ and $\mathcal{A}$ is a PPT algorithm with running time in $\rho$. \end{definition} \begin{definition}(${\sf NIZK}$) The advantage $${\sf Adv}^{{\sf NIZK}}_{\mathbb{G}, \mathbb{G}_T}(\mathcal{A}) = {\sf Pr}[\mathcal{A}({\sf GS'}, \mathcal{S}, \mathcal{W}) = \pi_0] - {\sf Pr}[\mathcal{A}({\sf GS'}, \mathcal{S}, {\sf t_{sim}}) = \pi_1]$$ of non-interactive zero-knowledge (${\sf NIZK}$) proof is negligible in $\rho$ under DLIN assumption, where ${\sf GS'}$ denotes witness indistinguishability setting's common reference string, $\mathcal{S}$ is a pairing product equation, $\mathcal{W}$ denotes witnesses satisfying pairing equation $\mathcal{S}$, $\pi_0$ denotes proof for equation $\mathcal{S}$, $\pi_1$ denotes simulated proof for equation $\mathcal{S}$ and $\mathcal{A}$ is a PPT algorithm with running time in $\rho$. \end{definition} \begin{theorem}\label{niwi}\cite{groth2008efficient} Under DLIN assumption, the Groth-Sahai proofs are composable ${\sf NIWI}$ and ${\sf NIZK}$ for satisfiability of a set of pairing product equation under a bilinear group . \end{theorem} \subsection{Access Structure for Ciphertext \cite{nishide2008attribute}}\label{access structure} Let $n$ be the total number of attributes, and the attributes are indexed as $\{\mathbb{A}_1, \mathbb{A}_2, \ldots, \mathbb{A}_n\}$. Each attribute $\mathbb{A}_{\ell}$ can take $n_{\ell}$ values, where $\ell=1, 2, \ldots, n$. Let $S_{\ell} = \{v_{\ell,1}, v_{\ell,2}, \ldots, v_{\ell, t}$, $\ldots, v_{\ell, n_{\ell}}\}$ be the set of possible values for $\mathbb{A}_{\ell}$. The notation $L = [L_1, L_2, \ldots, L_n]$ is used to denote attribute list or attribute set for a receiver $R$, where $L_{\ell} \in S_{\ell}$. Each receiver has exactly one attribute value from each $S_{\ell}$, $\ell=1, 2, \ldots, n$. The notation $W = [W_1, W_2, \ldots, W_n]$ is used to specify the access policy associated with a message, where $W_{\ell} \subseteq S_{\ell}$. The attribute list $L$ satisfies $W$ if and only if $L_{\ell} \in W_{\ell}$, for $\ell=1, 2, \ldots, n$. The notation $L\models W$ is used to denote $L$ satisfies $W$. For instance, let $n=3$ and $m'$ be a message with access policy $W = [W_1 = \{v_{1, 1} , v_{1, 3}\}, W_2 = \{v_{2, 2}\}, W_3 = \{v_{3, 1}, v_{3, 2}, v_{3, 3}\}]$. Consider two attribute lists $L = [v_{1, 1}, v_{2, 2}, v_{3, 1}], \widetilde{L} = [v_{1, 2}, v_{2, 2}, v_{3, 3}]$. The attribute list $L\models W$, but $\widetilde{L}$ does not satisfy $W$ as $v_{1, 2} \notin W_1$. \section{Security Model}\label{securitymodel} \subsection{Syntactic of ${\sf AOT\mhyphen HAP}$} The adaptive oblivious transfer with hidden access policy (${\sf AOT\mhyphen HAP}$) is a tuple of PPT algorithms and interactive protocols, ${\sf AOT\mhyphen HAP} = ({\sf CRSSetup}, {\sf DBSetup}, {\sf Issue}, {\sf Transfer})$ between a sender and multiple receivers. \begin{description} \item {\textbf{-- ${\sf CrsSetup}$:}} This randomized algorithm generates common reference string ${\sf crs}$ which is made public for everyone. \item \textbf{-- ${\sf DBsetup}$:} This randomized algorithm is run by the sender $S$ who holds a database ${\sf DB} = \{m_i, W_i\}_{1 \leq i \leq N}$ of $N$ messages. Each message $m_i$ is associated with access policy $W_i = [W_{i, 1}, W_{i,2}, \ldots, W_{i,n}]$, where $W_{i, \ell} \subseteq S_{\ell}$, $i = 1, 2, \ldots, N, \ell = 1, 2, \ldots, n$. The algorithm generates public/secret key pair $({\sf pk}, {\sf sk})$, encrypts each $m_i$ under access policy $W_i$ and generates ciphertext $\phi_i$. The access policy $W_i$ is embedded implicitly in $\phi_i$. The algorithm outputs $({\sf pk}, {\sf sk}, \psi, {\sf cDB})$ to $S$. The proof $\psi$ guarantees that $({\sf pk}, {\sf sk})$ is a valid public/secret key pair. The sender $S$ publishes public key ${\sf pk}$, proof $\psi$ and cihertext database ${\sf cDB} = \{\phi_i\}_{1 \leq i \leq N}$ to all the receivers. The secret key ${\sf sk}$ and access policies $\{W_i\}_{1 \leq i \leq N}$ are kept hidden. \item \textbf{-- ${\sf Issue}$ Protocol:} A receiver $R$ with attributes list $L = [L_1, L_2, \ldots, L_n]$, $L_{\ell} \in S_{\ell}$, engages in ${\sf Issue}$ protocol with $S$ to get attribute secret key ${\sf ASK}$ from $S$, where $\ell = 1, 2, \ldots, n$. At the end, $R$ gets ${\sf ASK}$ without leaking any information about its attribute list $L$ to $S$. \item \textbf{-- ${\sf Transfer}$ Protocol:} The receiver $R$ invokes ${\sf transfer}$ protocol with $S$ to decrypt the ciphertext $\phi_{\sigma_j}$, where $\sigma_j \in \{1, 2, \ldots, N\}, j= 1, 2, \ldots, k$. The receiver $R$ recovers the message $m_{\sigma_j}$ if its attribute list $L$ satisfies the access policy $W_{\sigma_j}$ associated with $m_{\sigma_j}$, otherwise, $R$ gets some random message. \end{description} \subsection{UC Framework}\label{ucmodel} The \emph{universal composable (UC)} framework was presented by Canetti \cite{canetti2002universally} is adapted in this paper. It includes two worlds-- (i) \emph{real world} and (ii) \emph{ideal world}. The members involved in real world are $M$ receivers $R_1, R_2, \ldots, R_M$, a sender $S$, an adversary $\mathcal{A}$ and an \emph{environment machine} $\mathcal{Z}$. The environment machine $\mathcal{Z}$ starts the protocol execution $\Pi$ by giving inputs to all the parties. In real protocol execution $\Pi$, parties interact with each other and $\mathcal{A}$. Respective outputs are given $\mathcal{Z}$ by all the parties. Finally, $\mathcal{Z}$ outputs a bit, which is output of real world. Ideal world includes an additional incorruptible trusted party $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$, called an ideal functionality. In ideal world, parties are dummy parties, $\widetilde{S}, \widetilde{R}_1, \widetilde{R}_2, \ldots, \widetilde{R}_M$. Dummy parties do not communicate with each other. Activated dummy party just forwards its input to $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ which is instructed to do the computation work. The ideal functionality $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ gives outputs to the dummy parties. At the end, $\mathcal{Z}$ outputs a bit, which is output of ideal world. Now, a protocol $\Pi$ is said to be securely realizes an ideal functionality $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ if $\mathcal{Z}$ is unable to distinguish its interaction with non-negligible probability between a real world adversary $\mathcal{A}$ and parties $S, R_1, R_2, \ldots, R_M$ running the protocol $\Pi$ in the real world $\mathcal{A}$ and an ideal world adversary $\mathcal{A'}$ and $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$, interacting with $\widetilde{S}, \widetilde{R}_1, \widetilde{R}_2, \ldots, \widetilde{R}_M$, in the ideal world. \begin{definition} A protocol $\Pi$ securely realizes the ideal functionality $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1} $ if for any $\mathcal{A}$, there exists $\mathcal{A'}$ such that for any $\mathcal{Z}$, $${ \sf IDEAL}_{\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}, \mathcal{A'}, \mathcal{Z}} \stackrel{c}{\approx} {\sf REAL}_{\Pi, \mathcal{A}, \mathcal{Z}},$$ where ${\sf IDEAL}_{\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}, \mathcal{A'}, \mathcal{Z}}$ and ${\sf REAL}_{\Pi, \mathcal{A}, \mathcal{Z}}$ are ideal world and real world outputs respectively. \end{definition} \noindent \textbf{$\mathcal{F}_{{\sf CRS}}^{\mathcal{D}}$ Hybrid Model:} As ideal functionalities can be UC-realized only in the presence of some trusted setup. One common form of setup is common reference string $({\sf CRS})$. Let us describe the $\mathcal{F}_{{\sf CRS}}^{\mathcal{D}}$-hybrid model \cite{canetti2002universally} that UC realizes a protocol parameterized by some specific distribution $\mathcal{D}$. Upon receiving a message $({\sf CRS})$, from any party, $\mathcal{F}_{{\sf CRS}}^{\mathcal{D}}$ first checks if there is a recorded value ${\sf crs}$. If not, $\mathcal{F}_{{\sf CRS}}^{\mathcal{D}}$ generates ${\sf crs} \leftarrow \mathcal{D}(1^{\rho})$ and records it. Finally, $\mathcal{F}_{{\sf CRS}}^{\mathcal{D}}$ sends ${\sf crs}$ to the party and the adversary. In the proposed construction, the distribution $\mathcal{D}$ is ${\sf CrsSetup}$ algorithm, i.e, ${\sf crs} \leftarrow {\sf CrsSetup}(1^{\rho})$. The algorithm outputs ${\sf crs}$ and is made public. Let us briefly discuss how the parties interacts in the real world and in the ideal world. We discuss an interaction between the sender $S$ and a receiver $R$. \noindent \textbf{Real World:} Upon receiving messages $({\sf initdb}, {\sf DB} = \{m_i$, $W_i\}_{1\leq i\leq N})$, $({\sf issue}, L)$ and $({\sf transfer}, \sigma_j)$ from $\mathcal{Z}$, how $S$ and $R$ behave are briefly discussed below, where $j = 1, 2, \ldots, k$. \begin{itemize} \item The sender $S$ upon receiving a message $({\sf initdb}, {\sf DB} = \{m_i, W_i\}_{1\leq i\leq N})$ runs ${\sf DBSetup}$ algorithm with input ${\sf crs}$ and ${\sf DB}$. In ${\sf DB} = \{m_i, W_i\}_{1\leq i\leq N}$, each $m_i$ is associated with access policy $W_i$. The algorithm outputs public/secret key pair $({\sf pk}, {\sf sk})$, a proof $\psi$ which shows the validity of public/secret key pair and ciphertext database ${\sf cDB} = \{\phi_i\}_{1\leq i\leq N}$. The access policy $W_i$ is embedded implicitly in $\phi_i$. The sender $S$ publishes ${\sf pk}, \psi, {\sf cDB}$ to all parties and keeps ${\sf sk}, \{W_i\}_{1\leq i\leq N}$ secret to itself. \item The receiver $R$ upon receiving a message $({\sf issue}, L)$ from $\mathcal{Z}$ engages in ${\sf Issue}$ protocol with $S$. The receiver $R$ first randomizes its attribute set $L$ and generates randomized attributes set $L'$. The sender $S$ upon receiving $L'$ from $R$, generates randomized attribute secret key ${\sf ASK'}$ and gives ${\sf ASK'}$ to $R$. The receiver $R$ extracts ${\sf ASK}$ from ${\sf ASK'}$ using random values which were used to randomize the set $L$. \item Upon receiving a message $({\sf transfer}, \sigma_j)$ from $\mathcal{Z}$, $R$ invokes ${\sf Transfer}$ protocol. The receiver $R$ takes ciphertext $\phi_{\sigma_j}$ from ${\sf cDB}$ and generates ${\sf Req}_{\sigma_j}, {\sf Pri}_{\sigma_j}, \pi_{\sigma_j}$. The request ${\sf Req}_{\sigma_j}$ is the randomization of $\phi_{\sigma_j}$ and is given to $S$. The proof $\pi_{\sigma_j}$ convinces $S$ that $R$ is decrypting a valid ciphertext $\phi_{\sigma_j} \in {\sf cDB}$. The receiver $R$ gives ${\sf Req}_{\sigma_j}, \pi_{\sigma_j}$ to $S$ and keeps ${\sf Pri}_{\sigma_j}$ secret to itself. The sender $S$ generates response ${\sf Res}_{\sigma_j}$ using secret key ${\sf sk}$. The response ${\sf Res}_{\sigma_j}$ and proof $\delta_{\sigma_j}$ is given to $R$ by $S$. The proof $\delta_{\sigma_j}$ convinces $R$ that $S$ has used the same secret key ${\sf sk}$ which was used in generating ${\sf cDB}$. Using ${\sf Res}_{\sigma_j}$, $R$ correctly decrypts $\phi_{\sigma_j}$ and recovers $m_{\sigma_j}$ if $R$'s attribute set $L$ implicitly satisfies $W_{\sigma_j}$. Otherwise, $R$ obtains a random message. \end{itemize} \noindent \textbf{Ideal World:} Upon receiving messages $({\sf initdb}, {\sf DB} = \{m_i$, $W_i\}_{1\leq i\leq N})$, $({\sf issue}, L)$ and $({\sf transfer}, \sigma_j)$ from $\mathcal{Z}$, the dummy parties forward them to the ideal functionality $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$. The ideal functionality $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ keeps an empty database ${\sf DB}$ and an empty attribute set $L_R$ for each receiver $R$. Let us briefly discuss the actions of $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$. \begin{itemize} \item Upon receiving message $({\sf initdb}, {\sf DB} = \{m_i, W_i\}_{1\leq i\leq N})$ from $S$, $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ updates ${\sf DB} = \{m_i, W_i\}_{1\leq i\leq N}$. \item The ideal functionality $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ upon receiving message $({\sf issue}, L)$ from $R$, sends $({\sf issue})$ to $S$ and receives a bit $b \in \{0, 1\}$ in return from $S$. If $b=1$, $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ updates $L_R = L$. Otherwise, $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ does nothing. \item Upon receiving message $({\sf transfer}, \sigma_j)$ from $R$, $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ sends $({\sf transfer})$ to $S$, who in turn returns a bit $b \in \{0, 1\}$. If $b=1$, $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ checks whether $\sigma_j \in \{1, 2$, $\ldots, N\}$ and $L_R$ satisfies $W_{\sigma_j}$. If yes, $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ gives $m_{\sigma_j}$ to $R$. Otherwise, $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ sends a null string $\bot$ to $R$. \end{itemize} \section{Protocol}\label{protocoldescription} Our ${\sf AOT\mhyphen HAP} = ({\sf CRSSetup}, {\sf DBSetup}, {\sf Issue}, {\sf Transfer})$ couples Boneh-Boyen (BB) signature \cite{boneh2004shortsignature} and ciphertext policy attribute based encryption (CP-ABE) of Nishide {\em et al.} \cite{nishide2008attribute}. In addition, Groth-Sahai \cite{groth2008efficient} proofs for non-interactive verification of pairing product equations are employed. The ${\sf AOT\mhyphen HAP}$ involves a sender and multiple receivers. We consider an interaction between a sender $S$ with a database ${\sf DB} = \{m_i, W_i\}_{1 \leq i\leq N}$ and a receiver $R$ with attribute list $L = [v_{1, t_1}, v_{2, t_2}, \ldots, v_{n, t_n}]$. As a protocol can be universal composable (UC) realized only under some trusted setup assumptions, algorithm ${\sf CRSSetup}$ is invoked to generate a common reference string ${\sf crs}$ which is made public to all parties. In initialization phase, S executes algorithm ${\sf DBSetup}$. The sender $S$ holds a database ${\sf DB} = \{m_i, W_i\}_{1 \leq i \leq N}$, where $W_i$ is the access policy associated with $m_i$, $1 \leq i \leq N$. Each message $m_i$ is encrypted to generate ciphertext $\phi_i$ using CP-ABE of \cite{nishide2008attribute} under $W_i$. The access policy $W_i$ is embedded implicitly in $\phi_i$ and is not made public. The BB signature is used to sign the random value which was used to encrypt $m_i$. In Issue phase, the receiver $R$ with the attribute list $L$ interacts with $S$ to get attribute secret key ${\sf ASK}$. To decrypt the ciphertext $\phi_{\sigma_j}$, $R$ engages in ${\sf Transfer}$ protocol with $S$ and extracts $m_{\sigma_j}$ if $L \models W_{\sigma_j}$, otherwise, $R$ gets a random message. Formal description of ${\sf CRSSetup}, {\sf DBSetup}, {\sf Issue}, {\sf Transfer}$ are given below. Let $\mathbb{A}_1, \mathbb{A}_2, \ldots, \mathbb{A}_n$ be $n$ attributes and each attribute $\mathbb{A}_{\ell}$ can take $n_{\ell}$ values, namely, $\{v_{\ell, 1}, v_{\ell, 2}, \ldots, v_{\ell, n_{\ell}}\}$ as explained in section \ref{access structure}, $\ell = 1, 2, \ldots, n$. \begin{algorithm}[H] \caption {~~${\sf CrsSetup}$}\label{} \begin{algorithmic}[1] \Require {$\rho, n, n_1, n_2, \ldots, n_n$.} \Ensure {${\sf crs}$.} \State {Generate ${\sf Assparams} \leftarrow {\sf AssBilinearSetup}(1^{\rho})$}; \For {$(\ell =0, 1, \ldots, n)$} \For {$(t =0, 1, \ldots, n_{\ell})$} \State {Choose $a_{\ell, t} \xleftarrow{\$} \mathbb{Z}_{p}$}; \State {Set $A_{\ell, t} = g_{1}^{a_{\ell, t}}, B_{\ell, t} = g_{2}^{a_{\ell, t}}$}; \EndFor \EndFor \State {Take $\xi_1, \xi_2, \widetilde{a}, \widetilde{b}, \widehat{\xi}_1, \widehat{\xi}_2, \widehat{a}, \widehat{b} \xleftarrow{\$} \mathbb{Z}_{p}$}; \State {Set $u_1 = (g_{1}^{\widetilde{a}}, 1, g_1), u_2 = (1, g_{1}^{\widetilde{b}}, g_1)$, $u_3 = u_{1}^{\xi_1}u_{2}^{\xi_2} = (g_{1}^{\widetilde{a}\xi_1}, g_{1}^{\widetilde{b}\xi_2}, g_{1}^{\xi_1+\xi_2})$}; \State {Set $v_1 = (g_{2}^{\widetilde{a}}, 1, g_{2}), v_2 = (1, g_{2}^{\widetilde{b}}, g_{2}),$ $v_3 = v_{1}^{\xi_1}v_{2}^{\xi_2} = (g_{2}^{\widetilde{a}\xi_1}, g_{2}^{\widetilde{b}\xi_2}, g_{2}^{\xi_1+\xi_2})$}; \State {$\widehat{u}_1 = (g_{1}^{\widehat{a}}, 1, g_1), \widehat{u}_2 = (1, g_{1}^{\widehat{b}}, g_1),$ $\widehat{u}_3 = \widehat{u}_{1}^{\widehat{\xi}_1}\widehat{u}_{2}^{\widehat{\xi}_2} = (g_{1}^{\widehat{a}\widehat{\xi}_1}, g_{1}^{\widehat{b}\widehat{\xi}_2}, g_{1}^{\widehat{\xi}_1+\widehat{\xi}_2})$}; \State {$\widehat{v}_1 = (g_{2}^{\widehat{a}}, 1, g_2), \widehat{v}_2 = (1, g_{2}^{\widehat{b}}, g_2),$ $\widehat{v}_3 = \widehat{v}_{1}^{\widehat{\xi}_1}\widehat{v}_{2}^{\widehat{\xi}_2} = (g_{2}^{\widehat{a}\widehat{\xi}_1}, g_{2}^{\widehat{b}\widehat{\xi}_2}, g_{2}^{\widehat{\xi}_1+\widehat{\xi}_2})$}; \State {${\sf GS}_S = (u_1, u_2, u_3, v_1, v_2, v_3), {\sf GS}_R = (\widehat{u}_1, \widehat{u}_2, \widehat{u}_3, \widehat{v}_1, \widehat{v}_2, \widehat{v}_3)$}; \State {${\sf crs} = ({\sf Assparams}, {\sf GS}_S, {\sf GS}_R, \{A_{\ell, t}, B_{\ell, t}\}_{1 \leq t \leq n_{\ell}, 1 \leq \ell \leq n})$}; \end{algorithmic} \end{algorithm} \noindent \textbf{-- ${\sf CrsSetup}(1^{\rho})$.} The algorithm generates common reference string ${\sf crs}$ on input security parameter $\rho$, number of attributes $n$, number of values $n_{\ell}$ that each attribute $\mathbb{A}_{\ell}$ can take. The algorithm generates bilinear pairing groups by invoking algorithm ${\sf BilinearSetup}$ given in section \ref{preliminaries}. For each attribute value $v_{\ell, t}$, the algorithm picks $a_{\ell, t}$ randomly from $\mathbb{Z}_{p}$, sets $A_{\ell, t} = g_{1}^{a_{\ell, t}}, B_{\ell, t} = g_{2}^{a_{\ell, t}}$, for $t = 1, 2, \ldots, n_{\ell}$, $\ell =1, 2, \ldots, n$. The algorithm generates the Groth-Sahai common reference strings ${\sf GS}_S = (u_1, u_2, u_3, v_1, v_2, v_3)$ for $S$ and ${\sf GS}_R = (\widehat{u}_1, \widehat{u}_2, \widehat{u}_3, \widehat{v}_1, \widehat{v}_2, \widehat{v}_3)$ for $R$ in perfectly sound setting by choosing $\xi_1, \xi_2, \widetilde{a}, \widetilde{b}, \widehat{\xi}_1, \widehat{\xi}_2, \widehat{a}, \widehat{b}$ randomly from $\mathbb{Z}_{p}$ and setting $u_1 = (g_{1}^{\widetilde{a}}, 1, g_1), u_2 = (1, g_{1}^{\widetilde{b}}, g_1), u_3 = (g_{1}^{\widetilde{a}\xi_1}, g_{1}^{\widetilde{b}\xi_2}, g_{1}^{\xi_1+\xi_2})$, $v_1 = (g_{2}^{\widetilde{a}}, 1, g_{2}), v_2 = (1, g_{2}^{\widetilde{b}}, g_{2}), v_3 = (g_{2}^{\widetilde{a}\xi_1}, g_{2}^{\widetilde{b}\xi_2}, g_{2}^{\xi_1+\xi_2})$, $\widehat{u}_1 = (g_{1}^{\widehat{a}}, 1, g_1), \widehat{u}_2 = (1, g_{1}^{\widehat{b}}, g_1), \widehat{u}_3 = (g_{1}^{\widehat{a}\widehat{\xi}_1}, g_{1}^{\widehat{b}\widehat{\xi}_2}, g_{1}^{\widehat{\xi}_1+\widehat{\xi}_2})$, $\widehat{v}_1 = (g_{2}^{\widehat{a}}, 1, g_2), \widehat{v}_2 = (1, g_{2}^{\widehat{b}}, g_2), \widehat{v}_3 = (g_{2}^{\widehat{a}\widehat{\xi}_1}, g_{2}^{\widehat{b}\widehat{\xi}_2}, g_{2}^{\widehat{\xi}_1+\widehat{\xi}_2})$. It outputs ${\sf crs} = ({\sf params}, {\sf GS}_S, {\sf GS}_R$, $\{\{A_{\ell, t}, B_{\ell, t}\}_{1 \leq t \leq n_{\ell}}\}_{1 \leq \ell \leq n})$ to all parties. \begin{algorithm}[H] \caption {~~${\sf DBSetup}$}\label{} \begin{algorithmic}[1] \Require {${\sf crs}, {\sf DB} = \{(m_i, W_i)\}_{1\leq i \leq N}$.} \Ensure {${\sf pk}, {\sf sk}, \psi, {\sf cDB} = \{\phi_i\}_{1\leq i \leq N}$.} \State {Randomly choose $w, x, \alpha, \beta, \gamma \xleftarrow{\$} \mathbb{Z}_{p}$}; \State {Compute $B = g_{1}^{\beta}, y = g_{1}^{x}, h = g_{2}^{\gamma}, Y = e(g_1, g_{2}^{w}),$ $H = e(B, h), B^{\gamma}, g_{1}^{w}, g_{1}^{\alpha}, g_{2}^{\alpha}$}; \State {Generate ${\sf Com'}(h)$ using ${\sf GS}_R$}; \State {Set public key ${\sf pk} = (B, y, Y, H$, ${\sf Com'}(h))$}; \State {Set secret key ${\sf sk} = (x,\alpha, \beta, \gamma, w, h, B^{\gamma}, g_{1}^{w}, g_{2}^{w}, g_{1}^{\alpha}, g_{2}^{\alpha})$}; \State {The proof $\psi = {\sf NIZK}_{{\sf GS}_S}\{(B^{\gamma}, g_{1}^{w}, g_{2}^{w}, h, g_{1}^{\alpha}, g_{2}^{\alpha}, g')~|~ $ $e(g_{1}^{-1}, g_{2}^{w})e(g_{1}^{w}, g') =1 \wedge e(B^{-1}, h)e(B^{\gamma}, g') = 1 \wedge e(g_{1}^{-1}, g_{2}^{\alpha})e(g_{1}^{\alpha}, g')=1 \wedge e(g_1, g') = e(g_1, g_2)\}$}; \For{$(i =1, 2, \ldots, N)$} \State {$s_{i, 1}, s_{i,2}, \ldots, s_{i,n} \xleftarrow{\$} \mathbb{Z}_{p}$}; \State {Set $r_i = s_{i, 1} + s_{i,2} + \ldots + s_{i,n} $}; \State {$c_{i}^{(1)} = g_{2}^{\frac{1}{x+r_i}}$}; \State {$c_{i}^{(2)} = B^{r_i}$}; \State {$c_{i}^{(3)} = m_i\cdot Y^{r_i}\cdot e(c_{i}^{(2)}, h)$}; \For {$(\ell =1, 2, \ldots, n)$} \State {$c_{i, \ell}^{(4)} = g_{1}^{s_{i, \ell}}$} \For {$(t =1, 2, \ldots, n_{\ell})$} \If {$(v_{\ell, t} \in W_{i, \ell})$} \State {$c_{i, \ell, t}^{(5)} = A_{\ell, t}^{\alpha s_{i, \ell}}$, where $A_{\ell, t}$ is extracted from ${\sf crs}$}; \Else \State {$c_{i, \ell, t}^{(5)} = g_{1}^{z_{i, \ell, t}}, z_{i, \ell, t}\xleftarrow{\$} \mathbb{Z}_{p}$}; \EndIf \EndFor \EndFor \State {Set $c_{i}^{(4)} = \{c_{i, \ell}^{(4)}\}_{1\leq \ell \leq n}, c_{i}^{(5)} = \{\{c_{i, \ell, t}^{(5)}\}_{1\leq t\leq n_{\ell}}\}_{1\leq \ell \leq n}$}; \State {$\phi_i = (c_{i}^{(1)}, c_{i}^{(2)}, c_{i}^{(3)}, c_{i}^{(4)}, c_{i}^{(5)})$}; \EndFor \State {${\sf cDB} = (\phi_1, \phi_2, \ldots, \phi_N)$}; \end{algorithmic} \end{algorithm} \noindent \textbf{-- ${\sf DBSetup}({\sf crs}, {\sf DB} = \{(m_i, W_i)\}_{1\leq i \leq N})$.} This algorithm is executed by $S$ with input ${\sf crs}$, database ${\sf DB} = \{(m_i$, $W_i)\}_{1\leq i \leq N}$, where $m_i \in \mathbb{G}_T$, $W_i = [W_{i,1}, W_{i,2}, \ldots, W_{i,n}]$ is the access policy associated with $m_i$, $W_{i, \ell} \subseteq S_{\ell}$ as discussed in section \ref{access structure}, $i =1, 2, \ldots, N$, $\ell =1, 2, \ldots, n$. The algorithm randomly takes $w, x, \alpha, \beta, \gamma$ from $\mathbb{Z}_{p}$, computes $B = g_{1}^{\beta}, y = g_{1}^{x}, h = g_{2}^{\gamma}, Y = e(g_1, g_{2}^{w}), H = e(B, h), B^{\gamma}, g_{1}^{w}$, $g_{1}^{\alpha}, g_{2}^{\alpha}$. Additionally, it computes ${\sf Com'}(h)$ using ${\sf GS}_R$ as in section \ref{noninteractive}. The algorithm sets public key ${\sf pk} = (B, y, Y, H$, ${\sf Com'}(h))$, secret key ${\sf sk} = (x,\alpha, \beta, \gamma, w, h, B^{\gamma}, g_{1}^{w}, g_{2}^{w}, g_{1}^{\alpha}, g_{2}^{\alpha})$. The proof \begin{gather*} \psi = {\sf NIZK}_{{\sf GS}_S} \{(B^{\gamma}, g_{1}^{w}, g_{2}^{w}, h, g_{1}^{\alpha}, g_{2}^{\alpha}, g')~|~ e(g_{1}^{-1}, g_{2}^{w})e(g_{1}^{w}, g') =1\wedge e(B^{-1}, h)e(B^{\gamma}, g') = 1 \wedge \\ (g_{1}^{-1}, g_{2}^{\alpha})e(g_{1}^{\alpha}, g')=1 \wedge e(g_1, g') = e(g_1, g_2)\} \end{gather*} is generated to convince $R$ that $S$ knows secrets $g_{2}^{w}, g_{2}^{\alpha}, h$. The proof $\psi$ consists of proof components of equations $e(g_{1}^{-1}, g_{2}^{w})e(g_{1}^{w}, g') =1\wedge e(B^{-1}, h)e(B^{\gamma}$, $g')$ $= 1 \wedge (g_{1}^{-1}, g_{2}^{\alpha})e(g_{1}^{\alpha}, g')=1\wedge e(g_1, g') = e(g_1, g_2)$ and commitments of witnesses $B^{\gamma}, g_{1}^{w}, g_{2}^{w}, h, g_{1}^{\alpha}, g_{2}^{\alpha}, g'$ generated using ${\sf GS}_S$ as explained in section \ref{noninteractive}. To encrypt $m_i$ under $W_i$, random values $s_{i, 1}, s_{i,2}, \ldots, s_{i,n}$ are selected from $ \mathbb{Z}_{p}$. The algorithm sets $r_i = s_{i, 1} + s_{i,2} + \ldots + s_{i,n}$ and signs $r_i$ using Boneh-Boyen (BB) signature \cite{boneh2004shortsignature} as shown in line 10 of Algorithm 2. After that, $m_i$ is encrypted using CP-ABE of \cite{nishide2008attribute}. \noindent In ciphertext $\phi_i = (c_{i}^{(1)}, c_{i}^{(2)}, c_{i}^{(3)}, c_{i}^{(4)}, c_{i}^{(5)})$, $c_{i}^{(1)} = g_{2}^{\frac{1}{x+r_i}}$ is the signature on $r_i$, $c_{i}^{(2)}, c_{i}^{(3)}, c_{i}^{(4)}, c_{i}^{(5)}$ is the CP-ABE of $m_i$, $i =1, 2, \ldots, N$. The algorithm outputs ${\sf pk}, {\sf sk}, \psi$ and ciphertext database ${\sf cDB} = \{\phi_i\}_{1\leq i \leq N}$ to $S$. The sender $S$ publishes ${\sf pk}, \psi, {\sf cDB} = \{\phi_i\}_{1\leq i \leq N}$ to all parties and keeps ${\sf sk}$ hidden. \noindent \textbf{Correctness of ${\sf cDB}$:} Anyone can verify the correctness of ${\sf cDB} = \{\phi_i\}_{1\leq i \leq N}$, $\phi_i = (c_{i}^{(1)}, c_{i}^{(2)}, c_{i}^{(3)}$, $c_{i}^{(4)} = \{c_{i, \ell}^{(4)}\}_{1\leq \ell \leq n}, c_{i}^{(5)})$ by verifying the following equations $$e\left(\prod_{\ell=1}^{n}c_{i, \ell}^{(4)}\cdot y, c_{i}^{(1)}\right) = e(g_1, g_2) ~~\forall~ i = 1, 2, \ldots, N.$$ \begin{figure} \centering \epsfig{file=issuephase.eps, height=4in, width=6in,} \caption{Issue Phase.} \label{fig1} \end{figure} \noindent \textbf{-- ${\sf Issue}$ Protocol$({\sf crs}, L, {\sf sk})$.} The receiver $R$ invokes ${\sf Issue}$ protocol, given in Figure \ref{fig1}, to obtain attribute secret key ${\sf ASK}$ for its attributes list $L = [v_{1, t_1}, v_{2, t_2}, \ldots, v_{n, t_n}]$, where $v_{\ell, t_{\ell}} \in S_{\ell}, \ell =1, 2, \ldots, n$ as given in section \ref{access structure}. To achieve this, $R$ randomizes $B_{\ell, t_{\ell}}$, extracted from ${\sf crs} = ({\sf params}, {\sf GS}_S$, ${\sf GS}_R$, $\{\{A_{\ell, t}, B_{\ell, t}\}_{1 \leq t \leq n_{\ell}}\}_{1 \leq \ell \leq n})$, corresponding to each $v_{\ell, t_{\ell}}$ from $L$ using random value $z_{\ell}$ and generates randomized set $L'$ as shown in Figure \ref{fig1}, where $t_{\ell} =1, 2, \ldots, n_{\ell}$, $\ell = 1, 2, \ldots, n$. The receiver $R$ sets proof $\varphi = {\sf NIWI}_{{\sf GS}_R}\{(\{A_{\ell, t_{\ell}}$, $T_{\ell}\}_{1 \leq \ell \leq n})~|~ e(g_1, \prod_{\ell=1}^{n}R_{\ell}) = e(\prod_{\ell=1}^{n}A_{\ell,t}$, $g_2)e(\prod_{\ell=1}^{n}T_{\ell}, g_2)\}$ to convince $R$ that $L'$ is the randomization of $L$. The randomized set $L'$ and proof $\varphi$ is given to $S$. If the proof $\varphi$ is valid, $S$ uses secret $w, \alpha, \beta$ from its secret key ${\sf sk} = (x,\alpha, \beta, \gamma, w, h, B^{\gamma}, g_{1}^{w}, g_{2}^{w}, g_{1}^{\alpha}, g_{2}^{\alpha})$ to generate $d_0 = g_{2}^{\frac{w+s}{\beta}}$, $d'_{\ell, 1} = g_{2}^{s}\cdot R_{\ell}^{\lambda_{\ell}}$, $d_{\ell, 2} = g_{2}^{\lambda_{\ell}}$, $d_{\ell, 3} = g_{2}^{\frac{\lambda_{\ell}}{\alpha}}$, where $s, \lambda_{\ell}$ are random values from $\mathbb{Z}_{p}$, $\ell =1, 2, \ldots, n$. The randomized attribute secret key ${\sf ASK'} = (d_0, \{d'_{\ell, 1}, d_{\ell, 2}, d_{\ell, 3}\}_{1\leq \ell \leq n})$ is given to $R$. The receiver $R$ extracts ${\sf ASK}$ from ${\sf ASK'}$ using random values $z_1, z_2, \ldots, z_n$ which were used by $R$ to randomize $L$ and checks the correctness of ${\sf ASK}$ by verifying the equation $$\frac{e(g_1, d_{\ell,1})\cdot Y}{e(B, d_0)} = e(A_{\ell, t_{\ell}}, d_{\ell,2}) ~~~~~\forall~\ell, 1\leq \ell \leq n.$$ \begin{figure*} \centering \epsfig{file=transferphase.eps, height=4in, width=6in,} \caption{Transfer Phase.} \label{fig2} \end{figure*} \noindent \textbf{-- ${\sf Transfer}$ Protocol$({\sf crs}, {\sf ASK}, {\sf pk}, {\sf sk}, \sigma_j, \phi_{\sigma_j})$.} The receiver $R$ engages in ${\sf Transfer}$ protocol with $S$ to decrypt the ciphertext $\phi_{\sigma_j} = (c_{\sigma_j}^{(1)}, c_{\sigma_j}^{(2)}, c_{\sigma_j}^{(3)}, c_{\sigma_j}^{(4)}, c_{\sigma_j}^{(5)})$ as shown in Figure \ref{fig2}. For this, $R$ randomizes $c_{\sigma_j}^{(2)} = B^{r_{\sigma_j}}$ using random value $v_{\sigma_j}$ from $\mathbb{Z}_{p}$ and generates request ${\sf Req}_{\sigma_j} = c_{\sigma_j}^{(2)}\cdot B^{v_{\sigma_j}} = B^{r_{\sigma_j} + v_{\sigma_j}}$, private ${\sf Pri}_{\sigma_j} = H^{v_{\sigma_j}}$ and sets proof $\pi_{\sigma_j}$ $= {\sf NIWI}_{{\sf GS}_R}\{(c_{\sigma_j}^{(1)}, \prod_{\ell=1}^{n}c_{\sigma_j, \ell}^{(4)}, g_{2}^{v_{\sigma_j}}, c_{\sigma_j}^{(2)})~|~$ $e\left(\prod_{\ell=1}^{n}c_{\sigma_j, \ell}^{(4)}\cdot y, c_{\sigma_j}^{(1)}\right)$ $= e(g_1, g_2) \wedge e({\sf Req}_{\sigma_j}, g_2) = e(c_{\sigma_j}^{(2)}, g_2)e(B, g_{2}^{v_{\sigma_j}})\}$. The first equation $e(\prod_{\ell=1}^{n}c_{\sigma_j, \ell}^{(4)}\cdot y, c_{\sigma_j}^{(1)})$ $= e(g_1, g_2)$ in $\pi_{\sigma_j}$ guarantees $S$ that $\phi_{\sigma_j}$ is a valid ciphertext, i.e, $\phi_{\sigma_j} \in {\sf cDB}$, and 2nd equation shows that ${\sf Req}_{\sigma_j}$ is the randomization of $c_{\sigma_j}^{(2)}$. The receiver also raises power $v_{\sigma_j}$ to ${\sf Com'}(h)$ which was given by $S$ in initialization phase generated using ${\sf GS}_R$. This extra component does not give any advantage to adversary and is useful only for simulator during security analysis. The sender $S$ checks the validity of $\pi_{\sigma_j}$. If invalid, $S$ outputs a null string $\bot$. Otherwise, $S$ uses secret $h, \gamma$ from secret key ${\sf sk} = (x,\alpha, \beta, \gamma, w, h, B^{\gamma}, g_{1}^{w}, g_{2}^{w}, g_{1}^{\alpha}, g_{2}^{\alpha})$ to compute response ${\sf Res}_{\sigma_j} = e({\sf Req}_{\sigma_j}, h)$. The proof $\delta_{\sigma_j} = {\sf NIZK}_{{\sf GS}_S}\{(h, {\sf Req}_{\sigma_j}^{\gamma}, B^{\gamma}, g')~|~ e({\sf Req}_{\sigma_j}^{-1}, h)e({\sf Req}_{\sigma_j}^{\gamma}$, $g') = 1 \wedge e(B^{-1}$, $h)e(B^{\gamma}, g') = 1 \wedge e(g_1, g') = e(g_1, g_2)\}$ is generated to guarantee $R$ that $S$ has used the same secrets $h, \gamma$ which were used in initialization phase to set ciphertext database ${\sf cDB}$. Upon receiving ${\sf Res}_{\sigma_j}, \delta_{\sigma_j}$, $R$ first verifies $\delta_{\sigma_j}$. If fails, $R$ outputs a null string $\bot$. Otherwise, $R$ sets $\widetilde{c}_{\sigma_j, \ell}^{(5)} = c_{\sigma_j, \ell, t_{\ell}}^{(5)}$ if $v_{\ell, t_{\ell}} \in L$ for $\ell =1, 2, \ldots, n$ and computes \vspace{-.3cm} \begin{equation}\label{eq4} \frac{c_{\sigma_j}^{(3)}\cdot \prod_{\ell=1}^{n} e(c_{\sigma_j, \ell}^{(4)}, d_{\ell, 1})\cdot H^{v_{\sigma_j}}}{e(c_{\sigma_j}^{(2)}, d_0)\cdot \prod_{\ell=1}^{n} e(\widetilde{c}_{\sigma_j, \ell}^{(5)}, d_{\ell, 3})\cdot{\sf Res}_{\sigma_j}} = m_{\sigma_j}. \end{equation} \noindent \textbf{Correctness of equation \ref{eq4}.} \begin{align*} &{\sf val}_1 = \prod_{\ell=1}^{n} e(c_{\sigma_j, \ell}^{(4)}, d_{\ell, 1}) = \prod_{\ell=1}^{n}e(g_{1}^{s_{\sigma_j, \ell}}, g_{2}^{s}\cdot B_{\ell, t_{\ell}}^{\lambda_{\ell}})\\ & = \prod_{\ell=1}^{n}e(g_{1}^{s_{\sigma_j, \ell}},g_{2}^{s + a_{\ell, t_{\ell}}\lambda_{\ell}})\\ &= e(g_1, g_2)^{s\sum_{\ell=1}^{n} s_{\sigma_j, \ell}}e(g_1, g_2)^{\sum_{\ell=1}^{n} s_{\sigma_j, \ell}a_{\ell, t_{\ell}}\lambda_{\ell}} \end{align*} \begin{align*} &= e(g_1, g_2)^{sr_{\sigma_j}}e(g_1, g_2)^{\sum_{\ell=1}^{n} s_{\sigma_j, \ell}a_{\ell, t_{\ell}}\lambda_{\ell}} {\rm ~as~} \sum_{\ell=1}^{n} s_{\sigma_j, \ell} = r_{\sigma_j} \\ &{\sf val}_2 = \prod_{\ell=1}^{n} e(\widetilde{c}_{\sigma_j, \ell}^{(5)}, d_{\ell, 3}) \\ &= \prod_{\ell=1}^{n} e(A_{\ell, t_{\ell}}^{\alpha s_{\sigma_j, \ell}}, g_{2}^{\frac{\lambda_{\ell}}{\alpha}}) = \prod_{\ell=1}^{n}e(g_{1}^{a_{\ell, t_{\ell}}s_{\sigma_j, \ell}}, g_{2}^{\lambda_{\ell}})\\ &=e(g_1, g_2)^{\sum_{\ell=1}^{n} s_{\sigma_j, \ell}a_{\ell, t_{\ell}}\lambda_{\ell}}\\ &\frac{{\sf val}_1}{e(c_{\sigma_j}^{(2)}, d_0){\sf val}_2} = \frac{e(g_1, g_2)^{sr_{\sigma_j}}}{e(g_1, g_2)^{r_{\sigma_j}(w+s)}} = \frac{1}{e(g_1, g_2)^{r_{\sigma_j}w}} = \frac{1}{Y^{r_{\sigma_j}}}\\ &\frac{c_{\sigma_j}^{(3)}\cdot {\sf val}_1\cdot H^{v_{\sigma_j}}}{e(c_{\sigma_j}^{(2)}, d_0)\cdot {\sf val}_2\cdot{\sf Res}_{\sigma_j}} = \frac{m_{\sigma_j} Y^{r_{\sigma_j}}e(B,h)^{r_{\sigma_j}}H^{v_{\sigma_j}}}{Y^{r_{\sigma_j}}e(B, h)^{r_{\sigma_j + v_{\sigma_j}}}} = m_{\sigma_j}, \end{align*} where $c_{\sigma_j}^{(3)} = m_{\sigma_j}\cdot Y^{r_{\sigma_j}}\cdot e(B, h)^{r_{\sigma_j}},H = e(B, h)$, ${\sf Res}_{\sigma_j} = e(B, h)^{r_{\sigma_j + v_{\sigma_j}}}$. \begin{theorem}\label{theorem1} The adaptive oblivious transfer with hidden access policy (${\sf AOT\mhyphen HAP}$) presented in section \ref{protocoldescription} securely realizes the ideal functionality $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ in the $\mathcal{F}_{\sf CRS}^{\mathcal{D}}$-hybrid model described in section \ref{securitymodel} under DLIN, DBDH and $q$-SDH assumptions. \end{theorem} \begin{proof} The ${\sf AOT\mhyphen HAP}$ is run between a sender $S$ and multiple receivers. We simulate an interaction between $S$ and a receiver $R$ with an attribute list $L = [L_1, L_2, \ldots, L_n]$, $L_{\ell} \in S_{\ell}$, $n$ is the total number of attributes, $\ell=1, 2, \ldots, n$, (see section \ref{access structure} for detail definition of $L_{\ell}, S_{\ell}$). The adversarial model adapted in this paper is static corruption model in which adversary decides which party to corrupt before the protocol starts. The corrupted parties remain corrupt, and honest parties remain honest throughout. Therefore, the case of malicious sender and malicious receiver are addressed separately. In both cases, an ideal world adversary $\mathcal{A'}$ corresponding to a real world adversary $\mathcal{A}$ is constructed such that no environment machine or distinguisher $\mathcal{Z}$ can distinguish with non-negligible probability, its interactions with $\mathcal{A}$ and parties running the protocol $\Pi$ in the real world from its interaction with $\mathcal{A'}$ and ideal functionality $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ in the ideal world. The protocol $\Pi$ is said to be secure if the real world is as secure as the ideal world. Since the ideal world consists of incorruptible trusted parties, it is impossible for an adversary to attack the ideal world. Therefore, all adversarial attack must fail in real world too. Technically, for each case we show ${ \sf IDEAL}_{\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}, \mathcal{A'}, \mathcal{Z}} \stackrel{c}{\approx} {\sf REAL}_{\Pi, \mathcal{A}, \mathcal{Z}},$ where ${ \sf IDEAL}_{\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}, \mathcal{A'}, \mathcal{Z}}, {\sf REAL}_{\Pi, \mathcal{A}, \mathcal{Z}}$ are as defined in section \ref{ucmodel}. \noindent The security proof is presented using sequence of hybrid games. Let ${\sf Pr}[{\sf Game}~ i]$ be the probability that $\mathcal{Z}$ distinguishes the transcript of ${\sf Game} ~i$ from the real execution. \noindent \textbf{(a) Security Against Malicious Receiver $R$}. In this case, the real world adversary $\mathcal{A}$ corrupts the receiver $R$ before the execution of the protocol. The ideal world adversary $\mathcal{A'}$ simulates the honest sender $S$ by invoking a copy of $\mathcal{A}$ as follows. \noindent $\bf{\underline{{\sf Game} ~0}}$: This game corresponds to the real world protocol interaction in which the receiver interacts with the honest sender $S$. So, ${\sf Pr}[{\sf Game}~ 0] = 0$. \noindent $\bf{\underline{{\sf Game} ~1}}$: This game is the same as ${\sf Game} ~0$ except that $\mathcal{A'}$ simulates ${\sf crs}$. The adversary $\mathcal{A'}$ simulates ${\sf crs} = ({\sf params}, {\sf GS}_S$, ${\sf GS}_R, \{\{A_{\ell, t}, B_{\ell, t}\}_{1 \leq t \leq n_{\ell}}\}_{1 \leq \ell \leq n})$ exactly as in the above game except that ${\sf GS}_S$ is in witness indistinguishability setting as explained below. The adversary $\mathcal{A'}$, first generates ${\sf params} = (p, \mathbb{G}_1, \mathbb{G}_2, \mathbb{G}_{T}, e, g_1, g_2) \leftarrow {\sf BilinearSetup}(1^{\rho})$ by invoking algorithm ${\sf BilinearSetup}$ discussed in section \ref{preliminaries}. The components $\{\{A_{\ell, t} = g_{1}^{a_{\ell,t}}, B_{\ell, t} = g_{2}^{a_{\ell,t}}\}_{1 \leq t \leq n_{\ell}}\}_{1 \leq \ell \leq n}$ are generated exactly as in ${\sf Game} ~0$. To generate ${\sf GS}_S$ in witness indistinguishability setting and ${\sf GS}_R$ in perfectly sound setting, the algorithm chooses $\xi_1, \xi_2, \widetilde{a}, \widetilde{b}, \widehat{\xi}_1, \widehat{\xi}_2, \widehat{a}, \widehat{b} \xleftarrow{\$} \mathbb{Z}_{p}$ and sets $u_1 = (g_{1}^{\widetilde{a}}, 1, g_{1})$, $u_2 = (1, g_{1}^{\widetilde{b}}, g_{1}), u_3 = u_{1}^{\xi_1}u_{2}^{\xi_2}(1, 1, g_{1}) = (g_{1}^{\widetilde{a}\xi_1}, g_{1}^{\widetilde{b}\xi_2}, g_{1}^{\xi_1+\xi_2+1})$, $v_1 = (g_{2}^{\widetilde{a}}, 1, g_{2}), v_2 = (1, g_{2}^{\widetilde{b}}, g_{2}), v_3 = v_{1}^{\xi_1}v_{2}^{\xi_2}(1, 1, g_{2}) = (g_{2}^{\widetilde{a}\xi_1}$, $g_{2}^{\widetilde{b}\xi_2}, g_{2}^{\xi_1+\xi_2+1})$, $\widehat{u}_1 = (g_{1}^{\widehat{a}}, 1, g_{1}), \widehat{u}_2 = (1, g_{1}^{\widehat{b}}, g_{1}), \widehat{u}_3 = \widehat{u}_{1}^{\widehat{\xi}_1}\widehat{u}_{2}^{\widehat{\xi}_2}$ $= (g_{1}^{\widehat{a}\widehat{\xi}_1}, g_{1}^{\widehat{b}\widehat{\xi}_2}, g_{1}^{\widehat{\xi}_1+\widehat{\xi}_2})$, $\widehat{v}_1 = (g_{2}^{\widehat{a}}, 1, g_{2})$, $\widehat{v}_2 = (1, g_{2}^{\widehat{b}}, g_{2}), \widehat{v}_3 = \widehat{v}_{1}^{\widehat{\xi}_1}\widehat{v}_{2}^{\widehat{\xi}_2} = (g_{2}^{\widehat{a}\widehat{\xi}_1}, g_{2}^{\widehat{b}\widehat{\xi}_2}, g_{2}^{\widehat{\xi}_1+\widehat{\xi}_2})$, ${\sf GS}_S = (u_1, u_2, u_3, v_1, v_2, v_3)$, ${\sf GS}_R = (\widehat{u}_1, \widehat{u}_2, \widehat{u}_3, \widehat{v}_1, \widehat{v}_2, \widehat{v}_3)$, ${\sf \widetilde{t}_{sim}} = (\widetilde{a}, \widetilde{b}, \xi_1, \xi_2)$, ${\sf \widehat{t}_{ext}} = (\widehat{a}, \widehat{b}$, $\widehat{\xi}_1, \widehat{\xi}_2)$. The adversary $\mathcal{A'}$ distributes ${\sf crs} = ({\sf params}, {\sf GS}_S$, ${\sf GS}_R, \{\{A_{\ell, t}$, $B_{\ell, t}\}_{1 \leq t \leq n_{\ell}}\}_{1 \leq \ell \leq n})$ to all parties and keeps ${\sf \widetilde{t}_{sim}}$, ${\sf \widehat{t}_{ext}}$, $\{\{a_{\ell, t}\}_{1 \leq t \leq n_{\ell}}\}_{1 \leq \ell \leq n}$ hidden. \noindent The common reference string ${\sf crs}$ simulated by $\mathcal{A'}$ and ${\sf crs}$ generated by ${\sf CrsSetup}$ in actual protocol run are computationally indistinguishable by Theorem \ref{crs} in section \ref{noninteractive}. Therefore, there exists a negligible function $\epsilon_{1}(\rho)$ such that $|{\sf Pr}[{\sf Game}~ 1] - {\sf Pr}[{\sf Game}~ 0]| \leq \epsilon_{1}(\rho)$. \noindent $\bf{\underline{{\sf Game} ~2}}$: This game is exactly the same as ${\sf Game} ~1$ except that $\mathcal{A'}$ extracts attribute list $L = [v_{1, t_1}, v_{2, t_2}, \ldots, v_{n, t_n}]$ from $L' = \{R_{\ell}\}_{1\leq\ell \leq n}, ~\varphi$ sent by $\mathcal{A}$ in ${\sf Issue}$ phase as follows. The proof $\varphi = {\sf NIWI}_{{\sf GS}_R}\{(\{A_{\ell, t_{\ell}}, T_{\ell}\}_{1\leq\ell \leq n})~|~ e(g_1, \prod_{\ell=1}^{n}R_{\ell})$ $= e(\prod_{\ell=1}^{n}A_{\ell,t}, g_2)e(\prod_{\ell=1}^{n}T_{\ell}, g_2)\}$ was constructed by $\mathcal{A}$ which consists of commitments to $\{A_{\ell, t_{\ell}}, T_{\ell}\}_{1\leq\ell \leq n}$ and proof components of pairing equation $e(g_1, \prod_{\ell=1}^{n}R_{\ell}) = e(\prod_{\ell=1}^{n}A_{\ell,t}, g_2)$ $e(\prod_{\ell=1}^{n}T_{\ell}, g_2)$ using ${\sf GS}_R = (\widehat{u}_1, \widehat{u}_2, \widehat{u}_3, \widehat{v}_1, \widehat{v}_2$, $\widehat{v}_3)$ extracted from ${\sf crs}$, simulated by $\mathcal{A'}$ in ${\sf Game} ~1$. The adversary $\mathcal{A'}$ uses trapdoor ${\sf \widehat{t}_{ext}} = (\widehat{a}, \widehat{b}, \widehat{\xi}_1, \widehat{\xi}_2)$ to extract $\{A_{\ell, t_{\ell}}, T_{\ell}\}_{1\leq\ell \leq n}$ from their respective commitments as explained below. For instance, let ${\sf Com}(A_{\ell, t_{\ell}}) = \mu_1(A_{\ell, t_{\ell}})\widehat{u}_{1}^{\widehat{r}_1}\widehat{u}_{2}^{\widehat{r}_2}\widehat{u}_{3}^{\widehat{r}_3} = (X$ $= g_{1}^{\widehat{a}(\widehat{r}_{1} + \widehat{\xi}_{1}\widehat{r}_{3})}$, $Y = g_{1}^{\widehat{b}(\widehat{r}_{2} + \widehat{\xi}_{2}\widehat{r}_{3})}, Z = A_{\ell, t_{\ell}}\cdot g_{1}^{\widehat{r}_{1} + \widehat{r}_{2} + \widehat{r}_{3}(\widehat{\xi}_{1} + \widehat{\xi}_{2})})$, where $\widehat{r}_1, \widehat{r}_2, \widehat{r}_3$ are random values from $\mathbb{Z}_p$ which were used to generate ${\sf Com}(A_{\ell, t_{\ell}})$. The adversary $\mathcal{A'}$ computes $$\frac{Z}{X^{\frac{1}{\widehat{a}}}Y^{\frac{1}{\widehat{b}}}} = \frac{A_{\ell, t_{\ell}}\cdot g_{1}^{\widehat{r}_{1} + \widehat{r}_{2} + \widehat{r}_{3}(\widehat{\xi}_{1} + \widehat{\xi}_{2})}}{(g_{1}^{\widehat{a}(\widehat{r}_{1} + \widehat{\xi}_{1}\widehat{r}_{3})})^{\frac{1}{\widehat{a}}}(g_{1}^{\widehat{b}(\widehat{r}_{2} + \widehat{\xi}_{2}\widehat{r}_{3})})^{\frac{1}{\widehat{b}}}} = A_{\ell, t_{\ell}}.$$ Similarly, $\mathcal{A'}$ extracts all the witnesses $\{A_{\ell, t_{\ell}}, T_{\ell}\}_{1 \leq \ell \leq n}$. Let $L$ be the set of attribute values which is initially empty. The adversary $\mathcal{A'}$ checks whether $A_{\zeta, \varsigma} = A_{\ell, t_{\ell}}$ for $\varsigma = 1$ to $n_{\zeta}$, $\zeta = 1$ to $n$. If yes, set $L = L \cup \{v_{\ell, t_{\ell}}\}$. In this way, $\mathcal{A'}$ recovers $L = [v_{1, t_1}, v_{2, t_2}, \ldots, v_{n, t_n}]$ and simulates $d_0$, $\{d'_{\ell, 1}, d_{\ell, 2}, d_{\ell, 3}\}_{1\leq\ell \leq n}$ exactly as in above game. \noindent The randomized attribute secret key ${\sf ASK'}$ generated in ${\sf Game} ~2$ by $\mathcal{A'}$ is distributed identically to ${\sf ASK'}$ honestly generated by the sender $S$. Hence, $|{\sf Pr}[{\sf Game}~ 2] - {\sf Pr}[{\sf Game}~ 1]| = 0$. \noindent $\bf{\underline{{\sf Game} ~3}}$: This game is exactly the same as ${\sf Game} ~2$ except that $\mathcal{A'}$ extracts the index $\sigma_j$ from the proof $\pi_{\sigma_j} = {\sf NIWI}_{{\sf GS}_R}\{(c_{\sigma_j}^{(1)}, \prod_{\ell=1}^{n}c_{\sigma_j, \ell}^{(4)}, g_{2}^{v_{\sigma_j}}, c_{\sigma_j}^{(2)})~|~ ~e(\prod_{\ell=1}^{n}c_{\sigma_j, \ell}^{(4)}\cdot y,c_{\sigma_j}^{(1)}) = e(g_1, g_2) \wedge e({\sf Req}_{\sigma_j}, g_2) = e(c_{\sigma_j}^{(2)}, g_2)e(B, g_{2}^{v_{\sigma_j}})\}$ by extracting witnesses ${\sf wit}_1 = c_{\sigma_j}^{(1)}$, ${\sf wit}_2 = \prod_{\ell=1}^{n}c_{\sigma_j, \ell}^{(4)}$, ${\sf wit}_3 = g_{2}^{v_{\sigma_j}}$, ${\sf wit}_4 = c_{\sigma_j}^{(2)}$ using ${\sf \widehat{t}_{ext}}$ exactly as $\mathcal{A'}$ has extracted $A_{\ell, t_{\ell}}$ in ${\sf Game} ~2$. The adversary $\mathcal{A'}$ checks whether $c_{i}^{(1)} = {\sf wit}_1$, $\prod_{\ell=1}^{n}c_{i, \ell}^{(4)} = {\sf wit}_2$, $c_{i}^{(2)} = {\sf wit}_4$ for $i =1, 2, \ldots, N$. Let $\sigma_j$ be the matching index. The adversary $\mathcal{A'}$ sends the message $({\sf transfer}, \sigma_j)$ to $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$. The ideal functionality returns $m_{\sigma_j}$ to $\mathcal{A'}$ if $L$ satisfies the access policy $W_{\sigma_j}$ associated with $m_{\sigma_j}$, otherwise, $\bot$ is given to $\mathcal{A'}$. If no matching index $\sigma_j$ found and the proof $\pi_{\sigma_j}$ is correct, then this means that $\sigma_j \notin \{1, 2, \ldots, N\}$ and $\pi_{\sigma_j}$ is constructed by $\mathcal{A}$ for the ciphertext $\phi_{\sigma_j} = (c_{\sigma_j}^{(1)}, c_{\sigma_j}^{(2)}, c_{\sigma_j}^{(3)}, c_{\sigma_j}^{(4)}, c_{\sigma_j}^{(5)})$. Eventually, this shows that $\mathcal{A}$ is able to construct a valid BB signature $c_{\sigma_j}^{(1)}$, thereby $\mathcal{A}$ outputs $c_{\sigma_j}^{(1)}$ as a forgery contradicting the fact that the BB signature is unforgeable assuming $q$-SDH problem is hard \cite{boneh2004shortsignature}. Therefore, there exists a negligible function $\epsilon_{3}(\rho)$ such that $|{\sf Pr}[{\sf Game}~ 3] - {\sf Pr}[{\sf Game}~ 2]| \leq \epsilon_{3}(\rho)$. \noindent $\bf{\underline{{\sf Game} ~4}}$: This game is the same as ${\sf Game} ~3$ except that the response ${\sf Res}_{\sigma_j}$ and proof $\delta_{\sigma_j}$ are simulated by $\mathcal{A'}$, for each transfer phase $j = 1, 2, \ldots, k$. To simulate response ${\sf Res}'_{\sigma_j}$, $\mathcal{A'}$ first extracts $h^{v_{\sigma_j}}$ from its commitment $({\sf Com'}(h))^{v_{\sigma_j}}$ using ${\sf \widehat{t}_{ext}}$ as done in ${\sf Game} ~2$. The commitment to $h^{v_{\sigma_j}}$ is generated by $\mathcal{A}$ by raising power $v_{\sigma_j}$ to the commitment of $h$ given by $S$ in initialization phase as \begin{align*} ({\sf Com'}(h))^{v_{\sigma_j}} &= (g_{2}^{\widehat{a}(\widehat{r}_1 + \widehat{r}_3\widehat{\xi}_1)}, g_{2}^{\widehat{b}(\widehat{r}_2 + \widehat{r}_3\widehat{\xi}_2)}, hg^{\widehat{r}_1+\widehat{r}_2+\widehat{r}_3(\widehat{\xi}_1+\widehat{\xi}_2)})^{v_{\sigma_j}}\\ &= (g_{2}^{\widehat{a}v_{\sigma_j}(\widehat{r}_1 + \widehat{r}_3\widehat{\xi}_1)}, g_{2}^{\widehat{b}v_{\sigma_j}(\widehat{r}_2 + \widehat{r}_3\widehat{\xi}_2)}, h^{v_{\sigma_j}}g^{v_{\sigma_j}(\widehat{r}_1+\widehat{r}_2+\widehat{r}_3(\widehat{\xi}_1+\widehat{\xi}_2))})\\ &= (g_{2}^{\widehat{a}(\widetilde{r}_1 + \widetilde{r}_3\widehat{\xi}_1)}, g_{2}^{\widehat{b}(\widetilde{r}_2 + \widetilde{r}_3\widehat{\xi}_2)}, h^{v_{\sigma_j}}g^{\widetilde{r}_1+\widetilde{r}_2+\widetilde{r}_3(\widehat{\xi}_1+\widehat{\xi}_2)}) = {\sf Com'}(h^{v_{\sigma_j}}), \end{align*} where $\widetilde{r}_1 = v_{\sigma_j}\widehat{r}_1, \widetilde{r}_2 = v_{\sigma_j}\widehat{r}_2, \widetilde{r}_3 = v_{\sigma_j}\widehat{r}_3$, $\widehat{r}_1, \widehat{r}_2, \widehat{r}_3 \xleftarrow{\$} \mathbb{Z}_{p}$. The adversary $\mathcal{A'}$ sets $\widetilde{c}_{\sigma_j, \ell}^{(5)} = c_{\sigma_j, \ell, t_{\ell}}^{(5)}$ if $v_{\ell, t_{\ell}} \in L$ for $\ell =1, 2, \ldots, n$ and $${\sf Res}'_{\sigma_j} = \frac{c_{\sigma_j}^{(3)}\cdot \prod_{\ell=1}^{n} e(c_{\sigma_j, \ell}^{(4)}, d_{\ell, 1})\cdot e(B, h^{v_{\sigma_j}})}{e(c_{\sigma_j}^{(2)}, d_0)\cdot \prod_{\ell=1}^{n} e(\widetilde{c}_{\sigma_j, \ell}^{(5)}, d_{\ell, 3})\cdot{m_{\sigma_j}}},$$ where $L = [v_{1, t_1}, v_{2, t_2}, \ldots, v_{n, t_n}]$ and $m_{\sigma_j}$ are extracted in ${\sf Game} ~2$ and ${\sf Game} ~3$, respectively. The proof \begin{gather*} \delta_{\sigma_j} = {\sf NIZK}_{{\sf GS}_S}\{(a_{1, \sigma_j}, a_{2, \sigma_j}, a_{3, \sigma_j}, a_{4, \sigma_j})~|~ e({\sf Req}_{\sigma_j}^{-1}, a_{1, \sigma_j})e(a_{2, \sigma_j}, a_{3, \sigma_j}) = 1 \wedge\\ e(B^{-1}, a_{1, \sigma_j})e(a_{4, \sigma_j}, a_{3, \sigma_j}) = 1 \wedge e(g_1, a_{3, \sigma_j}) = e(g_1, g_2)\}, \end{gather*} where $a_{1, \sigma_j} = h$, $a_{2, \sigma_j} = {\sf Req}_{\sigma_j}^{\gamma}$, $a_{3, \sigma_j} = g_2$, $a_{4, \sigma_j} = B^{\gamma}$. To simulate the proof $\delta_{\sigma_j}$, $\mathcal{A'}$ sets $a_{1, \sigma_j} = a_{2, \sigma_j} = a_{3, \sigma_j} = a_{4, \sigma_j} = g_{2}^{0}$ and generate commitment to all these values. As $\mathcal{A'}$ knows the simulation trapdoor ${\sf \widetilde{t}_{sim}}$, it can open the commitment to any value of its choice as explained in Example \ref{exmp1} in section \ref{noninteractive}. The adversary $\mathcal{A'}$ opens commitment of $a_{3, \sigma_j}$ to $g_{2}^{0}$ in first and 2nd equation of $\delta_{\sigma_j}$ and to $g_{2}^{1}$ in 3rd equation of of $\delta_{\sigma_j}$. Thus, all the equations in $\delta_{\sigma_j}$ are simulated by $\mathcal{A'}$. \noindent \emph{Claim 1. Under the DLIN assumption, the response ${\sf Res}_{\sigma_j}$ and the proof $\delta_{\sigma_j}$ honestly generated by the sender $S$ are computationally indistinguishable from the response ${\sf Res}'_{\sigma_j}$ and proof $\delta'_{\sigma_j}$ simulated by the simulator $\mathcal{A'}$.} \noindent \emph{Proof of Claim 1.} The simulated response ${\sf Res}'_{\sigma_j}$ is \vspace{-.3cm} \begin{eqnarray*} {\sf Res}'_{\sigma_j} &=& \frac{c_{\sigma_j}^{(3)}\cdot \prod_{\ell=1}^{n} e(c_{\sigma_j, \ell}^{(4)}, d_{\ell, 1})\cdot e(B, h^{v_{\sigma_j}})}{e(c_{\sigma_j}^{(2)}, d_0)\cdot \prod_{\ell=1}^{n} e(\widetilde{c}_{\sigma_j, \ell}^{(5)}, d_{\ell, 3})\cdot{m_{\sigma_j}}}\\ &=& \frac{m_{\sigma_j}\cdot Y^{r_{\sigma_j}}\cdot e(B^{r_{\sigma_j}}, h)\cdot e(B, h^{v_{\sigma_j}})}{m_{\sigma_j}\cdot Y^{r_{\sigma_j}}}\\ &=& e(B, h)^{r_{\sigma_j} + v_{\sigma_j}}, \end{eqnarray*} as $\frac{\prod_{\ell=1}^{n} e(c_{\sigma_j, \ell}^{(4)}, d_{\ell, 1})}{e(c_{\sigma_j}^{(2)}, d_0)\cdot \prod_{\ell=1}^{n} e(\widetilde{c}_{\sigma_j, \ell}^{(5)}, d_{\ell, 3})} = \frac{1}{Y^{r_{\sigma_j}}}$and, the honestly generated response ${\sf Res}_{\sigma_j}$ is $${\sf Res}_{\sigma_j} = e({\sf Req}_{\sigma_j}, h) = e(B^{r_{\sigma_j} + v_{\sigma_j}}, h) = e(B, h)^{r_{\sigma_j} + v_{\sigma_j}}.$$ The simulated response ${\sf Res}'_{\sigma_j}$ is distributed identically to honestly generated response ${\sf Res}_{\sigma_j}$. \noindent As Groth-Sahai proofs are composable ${\sf NIZK}$ by Theorem \ref{niwi}, the simulated proof $\delta'_{\sigma_j}$ is computationally indistinguishable from the honestly generated proof $\delta_{\sigma_j}$. Therefore, we have $|{\sf Pr}[{\sf Game}~ 4] - {\sf Pr}[{\sf Game}~ 3]| \leq \epsilon_{4}(\rho)$, where $\epsilon_{4}(\rho)$ is a negligible function. \noindent $\bf{\underline{{\sf Game} ~5}}$: This game is the same as ${\sf Game} ~4$ except that $\mathcal{A'}$ replaces perfect database ${\sf DB} = \{(m_i, W_i)\}_{1 \leq i \leq N}$ by a random database ${\sf DB'} = \{(\widehat{m}_i, \widehat{W}_i)\}_{1 \leq i \leq N}$. The adversary $\mathcal{A'}$ generates $({\sf pk'}, \psi', {\sf cDB'})$, computes response ${\sf Res}_{\sigma_j}$, proof $\delta_{\sigma_j}$. But, in this game the response ${\sf Res}_{\sigma_j}$ is computed on an invalid statement. The proof $\psi'$ is simulated exactly in the same way as $\mathcal{A'}$ simulates $\delta_{\sigma_j}$ in ${\sf Game} ~4$. In ${\sf Game} ~4$, ${\sf cDB}$ is the encryption of perfect database ${\sf DB}$ whereas in ${\sf Game} ~5$, ${\sf cDB'}$ is the encryption of random database ${\sf DB'}$. If the distinguisher $\mathcal{Z}$ can distinguish between the transcript of ${\sf Game} ~5$ from the transcript of ${\sf Game} ~4$, we can construct a solver for DLIN and DBDH using $\mathcal{Z}$ as a subroutine -- a contradiction as CP-ABE of \cite{nishide2008attribute} is semantically secure assuming the hardness of DLIN and DBDH problems. Therefore, ${\sf Game}~ 4$ and ${\sf Game}~ 5$ are computationally indistinguishable and $|{\sf Pr}[{\sf Game}~ 5] - {\sf Pr}[{\sf Game}~ 4]| \leq \epsilon_{5}(\rho)$, where $\epsilon_{5}(\rho)$ is a negligible function. \noindent Thus ${\sf Game} ~5$ is the ideal world interaction whereas ${\sf Game} ~0$ is the real world interaction. Now\\ $ |{\sf Pr}[{\sf Game}~ 5] - [{\sf Game}~ 0]| \leq \sum_{t=1}^{5}|{\sf Pr}[{\sf Game}~ t] - [{\sf Game}~ (t-1)]|\leq \epsilon_{6}(\rho), $ where $\epsilon_{6}(\rho) = \sum_{t=1}^{5}\epsilon_{t}(\rho)$ is a negligible function. Hence, ${\sf IDEAL}_{\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}, \mathcal{A'}, \mathcal{Z}} \stackrel{c}{\approx} {\sf REAL}_{\Pi, \mathcal{A}, \mathcal{Z}}$. \noindent \textbf{(b) Security Against Malicious Sender $S$}. In this case, $\mathcal{A}$ corrupts the sender $S$ and delivers all messages on behalf of $S$. The adversary $\mathcal{A'}$ simulates the actions of $R$ as follows. \noindent $\bf{\underline{{\sf Game} ~0}}$: This game corresponds to the real world protocol interaction in which $S$ communicates with honest $R$. So, ${\sf Pr}[{\sf Game}~ 0] = 0$. \noindent $\bf{\underline{{\sf Game} ~1}}$: This game is the same as ${\sf Game} ~0$ except that $\mathcal{A'}$ simulates ${\sf crs} = ({\sf params}, {\sf GS}_S, {\sf GS}_R$, $\{\{A_{\ell, t}, B_{\ell, t}\}_{1 \leq t \leq n_{\ell}}\}_{1 \leq \ell \leq n})$. The adversary $\mathcal{A'}$, first generates ${\sf params} = (p, \mathbb{G}_1, \mathbb{G}_2, \mathbb{G}_{T}, e$, $g_1, g_2)$ by invoking algorithm ${\sf BilinearSetup}$ discussed in section \ref{preliminaries}. The components $\{A_{\ell, t} = g_{1}^{a_{\ell,t}}, B_{\ell, t} = g_{2}^{a_{\ell,t}}\}_{1 \leq t \leq n_{\ell}, 1 \leq \ell \leq n}$, are generated exactly as in ${\sf Game} ~0$. To generate ${\sf GS}_S$ and ${\sf GS}_R$ in perfectly sound setting, the algorithm, takes $\xi_1, \xi_2, \widetilde{a}, \widetilde{b}$, $\widehat{\xi}_1, \widehat{\xi}_2, \widehat{a}, \widehat{b}$ $\xleftarrow{\$} \mathbb{Z}_{p}$ and sets $u_1 = (g_{1}^{\widetilde{a}}, 1, g_1), u_2 = (1, g_{1}^{\widetilde{b}}, g_1), u_3 = (g_{1}^{\widetilde{a}\xi_1}$, $g_{1}^{\widetilde{b}\xi_2}, g_{1}^{\xi_1+\xi_2})$, $v_1 = (g_{2}^{\widetilde{a}}, 1, g_2), v_2 = (1, g_{2}^{\widetilde{b}}, g_2), v_3 = (g_{2}^{\widetilde{a}\xi_1}$, $g_{2}^{\widetilde{b}\xi_2}, g_{2}^{\xi_1+\xi_2})$, $\widehat{u}_1 = (g_{1}^{\widehat{a}}, 1, g_1), \widehat{u}_2 = (1, g_{1}^{\widehat{b}}, g_1)$, $\widehat{u}_3 = (g_{1}^{\widehat{a}\widehat{\xi}_1}$, $g_{1}^{\widehat{b}\widehat{\xi}_2}, g_{1}^{\widehat{\xi}_1+\widehat{\xi}_2})$, $\widehat{v}_1 = (g_{2}^{\widehat{a}}, 1, g_2), \widehat{v}_2 = (1, g_{2}^{\widehat{b}}, g_2), \widehat{v}_3 = (g_{2}^{\widehat{a}\widehat{\xi}_1}$, $g_{2}^{\widehat{b}\widehat{\xi}_2}, g_{2}^{\widehat{\xi}_1+\widehat{\xi}_2})$, ${\sf GS}_S = (u_1, u_2$, $u_3, v_1, v_2, v_3), {\sf GS}_R = (\widehat{u}_1, \widehat{u}_2, \widehat{u}_3$, $\widehat{v}_1, \widehat{v}_2, \widehat{v}_3)$, ${\sf \widetilde{t}_{ext}} = (\widetilde{a}, \widetilde{b}, \xi_1, \xi_2)$, ${\sf \widehat{t}_{ext}} = (\widehat{a}, \widehat{b}, \widehat{\xi}_1, \widehat{\xi}_2)$. The adversary $\mathcal{A'}$ distributes ${\sf crs}$ to all parties and keeps ${\sf \widetilde{t}_{ext}}$, ${\sf \widehat{t}_{ext}}$, $\{\{a_{\ell, t}\}_{1 \leq t \leq n_{\ell}}\}_{1 \leq \ell \leq n}$ secret to itself. \noindent The common reference string ${\sf crs}$ generated in ${\sf Game} ~1$ by $\mathcal{A'}$ is distributed identically to ${\sf crs}$ generated in ${\sf Game} ~0$ by ${\sf CrsSetup}$. Hence, $|{\sf Pr}[{\sf Game}~ 1] - {\sf Pr}[{\sf Game}~ 0]| = 0$. \noindent $\bf{\underline{{\sf Game} ~2}}$: This game is the same as ${\sf Game} ~1$ except that $\mathcal{A'}$ upon receiving ${\sf pk}, \psi, {\sf cDB}$ from $\mathcal{A}$ extracts $h, g_{2}^{w}, g_{1}^{\alpha}, g_{2}^{\alpha}$ using ${\sf \widetilde{t}_{ext}}$ in the same way as $\mathcal{A'}$ extracts witnesses in ${\sf Game} ~2$ of Case (a). Using extracted $h, g_{2}^{w}, g_{2}^{\alpha}$, $\mathcal{A'}$ extracts $(m_i, W_i)$ from ciphertext $\phi_i$ for each $i =1, 2, \ldots, N$ as follows. The adversary $\mathcal{A'}$ extracts $c_{i}^{(2)} = B^{r_i}$, $ c_{i}^{(3)} = m_i\cdot Y^{r_i}\cdot e(c_{i}^{(2)}, h)$, $c_{i}^{(4)} = \{c_{i, \ell}^{(4)}\}_{1\leq \ell \leq n} =\{g_{1}^{s_{i, \ell}}\}_{1\leq \ell \leq n}$, $c_{i}^{(5)} = \{\{c_{i, \ell, t}^{(5)}\}_{1\leq t\leq n_{\ell}}\}_{1\leq \ell \leq n}$ from $\phi_i = (c_{i}^{(1)}, c_{i}^{(2)}, c_{i}^{(3)}, c_{i}^{(4)}, c_{i}^{(5)})$ and computes \vspace{-.3cm} \begin{eqnarray*} \frac{c_{i}^{(3)}}{e\left(\prod_{\ell=1}^{n}c_{i, \ell}^{(4)}, g_{2}^{w}\right)e(c_{i}^{(2)}, h)} = \frac{m_i\cdot Y^{r_i}e(B^{r_i}, h)}{e\left(\prod_{\ell=1}^{n}g_{1}^{s_{i, \ell}}, g_{2}^{w}\right)e(B^{r_i}, h)}\\ = \frac{m_i\cdot Y^{r_i}e(B^{r_i}, h)}{e\left(g_{1}^{\sum_{\ell=1}^{n}s_{i, \ell}}, g_{2}^{w}\right)e(B^{r_i}, h)} = m_i \end{eqnarray*} as $\sum_{\ell=1}^{n}s_{i, \ell} = r_i, Y = e(g_1, g_{2}^{w})$. Let $W_i = [W_{i, 1}, W_{i, 2}$, $\ldots, W_{i, n}]$ be the access policy associated with $m_i$ which is initially empty. The adversary $\mathcal{A'}$ checks whether $e((c_{i, \ell, t}^{(5)})^{\frac{1}{a_{\ell, t}}}$, $g_2)$ $= e(c_{i, \ell}^{(4)}, g_{2}^{\alpha})$, if yes, $W_{i, \ell} = W_{i, \ell}\cup \{v_{\ell, t}\}$ for $t = 1, 2, \ldots, n_{\ell}$, for $\ell=1, 2, \ldots, n$. In this way, $W_i$ is recovered by $\mathcal{A'}$. The adversary $\mathcal{A'}$ sends $({\sf initdb}, {\sf DB} = \{m_i, W_i\}_{1 \leq i\leq N})$ to $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$. Upon receiving message $({\sf issue}, L)$, $\mathcal{A'}$ simulates $R$'s side of the ${\sf Issue}$ phase. The adversary $\mathcal{A'}$ picks attributes, constructs attribute list $\widetilde{L} = [v_{1, t_1}, v_{2, t_2}, \ldots, v_{n, t_n}]$ such that $\widetilde{L}\models W_1$ and simulates $R_{\ell}$ by taking $B_{\ell, t_{\ell}}$, from ${\sf crs}$, corresponding to $v_{\ell, t_{\ell}}$ from $\widetilde{L}$ and sets $L'= \{R_{\ell}\}_{1\leq \ell \leq n}$, where $R_{\ell} = B_{\ell, t_{\ell}}\cdot g_{2}^{z_{\ell}}$. The proof $\varphi$ is simulated with the witnesses $\{A_{\ell, t_{\ell}}, T_{\ell}\}_{1\leq \ell \leq n}$, where $T_{\ell} = g_{1}^{z_{\ell}}$. The proof $\varphi$ and $L'$ is given to $\mathcal{A}$ who in turn gives randomized attribute key ${\sf ASK'} = (d_0 = g_{2}^{\frac{w+s}{\beta}}, \{d'_{\ell, 1} = g_{2}^{s}\cdot R_{\ell}^{\lambda_{\ell}}, d_{\ell, 2} = g_{2}^{\lambda_{\ell}}, d_{\ell, 3} = g_{2}^{\frac{\lambda_{\ell}}{\alpha}}\}_{1\leq\ell \leq n})$. The adversary $\mathcal{A'}$ checks whether $\frac{e(g_1, d'_{\ell, 1}\cdot d_{\ell, 2}^{-z_{\ell}})e(g_{1}^{\alpha}, d_{\ell, 3})}{e(B, d_0)e(g_{1}^{-1}, g_{2}^{w})} = e(A_{\ell, t_{\ell}}\cdot g_1, d_{\ell, 2})$ for $\ell=1, 2, \ldots, n$. If yes, a bit $b=1$ to given to $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ by $\mathcal{A'}$. Otherwise, $\mathcal{A'}$ gives $b=0$ to $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$. \noindent As Groth-Sahai proofs are composable ${\sf NIWI}$ by Theorem \ref{niwi}, the simulated $L', \varphi$ are computationally indistinguishable from the honestly generated $L', \varphi$. Therefore, we have $|{\sf Pr}[{\sf Game}~ 2] - {\sf Pr}[{\sf Game}~ 1]| \leq \epsilon_{1}(\rho)$, where $\epsilon_{1}(\rho)$ is a negligible function. \noindent $\bf{\underline{{\sf Game} ~3}}$: This game is the same as ${\sf Game} ~2$ except that $\mathcal{A'}$ upon receiving the message $({\sf transfer}, {\sf \sigma}_j)$, runs the receiver $R$'s side of the ${\sf Transfer}$ phase. The adversary $\mathcal{A'}$ simulates ${\sf Req}_{\sigma_j}, \pi_{\sigma_j}$ as follows. The adversary $\mathcal{A'}$ replaces ${\sf Req}_{\sigma_j}$ by ${\sf Req}_{1} = c_{1}^{(2)}\cdot B^{v_{1}}$ and proof $\pi_{\sigma_j}$ by $\pi_{1} = {\sf NIWI}_{{\sf GS}_R}\{(c_{1}^{(1)}$, $\prod_{\ell=1}^{n}c_{1, \ell}^{(4)}, g_{2}^{v_{1}}, c_{1}^{(2)})~|~$ $e(\prod_{\ell=1}^{n}c_{1, \ell}^{(4)}\cdot y, c_{1}^{(1)}) = e(g_1, g_2) \wedge e({\sf Req}_{1}, g_2)$ $= e(c_{1}^{(2)}, g_2)e(B, g_{2}^{v_{1}})\}$, ${\sf Pri}_{1} = H^{v_{1}}$. The adversary $\mathcal{A'}$ gives $\pi_1$ and ${\sf Req}_{1}$ to $\mathcal{A}$. Upon receiving response ${\sf Res}_1, \delta_1$ from $\mathcal{A}$, $\mathcal{A'}$ checks whether $\frac{c_{1}^{(3)}\cdot \prod_{\ell=1}^{n} e(c_{1, \ell}^{(4)}, d_{\ell, 1})\cdot H^{v_{1}}}{e(c_{1}^{(2)}, d_0)\cdot \prod_{\ell=1}^{n} e(\widetilde{c}_{1, \ell}^{(5)}, d_{\ell, 3})\cdot{\sf Res}_{1}} = m_{1}$. If yes, a bit $b=1$ to given to $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$ by $\mathcal{A'}$. Otherwise, $\mathcal{A'}$ gives $b=0$ to $\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}$. As Groth-Sahai proofs are composable ${\sf NIWI}$ by Theorem \ref{niwi}, the simulated request ${\sf Req}_{1}$ is computationally indistinguishable from the honestly generated ${\sf Req}_{\sigma_j}$. Therefore, we have $|{\sf Pr}[{\sf Game}~ 3] - {\sf Pr}[{\sf Game}~ 2]| \leq \epsilon_{2}(\rho)$, where $\epsilon_{2}(\rho)$ is a negligible function \noindent Thus ${\sf Game} ~3$ is the ideal world interaction whereas ${\sf Game} ~0$ is the real world interaction. Now\\ $ |{\sf Pr}[{\sf Game}~ 3] - [{\sf Game}~ 0]| = \sum_{t=1}^{3}|{\sf Pr}[{\sf Game}~ t] - [{\sf Game}~ (t-1)]|, $ where $\epsilon_{3}(\rho) = \epsilon_{2}(\rho) + \epsilon_{1}(\rho) + 0$ is a negligible function. Hence, ${\sf IDEAL}_{\mathcal{F}_{\sf AOT\mhyphen HAP}^{N \times 1}, \mathcal{A'}, \mathcal{Z}} \stackrel{c}{\approx} {\sf REAL}_{\Pi, \mathcal{A}, \mathcal{Z}}$. \end{proof} \section{Comparison} We compare our proposed issuer-free adaptive oblivious transfer with hidden access policy (${\sf AOT\mhyphen HAP}$) with \cite{abe2013universally}, \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious}, \cite{guleria2014adaptive}. The subtle differences between our scheme and \cite{abe2013universally}, \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious}, \cite{guleria2014adaptive} are listed below. \begin{enumerate} \item The schemes of \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious} and \cite{guleria2014adaptive}, to the best of our knowledge, are the only existing ${\sf AOT\mhyphen HAP}$, but they are not secure in UC framework. \item Abe {\em et al.} \cite{abe2013universally} introduced the first adaptive oblivious transfer with access policy in UC framework, but access policies are not hidden. \item Guleria and Dutta \cite{guleria2015universally} presented the concept of issuer-free adaptive oblivious transfer with public access policies. \item The schemes \cite{abe2013universally}, \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious}, \cite{guleria2014adaptive} assume an issuer apart from a sender and multiple receivers in their constructions and are secure under the restriction that the issuer never colludes with a set of receivers. \end{enumerate} \begin{table}[htbp] \scriptsize \centering \begin{tabular}{|c|c|c|c|c|c|c|} \hline UC Secure & \multicolumn{2}{|c|}{Pairing}&\multicolumn{3}{|c|}{Exponentiation}& Hidden \\ Schemes&\multicolumn{2}{c|}{${\sf PO}$}&\multicolumn{3}{c|}{${\sf EXP}$}& ${\sf AP}$\\ \cline{2-7} & ${\sf Transfer}$ & ${\sf DBSetup}$ & ${\sf Transfer}$ & ${\sf DBSetup}$ & ${\sf CRSG}$& \\ \cline{1-7} \cite{abe2013universally} & $(100n + 199)k$ & $(n+21)N$ & $(138n+237)k$ &$(2n+21)N + 2n+20$ & $n+26$& $\times$\\ \hline \cite{guleria2015universally}& $2\displaystyle\sum_{j=1}^{k}n_{\sigma_j}+86k$ & $N + 2$ & $96k$ &$3\displaystyle\sum_{i=1}^{N}n_i+3N + 59$ & 10& $\times$\\ \hline Ours & $(2n+149)k$ & $N + 2$ & $112k$ &$\leq((m+n+3)N + 112)$ & $2m+20$ &$\surd$\\ \hline \end{tabular} \vspace{2mm} \caption{Comparison Summary of computation cost in $k$ transfer phases and initialization phase (${\sf PO}$ stands for number of pairing operations, ${\sf EXP}$ for number of exponentiation operations, ${\sf CRSG}$ for ${\sf crs}$ generation, and ${\sf AP}$ for access control, $n$ is the number of attributes, $m$ is total number of values which $n$ attributes can take, $N$ is the database size).} \label{tab6} \end{table} \begin{table}[htbp] \scriptsize \centering \begin{tabular}{|c|c|c|c|c|c|} \hline UC Secure & Communication &\multicolumn{2}{|c|}{Storage}& Security Assumptions& Issuer-Free \\ Schemes & & \multicolumn{2}{c|}{} & &\\ \cline{2-4} &Request + Response & ${\sf crs}$-Size & (${\sf cDB}$ + ${\sf pk}$)Size &-- &\\ \cline{1-6} \cite{abe2013universally} & $(125+64n)k$& $n+28$ & $16N + n + 20$& SXDH, XDLIN& $\times$\\ \hline \cite{guleria2015universally}& $47\mathbb{G}k$& $11\mathbb{G}$ & $\left(\displaystyle\sum_{i=1}^{N}n_i+3N + m+11\right)\mathbb{G}$&DLIN, $q$-SDH , $q$-DBDHE &$\surd$\\ \hline Ours & $62k$& $2m+22$ & $\leq((m+n+3)N + 7)$&$q$-SDH, DBDH, DLIN&$\surd$ \\ \hline \end{tabular} \vspace{2mm} \caption{Comparison summary of communication cost in $k$ transfer phases and initialization phase (${\sf cDB}$ stands for ciphertext database, ${\sf pk}$ for public key, $n$ is the number of attributes, $m$ is total number of values which $n$ attributes can take, $N$ is the database size, SXDH - symmetric external Diffie-Hellman assumption, XDLIN -external decision Linear assumption, $q$-SDH - $q$-strong Diffie-Hellman assumption, DBDH-decision bilinear Diffie-Hellman assumption, DBDHE-decision bilinear Diffie-Hellman exponent assumption, DLIN-decision Linear assumption).} \label{tab2} \end{table} \noindent In contrast to \cite{abe2013universally}, \cite{camenisch2012oblivious}, \cite{camenisch2011oblivious}, \cite{guleria2014adaptive}, we introduce the \emph{first} issuer-free ${\sf AOT\mhyphen HAP}$ in UC framwork. Our ${\sf AOT\mhyphen HAP}$ is issuer-free, realizes hidden access policy and achieves UC security. Table \ref{tab6} provides the computation cost involved in $k$ ${\sf Transfer}$ phases, in algorithms ${\sf DBSetup}$ and ${\sf CRSSetup}$. In addition, ${\sf issue}$ phase takes $20n+8$ exponentiations and $2n+15$ pairing. Table \ref{tab2} exhibits (i) communication cost in $k$ transfer phases, (ii) storage complexity of common reference string ${\sf crs}$, (iii) storage complexity of ciphertext database and public key size and (iv) complexity assumption. As illustrated in Table \ref{tab6} and \ref{tab2}, our protocol is more efficient as compared to the only existing UC secure adaptive oblivious without hidden access policy. \section{Conclusion} We have proposed the first issuer-free adaptive oblivious transfer with hidden access policy (${\sf AOT\mhyphen HAP}$) in universal composable (UC) framework. In issuer-free ${\sf AOT\mhyphen HAP}$, the sender publishes the ciphertext database encrypted under associated access policies. The receiver interacts with the sender in order to decrypt the messages of its choice without revealing its identity, attribute set and choice of messages. The receiver either recovers the correct message or a garbage one depending upon whether the receiver's attribute set satisfies the access policy associated with the message. The proposed ${\sf AOT\mhyphen HAP}$ has been proved UC secure under the hardness of $q$-strong Diffie-Hellman (SDH), decision Linear (DLIN) and decision bilinear Diffie-Hellman (DBDH) problems in the presence of malicious adversary. Moreover, the protocol is efficient as compared to the existing similar schemes.

\section{Background and objectives} Breast cancer is the most common invasive disease among women \citep{siegel2014cancer} Optimistically, an early diagnosis of the disease increases the chances of recovery dramatically and as such, makes the early detection crucial. Mammography is the recommended screening technique, but it is not enough, we also need the radiologist expertise to check the mammograms for lesions and give a diagnosis, which can be a very challenging task\citep{kerlikowske2000performance}. Radiologists often resort to biopsies and this ends up adding exorbitant expenses to an already burdened patient and health care system \citep{sickles1991periodic}. We propose a Computer Assisted Diagnosis (CAD) system, based on a deep Convolutional Neural Network (CNN) model, designed to be used as a "second-opinion" to help the radiologist give more accurate diagnoses. Deep Learning requires large datasets to train networks of a certain depth from scratch, which are lacking in the medical domain especially for breast cancer. Transfer learning proved to be efficient to deal with little data, even if the knowledge transfer is between two very different domains \citep{shin2016deep}. But still using the technique can be tricky, especially with medical datasets that tend to be unbalanced and limited. And when using the state-of-the art CNNs which are very deep, the models are highly inclined to suffer from overfitting even with the use of many tricks like data augmentation, regularization and dropout. The number of layers to fine-tune and the optimization strategy play a substantial role on the overall performance \citep{yosinski2014transferable}. This raises few questions: \begin{itemize} \item Is Transfer Learning really beneficial for this application? \item How can we avoid overfitting with our small dataset ? \item How much fine-tuning do we need? and what is the proper way to do it? \end{itemize} We investigate the proper way to perform transfer learning and fine-tuning, which will allow us to take advantage of the pre-trained weights and adapt them to our task of interest. We empirically analyze the impact of the fine-tuned fraction on the final results, and we propose to use an exponentially decaying learning rate to customize all the pre-trained weights from ImageNet\citep{deng2009imagenet} and make them more suited to our type of data. The best model can be used as a baseline to predict if a new "never-seen" breast mass lesion is benign or malignant. \section{Methodology} We use BCDR-F03 dataset \citep{lopez2012bcdr}from the Breast Cancer Digital Repository (BCDR-FM) which contains digitized mammograms from 344 patients. It is a binary class dataset composed of benign and malignant findings. To get a balanced dataset, we used a subset of the data comprised of 300 cases of patients which gave us 600 images, 300 images showing benign lesions and 300 showing malignant lesions. We can later use the remaining cases of image lesions, from the dataset, for prediction to test the performance of the proposed CAD system, since they can be considered as "never-seen" images. Each patient's case includes mammograms with the associated coordinates of the lesions' contours. We followed the same pre-processing steps of \citep{arevalo2016representation} and \citep{perre2017lesion} except for the size of the cropped ROIs (Regions Of Interest) which were 299 x 299 instead of 150 x 150, so that the resulting image lesions are compatible with the ImageNet images that were used to train the original Inception-v3 CNN model. All the images were preprocessed as follows: \begin{enumerate} \item First we cropped fixed sized regions of interest using the ground truth provided with the dataset.\label{item:1} \item Then, we used global contrast normalization where every image $\mathbf{x}_{i,j}$ was normalized by subtracting the mean and dividing by the standard deviation of its elements.\label{item:2} \item Then we used data augmentation by applying series of random transformations to the images i.e. width and height shifts by a fraction of 0.25, random rotation that ranges from 0-40 degrees and random horizontal flip. \label{item:3} \end{enumerate} Transfer learning and fine-tuning aim at reusing the pre-trained weights of a CNN as an initialization for a new task of interest. The CNN model Inception-v3 \citep{szegedy2016rethinking} devised a new module named "The inception module" which is a 4 parallel pathway of 1x1, 3x3 and 5x5 convolution filters. The architecture of this module allows the model to recover both local features via smaller convolutions and high abstracted features via large ones. And because of the parallel network implementation in addition to the down sampling layers in each block, the model's execution time beats other state-of-the-art CNNs. The Inception-v3 CNN consists of two parts: (1) The convolutional neural network part for feature extraction and (2) the classification part with the fully-connected and softmax layers. We removed the fully-connected part of the model and built one customized to our number of classes (i.e. Benign and Malignant. See figure ~\ref{}). The resulting model is composed of 221 layers. The top convolutional part, which we kept unchanged, has 5 convolutional layers each one followed by batch normalization, 2 pooling layers and 11 inception modules. While the customized fully-connected part of the model we added, is composed of a flatten layer, 2 fully-connected dense layers and a dropout layer to randomly turn off the activations at training time with a probability of 0.5. Finally, the softmax layer is used to give normalized class probabilities for the output being "Benign" or "Malignant". \begin{figure}[h] \centering \includegraphics[width=\linewidth]{image03.png} \caption{The customized inception v3} \label{} \end{figure} We conducted four main experiments: \begin{itemize} \item (i) Inceptionv3-RI: we used randomly initialized weights vs pre-trained weights. \item (ii) Inceptionv3-FE: we used the pre-trained network as a fixed feature extractor. \item (iii) Inceptionv3-1FT, Inceptionv3-2FT, Inceptionv3-3FT, Inceptionv3-AllFT: we fine-tuned the weights of the pre-trained network by gradually unfreezing the bottom layers of the convolutional part of the model (the last block, the last two blocks, the last three blocks and all the convolutional blocks). \item (iv) Inceptionv3-FTED: we fine-tuned all the weights by resuming back-propagation on all the layers with a per-layer exponentially decaying learning rate. \end{itemize} According to \citep{yosinski2014transferable}, the first layers of a CNN learn generic features, while the last layers tend to be more specific to the data. The assumption is that when we froze all top layers and fine-tuned only the last two, we helped the model learn more data-specific features. In other words, when fine-tuning, the weights of the last convolutional layers need to be tweaked as much as possible, while the ones from the top layers remain nearly untouched. In accordance with this, we propose to fine-tune all the layers of the network using a per-layer exponentially decaying learning rate \citep{walter1990nonlinear} (Inceptionv3-FTED). We set the learning rate to be a bit high (1E-3) for the last convolutional layers and we decrease its value gradually per-layer in an exponentially-decaying manner as we move backwards to the top layers. We gradually update the layers starting from the last ones going towards the top layers with a corresponding reduction in the learning rate. The learning rate $\mathbf t_{l}^{(k)} $ decays exponentially such as: $\mathbf t_{l}^{(k)}= t_{0} \ast \exp (-\lambda \ast l)$ with $\mathbf \lambda $ a hyperparameter set to -3 and $\mathit{l}$ the number of layers. The update equation for SGD with momentum is then: $\mathbf \Delta x_{l}^{(k)}= \mu \Delta x_{l}^{(k-1)}-t_{l}^{(k)} g^{(k)} $ The update rule of the gradient descent searches the direction given the negative gradient $\mathbf g^{(k)} $. $\mathbf \Delta x_{l}^{(k)} $ denotes the update in the parameters of the $\mathit{l}$ -th layer and $\mathit{k}$ -th epoch and $\mathbf \mu $ is the momentum coefficient. \section{Results and discussion} Experiments were run 5 times using 80\% of the data as training/validation data and 20\% as testing data. We trained the models for 90 epochs with a batch size of 128. We used Adam optimizer to train from scratch and SGD with momentum of 0.9 when fine-tuning , with a learning rate set to 1E-4; which was then divided by 10 each time the validation error stopped improving. The initial learning rate for Inceptionv3-FTED is 1E-3 which will be annealed at each layer. We adopted an early stopping strategy to monitor the validation loss with a patience set to 15 epochs. We also used data augmentation, regularization and dropout to avoid overfitting. Table~\ref{res} gives the mean accuracy, standard deviation and time elapsed for each model and figure~\ref{} gives the ROC curve plot of the classification of the test data by the model variants. \begin{table}[h] \caption{Comparison summary between the different model implementations.} \label{res} \centering \begin{tabular}{llll} \toprule Model & Accuracy(\%) & Std(\%) &Time elapsed(min) \\ \midrule Inceptionv3-RI & 75.83 & $\pm$1.53 &176.2 \\ Inceptionv3-FE & 92.67 & $\pm$1.79 &59.1 \\ Inceptionv3-1FT & 94.17 & $\pm$0.92 &61.6 \\ Inceptionv3-2FT & 96.67 & $\pm$0.85 &63.9 \\ Inceptionv3-3FT & 95.17 & $\pm$1.18 &67.1 \\ Inceptionv3-AllFT & 92.67 & $\pm$1.34 &158.9 \\ Inceptionv3-FTED & \textbf{97.50} & $\pm$1.26 &161.5 \\ \bottomrule \end{tabular} \end{table} \begin{figure} \centering \includegraphics[scale=0.6]{roc2.png} \caption{The ROC curve plot of the classification of the test data using the model variants.} \label{} \end{figure} Experiments showed that transfer learning is advantageous to our task of interest and this in terms of time and accuracy (Inceptionv3-RI vs Inceptionv3-FE). Even if the transfer is between two unrelated tasks, the use of pre-trained weights for initialization is a good way to start the learning process. The next step is to start fine-tuning the loaded weights so that we train the network to better fit the new dataset. This is achieved by resuming backpropagation of the layers with a small learning rate. We gradually unfroze the last convolutional blocks of the model and tested the different model variants on the data. The performance increased when we started fine-tuning the last convolutional layers up until 2 convolutional blocks (Inceptionv3-2FT), when 3 or more blocks were fine-tuned the accuracy dropped. The idea here is that we need to make the last convolutional layers learn more data-specific features, but there is no need for us to change the weights of the first convolutional layers as much; they're already well-tuned to learn generic features especially if we don't have enough data to train on. The middle convolutional layers might be responsible of learning somewhat more complex features compared to the first layers and thus they need to be slightly modified. The exponentially decaying learning rate helps us in this regard. The proposed method for fine-tuning (Inceptionv3-FTED) showed better performance compared to other proposals that classified the same dataset(see table~\ref{comp}). The per-layer decaying learning rate helps us control the rate at which weights change for each part of the network. When fine-tuning, we want some of the layers to be more or less receptive to change. Transfer learning with a per-layer exponentially decaying learning rate yielded better results. The model Inceptionv3-FTED demonstrated an accuracy of 97.50\% and an AUC of 0.96 , thus outperforming human-level performance \citep{elmore2009variability} and hence the model can be used as a baseline to build a powerful tool for assisting radiologists in improving the efficiency, consistency, and accuracy of breast cancer diagnosis. \begin{table}[h] \caption{Comparison with other proposals that classified the same datastet.} \label{comp} \centering \begin{tabular}{lll} \toprule Model & Accuracy(\%) & AUC(Area Under the Curve) \\ \midrule \citep{perez2014improving} & - &0.8562 \\ \citep{arevalo2016representation}& - &0.826 \\ \citep{padillabreast} &- &0.950 \\ \citep{jadoon2017three} &83.74\% &- \\ \citep{perre2017lesion} &- &0.793 \\ Ours & \textbf{97.50} &\textbf{0.96} \\ \bottomrule \end{tabular} \end{table} \subsubsection*{Acknowledgments} The authors would like to thank Stephan Gouws and Nyalleng Moroosi for their valuable and constructive suggestions during the presentation of the first draft of this research at "Deep Learning Indaba 2017".\\ The database used in this work was a Courtesy of MA Guevara Lopez and coauthors, Breast Cancer Digital Repository Consortium. \bibliographystyle{plainnat}

\section{Introduction} \label{intro} In many situations in engineering and other fields, it is of interest to find the roots of some unknown function $h\,:\,\mathbb{R}^n \to \mathbb{R}^n$ to which we have access to only through noisy measurements. Stochastic Approximation, originally introduced by Robbins and Monro in \cite{Robbins}, provides an incremental scheme for doing this. It involves running the following iteration $$x_{n+1} = x_n + a_n (h(x_n) +M_{n+1}), $$ where stepsizes $a(n) > 0$ satisfy the assumptions: $$\sum_na(n) = \infty, \ \sum_na(n)^2 < \infty,$$ and $\{M(n)\}$ is the noise term, usually assumed to be a martingale difference sequence. A generalizaiton of this is when the relevant function is defined on a manifold. There has been considerable interest in optimization and related algorithms on manifolds, particularly on matrix manifolds \cite{Absil}, \cite{Helmke}. But the corresponding development for such algorithms in presence of noisy measurements has been lacking. A notable exception is \cite{Bonn} which analyzed stochastic gradient schemes on Riemannian manifolds. The contributions of this work are as follows : First and foremost, we extend the Ordinary Differential Equations (ODE) method for analyzing stochastic approximation schemes (\cite{BorkarBook}, \cite{Kushner}, \cite{Ljung}) to Riemannian manifolds. The usual addition operation of the Euclidean Robbins-Monro scheme is modified using a retraction (see Definition 1,2). We consider a much more general model than \cite{Bonn} and furthermore, our approach is quite distinct from that of \textit{ibid.}, which is inspired from \cite{Bot}. Studying the convergence properties using the associated ODE connects the stochastic approximation theory on manifolds with that of ODEs on manifolds. The upside here is that ODEs on manifolds are very well studied and have a rich theory (see \cite{Hairer}). This opens up the possibility of adapting analytic and computational techniques in the latter domain for stochastic approximation on manifolds in future. The second contribution is in the context of projected SA for submanifolds embedded in a Euclidean space as well as for general constraint sets which may not have a manifold structure. In many situations the iterates perforce are executed in an ambient space and a suitable correction at each iterate is required to pull it back to the manifold. For example, for an iteration over the unit sphere, the noisy measurement or even the discretization error may put the next iterate off the sphere and one would typically normalize it to bring it back to the sphere. Such issues also arise, e.g., in gradient projection algorithms. We approach projected SA by taking recourse to ideas developed in \cite{Mathkar} which studied an iteration which used a fixed nonlinear map perturbed by a stochastic approximation like expression with decreasing stepsizes, i.e., $$x(n+1) = F(x(n)) + a(n)(h(x(n)) + M(n+1)),$$ $F, h, \{M(n)\}$ satisfy suitable technical hypotheses. The main claim was that asymptotically one recovers the limiting behavior of the projected o.d.e.\ $$\dot{x}(t) = \Gamma(h(x(t))),$$ where $\Gamma$ is the Frechet differnetial of the associated projection map to the set of fixed points of $F$. These results were somewhat restrictive in the regularity assumptions imposed on $F$. This lacuna was worked around in \cite{SuhBor} where $F$ was replaced by a suitable time dependent sequence of maps $\{F_n\}$ in a manner that ensured the above under weaker technical conditions, albeit for the special case when the projection is to the intersection of compact convex sets. In the present work, we develop an analogous scheme wherein the iteration sits in an ambient Euclidean space, but the choice of $\{F_n\}$ ensures the correct asymptotic behavior on the manifold. It is worth noting that both \cite{Mathkar} and \cite{SuhBor} consider distributed algorithms where multiple processors coordinate their computations for a common goal. As this is not our concern here, we downplay this aspect of \cite{Mathkar}. The rest of the paper is organized as follows. In Section 2 we prove the the main convergence result pertaining to SA on manifolds, where an ODE approximation is established. Section 3 studies retractions in the framework of submanifolds. We conclude the paper in Section 4 with some comments and future directions. Before proceeding further we recall some basic facts and definitions about Riemannian geometry which we will use throughout the rest of the paper. Throughout the rest of the paper we let $\mathcal{M}$ denote a connected Riemannian manifold. A smooth $n$-dimensional manifold is a pair $(\mathcal{M},\mathcal{A})$, where $\mathcal{M}$ is a Haussdorf second countable topological space and $\mathcal{A}$ is a collection of charts $\{\mathcal{U}_\alpha,\psi_{\alpha}\}$ of the set $\mathcal{M}$, i.e., the following holds : \begin{enumerate} \item The collection $\{\mathcal{U}_\alpha\}$ of open sets in $\mathcal{M}$ covers $\mathcal{M}$, i.e., $$ \cup_{\alpha} \mathcal{U_\alpha} = \mathcal{M}.$$ \item Each $\psi_{\alpha}$ is a bijection of $\mathcal{U}_{\alpha}$ onto an open set of $\mathbb{R}^n$. \item for any pairs $\alpha, \beta$ with $\mathcal{U}_{\alpha} \cap \mathcal{U}_{\beta} \neq \emptyset $, the set $\psi_\alpha(\mathcal{U}_{\alpha} \cap \mathcal{U}_{\beta} )$ and $\psi_\beta(\mathcal{U}_{\alpha} \cap \mathcal{U}_{\beta})$ are open sets in $\mathbb{R}^n$ and the change of coordinates $$ \psi_{\beta} \circ \psi^{-1}_\alpha : \mathbb{R}^n \to \mathbb{R}^n$$ is smooth on its domain $\psi_\alpha(\mathcal{U}_{\alpha} \cap \mathcal{U}_{\beta} )$. \end{enumerate} By a Riemannian manifold we mean a manifold whose tangent spaces are endowed with a smoothly varying inner product $\langle \cdot,\cdot \rangle_x$ called the Riemannian metric. The tangent space at any point $x \in \mathcal{M}$ is denoted by $\mathcal{T}_x \mathcal{M}$ (for definition see \cite{Absil} or \cite{dcarmo}). We recall that a tangent space admits a structure of a vector space and for a Riemannian manifold it is a normed vector space. The tangent bundle $\mathcal{T}\mathcal{M}$ is defined to be the disjoint union $\cup_{x \in \mathcal{M}}\{x\}\times\mathcal{T}_x\mathcal{M}$. The normal space at the point $x$ denoted by $\mathcal{N_{M}}(x)$ is the set of all vectors orthogonal (w.r.t to $\langle \cdot,\cdot \rangle_x$) to the tangent space at $x$. Using the norm, one can also define the arc length of a curve $\gamma : [a,b] \to \mathcal{M}$ as $$L(\gamma)= \int_a^b \sqrt{\langle \dot{\gamma}(t),\dot{\gamma}(t) \rangle_{\gamma(t)}} \, dt.$$ \begin{defn} A geodesic on $\mathcal{M}$ is a curve that locally minimizes the arc length (equivalently, these are the curves that satisfy $\gamma''\in\mathcal{N_{M}}(\gamma(t))$ for all $t$). The exponential of a tangent vector $u$ at $x$, denoted by $\text{exp}_x(u)$, is defined to be $\Gamma(1,x,u)$, where $t\to\Gamma(t,x,u)$ is the geodesic that satisfies \[ \Gamma(0,x,u)=x\;\text{ and }\;\frac{d}{dt}\Gamma(0,x,u)\rvert_{t=0}=u. \] \end{defn} We let $d(\cdot,\cdot)$ denote the Riemannian distance between any two points $x,y\in \mathcal{M}$, i.e., $$ d \,:\, \mathcal{M} \times \mathcal{M} \to \mathbb{R}\, : \, d(x,y) = \inf_{\Gamma}L(\gamma),$$ where $\Gamma$ is the set of all curves in $\mathcal{M}$ joining $x$ and $y$. We recall that $d(\cdot,\cdot)$ defines a\textit{ metric }on $\mathcal{M}$. By the neighborhood $\mathcal{U}_{x}$ of a point $x$ we mean the normal neighborhood or the geodesic ball centered at $x$. The coordinate chart for this neighborhood is provided by $$ \psi_x = E^{-1}\circ \text{exp}^{-1}_x : \mathcal{U}_{x} \to \mathbb{R}^n, $$ where $E:\mathbb{R}^n \to \mathcal{T}_x\mathcal{M}$ is the isomorphism mapping a point $x=(x(1),...,x(n))\in \mathbb{R}^n$ to a point in $\mathcal{T}_x\mathcal{M}$ expressed in the orthonormal basis $\{E^i\}$ for $\mathcal{T}_x\mathcal{M}$, i.e $E(x)=\sum_i x(i)E^i$. The coordinates of the point $x$ under $\psi_x$ are $0$. \section{Stochastic Approximation on Manifolds} Suppose we have a smooth vector field $H\,:\,\mathcal{M} \to \mathcal{T}\mathcal{M} $ assigning to each point $x \in \mathcal{M}$ a tangent vector $H(x) \in \mathcal{T}_x \mathcal{M}$. We want to find the zeroes of $H(\cdot)$ based on its samples corrupted by a martingale difference noise $M_{x}\in\mathcal{T}_{x}\mathcal{M}$. A common scenario where this happens is when we have to minimize a smooth function : \[ \min_{x\in\mathcal{M}} \big\{ L(x) \doteq \mathbb{E}_{z}Q(x,z)=\int Q(x,z)dP(z) \big\}\] for some function $Q : \mathcal{M}\times \mathcal{Z} \mapsto \mathbb{R},$ where $x\in\mathcal{M}$ is the minimization parameter and $P$ is a probability measure on a measurable space $\mathcal{Z}$. The minimization is to be performed based on noisy measurements of a black box that outputs at time $n$ on input $x_n$, the quantity $Q(x_n, \xi_n)$ where $\xi_n$ is an independent sample with law $P$. In this case one sets $H(x) = E\left[Q(x, \xi)\right]$ and $M_{n+1} = Q(x_n, \xi_n) - H(x_n)$. The most obvious way to do this would be to take the noisy sample of the vector field and perform a simple iterative gradient descent like scheme. However, such an iteration involving the addition of two points would not be well defined for manifolds. A way to remedy this is to use the notion of a geodesic and the associated concept of the exponential map. The natural generalization of the SA scheme to the case of manifolds using the exponential map would yield\footnote{For ease of notation we drop hereafter the superscript $x_k$ on the noise term $M_{k+1}$ which is understood to belong to the tangent space at $x_k$.} : \begin{equation}\label{expiter} x_{k+1}=\text{exp}_{x_k}\big(a_{k}\{H(x_{k})+M^{x_k}_{k+1}\}\big). \end{equation} We have to be careful while defining the above update because of the \textit{injectivity radius}. We assume throughout the paper that this quantity is bounded below by some $K>0$. Informally, the injectivity radius at $x\in\mathcal{M}$ is the least distance to the cut locus which is where the $\text{exp}_x(\cdot)$ ceases to be the path of minimizing length (see \cite{dcarmo} for details). But since the time step $a_k \to 0$ (see assumption (A2)), this is not a problem as long as the stepsizes are sufficiently small. The downside of the above scheme is that computing the exponential updates requires solving an ordinary differential equation which defines the geodesic and is usually computationally expensive. Even for the simple case of a spherical constraint $S^{n-1}$, the geodesic $t\to x(t)$ expressed as a function of $x_0 \in S^{n-1}$ and $\dot{x}_0 \in \mathcal{T}_{x_0}S^{n-1}$ is given by (Example 5.4.1, \cite{Absil}) : $$x(t)=x_0\text{cos}(\|\dot{x}_0t\|) +\dot{x}_0\frac{1}{\| \dot{x}_0\|}\text{sin}(\|\dot{x}_0\|t).$$ An alternative is provided by approximating the geodesics using the concept of retractions defined next. \begin{defn} A retraction on $\mathcal{M}$ is a smooth mapping $\mathcal{R}:\mathcal{TM}\to\mathcal{M}$, where $\mathcal{TM}$ is the tangent bundle, with the following properties : i) $\mathcal{R}_{x}(0_{x})=x$, where $\mathcal{R}_{x}$ is the restriction of the retraction to $\mathcal{T}_{x}\mathcal{M}$ and $0_{x}$ denotes the zero element of $\mathcal{T}_{x}\mathcal{M}$. ii) {[}Local rigidity{]} With the canonical identification $\mathcal{T}_{0_{x}}\mathcal{T}_{x}\mathcal{M}\simeq\mathcal{T}_{x}\mathcal{M}$, $\mathcal{R}_{x}$ satisfies \[ D\mathcal{R}_{x}(0_{x})=\text{id}_{\mathcal{T}_{x}\mathcal{M}}, \] where $\text{id}_{\mathcal{T}_{x}\mathcal{M}}$ denotes the identity mapping on $\mathcal{T}_{x}\mathcal{M}$. \end{defn} The update (\ref{expiter}) then becomes: \begin{equation}\label{ret} x_{k+1}=\mathcal{R}_{x_{k}}\big(a_{k}\{H(x_{k})+M_{k+1}\}\big). \end{equation} Again, in the above iteration it is assumed that $M_{k+1}$ belongs to $\mathcal{T}_{x_k} \mathcal{M}$. To compare for the example of the spherical constraint, a possible retraction would be simply to normalise (i.e., divide by the norm) $x_n$ at each step. \begin{rem} An important thing to note here is that we do not prescribe the retraction mapping in (\ref{ret}). It is not unique and depending upon the manifold there may be more than one possible choice. In fact the exponential map itself is a retraction (Proposition 5.4.1, \cite{Absil}). Keeping the latter fact in mind, we perform the convergence analysis only for (\ref{ret}). All the results presented also hold for (\ref{expiter}). The specific retraction will usually be chosen based on computational ease. \end{rem} We give some examples (Section 3.2, 3.3 \cite{Malick}) of retractions on some matrix manifolds, which will help highlight the computational appeal of using general retractions as opposed to the exponential map.\\ \noindent{\textbf{Example 1 }(\textit{Projection on Fixed Rank Matrices and Stiefel Manifolds) }}: Let \[ \mathcal{R}_{r}=\{X\in\mathbb{R}^{n\times m}:\,\text{rank}(X)=r\} \] be the set of matrices with rank $r$ which is known to be a smooth manifold of $\mathbb{R}^{n\times m}$. Let the singular value decomposition of any $X \in \mathbb{R}^{n\times m}$ be given by \[ X=U\Sigma V^{T} \] with $U=[u_{1},....,u_{n}]$ and $V=[v_{1},....,v_{m}]$ being orthogonal matrices and $\Sigma$ being a diagonal matrix having the singular values of $X$ on its diagonal in the non-increasing order $\big(\sigma_{1}(X)\geq\sigma_{2}(X)....\geq\sigma_{\min\{n.m\}}(X)\geq0 \big)$. With a fixed $\tilde{X} \in \mathcal{R}_r$, for any $X$ such that $\|X-\tilde{X}\|<\sigma_{r}(\tilde{X})/2$, the projection of $X$ onto $\mathcal{R}_{r}$ exists, is unique and is given by \[ P_{\mathcal{R}_{r}}(X)=\sum_{i=1}^{r}\sigma_{i}(X)u_{i}v_{i}^{T}. \] The Stiefel manifold is defined for $m\leq n$ by \[ \mathcal{S}_{n.m}=\{X\in\mathbb{R}^{n\times m}\,:\,X^{T}X=I_{m}\}. \] Along the same lines as above, let $\tilde{X}\in\mathcal{S}_{n,m}$; then for any $X$ such that $\|X-\tilde{X}\|<\sigma_{m}(\tilde{X})/2$, the projection of $X$ onto $\mathcal{S}_{n,m}$ exists, is unique and is given by \[ P_{\mathcal{S}_{n,m}}(X)=\sum_{i=1}^{m}u_iv_{i}^{T}. \] In particular, this is a retraction.\\ Since we will be using the ODE method to analyze the algorithm (\ref{ret}), we first recall the notion of an ODE on a manifold (\cite{Hairer}, Theorem 5.2, Chapter 4): \begin{defn} The dynamics \begin{equation} \label{manODE} \dot{x}=H(x) \end{equation} defines a differential equation on the manifold when $$H(x)\in\mathcal{T}_{x}\mathcal{M}\:\text{ for all }x\in\mathcal{M}$$ is a smooth vector field. The existence and uniqueness theorem for ordinary differential equations guarantees that there exists a unique smooth function $\Phi\,:\,\mathbb{R}\times\mathcal{M} \to \mathcal{M}$ such that i) $\Phi_\cdot (t)\,:\, \mathcal{M} \to \mathcal{M}$ is a diffeomorphism for each $t \in \mathbb{R}$; ii) $\Phi_x(t+s) = \Phi_{\Phi_x(t)}(s) $ ; and iii) for each $x\in \mathcal{M},$ $$\frac{d}{dt}\bigg\rvert_{t=0}\Phi_x(t) = H(x).$$ \end{defn} The proof of the existence of a unique smooth function satisfying the above prooperties is given in Chapter 5, \cite{Arnold}. The notion of convergence on manifolds is defined by generalizing that for the standard Euclidean space using local charts : \begin{defn}\label{mancon} An infinite sequence $\{x_{k}\}_{k\geq0}$ of points of a manifold $\mathcal{M}$ is said to be convergent to a point $x^{*}\in\mathcal{M}$ if there exists a chart ($\mathcal{U},\psi$) of $\mathcal{M}$, and a $K>0$ such that $x^* \in \mathcal{U}$, $x_k \in \mathcal{U}$ for $k \geq K$, and $\{\psi(x_{k})\}_{k\geq K}$ converges to the point $\psi(x^{*})$. \end{defn} We make the following important assumption:\\ \begin{enumerate} \item[(A0)] The injectivity radius at all points in $\mathcal{M}$ is bounded away from zero by some $r_0 > 0$.\\ \item[(A1)] The map $H\,:\,\mathcal{M} \to \mathcal{T}\mathcal{M}$ is a smooth vector field.\\ \item[(A2)] Step-sizes $\{a(n)\}$ are positive scalars satisfying : \begin{equation} \label{summability} \sum_{n}a_n=\infty\,,\sum_{n}a_n^{2} < \infty. \end{equation} \item[(A3)] $\{M_{n}\}$ is a martingale difference sequence with respect to the increasing $\sigma$-fields \begin{equation}\label{sigma-field-1} \mathcal{F}_{n}\doteq\sigma(x_0,M_{m},m \leq n),\,n\geq0 \end{equation} so that $$\mathbb{E}[M_{n+1}|\mathcal{F}_{n}]=0\,\text{ a.s.}$$ Furthermore, we assume that \begin{equation} \sup_n\mathbb{E}[\|M_{n+1}\|^{2}|\mathcal{F}_{n}] < \infty. \label{mgbound} \end{equation} \item[(A4)] The iterates of (\ref{ret}) remain in a (possibly sample point dependent) compact subset of $\mathcal{M}$, a.s. \end{enumerate} \begin{rem} Note that the norm $\|.\|$ used in (A3) is derived from the Rimeannian metric $\langle \cdot,\cdot\rangle$. Condition (A4) is a stability condition that needs to be separately verified, see \cite{BorkarBook}, Chapter 3 for some tests for stability in the Euclidean case. \end{rem} \subsection{Convergence Analysis} The analysis presented here builds upon the proof for the Euclidean case (Lemma 1, Theorem 2, Chapter 2, \cite{BorkarBook}). Let $t_{n}=\sum_{m=0}^{n-1}a_{m}$ with $t_{0}=0$. Let $x(t)$ be a continuous time trajectory evolving on the manifold defined by setting $x(t_{n})=x_{n}$, where $x_n$ is the iterate produced by (\ref{ret}), and then joined by a \textit{geodesic} on the interval $[t_{n},t_{n+1}).$ Also let $x^{s}(t),\,t\geq s,$ denote the solution of (\ref{manODE}) starting at the point $x(s)$, i.e. $x^s(s)=x(s)$ , so that $x^{t_n}(t), t \geq t_n,$ is the solution to the ODE (\ref{manODE}) with $x^{t_n}(t_n)=x(t_n)=x_{n}$. \begin{thm}\label{manthrm} Suppose (A0)-(A4) hold. Then for any $T>0$, \[ \lim_{s\to\infty}\sup_{t\in[s,s+T]} d(x(t),x^{s}(t)) \to0\, \textrm{ a.s. }, \] where $d(\cdot,\cdot)$ is the Riemannian distance. \end{thm} \begin{proof} We first do the following construction : Cover $\mathcal{M}$ by a countable collection of open geodesic balls of radii $\frac{r_0}{2}$ (i.e., points whose Riemannian distance from a `center' is $< \frac{r_0}{2}$) with the additional properties : \begin{itemize} \item [\textbullet] any point in $\mathcal{M}$ is at a Riemannian distance $< \frac{r_0}{4}$ from the center of at least one such ball, \item [\textbullet] any compact subset of $\mathcal{M}$ has a finite subcover of such balls. \end{itemize} It is easy to see that under (A0), this is always possible. Take a finite subcover from the given collection to cover the compact set in which $x_n$ lie. Let $\{\mathcal{U}_n,\psi_n\}$ denote \textit{a neighborhood} in which the point $x_n$ lies. Under the above construction, the following hold: \begin{enumerate} \item Given a pair $(\mathcal{U}, \psi)$ from the chart, without any loss of generality, both $\psi(\cdot)$ and its differential $D\psi(\cdot)$ viewed as a map $ D\psi(\cdot): \mathcal{U} \mapsto$ the space of bounded linear operators $\mathcal{T}_{\cdot}\mathcal{M} \to \mathbb{R}^n$ with operator norm, are uniformly bounded on $\mathcal{U}$. If $\psi(\mathcal{U})$ does not map to a bounded set in $\mathbb{R}^n$ we can take a smaller neighborhood $\mathcal{U'}\subset \mathcal{U}$ with $\overline{\mathcal{U'}} \subset \mathcal{U}$ where $\overline{\mathcal{U'}}$ denotes the closure of $\mathcal{U}'$. The difference $\mathcal{U}\,/\,\overline{\mathcal{U}'}$ can again be covered by a finite number of neighborhoods to make sure $\cup_{\alpha}\mathcal{U}_{\alpha} = \mathcal{M}$. The same argument applies to $D\psi(\cdot)$. Note that all of this is possible since we are dealing with geodesic neighborhoods and shrinking such a neighborhoods radius does not violate its diffeomorphic property. \\ \item For any $x_n$, $d(x_n, y) \geq \frac{r_0}{4} \ \forall \ y \in \partial\mathcal{U}_{n}$, where $\partial\mathcal{U}_{n}$ denotes the boundary of $\mathcal{U}_n$, for at least one such geodesic ball $\mathcal{U}_n$. This property ensures that each $x_n$ lies in the interior of a geodesic ball. Also, this is again without any loss of generality since if this property did not hold true, we would have an infinite number of $x_n$ with $d(x_n, y) < \frac{r_0}{4} \text{ for some} \ y \in \partial\mathcal{U}_{n}$ and all such $\mathcal{U}_n$ which contain $x_n$ (for a finite number of such $x_n$ we can just add the geodesic balls centered at them to the chart). But then again, for all such $\mathcal{U}_n$ which contain these $x_n$ (which are finite in number from compactness), we can cover the points $y \in \mathcal{U}_n $ with $d(x_n, y) \geq \frac{r_0}{4}$ with a finite number of geodesic balls of radii $\frac{r_0}{2}$ so that the property is restored. \end{enumerate} The motivation behind this construction will be discussed later. The following notation is used to denote the coordinate expressions : \[ \hat{x}_{n} =\psi_n(x_{n}), \] \[ \hat{H}(\hat{x}_{n}) = D\psi_n(x_{n})[H(x_{n})], \] \[ \hat{M}_{n} = D\psi_n(x_{n})[M_n], \] where $D\psi_n(x_n) : \mathcal{T}_{x_n}\mathcal{M} \to \mathbb{R}^n$ is the differential of the coordinate mapping $\psi_n$ and is a linear map. Let $\bar{x}(t)$ be a piecewise linear trajectory evolving in $\mathbb{R}^{n}$ defined by setting $\bar{x}(t_{n})=\hat{x}_{n}$ and then doing a \textit{linear interpolation} on the interval $[t_{n},t_{n+1}).$ Consider the ODE on the manifold expressed in the local chart $\hat{x}^{t_n}(t)=\psi_n(x^{t_n}(t))$, where $\hat{x}^{t_n}(t)$ denotes the solution to (\ref{newmanODE}) below with $\hat{x}^{t_{n}}(t_{n})=\hat{x}_{n}$, well-defined for $t$ in a sufficiently small interval : \[ \dot{x}(t)=H(x(t)) \] \[ \Rightarrow(D\psi(x(t)))\dot{x}(t)=(D\psi(x(t)))H(x(t)) \] \begin{equation} \label{newmanODE} \Rightarrow\dot{\hat{x}}(t)=\hat{H}(\hat{x}(t)). \end{equation} It is sufficient to show here that $$\lim_{s\to\infty}\sup_{t\in[s,s+T]} \|\bar{x}(t)-\hat{x}^{t_n}(t)\| \to 0,$$ so that we deal with iterates on the manifold using local parameterization. This is where Properties (a) and (b) help us. Property (a) helps in making sure that $\hat{x}_n,\,n \geq 0$ remain bounded. We will be comparing trajectories $x(\cdot)$ and $x^{t_n}(\cdot)$ in the interval $[t_n, t_n+T]$ by mapping them to $\bar{x}(\cdot)$ and $\hat{x}^{t_n}(\cdot)$ respectively. So we may have to switch the coordinate charts multiple times while mapping $\{x(t)\,,\,x^{t_n}(t):\,t\in[t_n, t_n+T)$\}. To circumvent this, let us define $$\bar{T}= \sup_{n\geq0 }\{ t : x(t_n + t ),x^{t_n}( t ) \in \psi_{n}(\mathcal{U}_n) \}. $$ From Property (b) $d(x_n, y) \geq \frac{r_0}{4} \ \forall \ y \in \partial\mathcal{U}_{n}$, so $\bar{T}>0$.\footnote{If $\bar{T} = \infty $ (an infinite injectivity radius) or $T < \bar{T}$, all these constructions are not required.} Note that for $\bar{T}$ thus defined, we do not need to switch coordinate charts while comparing the trajectories $x(\cdot)$ and $x^{t_n}(\cdot)$ in the interval $[t_n, t_n+\bar{T}]$.\\ \noindent{\textit{Claim :}} If the theorem holds for $\bar{T}$, then it does so for any $T\geq\bar{T}$.\\ \noindent{\textit{Proof :}} Let the theorem hold for $T=\bar{T}$, we show that it also holds for $T=2\bar{T}$. We have, $$\sup_{t\in[s,s+2\bar{T}]} d(x(t),x^{s}(t)) \leq \sup_{t\in[s,s+\bar{T}]} d(x(t),x^{s}(t)) + \sup_{t\in[s+\bar{T},s+2\bar{T}]} d(x(t),x^{s}(t)).$$ The first term in the above goes to zero in the limit from the assumption. The second term can be handled as follows : \begin{align*} \sup_{t\in[s+\bar{T},s+2\bar{T}]} d(x(t),x^{s}(t)) & \leq \sup_{t\in[s+\bar{T},s+2\bar{T}]} d(x(t),x^{s+\bar{T}}(t)) +\\ & \qquad \qquad \sup_{t\in[s+\bar{T},s+2\bar{T}]} d(x^{s+\bar{T}}(t),x^{s}(t)), \end{align*} where $x^{s+\bar{T}}(t)$ satisfies $x^{s+\bar{T}}(s+\bar{T})= x(s+\bar{T})$. The first term in the RHS of the above inequality again goes to zero by assumption (note that the time window is of length $\bar{T}$ and $x^{s+\bar{T}}(t)$ starts at the point $x(s+\bar{T})$). The second term goes to zero because $$d(x^{s+\bar{T}}(s+\bar{T}),x^{s}(s+\bar{T}))=d(x(s+\bar{T}),x^{s}(s+\bar{T})) \to 0. $$ By the uniqueness and smoothness of the flow, the claim follows. \qed \\ For the rest of the proof we assume that $\bar{T}=T$. Let $t_{n+m}\in[t_{n},t_{n}+T]$, then we have from (\ref{ret}) : \begin{equation} \label{mantemp} x_{n+m}=\mathcal{R}_{x_{n+m-1}}\{a_{n+m-1}(H(x_{n+m-1})+M_{n+m})\}. \end{equation} The coordinate expression of the above can be written as : \begin{equation} \label{cordin-expre} \psi_{n}(x_{n+m}) = \psi_{n}(\mathcal{R}_{x_{n+m-1}}\{a_{n+m-1}(H(x_{n+m-1})+M_{n+m})\}). \end{equation} As shown earlier, $x_{n+m}$ belongs to $\mathcal{U}_{{n}}$ without loss of generality, by our choice of $\bar{T}$. A Taylor expansion of the term on the RHS of (\ref{cordin-expre}) gives, \begin{multline}\label{newmaneq} \psi_{n}\big(\mathcal{R}_{x_{n+m-1}}\{a_{n+m-1}(H(x_{n+m-1})+M_{n+m})\} \big)=\\ \begin{aligned} \psi_{n}(\mathcal{R}_{x_{n+m-1}}(0_{x_{n+m-1}})) \ + D\psi_{n}(\mathcal{R}_{x_{n+m-1}}(0_{x_{n+m-1}})) \ \times \\ D\mathcal{R}_{x_{n+m-1}} (0_{x_{n+m-1}}) (a_{n+m-1}(H(x_{n+m-1}) + \ M_{n+m}) \ +\\ \ \text{Rd}(\hat{x}_{n+m-1}). \end{aligned} \end{multline} Here the remainder $\text{Rd}(\hat{x}_{n+m-1})=\mathcal{O}(a_{n+m-1}^{2})$ because the iterates remain in a compact set and $\mathcal{R}(\cdot)$, $H(\cdot)$ are smooth. Thus using the local rigidity property of a retraction ($DR_{x}(0_{x})=\text{id}_{\mathcal{T}_{x}\mathcal{M}}$) and the fact that $\mathcal{R}_{x}(0_{x})=x$, we can write (\ref{newmaneq}) as \begin{multline*} \psi_{n}(\mathcal{R}_{x_{n+m-1}}\{a_{n+m-1}(H(x_{n+m-1})+M_{n+m})\})= \psi_{n}(x_{n+m-1}) + \\ D\psi_{n}(x_{n+m-1})(a_{n+m-1}(H(x_{n+m-1})+M_{n+m})+ \mathcal{O}(a_{n+m-1}^{2}), \end{multline*} which gives \begin{multline*} \psi_{n}(\mathcal{R}_{x_{n+m-1}}\{a_{n+m-1}(H(x_{n+m-1})+M_{n+m})\})= \hat{x}_{n+m-1} \ +\\ a_{n+m-1}\hat{H}(\hat{x}_{n+m-1}) +a_{n+m-1}\hat{M}_{n+m}+\mathcal{O}(a_{n+m-1}^{2}), \end{multline*} where we used the linearity of the map $D\psi_{n}(\cdot)$ in the last equation. Using this in (\ref{cordin-expre}), we get \begin{multline*} \hat{x}_{n+m}=\hat{x}_{n+m-1}+a_{n+m-1}\hat{H}(\hat{x}_{n+m-1})+ a_{n+m-1}\hat{M}_{n+m}+\mathcal{O}(a_{n+m-1}^{2}). \end{multline*} Doing a recursion on the above we have \begin{equation}\label{man-x} \hat{x}_{n+m}=\hat{x}_{n}+\sum_{k=0}^{m-1}a_{n+k}\hat{H}(\hat{x}_{n+k})+\delta_{n,n+k}, \end{equation} where \[ \delta_{n,n+k}=\sum_{k=0}^{m-1}a_{n+k}\big\{\hat{M}_{n+k+1}+\mathcal{O}(a_{n+k})\big\}. \] \begin{figure}[H] \begin{center} \includegraphics[width=11cm,height=8cm]{fig} \end{center} \caption{This figure depicts an illustration of the local parametrization at $t=t_n$, $t=t_{[n+T]}$} where $[n+T]$ $=\min\{ k : t_k \geq t_n + T \}$. Note that the piecewise linear trajectory corresponds to the iterates $\{\hat{x}_n\}$. \end{figure} Note that for $ k \geq 0$, $\hat{M}_{k+1} \in \mathbb{R}^n$ is a martingale since $D\psi(\cdot)$ is a linear operator. Furthermore, we have for any $n \geq 0$, \begin{align*} \mathbb{E}[\|D\psi_n(x_n)M_{n+1}\|^{2}|\mathcal{F}_{n}] \leq \|D\psi_n(x_n)\|^2 \mathbb{E}[\|M_{n+1}\|^{2}|\mathcal{F}_{n}], \end{align*} where $\|D\psi_n(x_n)\|$ denotes the norm of $D\psi_n(x_n)$. To show that the above quantity is bounded we note that the expectation term in the RHS is bounded by (A3). Also, by (A4) and the judicious choice of charts we have made, $\psi_n, D\psi_n$ vary over a (possibly sample path dependent) finite family and therefore are bounded uniformly in $n$ as maps $\mathcal{U}_{n} \mapsto \mathbb{R}^n$ and $x \in \mathcal{U}_{n} \mapsto$ the space of bounded linear operators $\mathcal{T}_{x}\mathcal{M} \mapsto \mathbb{R}^n$ with operator norm. So by (A3), (A4), we have $$\sup_n\mathbb{E}[\|D\psi_n(x_n)M_{n+1}\|^{2}|\mathcal{F}_{n}] < \infty.$$ By the martingale convergence theorem (Appendix C, \cite{BorkarBook}), since $\sum_{k\geq0} a_{n+k}^{2}<\infty$, it then follows that $\sum_k a_k \hat{M}_k $ is convergent a.s. This in turn implies that $\sup_{k}\delta_{n,n+k}$ $\to 0$ a.s.\ as $n\to\infty$. Next we establish similar bounds for the ODE considered in the coordinate expression. Integrating (\ref{newmanODE}) between the limits $t_{n}$ to $t_{n+m}$ we have : \begin{align} \hat{x}^{t_{n}}(t_{n+m}) & =\hat{x}^{t_{n}}(t_{n})+\int_{t_{n}}^{t_{n+m}}\hat{H}(\hat{x}^{t_{n}}(t))dt,\nonumber \\ & =\hat{x}^{t_{n}}(t_{n})+\sum_{k=0}^{m-1}a_{n+k}\hat{H}(\hat{x}^{t_{n}}(t_{n+k})) \nonumber \\ & +\int_{t_{n}}^{t_{n+m}}\big\{\hat{H}(\hat{x}^{t_{n}}(t))-\hat{H}(\hat{x}^{t_{n}}([t]))\big\} dt, \label{manintbound} \end{align} where $[t]=\max\{t_{n}:t_{n}\leq t\}$. We now establish a bound on the last integral on the right hand side. Since $H(\cdot)$ is smooth we have : \begin{equation}\label{Hdiff} \|\hat{H}(\hat{x})-\hat{H}(\hat{y})\|\leq\gamma\|\hat{x}-\hat{y}\| \end{equation} for some constant $\gamma$. Using this, we have from integrating (\ref{newmanODE}) between $t_n$ to $t\in [t_n, t_{n}+T)$, \begin{align*} \|\hat{x}^{t_{n}}(t)\| & \leq\|\hat{x}^{t_{n}}(t_{n})\|+\int_{t_n}^{t}\|(\hat{H}(\hat{x}^{t_{n}}(t))\|dt,\\ & \leq \underbrace{\|\hat{x}^{t_{n}}(t_{n})\|+ \| \hat{H}(0)T \|}_{c_T} +\int_{t_n}^{t}\|(\hat{H}(\hat{x}^{t_{n}}(t))-\hat{H}(0)\|dt, \\ & \leq c_{T}+\gamma\int_{t_n}^{t}\|\hat{x}^{t_{n}}(t)\|dt, \,\,\, \text{ (from (\ref{Hdiff}))} \end{align*} so that by the Gronwall inequality we have \[ \|\hat{x}^{t_{n}}(t)\|\leq c_{T}e^{\gamma T},\,\,t\in [t_n, t_{n}+T). \] Let $C_T \doteq c_{T}e^{\gamma T} + \| \hat{H}(0)\|$. Integrating (\ref{newmanODE}) between $t_{n+k}$ to $t\in [t_{n+k},t_{n+k+1})$ with $0\leq k \leq m-1$, \begin{align*} \|\hat{x}^{t_{n}}(t)-\hat{x}^{t_{n}}(t_{n+k})\| &\leq\|\int_{t_{n+k}}^{t}\hat{H}(\hat{x}^{t_{n}}(s))ds\|,\\ & \leq C_T (t-t_{n+k})\\ & \leq C_T a_{n+k}. \end{align*} This gives the required bound : \begin{align*} \|\int_{t_{n}}^{t_{n+m}}\big\{\hat{H}(\hat{x}^{t_{n}}(t))-\hat{H}(\hat{x}^{t_{n}}([t]))\big\} dt\| & \leq \int_{t_{n}}^{t_{n+m}}\|\hat{H}(\hat{x}^{t_{n}}(t))-\hat{H}(\hat{x}^{t_{n}}([t]))\|dt,\nonumber \\ & \leq\gamma\int_{t_{n}}^{t_{n+m}}\|\hat{x}^{t_{n}}(t)-\hat{x}^{t_{n}}([t])\|dt \,\,\, \text{ (from (\ref{Hdiff}))} \nonumber \\ & \leq\gamma\sum_{k=0}^{m-1}\int_{t_{n+k}}^{t_{n+k+1}}\|\hat{x}^{t_{n}}(t)-\hat{x}^{t_{n}}(t_{n+k})\|dt,\nonumber \\ & \leq\gamma C_{T}\sum_{k=0}^{m-1}a_{n+k}^{2},\\ \end{align*} so that \begin{equation}\label{maninterpol} \|\int_{t_{n}}^{t_{n+m}}\big\{\hat{H}(\hat{x}^{t_{n}}(t))-\hat{H}(\hat{x}^{t_{n}}([t]))\big\} dt\| \leq\gamma C_{T}\sum_{k=0}^{\infty}a_{n+k}^{2}. \end{equation} Subtracting (\ref{man-x}) and (\ref{manintbound}) and taking norms, we get : \begin{multline*} \|\hat{x}_{n+m}-\hat{x}^{t_{n}}(t_{n+m})\|\leq\|\hat{x}_{n}-\hat{x}^{t_{n}}(t_{n})\|+\|\delta_{n,n+m}\|\\ \begin{aligned} &+\|\int_{t_{n}}^{t_{n+m}}\big\{\hat{H}(\hat{x}^{t_{n}}(t))-\hat{H}(\hat{x}^{t_{n}}([t]))\big\} dt\| \\ &+ \sum_{k=0}^{m-1}a_{n+k}\|\hat{H}(\hat{x}_{n+k})-\hat{H}(\hat{x}^{t_{n}}(t_{n+k}))\|. \end{aligned} \end{multline*} Using (\ref{maninterpol}) in the above equation, we get \begin{multline*} \|\hat{x}_{n+m}-\hat{x}^{t_{n}}(t_{n+m})\|\leq\|\hat{x}_{n}-\hat{x}^{t_{n}}(t_{n})\|+\sup_{k\geq0}\delta_{n,n+k} +\gamma C_{T}\sum_{k=0}^{\infty}a_{n+k}^{2}\\ \begin{aligned} & +\sum_{k=0}^{m-1}a_{n+k}\|\hat{H}(\hat{x}_{n+k})-\hat{H}(\hat{x}^{t_{n}}(t_{n+k}))\|.\\ \end{aligned} \end{multline*} Set $K_{T,n}\doteq\sup_{k\geq0}\delta_{n,n+k}+\gamma C_{T}\sum_{k=0}^{\infty}a_{n+k}^{2}$. Since $\hat{x}_{n}=\hat{x}^{t_{n}}(t_{n})$ and using (\ref{Hdiff}), \begin{multline}\label{K-T} \|\hat{x}_{n+m}-\hat{x}^{t_{n}}(t_{n+m})\|\leq K_{T,n}+ \gamma \sum_{k=0}^{m-1}a_{n+k}\|\hat{x}_{n+k}-\hat{x}^{t_{n}}(t_{n+k})\|. \end{multline} Applying the discrete Gronwall inequality (\cite{BorkarBook}, p.\ 146) to the above we get, \[ \|\hat{x}_{n+m}-\hat{x}^{t_{n}}(t_{n+m})\|\leq K_{T,n}\exp\{\gamma T\}\;\text{a.s.} \] Since $K_{T,n}\to0$, the result follows. Note that this proves the theorem only for $t=t_n\uparrow\infty$. However, we can extend it for general $t\uparrow\infty$ by exploiting the fact that $a_n \to 0$, so that $t_{n+1} -t_n \to 0$. \qed \end{proof} \begin{defn} (Definition 1.4.1, \cite{Alongi}) A set $A$ is said to be invariant under the flow $\Phi_x(\cdot)$ if $x\in A$ implies that $\Phi_x(t) \in A \, \ \forall t \in \mathbb{R}$. \end{defn} \begin{defn} \label{chaintrans}(Definition 2.7.1, \cite{Alongi}) Let $\Phi_x(\cdot)$ be a flow on a metric space $(\mathcal{M},d(\cdot,\cdot))$. Given $\epsilon>0,\,T>0$, and $x,\,y\in\mathcal{M}$, an $(\epsilon,T)$ chain from $x$ to $y$ with respect to $\Phi_x(t)$ is a pair of finite sequences $x_1=x,x_2,...,x_{n-1},x_n=y$ and $t_1,...,t_{n-1} \in [T,\infty)$ such that $$ d(\Phi_{x_i}(t_i), x_{i+1} ) < \epsilon$$ for $i=1,..,n-1$. A set $A$ is internally chain transitive for the flow if for any choice of $x, y$ in this set and any $\epsilon, T$ as above, there exists an $(\epsilon, T)$ chain in $A$. \end{defn} The following result is a variant of the well known theorem due to Benaim \cite{Benaim0}, a simple exposition of which appears in pp.\ 15-16 of \cite{BorkarBook}, Chapter 2. Taking $\epsilon < \frac{r_0}{2}$ above without loss of generality, an identical argument serves to prove it and is omitted.\\ \begin{thm}\label{manthrm2} Almost surely, the sequence $\{x_n\}$ generated by (\ref{ret}) converges to a (possibly path dependent) internally chain transitive invariant set of (\ref{manODE}). \end{thm} The above theorem establishes the ODE method for manifolds. If $H(\cdot)$ is derived from the Riemannian gradient of a smooth function $\Psi:\mathcal{M} \to \mathbb{R}$ (see Example 2 for definition), then the behavior of the associated ODE, sometimes called the gradient flow, is easy to characterize. Specifically, the solution converges to a connected component of the set of critical points of $\Psi$ (Appendix C.12, Proposition 12.1, \cite{Helmke}). Since these are the only internally chain transitive sets of the gradient flow, (\ref{ret}) has the same asymptotic behavior. We conclude this section with some illustrative examples to justify the usefulness of Theorem \ref{manthrm}.\\ \noindent{\textbf{Example 2}\textit{ (Constrained Regression)} :} Suppose we have a stream of data pairs $\{X_n,Y_n\}_{n \geq 1}$ and a measurement model, $$ Y_n = f_w(X_n) + \epsilon_n, \, n \geq 1, $$ where $X_n \in \mathbb{R}^m $ is the input, $w \in \mathbb{R}^d$ is a parameter which is to be estimated, $Y_n \in \mathbb{R}^k $ is the measurement or output and $\epsilon_n$ is some measurement noise. The family of functions $\{f_w : \mathbb{R}^m \to \mathbb{R}^k : w \in \mathbb{R}^d \}$ parameterized by $w$ is chosen as a `best fit' so as to minimize a suitable scalar penalty on the error $\epsilon$. The outcome depends upon the error criterion selected and one potential option is to use the mean square error which gives the following optimization problem : $$\min_w \Big\{\Psi(w) \doteq \mathbb{E}[\Phi(w)] \doteq \frac{1}{2}\mathbb{E}[\| Y_n -f_w(X_n) \|^2 ] \Big\}. $$ Suppose in addition to this, we are constrained to stay inside a submanifold $\mathcal{S}$ of $\mathbb{R}^d$. The constraint can be linear, say \begin{equation}\label{affine} Lw=c, \end{equation} where $L \in \mathbb{R}^{p \times d} $ and $c \in \mathbb{R}^p$, or it could be nonlinear, for instance, say the $d-1$ dimensional sphere $S^{d-1}$. Note that for this problem the vector field will be the Riemannian gradient of $\Psi\, : \, \mathcal{S} \to \mathbb{R} $. By the Riemannian gradient, $\text{grad }\Psi(x)$ of $\Psi$ at $x$, we mean a vector that satisfies the following two properties :\\ \noindent (a) Tangency Condition : $$\text{grad }\Psi(x) \in \mathcal{T}_x\mathcal{S}, $$ \noindent (b) Compatibility Condition : $$ D\Psi(x) \cdot \xi = \langle \text{grad }\Psi(x), \xi \rangle \,\text{ for all } \xi \in \mathcal{T}_x\mathcal{S}. $$ Since the expectation cannot be evaluated, the only recourse is to replace the exact gradient by the argument of the expetation evaluated at the current guess $w_n$. For the constraint (\ref{affine}) we have (Section 1.6, \cite{Helmke}), \begin{align*} \text{grad }\Phi(w)&=(I_d-L^{\dagger}L)\nabla_w \Phi(w)\\ & =(I_d-L^{\dagger}L)\langle Y_n -f_w(X_n), \nabla _w f_w(X_n) \rangle, \end{align*} where $L^\dagger = L^T(LL^T)^{-1}$ denotes the pseudo-inverse with $L^T$ denoting the transpose and $\nabla_w (\cdot) $ is the Euclidean gradient. Then iteration (\ref{ret}) becomes \begin{equation} w_{n+1}=\mathcal{R}_{w_{n}}\big(-a_{n}\text{grad }\Phi(w_n) \big). \label{stochgradconstr} \end{equation} For the spherical constraint, we have (Section 1.6, \cite{Helmke}), \begin{align*} \text{grad }\Phi(w)& = \nabla_w \Phi(w) - \langle w,\nabla_w \Phi(w) \rangle w, \end{align*} where $\nabla_w \Phi(w) = \langle Y_n -f_w(X_n), \nabla _w f_w(X_n) \rangle$ as before. Assuming interchange of expectation and differentiation, a simple calculation shows that the iteration (\ref{stochgradconstr}) is of the form $$w_{n+1}=\mathcal{R}_{w_{n}}\big(a_{n}(-\text{grad }\mathbb{E}\left[\Phi(w_n)\right] + M_{n+1}) \big),$$ where $\{M_n\}$ is a martingale difference sequence. So by Theorem \ref{manthrm}, we have the limiting ODE for either case as $$ \dot{w} = - \mathbb{E}\,[\text{grad }\Phi(w)] = - \text{grad }\Psi(w).$$ Note that the retraction in either case can just be the projection map.\\ \noindent{\textbf{Example 3}\textit{ (Principal Component Analysis)} :} This problem entails the computation of the $r$ principal eigenvectors of an $n\times n$ covariance matrix $A = \mathbb{E}[z_{k}z_{k}^{T}]$ with $z_{1},...,z_{k},...$ being a stream of uniformly bounded $n$-dimensional data vectors. Define the cost function : \[ C(W)=-\frac{1}{2}\mathbb{E}[z^{T}W^{T}Wz]=-\frac{1}{2}\text{Tr}(W^{T}AW) \] with $\text{Tr}(\cdot)$ denoting the trace of matrix and $W\in\mathcal{S}_{n,r}$ which becomes a basis of the dominant $r$-dimensional invariant subspace of $A$ when $C(W)$ is minimized. The constraint space of the above minimization problem is the Grassmann Manifold\footnote{This is because $C(W)$ is invariant to rotations $W\to WO$ for any orthogonal matrix $O$.}, $\mathcal{G}(n,r)$ of $r$-dimensional subspaces in a $n$-dimensional ambient space, i.e., it is the quotient manifold $$ \mathcal{G}(n,r) = \big\{ \mathcal{S}_{n,r} / O(r) \big\}, $$ where $O(r)$ is set of orthogonal $r \times r$ matrices. The noisy Riemannian gradient of $C(W)$ under the sample $z$ is given by $H(z,W)=(I_n-WW^{T})zz^{T}W$, so that iteration (\ref{ret}) becomes : \begin{equation}\label{Oja} W_{k+1}=\mathcal{R}_{W_{k}}\{-a_{k}(I_{n}-W_{k}W_{k}^{T})z_{k}z_{k}^{T}W_{k}\}. \end{equation} Note that the ODE vector field here is $H(W)=(I_{n}-WW^{T})AW$. From Theorem \ref{manthrm} the relevant ODE is \[ \dot{W}=(I_{n}-WW^{T})AW. \] This is the celebrated Oja's algorithm \cite{Oja}. To see how the above can be interpreted as an ODE on $\mathcal{G}_{n,r}$, see (\cite{Hairer}, Example 9.1). The ODE equilibrium points correspond to \begin{align*} AW&=WW^{T}AW \\ \Rightarrow AW &= WM, \end{align*} for $M=W^TAW$, proving that any limit $W$ is an invariant subspace of $A$. A retraction on $\mathcal{S}_{r,n}$ which could be used in (\ref{Oja}) is given by (Example 4.1.3, \cite{Absil}) : $$ \mathcal{R}_{W}(a H) = \text{qf}(W + aH),$$ where $\text{qf}(\cdot)$ gives the orthogonal factor in the QR-decomposition of its argument. \section{Approximate Retractions on Submanifolds} \subsection{ Sub-manifolds as constraint sets} In this subsection we consider the case of the constraint being an embedded submanifold of a Euclidean space. The setup is as follows : The constraint space is a submanifold $\mathcal{S}$ of class $\mathcal{C}^{k}$ ($k\geq2$) of a Euclidean space $\mathcal{E}$ of dimension $n$. A submanifold of class $\mathcal{C}^{k}$ and dimension $d$ here means that $\mathcal{S}$ is locally a coordinate slice, that is, for all $x\in\mathcal{S}$, there exits a neighborhood $\mathcal{U}_{\mathcal{E}}$ of $x$ in $\mathcal{E}$ and a $\mathcal{C}^{k}$ diffeomorphism $\psi$ on $\mathcal{U}_{\mathcal{E}}$ into $\mathbb{R}^{n}$ such that \[ \mathcal{S}\cap \mathcal{U}_{\mathcal{E}}=\{x\in\mathcal{U}_{\mathcal{E}}\,:\,\psi_{d+1}(x)=\cdot\cdot\cdot=\psi_{n}(x)=0\}. \] As before, assume (A0)-(A4). The only modification now is that the vector field is now defined on the entire embedding space $\mathcal{E}$, i.e. $H\,:\,\mathcal{E} \to \mathcal{E}$ and instead of being smooth, we only require it to be Lipschitz. Also, instead of using coordinate charts to prove the relevant results, we can directly use the norm $\|\cdot\|$ inherited from the embedding space $\mathcal{E}$. In addition, we assume the following. (A5) There exists an $f:\mathbb{R}^{n}\to\mathbb{R}^{n}$ such that for $f^n \doteq f\circ f\circ \cdots \circ f$ ($n$ times), $P(x)\doteq\lim_{n\to\infty}f^{n}(x$) exists for all $x\in\mathbb{R}^{n}$. We assume the following conditions on $f$ and $P$ : \begin{enumerate} \item[(I)] $f$ is continuous, \item[(II)] $f^{n}\to P$ uniformly on compacts, so that $P$ is continuous. \item[(III)] $P$ is the Euclidean projection onto the submanifold $\mathcal{S}$, i.e. \[ \mathbf{}P(y)=\textrm{arg}\min_{x\in \mathcal{S} }\|y-x\|. \] \end{enumerate} The aim of this subsection is to provide a framework to do projected SA on submanifolds. The idea is to use $f$ to effect a projection to a differentiable submanifold. This will enable us to execute stochastic approximation versions of algorithms on constraint sets such as matrix manifolds. We will provide some examples later in the section on how to choose such a $f$ for certain sub-manifolds. In case such a $f$ does not exist, one can replace $f$ with the exact retraction operation (see Lemma \ref{manproj} below), but the main point of the presented framework is to show that one need not use exact projections while doing projected SA. We first note an interesting connection between retractions and projections for submanifolds, which we recall in a lemma borrowed from \cite{Malick}. For a submanifold $\mathcal{S}$, the projection is interpreted in the usual way with the metric taken to be the standard Euclidean metric of the ambient space $\mathcal{E}$. We state the following result from (\cite{Malick}, Lemma 4) : \begin{lem}\label{manproj} Let $\mathcal{S}$ be a submanifold of class $\mathcal{C}^{k}$ around $x\in\mathcal{S}$ and $P_{\mathcal{S}}$ denote the projection onto $\mathcal{S}$. Then $P_{\mathcal{S}}$ is well defined (always exists and is unique) locally and the function $P_{\mathcal{S}}$ is of class $\mathcal{C}^{k-1}$ around $x$ with \[ DP_{\mathcal{S}}(x)=P_{\mathcal{T}_x\mathcal{S}}, \] where $P_{\mathcal{T}_x\mathcal{S}}$ is the orthogonal projection onto $\mathcal{T}_x\mathcal{S}$, the tangent space at point $x$. Also, the function $\mathcal{R}:\mathcal{TS}\to\mathcal{S}$ defined by $(x,u)\longmapsto P_{\mathcal{S}}(x+u)$ is a retraction around x. \end{lem} We propose the following iteration to do (projected) stochastic approximation on embedded submanifolds: \begin{equation}\label{nongos} x_{k+1} = f(x_k) + a_k(H(x_k) + M_{n+1}). \end{equation} Note that in the above iteration, feasibility is satisfied only asymptotically as proved in Lemma 10, \cite{Mathkar}. The analysis of this algorithm is easy for submanifolds (of class $\mathcal{C}^k,\,k\geq2$) as we can simply take the Taylor expansion of (\ref{nongos}) after projecting it : $$ P(x_{k+1}) = P(f(x_k) + a_k(H(x_k) + M_{n+1}) ).$$ Setting $\tilde{x}_k = P(x_k)$ we have, $$ \tilde{x}_{k+1} = P(f(x_k)) + a_k DP_{f(x_k)}(H(x_k) + M_{n+1}) ) +\mathcal{O}(a^2_k).$$ The rest of the analysis is similar to \cite{Mathkar}, and noting from Lemma \ref{manproj} that $DP_{\mathcal{S}}(x)=P_{\mathcal{T}_x\mathcal{S}}$ and $f(x)=x$ for any limit point, the limiting behaviour of (\ref{nongos}) is similar to the following projected ODE : \begin{equation}\label{Fret} \dot{x} = P_{\mathcal{T}_x\mathcal{S}}(H(x(t))). \end{equation} We give some example's that fit this framework :\\ \noindent{\textbf{Example 4}\textit{ (Affine Manifold)} :} We first examine the simple case of an affine manifold, i.e., $$\mathcal{S} = \{x \in \mathbb{R}^n \,:\,Ax=b\},$$ where $A \in \mathbb{R}^{m\times n}$ and $b\in \mathbb{R}^m$. Note that this manifold can be obtained as $$\mathcal{S} = \bigcap_{i=1}^m \mathcal{S}_i, $$ where for $a_i \doteq$ the $i$ th row of $A$, $$ \mathcal{S}_i \doteq \{ x \in \mathbb{R}^n \,:\, a_i^T x=b_i \}. $$ One way to project onto this set is to use alternating projections. Then if $P^i$ denotes the projection onto $\mathcal{S}_i$, a single full iteration of alternating projections can be used as a substitute for $f$, so that $$ f(\cdot) = P^m \circ \cdots \circ P^1(\cdot).$$\\ \noindent{\textbf{Example 5}\textit{ (Manifold of Correlation Matrices)} :} Suppose we have a constraint minimization problem of the following kind : \begin{align*} \min_{C \in \mathbb{R}^{n\times n}} & \quad f(C) \\ \text{s.t.} \,\,\,\,\,\,& C=C^T,\, C \succeq 0,\,C_{ii}=1\,(i=1, \cdots, N), \end{align*} where $C \succeq 0$ for a symmetric $C$ denotes that $C$ is positive semidefinite. The constraint set is the set of symmetric correlation matrices, i.e., the set of symmetric positive semidefinite matrices with diagonal elements equal to one. Note that the constraint can be obtained as the intersection of the following two convex sets : $$ S \doteq \{ C=C^T \in \mathbb{R}^{n \times n} : C \succeq 0 \}, $$ $$U \doteq \{ C=C^T \in \mathbb{R}^{n \times n} : C_{ii} = 1, \, i=1, \cdots, n \}.$$ The projection on both of these sets is well defined (Theorem 3.1, Theorem 3.2, \cite{Higham}). One can use the Boyle-Dykstra-Han algorithm (see Algorithm 3.3, \cite{Higham}) to project onto the above constraint set. As in the previous example, one can use a single full iteration of this algorithm as a substitute for $f$ in (\ref{non}).\\ \noindent{\textbf{Example 6} : Suppose we have a constraint set of the following form which arises quite often in control theory (see \cite{Lew}) : $$ \mathcal{R}_r \cap \mathcal{M}_\mathcal{A},$$ where $\mathcal{R}_r$ is as in Example 1 and $$ \mathcal{M}_A \doteq \{ X \in \mathbb{R}^{n \times m} \,:\, \mathcal{A}(X) =b\} $$ for a given linear map $\mathcal{A} :\mathbb{R}^{n\times m} \to \mathbb{R}^d$ and vector $b \in \mathbb{R}^d $. Note that the projection on both sets is well defined and the intersection is a manifold since $\mathcal{S}_r$ and $\mathcal{M}_A$ intersect transversally, i.e., $$\mathcal{T}_{\mathcal{R}_r} + \mathcal{T}_{\mathcal{M}_\mathcal{A}} = \mathbb{R}^{n \times m}.$$ Again, one could adapt the subroutine suggested in the previous examples. \begin{rem} One of the pitfalls of this scheme is that the ODE approximation to (\ref{Fret}) requires Frechet differentiability of the projection map (see assumption (B3), \cite{Mathkar}) which, although satisfied for submanifolds of class $\mathcal{C}^k,\,k\geq2$, is not satisfied for constraint sets with nondifferentiable boundaries. This motivates an alternative which we explore in the next subsection. \end{rem} \subsection{General Constraint Sets} In continuation of the framework presented in the previous subsection we extend it to general constraint sets which may not be differentiable manifolds. The justification for this is that the projection map is frequently not differentiable for constraints in optimization or stochastic approximation, so the results of the previous subsection will not hold. Here we consider more general sets such as polyhedra, intersection of convex sets which have boundaries that may be nonsmooth or nonconvex sets. We propose the following iteration to do (projected) stochastic approximation : \begin{equation}\label{non} x_{k+1} = (1 - \gamma_k)x_k + \gamma_k f(x_k) + a_k (H(x_k) + M_{k+1}), \end{equation} where $ a_k = o(\gamma_k)$ and both satisfy Assumption (A2). The rest of the assumptions are the same as in subsection 3.1 with one addition, viz., we require $f$ to be non-expansive, i.e. for any $x,y\in\mathcal{E}$ $$\|f(x)-f(y)\|\leq\|x-y\|.$$ We first provide some examples that fit the above framework.\\ \noindent{\textbf{Example 7} \textit{ (Polyhedral Set)} :} The most important example of a constraint set that fits this setup is the polyhedral set given by : $$\mathcal{X} = \bigcap_{i=1}^m \mathcal{X}_i, $$ where $$ \mathcal{X}_i \doteq \{ x \in \mathbb{R}^n \,:\, a_i^T x \leq b_i \}. $$ For this constraint set, the Boyle-Dykstra-Han algorithm produces an iterate which lies in the convex hull of the previous iterate and its projection (Theorem 2.4, \cite{Deut2}), i.e., \begin{equation} f^{i}(x)=(1-\lambda)x +\lambda P^{i} (x), \end{equation} where $\lambda \in (0,1)$. We can take $f(\cdot)=f^1\circ \cdots\circ f^m (\cdot)$ as in Example 4. Note that each $f^i$ is non-expansive here.\\ \noindent{\textbf{Example 8} \textit{ (Two time scale projection)} :} Another approach to tackle the projection which falls under this setup was proposed in \cite{SuhBor}. We mention it briefly here. It involves running the following iteration : $$x_{n+1} = (1-a_k)x_k + a_k z_k + a_k (H(x_k) +M_{k+1}), $$ where $z_k$ is obtained from a sub-routine running on a faster time scale and satisfies $\|z_k -P(x_k)\|\to 0 $. For more details on this algorithm see \cite{SuhBor}.\\ Let $d(x,A)=\inf_{y \in \mathcal{S}}\|x-y\|$ be the distance of a point $x$ from the constraint set $\mathcal{S}$ which we assume to be closed. Since we are no longer dealing with smooth constraints, we have to deal with more general notions of tangent and normal spaces. We define the tangent cone and the normal cone for our constraint set $\mathcal{S}$ as : \begin{defn} For $x \in \mathcal{S}$, we define by $$T_{\mathcal{S}}(x) \doteq \{v \in \mathbb{R}^n \,|\, \lim_{h\to 0^+}\inf \text{d}(x + hv, \mathcal{S})/h = 0\},$$ the tangent cone to $\mathcal{S}$ at $x$ and by \begin{equation}\label{cone} N_{\mathcal{S}}(x) \doteq \{p \in \mathbb{R}^n \,|\, \langle p,v \rangle \leq 0 \, \forall \, v \in T_{\mathcal{S}}(x) \}, \end{equation} the normal cone to $\mathcal{S}$ at $x$. \end{defn} $T_{\mathcal{S}}(x)$ is sometimes also called the contingent cone. Also, for later use, we define the following set, \begin{equation*} N_{\text{prox}(\mathcal{S})}(x) \doteq \{v \in \mathbb{R}^n \,|\, v \in P^{-1}(x) -x \}. \end{equation*} This set is referred to as the proximal normal cone. It contains vectors $v \in N_{\text{prox}(\mathcal{S})}(x) $, such that $x \in P(x+ \tau v)$ for some $\tau >0$. We have (Lemma 6.1.1 \cite{Aubin1}), \begin{equation}\label{proxcone} N_{\text{prox}(\mathcal{S})}(x) \subset N_{\mathcal{S}}(x). \end{equation} The inclusion can be strict, even for sets defined as smooth inequalities (Section E, Chapter 6, \cite{Rock}). The relevant differential inclusion for our problem is \begin{equation} \dot{x}\in H(x(t))-N_{\mathcal{S}}(x(t)), \label{inclusion2} \end{equation} \[ x(t)\in\mathcal{S}\, \ \ \forall t\in[0,T], \] where $H(\cdot)$ is the vector field under consideration. For convex sets, this inclusion is identical to the well known `` Projected Dynamical System" considered in \cite{Nagurney} : \begin{equation*} \dot{x}=\Pi_{T_{\mathcal{S}}(x)}(x,H(x)), \end{equation*} where $\Pi_{T_{\mathcal{S}}(x)}(x,H(x))$ is defined to be the following limit for any $x\in\mathcal{S}$ : \[ \Pi_{T_{\mathcal{S}}(x)}(x,H(x))\doteq\lim_{\delta\to0}\frac{P(x+\delta H(x))-x}{\delta}. \] This map gives the projection of $H(x)$ on the tangent cone $T_{\mathcal{S}}(x)$. For arbitrary closed sets, $T_{\mathcal{S}}(x)$ is replaced by its convex hull, $\text{conv}\{T_{\mathcal{S}}(x)\}$. The proof of the fact that the operator $\Pi_{T_{\mathcal{S}}(x)}(\cdot,\cdot)$ is identical to the RHS of (\ref{inclusion2}) is provided in (\cite{Dupuis}, Lemma 4.6) for convex sets and in (\cite{Serea}, Proposition 5) for arbitrary closed sets. We recall the following notion from \cite{Benaim} where more general differential inclusions are considered: \begin{defn}(\cite{Benaim}) Suppose $F$ is a closed set valued map such that $F(x)$ is a non-empty compact convex set for each $x$. Then a perturbed solution $y$ to the differential inclusion \begin{equation} \dot{x}\in F(x) \label{DifInc} \end{equation} is an absolutely continuous function which satisfies:\\ \noindent i) $\exists$ a locally integrable function $t\to U(t)$ such that for any $T>0$, $$\lim_{t\to\infty}\sup_{0\leq v\leq T}\big|\int_{t}^{t+v}U(s)ds\big|=0,$$ ii) $\exists$ a function $\delta:[0,\infty)\to[0,\infty)$ with $\delta(t)\to0$ as $t \to \infty$ such that \[ \dot{y}-U(t)\in F^{\delta(t)}(y), \] where $F^{\delta}( \cdot )$ denotes the the following set : \[ F^{\delta}(x)=\big\{ z\in\mathbb{R}^{n}\,:\,\exists x' \,\textrm{such that}\, \ \|x-x'\|<\delta. \ \text{d}(z,F(x'))<\delta\big\} \] \end{defn} There is no guarantee that the perturbed solution remains close to a solution of (\ref{DifInc}), however, the following assumption helps in establishing some form of convergence. Let $\Lambda$ denote the equilibrium set of (\ref{inclusion2}), assumed to be non-empty. Then : \textit{ \[ \Lambda \subset \{x\,:\,H(x)\in N_{\mathcal{S}}(x)\}. \] } Assume: \noindent (A6) There exists a Lyapunov function for the set $\Lambda$, i.e., a continuously differentiable function $V:\mathbb{R}^{n}\to\mathbb{R}$ such that any solution $x$ to (\ref{inclusion2}) satisfies \[ V(x(t))\leq V(x(0)) \ \forall \\ t > 0 \] and the inequality is strict whenever $x(0)\notin \Lambda$.}\\ (A6) is satisfied, for instance if $H=-\nabla g$ for some continuously differentiable $g$. Then $\Lambda$ represents the KKT points and the function $g$ itself will serve as a Lyapunov function for the set $\Lambda$. If (A6) does not hold, one can still say the following. The asymptotic behavior is analogous to Theorem \ref{manthrm2}. Specifically, we already know that under reasonable assumptions, an SA scheme converges a.s.\ to an internally chain transitive invariant set of the limiting o.d.e. This behavior extends to differential inclusions as well (see \cite{Benaim} or Chapter 5, \cite{BorkarBook}) and is what we would expect for this algorithm if we remove (A6). It is important, however, that an `invariant' set should be defined in the weak sense, i.e., any point in it should have at least one trajectory (as opposed to \textit{all} trajectories) of the differential inclusion passing through it, that remains in the set for all time. The following result is from (\cite{Benaim}, Prop. 3.27) : \begin{prop} \label{Perturbed} Let $y$ be a bounded perturbed solution to (\ref{inclusion2}) and there exist a Lyapunov function for a set $\Lambda$ with $V(\Lambda)$ having an empty interior. Then \[ \bigcap_{t\geq0}\overline{y\big([t,\infty)\big)}\subset\Lambda. \] \end{prop} \noindent (A7) The following condition is satisfied $$\lim_n\sum_{j=n}^{n+m(n)} \gamma_j \|f(x_j)-x_j \| \to 0$$ with, $$m(n) := \min\{k \geq n : \sum_{j=n}^{n+k}a_j \geq T\}, n \geq 0.$$ Before, proceeding further, we discuss assumption (A7) since it seems a bit restrictive. To get some intuition regarding this assumption we first note that (\ref{non}) draws a parallel with the Tsitsiklis scheme \cite{Tsitsiklis}. The fixed points ($f(x)=x$) in the Tsitsiklis scheme belong to the consensus subspace which is the set of fixed points of some doubly stochastic matrix (say $Q$), i.e. the set $Qx=x$ (see \cite{Nedic}). Convergence to this set takes place at a linear rate which is a direct consequence of the fact that the product of doubly stochastic matrices converges to its stationary average at a linear rate. This helps in ensuring that the error terms $\|Qx_k-x_k\|$ stay sufficiently small (see Lemma 1, \cite{Mathkar}). However, we do not have such a condition here, in particular the convergence rate of $f^n (\cdot) \to P (\cdot)$ is not assumed to be linear. But all we need is (A7) and it can be ensured by making sure that the term $\|f(x_k)-x_k\|$ becomes sufficiently small. For instance, for Example 7 this can be done by replacing $f$ with a time dependent $f_n$, $$f_{n}(\cdot) = \underbrace{f\circ..\circ f}_{i_n \text{ times}}(\cdot), $$ where $i_n$ denotes a sufficient number of Boyle-Dykstra iterations, so that the term $\|f(x_k)-x_k\|$ stays small \cite{SuhBor}. We have to also make sure that there is at least one solution to (\ref{inclusion2}). A class of sets for which this holds is (bounded) proximal retracts which include convex sets, sets with continuously differentiable boundaries etc. A proximal retract set is defined as follows : \begin{defn} \cite{Serea} A closed set $\mathcal{S}\subset\mathbb{R}^{n}$ is called a proximal retract if there is a neighborhood $N$ of $\mathcal{S}$ such that the projection $P(\cdot)$ is single valued in $N$. \end{defn} \noindent (A8) The set $\mathcal{S}$ is a bounded proximal retract.\\ \begin{rem}\label{remark1} To prove the main result, we shall show that the suitably interpolated iterates generated by (\ref{non}) form a perturbed solution to the differential inclusion (\ref{inclusion2}), so that using Proposition \ref{Perturbed} (along with assumption (A6)), the algorithm is shown to converge to its equilibrium set.\\ \end{rem} \begin{thm}\label{pert.} If (A1)-(A8) hold, we have \[ x_n\to\{x\,:\,x\in\Lambda\} \ \mbox{a.s.}, \] where $\Lambda$ denotes the set : \[ \Lambda =\{x\in\mathcal{S}\,:\,H(x)\in\mathcal{N}_{\mathcal{S}}(x)\}. \] \end{thm} \begin{proof} The proof presented here uses the techniques of \cite{Benaim}, \cite{Bianchi}. Write (\ref{non}) as : \begin{equation}\label{thrmeq} x_{k+1} = (1 - \gamma_k)x_k + \gamma_k (f(x_k) + \epsilon_k + \hat{M}_{k+1}), \end{equation} where, $$\epsilon_k = \frac{a_k}{\gamma_k}(H(x_k)) \text{ and } \hat{M}_{k+1}= \frac{a_k}{\gamma_k}M_{k+1}.$$ Note that $\epsilon_k \to 0$ since $H(x)$ is bounded from (A1),(A4) and $a_k=o(\gamma_k)$. Also, $\hat{M}_{k+1}$ is just a scaled martingale. We first prove that $x_n \to \mathcal{S}$.\\ \noindent \textit{Claim :} $\lim_k \inf_{z\in \mathcal{S}}\|x_{k}-z\| = 0\, $\\ \noindent \textit{Proof.} Let us first consider the following fixed point iteration : \begin{equation}\label{xtildeeq} \tilde{x}_{k+1}=(1-\gamma_{k})\tilde{x}_{k}+ \gamma_{k}f(\tilde{x}_{k}). \end{equation} By the arguments of \cite{BorkarBook}, Chapter 2, this has the same asymptotic behavior as the o.d.e. \[ \dot{\bar{x}}=f(\bar{x})-\bar{x}. \] Consider the Lyapunov function $V(\bar{x})=\frac{1}{2}\|\bar{x}\|^{2}$. W.l.o.g. assume that $0$ is a fixed point of $f(\cdot)$.\footnote{If this does not hold, we use the Lyapunov candidate $V(\bar{x})=\frac{1}{2}\|\bar{x}-x^*\|^{2}$ for some fixed point $x^*$ of $f$.} Then \begin{align*} \frac{d}{dt}V(\bar{x}(t)) & =\bar{x}(t){}^{T}(f(\bar{x}(t))-\bar{x}(t))\\ & =\bar{x}(t){}^{T}f(\bar{x}(t))-\|\bar{x}(t)\|^{2}. \end{align*} For any $v\in\mathbb{R}^{n}$, using the non-expansive property of $f$ and the fact that $0$ is a fixed of $f$ leads to \[ \|f(v)\|\leq\|v\|. \] By the Cauchy-Schwartz inequality, \[ \frac{d}{dt}V(\bar{x}(t))\leq 0. \] By Lasalle's invariance principle we have any trajectory $\bar{x}(\cdot)$ converging to the largest invariant set where $\frac{d}{dt}V(\bar{x}(t))=0$, which is precisely the set of fixed points of $f$. The claim for $\{\tilde{x}_k\}$ now follows by a standard argument as in Lemma 1 and Theorem 2, pp.\ 12-16, \cite{BorkarBook}. To prove the claim, we simply note that (\ref{thrmeq}) and (\ref{xtildeeq}) only differ by an $o(1)$ perturbation and the noise term $\hat{M}_k$, which does not affect the convergence to the fixed ponts of $f$ (see remark, p.\ 17, \cite{BorkarBook}). This implies in particular that any limit point of (\ref{xtildeeq}), say $x^*$, satisfies $$ f(x^*)=x^*,$$ $$\Rightarrow P(x^*)=\lim_nf^n(x^*)=x^* .$$ Let $t_{0}=0$ and $t_{k}=\sum_{i=0}^{k-1}a_{i}$ for any $k\geq1$, so that $t_{k+1}-t_{k}=a_{k}$. Define the interpolated trajectory $\Theta:[0,\infty)\to\mathbb{R}^{n}$ as : \[ \Theta(t)=x_{k}+(t-t_{k})\,\frac{x_{k+1}-x_{k}}{t_{k+1}-t_{k}},\ t\in[t_{k},t_{k+1}],\ k\geq1 \] By differentiating the above we have \begin{align}\label{der.} \frac{d\Theta(t)}{dt}= & \frac{x_{k+1}-x_{k}}{a_{k}}\,\,\,\,\forall\,t\in[t_{k},t_{k+1}], \end{align} where we use the right, resp., left derivative at the end points. Thus by setting $Z_{k} \doteq \frac{\gamma_k}{a_k} (f(x_{k})-x_{k})$ and $W_{k}\doteq x_{k}-P(x_{k})$, we get the following from (\ref{non}) and (\ref{der.}) : \begin{equation} \frac{d\Theta(t)}{dt} \in H(x_{k})-W_{k}+W_k+Z_{k}+M_{k+1}. \label{theta1} \end{equation} We note here that $W_k \in N_{\text{prox}(\mathcal{S})}(x) (P(x_k)) \subset N_\mathcal{S}(P(x_k))$. Let $F(\cdot)$ denote the following set valued map \begin{equation} F(x)=\{H(x)-W\,:\, W \in N_\mathcal{S}(x)\,, \ \|W\|\leq K\},\label{eff} \end{equation} where $0<K<\infty$ is a suitable constant. Note that, $$ F(x_k)=H(x_{k})-W_{k} \subset F^\delta(P(x_k)) \subset F^{\delta +\delta' } (\Theta(t)), $$ where $\delta = \|x_k -P(x_k) \|$ and $\delta' = \|P(x_k)-\Theta(t) \| $. We have used the following fact in the above - for any set valued map $\tilde{H}(\cdot)$ we have \[ \forall\,(x,\hat{x})\ \in\mathbb{R}^{n}\times\mathbb{R}^{n},\;\!\ \tilde{H}(x)\subset \tilde{H}^{\|x-\hat{x}\|}(\hat{x}). \] To finish the proof, $\Theta(\cdot)$ is first shown to be a perturbed solution of a projected dynamical system. Let $\eta(t)\doteq\|x_k -P(x_k) \|+\|\Theta(t)-P(x_{k})\|,\,t\in[t_{k},t_{k+1}),k\geq0$. Also define \[ U(t)=U_{k}:=Z_{k}+W_{k+1}+M_{k+1}, \] for $t\in[t_{k},t_{k+1}),k\geq0$. Then we have from (\ref{theta1}), \begin{equation} \frac{d\Theta(t)}{dt}-U(t)\in F^{\eta(t)}(\Theta(t)).\label{incl1} \end{equation} If we show that $\eta(t)\to0$ and $\sum_{k}a_{k}U_{k}<\infty$, we are done by Proposition \ref{Perturbed} and Remark \ref{remark1}. Then, by (\ref{incl1}), $\Theta(\cdot)$ can be interpreted as a perturbed solution of the differential inclusion \[ \dot{\Psi}(t)\in F(\Psi(t)). \] Convergence to the set $\Lambda$ then follows as discussed in Remark \ref{remark1}. We first prove that : \[ \lim_{t\to\infty}\sup_{0\leq v\leq T}\big|\int_{t}^{t+v}U(s)ds\big|=0. \] We have \[ \int_{t_k}^{t}U(t)=a_{k}\big[\underbrace{M_{k+1}}_{I}+\underbrace{Z_{k}+W_k}_{II}\big] \] for $t=t_{k+1}$. For $T > 0$, let $$m(n) := \min\{k \geq n : \sum_{j=n}^{n+k}a_j \geq T\}, \ n \geq 0.$$\\ I: This term is the error induced by the noise. The process $\sum_{m=0}^{k-1}a_m\times$ $M_{m+1}, k \geq 1,$ is a zero mean square integrable martingale w.r.t.\ the increasing $\sigma$-fields defined in (A3), with $\sum_{m=0}^{\infty}a_m^2\mathbb{E}\left[\|M_{m+1}\|^2 | \mathcal{F}_m\right] < \infty$ from (A3) and the square-summability of $\{a_m\}$. It follows from the martingale convergence theorem (Appendix C, \cite{BorkarBook}), that this martingale converges a.s. Therefore \[ \sup_{\ell \leq m(n)}\|\sum_{k=n}^{\ell}a_{k}M_{k+1}\|\to0\,\,\,\text{a.s. } \] II: The fact that $a_kZ_k \to 0$ is simply assumption (A7). Note that \begin{align*} \sup_{\ell \leq m(n)}\|\sum_{k=n}^{\ell} a_kZ_k\| &= \sup_{\ell \leq m(n)}\|\sum_{k=n}^{\ell} \gamma_k (f(x_k) -x_k)\|,\\ & \to 0. \end{align*} From the claim proved previously we have $\|x_k -P(x_k)\|\to0$, so that \begin{align*} \sup_{\ell \leq m(n)}\|\sum_{k=n}^{\ell} a_k W_k\| &= \sup_{\ell \leq m(n)}\|\sum_{k=n}^{\ell} a_k (x_k-P(x_k) )\|,\\ & \leq \sup_{\ell \leq m(n)} \|x_{\ell} -P(x_{\ell}) \| \cdot \Big( \sup_{\ell \leq m(n)}\sum_{k=n}^{\ell} a_k \Big) ,\\ & \leq \sup_{\ell \leq m(n)} \|x_{\ell} -P(x_{\ell}) \|\cdot T\\ & \to 0. \end{align*} All that is left to prove that the theorem holds is that $\eta(t) \to 0$. By the definition of $\Theta(\cdot)$, for any $t\in[t_{k},t_{k+1}]$ : \begin{align*} \|\Theta(t)-P(x_{k})\| & \leq \|x_{k}-P(x_{k})\| \,+\|x_{k+1}-x_k\|\frac{\|t-t_{k}\|}{\|t_{k+1}-t_{k}\|},\\ & = \|x_{k}-P(x_{k})\|+ \mathcal{O}(\gamma_{k}), \\ & \to 0, \end{align*} using the claim proved above. Thus $\eta(t) = \|x_{k}-P(x_{k})\|+ \|\Theta(t)-P(x_{k})\| \to 0 $. This completes the proof. \qed \end{proof} \section{Conclusions} To conclude we note some important connections with related works in order to suggest some possible extensions and future research directions. \begin{enumerate}[label=(\roman*)] \item \textit{Newton like schemes for stochastic optimization } : Second order methods have attracted a lot of attention recently for stochastic optimization, particularly in the field of machine learning. The same trend has been observed in the field of Riemannian optimization, most notably in \cite{Qi}, \cite{Ring} and \cite{Wen} where quasi-Newton methods are studied. The same level of attention as the first order methods, however, has not been devoted to second order stochastic optimization methods and one would like to see how these algorithms fare here.\\ \item \textit{ODE computational methods } : The Euclidean Robbins-Monroe scheme can be thought of as a noisy discretization of the limiting ODE. In fact the SA scheme differs from the standard Euler scheme only in using a decaying time step and the presence of noise. The computational methods for solving ODEs on manifolds offer a lot more diversity (see Chapter 4, \cite{Hairer}). In fact the retraction scheme itself seems to draw a parallel with the projection methods used to solve ODEs on manifolds. One could perhaps adapt the various techniques there to the present case and study the properties of the resultant algorithms. \\ \end{enumerate}

\section{Introduction} \label{} Network analysis has attracted considerable research efforts in many areas of artificial intelligence, as networks are able to encode rich and complex data, such as human relationships and interactions. One major challenge of network analysis is how to represent network properly so that the network structure can be preserved. The most straightforward way is to represent the network by its adjacency matrix. However, it suffers from the data sparsity. Other traditional network representation relies on handcrafted network feature design (e.g., clustering coefficient), which is inflexible, non-scalable, and requires hard human labor. In recent years, network representation learning, also known as network embedding, has been proposed and aroused considerable research interest. It aims to automatically project a given network into a low-dimensional latent space, and represent each vertex by a vector in that space. For example, a number of recent works apply advances in natural language processing (NLP), most notably models known as word2vec~\cite{mikolov2013efficient}, to network embedding and propose word2vec-based algorithms, such as DeepWalk~\cite{perozzi2014deepwalk} and node2vec~\cite{grover2016node2vec}. Besides, researchers also consider network embedding as part of dimensionality reduction techniques. For instance, Laplacian Eigenmap (LE)~\cite{belkin2003laplacian} aims to learn the low-dimensional representation to expand the manifold where the data lie. Essentially, these methods mainly focus on preserving microscopic structure of network, like pairwise relationship between vertexes. Nevertheless, scale-free property, one of the most fundamental macroscopic properties of networks, is largely ignored. \begin{figure}[t] \centering \subfigure[Original]{ \label{} \includegraphics{figs/citation.pdf} } \subfigure[LE]{ \label{} \includegraphics{figs/citation_recover_unmod} } \subfigure[DP-Walker]{ \label{} \includegraphics{figs/citation_recover_mod.pdf} } \caption{Scale-free property of real-world networks. (a) is the degree distribution of an academic network. (b) and (c) are respectively the degree distribution of of the network reconstructed based on vertex representations learned by Laplacian Eigenmap (LE) and our proposed method (DP-Walker). \normalsize } \label{} \end{figure} Scale-free property depicts that the vertex degrees follow a power-law distribution, which is a common knowledge for many real-world networks. We take an academic network as the example, where each edge indicates if a vertex (researcher) has cited at least one publication of another vertex (researcher). Figure~\ref{} demonstrates the degree distribution of this network. The linearity on log-log scale suggests a power-law distribution: the probability decreases when the vertex degree grows, with a long tail tending to zero~\cite{faloutsos1999power,adamic1999nature}. In other words, there are only a few vertexes of high degrees. The majority of vertexes connected to a high-degree vertex is, however, of low degree, and not likely connected to each other. Moreover, compared with the microscopic structure, the macroscopic scale-free property imposes a higher level constraint on the vertex representations: only a few vertexes can be close to many others in the learned latent space. Incorporating scale-free property in network embedding can reflect and preserve the sparsity of real-world networks and, in turn, provide effective information to make the vertex representations more discriminative. In this paper, we study the problem of learning scale-free property preserving network embedding. As the representation of a network, vertex embedding vectors are expected to well reconstruct the network. Most existing algorithms learn network embedding in the Euclidean space. However, we find that most traditional network embedding algorithms will overestimate the number of higher degrees. Figure~\ref{} gives an example, where the representation is learned by Laplacian Eigenmap. We analyze and try to understand this theoretically, and study the feasibility of recovering power-law distributed vertex degree in the Euclidean space, by converting our problem to the Sphere Packing problem. Through our analysis we find that theoretically, moderately increasing the dimension of embedding vectors can help to preserve the scale-free property. See details in Section~\ref{}. Inspired by our theoretical analysis, we then propose the \textit{degree penalty} principle for designing scale-free property preserving network embedding algorithms in Section~\ref{}: punishing the proximity between vertexes with high degrees. We further introduce two implementations of our approach based on the spectral techniques~\cite{belkin2003laplacian} and the skip-gram model~\cite{mikolov2013efficient}. As Figure~\ref{} suggests, our approach can better preserve the scale-free property of networks. In particular, the Kolmogorov-Smirnov (aka. K-S statistic) distance between the obtained degree distribution and its theoretical power-law distribution is $0.09$, while the value for the degree distribution obtained by LE is $0.2$ (the smaller the better). To verify the effectiveness of our approach, we conduct experiments on both synthetic data and five real-world datasets in Section~\ref{}. The experimental results show that our approach is able to not only preserve the scale-free property of networks, but also outperform several state-of-the-art embedding algorithms in different network analysis tasks. We summarize our contribution of this paper as follows: \begin{itemize} \item We analyze the difficulty and feasibility of reconstructing a scale-free network based on learned vertex representations in the Euclidean space, by converting our problem to the Sphere Packing problem. \item We propose the \textit{degree penalty} principle and two implementations to preserve the scale-free property of networks and improve the effectiveness of vertex representations. \item We validate our proposed principle by conducting extensive experiments and find that our approach achieves a significant improvement on six datasets and three tasks compared to several state-of-the-art baselines. \end{itemize} \hide{ The rest of the paper is organized as follows. We give theoretical analysis in Section~\ref{}, based on which we propose and introduce our method in Section~\ref{}. In Section~\ref{}, we report the experimental results and then review the related work in Section~\ref{}. At last, Section~\ref{} concludes the paper. } \section{Theoretical Analysis} \label{} In this section, we try to study why most network embedding algorithms will overestimate higher degrees theoretically, and analyze if there exists a solution of scale-free property preserving network embedding in the Euclidean space. This section also provides intuitions of our approach in Section~\ref{}. \subsection{Preliminaries} \label{} \vpara{Notations.} We consider an undirected network $G=(V, E)$, where $V=\{v_1, \cdots, v_n\}$ is the vertex set containing $n$ vertexes and $E$ is the edge set. Each $e_{ij} \in E$ indicates an undirected edge between $v_i$ and $v_j$. We define the adjacency matrix of $G$ as $\mathbf{A}=[A_{ij}] \in \mathcal{R}^{n \times n}$, where $A_{ij} = 1$ if $e_{ij}\in E$ and $A_{ij} = 0$ otherwise. Let $\mathbf{D}$ be a diagonal matrix where $D_{ii}=\sum_{j}A_{ij}$ is the degree of $v_i$. \vpara{Network embedding.} In this paper, we focus on the representation learning for undirected networks. Given an undirected graph $G=(V, E)$, the problem of graph embedding aims to represent each vertex $v_i \in V$ into a low-dimensional space $R^k$, i.e., learning a function $\mathbf{f}:V \mapsto\mathbf{U}^{n\times k}$, where $\mathbf{U}$ is the embedding matrix, $k \ll n$ and network structures can be preserved in $\mathbf{U}$. \hide{ There are various ways to quantify to what extent network structures have been preserved in $\mathbf{U}$. For instance, some methods like DeepWalk~\cite{perozzi2014deepwalk} and node2vec~\cite{grover2016node2vec} take into account paths in $G$ by assuming vertexes appearing in the same path tend to be more similar. Besides, Laplacian Eigenmaps (LE)~\cite{belkin2003laplacian} aims to learn the low-dimensional representation to expand the manifold where vertexes lie. Few work study a more fundamental perspective of network structure: degree distribution, which we will introduce later. } \vpara{Network reconstruction.} As the representation of a network, the learned embedding matrix is expected to well reconstruct the network. In particular, one can reconstruct the network edges based on distances between vertexes in the latent space $\mathcal{R}^k$. For example, the probability of an edge existing between $v_i$ and $v_j$ can be defined as \beq{ \label{} p_{i,j} = \frac{1}{1+e^{\Vert \mathbf{u_i} - \mathbf{u_j} \Vert}} } \noindent where the Euclidean distance between embedding vectors $\mathbf{u_i}$ and $\mathbf{u_j}$ of the vertex $v_i$ and $v_j$, in respective, is considered. In practice, a threshold $\varepsilon \in [0, 1]$ is chosen and an edge $e_{ij}$ will be created if $p_{i, j} \geq \varepsilon$. We call the above method as $\varepsilon$-NN, which is geometrically informative and a common method used in many existing work \cite{shaw2009spe,belkin2003laplacian,alanis2016labne}. \begin{figure*}[t] \centering \includegraphics{figs/example.pdf} \caption{An illustration of an ego-network centered by a hub vertex, a potential embedding solution, which is equivalent to a sphere packing solution, and leads to a failed reconstruction, where higher degrees are overestimated. } \label{} \end{figure*} \subsection{Reconstructing Scale-free Networks} Given a network, we call it as a scale-free network, when its vertex degrees follow a power-law distribution. In other words, there are only a few vertexes of high degrees. The majority of vertexes connected to a high-degree vertex is, however, of low degree, and not likely connected to each other. Formally, the probability density function of vertex degree $D_{ii}$ has the following form: \beq{ \label{} p_{D_{ii}}(d) = C d^{-\alpha} , \alpha > 1, d>d_{\min{}} > 0 } \noindent where $\alpha$ is the exponent parameter and $C$ is the normalization term. In practice, the above power-law form only applies for vertexes with degrees greater than a certain minimum value $d_{\min}$~\cite{aaron2009}. However, it is difficult to reconstruct a scale-free network in the Euclidean space by $\varepsilon$-NN. As we see in Figure~\ref{}, the higher degrees will be overestimated. We aim to explain this theoretically. \vpara{Intuition.} While reconstructing the network by $\varepsilon$-NN, we select a certain $\varepsilon$, and for a vertex $v_i$ with embedding vector $\mathbf{u_i}$, we regard all points that fall in the closed ball of radius $\varepsilon$ centered at $\mathbf{u}_i$, denoted by $B(\mathbf{u_i}, \varepsilon)$, as ones having edges with $v_i$. When $v_i$ is a high-degree vertex, there will be many points in $B(\mathbf{u_i}, \varepsilon)$. We expect these points are far away from each other, keeping more vertexes with low degree and thus in turn keep the power-law distribution. However, intuitively, as these points are placed in the same closed ball $B(\mathbf{u_i}, \varepsilon)$, it will be more likely that their distances from each other are less than $\varepsilon$. As a result, there will be many edges created among those points, which violates the assumption of the scale-free property. Following the above idea, we introduce a theorem below, which is discussed in $\mathcal{R}^k$, and without loss of generality, we set $\varepsilon$ as $1$. \hide{ \subsubsection{Scale-free property and power-law distribution} In probability theory, we say a positive real-valued random variable $X$ follows power-law distribution if the probability density function (p.d.f.) $p_X$ has the following form: \begin{equation} p_X(t) = C t^{-\alpha} , \alpha > 1, x_{\min{}} > 0 \label{} \end{equation} $X$ cannot be defined on the whole range of $[0, \infty)$, because $\int_{0}^{\infty} t^{-\alpha}dt = \infty$. This does not damage the applicably of power-law, because in practice one often finds that power-law only applies for values greater than a certain minimum $x_{\min{}}$. \cite{aaron2009} The constant $C$ in (\ref{}) is provided by normalization: $$ \int_{x_{\min{}}}^\infty p_X(t)dt = 1 $$ We say graph (aka. network) is scale-free, is the degree distribution of the graph follows a power-law distribution. \subsubsection{Graph Embedding and Reconstruction} Let $G=(V,E)$ be a graph. The problem of graph embedding is to find a mapping $f:G\rightarrow \mathcal{R}^d$ for some positive integer $d << |V|$, which preserves proximity between nodes of the original graph. $\mathcal{R}^d$ is called the feature space. \subsubsection{Graph Reconstruction} \label{} In this paper we use the Euclidean distance as our measurement of similarities. Specifically, we use \begin{equation} p_{x,y} = \frac{1}{1+e^{\Vert x-y \Vert}} \end{equation} as the probability of the link between $x, y$. To construct the desired graph, we choose a threshold $\varepsilon$, and we assign link between $x, y$ if $p(x,y) \geq \varepsilon$. \subsection{Desired Result} We hope that the embedding implicitly encodes as much informations as possible about the original degree distributions. To this end, it is desirable to reconstruct a graph, whose degrees can be expressed as a linear or monotone function of the original degrees. The reconstructed degrees would then preserve the ranks of the original degrees. } \hide{ \subsection{Graph Reconstruction of Scale-free Property} \subsubsection{Problem Formation} We discuss the ability of graph reconstruction using $\varepsilon NN$ to recover power-law. In a power-law graph, there are only few vertexes of high degrees. The majority of vertexes connected to a high-degree vertex is, however, of low degree, and not likely connected to each other. While reconstruction, we select a certain $\varepsilon$, and we regard all points that fall in the closed ball of radius $\varepsilon$ centred at $x$, denoted by $B(x, \varepsilon)$, as having an edge with the embedded representation $x$. However, intuitively, there can't be as many points in $B(x, \varepsilon)$ as we desire, whose distances from each other are larger than $\varepsilon$. We formalize this idea in the following discussion. We shall discuss in $\mathcal{R}^n$. Without loss of generality, we shall set $\varepsilon$ as $1$. We start by giving a simple theorem. } \begin{theorem}[Sphere Packing] \label{} There are $m$ points in $B(0, 1)$ whose distances from each other are larger than or equal to $1$, if and only if, there exists $m$ disjoint spheres of radius $1/2$ in $B(0,3/2)$. \label{sp_thm1} \end{theorem} \begin{proof} The center of any sphere of radius $1/2$ in $B(0, 3/2)$ falls in $B(0, 1)$ and the distance between any two centers of two different spheres of radius $1/2$ is larger than or equal to 1. \end{proof} \begin{remark} Theorem~\ref{} converts our problem of reconstructing a scale-free network to the \textit{Sphere Packing Problem}, which seeks to find the optimal packing of spheres in high dimensional spaces~\cite{cohn2003new,vance2011sphere,venkatesh2012sphere}. Figure~\ref{} gives an example, where a network centered by a high-degree vertex is embedded into a two-dimensional space. The embedding result corresponds to an equivalent Sphere Packing solution, which fails to place enough disjoint spheres and leads to a failed network reconstruction (many nonexistent edges are created). Formally, we define the packing density as follows: \end{remark} \begin{definition}[Packing Density] The packing density is the fraction of the space filled by the spheres making up the packing. For convenience, we define the optimal sphere packing density in $\mathcal{R}^k$ as $\Delta_k$, which means no packing of spheres in $\mathcal{R}^k$ achieves a packing density larger than $\Delta_k$. \end{definition} As Theorem~\ref{} suggests, we aim to find a packing solution with large density so that more points in a closed sphere can keep their distance larger than 1. However, in general cases, finding the optimal packing density remains an open problem. Still, we are able to derive the upper and lower bounds of $\Delta_k$ for sufficiently large $k$. \begin{theorem}[Upper and lower bounds for $\Delta_k$] $$ -1\leq \lim_{k\rightarrow\infty} \frac{1}{k}\log_2\Delta_k\leq-0.599 $$ \normalsize Specifically, we have \begin{equation} \Delta_k \geq 2^{-k}, k \geq 1 \label{lower_bound} \end{equation} And the following inequality holds for sufficiently large $k$: \begin{equation} \Delta_k \leq 2^{-0.599k} \label{upper_bound} \end{equation} \end{theorem} \noindent Eq.~\ref{upper_bound} is one of the best upper bound when $k \geq 115$~\cite{cohn2014sphere}. The proof of the above theorem can be found in the work of~\citeauthor{kabatiansky1978bounds}. \begin{theorem} Suppose $x$ is in $\mathcal{R}^k$, $\varepsilon > 0$, and there can be at most $M_k$ points in $B(x, \varepsilon)$ whose distances from each other are larger than $\varepsilon$, then \begin{eqnarray} \left\lfloor\left(\frac{3}{2}\right)^k\right\rfloor \leq M_k \leq \left\lfloor 3^k 2^{-0.599k}\right\rfloor \label{bound_M} \end{eqnarray} where $\left\lfloor\cdot\right\rfloor$ means taking the integer part. The upper bound holds for sufficiently large \end{theorem} \begin{proof} By Theorem~\ref{sp_thm1} we only need to estimate the number of disjoint spheres of radius $\frac{1}{2}\varepsilon$ that can be fitted in $B(x, \frac{3}{2}\varepsilon)$. The volume of a $k$-dimensional ball of radius $R$ is given by $\frac{\pi^{k/2}}{\Gamma(k/2+1)}R^k$. Plugging in the radius, the volumes of $B(x, \frac{3}{2}\varepsilon)$ and a sphere of radius $\frac{1}{2}\varepsilon$ are respectively $V_1 = \frac{\pi^{k/2}}{\Gamma(k/2+1)}(\frac{3}{2}\varepsilon)^k$ and $V_2 = \frac{\pi^{k/2}}{\Gamma(k/2+1)}(\frac{1}{2}\varepsilon)^k$. Since the optimal packing density is given by $\Delta_k$, we have \begin{eqnarray*} M_k = \left\lfloor \Delta_k \frac{V_1}{V_2} \right\rfloor = \left\lfloor 3^k \Delta_k\right\rfloor \end{eqnarray*} \noindent Combing with Eq.~\ref{lower_bound} and Eq.~\ref{upper_bound}, we obtain the inequality as desired. \end{proof} \vpara{Discussion.} The lower bound of $M_k$ in Eq.~\ref{bound_M} suggests the feasibility to reconstruct scale-free network by $\varepsilon$-NN in the Euclidean space, when $k$, the dimension of embedding vector, is sufficiently large. For instance, when $k=100$, we have $M_k > 4.06\times 10^{17}$, which is enough to keep scale-free property holds for most real-world networks. \section{Our Approach} \label{} \vpara{General idea.} Inspired by our theoretical analysis, we propose a principle of \textit{degree penalty} for designing scale-free property preserving embedding algorithms: \textit{while preserving first- and second-order proximity, the proximity between vertexes that have high degrees shall be punished}. We give the general idea behind this principle below. Scale-free property is featured by the ubiquitous existence of ``big hubs'' which attract most edges in the network. Most of existing network embedding algorithms, explicitly or implicitly, attempt to keep these big hubs close to their neighbors by preserving first-order and/or second-order proximity~\cite{belkin2003laplacian,tang2015line,perozzi2014deepwalk}. However, connecting to big hubs does not imply proximity as strong as connecting to vertexes with mediocre degrees. Taking social network as an example, where a famous celebrity may receive a lot of followers. However, the celebrity may not be familiar or similar to her followers. Besides, two followers of the same celebrity may not know each other at all and can be totally dissimilar. As a comparison, a mediocre user is more likely to known and to be similar to her followers. From another perspective, a high-degree vertex $v_i$ is more likely to hurt the scale-free property, as placing more disjoint spheres in a closed ball is more difficult (Section~\ref{}). The \textit{degree penalty} principle can be implemented by various methods. In this paper, we introduce two proposed models based on our principle, implemented by spectral techniques and skip-gram models respectively. \subsection{Model I: DP-Spectral} Our first model, Degree Penalty based Spectral Embedding (DP-Spectral), mainly utilizes graph spectral techniques. Given a network $G=(V, E)$ and its adjacency matrix $\mathbf{A}$, we define a matrix $\mathbf{C}$ to indicate the common neighbors of any two vertexes in $G$: \beq{ \label{} \mathbf{C} \triangleq \mathbf{A}^T\mathbf{A} - \diag{(\mathbf{A}^T\mathbf{A})} } \noindent where $C_{ij}$ is the number of common neighbors of $v_i$ and $v_j$. $\mathbf{C}$ can be also regarded as a measurement of second-order proximity, and can be easily extended to further consider first-order proximity by \beq{ \label{} \mathbf{C}' \triangleq \mathbf{C} + \mathbf{A} } \noindent As we aim to model the influence of vertex degrees in our model, we further extend $\mathbf{C}'$ to consider degree penalty as \beq{ \label{} \mathbf{W} \triangleq (\mathbf{D}^{-\beta})^T \mathbf{C}' \mathbf{D}^{-\beta} } \noindent where $\beta \in \mathcal{R}$ is the model parameter used to indicate the strength of degree penalty, and $\mathbf{D}$ is a diagonal matrix where $D_{ii}$ is the degree of $v_i$. Thus $\mathbf{W}$ is proportional to $\mathbf{C}'$ and is inversely proportional to vertex degrees. \vpara{Objective.} Our goal is to learn vertex representations $\mathbf{U}$, where $\mathbf{u_i} \in \mathcal{R}^k$ is the $i$th row of $\mathbf{U}$ and represents the embedding vector of vertex $v_i$, and minimize \beq{ \label{} \sum_{i,j} \Vert\mathbf{u}_i - \mathbf{u}_j\Vert^2 W_{ij} } \noindent under the constraint \beq{ \label{} \mathbf{U}^T\mathbf{DU}=\mathbf{I} } \noindent Eq.~\ref{} is provided such that the embedding vectors will not collapse onto a subspace of dimension less than $k$~\cite{belkin2003laplacian}. \vpara{Optimization.} In order to minimize Eq.~\ref{}, we utilize graph Laplacian, which is an analogy to the Laplacian operator in multivariate calculus. Formally, we define graph Laplacian, $\mathbf{L}$, as follows: \beq{ \label{} \mathbf{L\triangleq D - W} } \hide{ The power of graph laplacian lies in the fact that for for any real-valued vector $\mathbf{x}\in\mathcal{R}^n$, we have \begin{equation} \mathbf{x}^TL\mathbf{x} = \sum_{\{i,j\}\in E} (x_i-x_j)^2 W_{i,j}\label{} \end{equation} Graph Laplacian provides a sense of "smoothness" of a function defined on vertices of the graph, the values of which are the components of $\mathbf{x}$, with respect to the connectivity of $G$.~\cite{kolaczyk2009statmodelsofnetwork} With Eq. (\ref{}), the optimal of the objective (\ref{}) under the constraint (\ref{}) is given by the $k+1$ eigenvectors of the lowest eigenvalues, discarding the one corresponding to $0$. } \noindent Observe that \beq{ \label{} \sum_{i,j} W_{ij}\Vert \mathbf{u}_i - \mathbf{u}_j\Vert^2 = \trace{(\mathbf{U^TLU})} } \noindent The desired $\mathbf{U}$ minimizing the objective function (\ref{}) is obtained by putting \begin{equation} \mathbf{U} = [\mathbf{t}_1, \ldots, \mathbf{t}_k] \end{equation} \noindent where $\mathbf{t}_i$ is an eigenvector of $\mathbf{L}$. In practice, we use normalized form of $\mathbf{W}$ and $\mathbf{L}$, (i.e., $\mathbf{D}^{-1/2}\mathbf{WD}^{1/2}$ and $\mathbf{I-D}^{-1/2}\mathbf{WD}^{1/2}$). \hide{ provided by \begin{eqnarray} \mathbf{W_0} & = & \mathbf{D}^{-1/2}\mathbf{WD}^{1/2} \\ \mathbf{L}_0 & = & \mathbf{D-W}_0 = \mathbf{I-D}^{-1/2}\mathbf{WD}^{1/2} \end{eqnarray} } \subsection{Model II: DP-Walker} Our second method, Degree Penalty based Random Walk (DP-Walker), utilizes a skip-gram model, word2vec. We start with a brief description of the word2vec model and its applications on network embedding tasks. Given a text corpus, Mikolov et al. proposed word2vec to learn the representation of each word by facilitating the prediction of its context. Inspired by it, some network embedding algorithms like DeepWalk~\cite{perozzi2014deepwalk} and node2vec~\cite{grover2016node2vec} define a vertex's ``context'' by co-occurred vertexes in random walk generated paths. Specifically, the Random Walk rooted at vertex $v_i$ is a stochastic process ${\mathcal{V}_{v_i}^k, k = 1,\ldots, m}$, where $\mathcal{V}_{v_i}^{k+1}$ is a vertex sampled from the neighbors of $\mathcal{V}_{v_i}^k$ and m is the path length. Traditional methods regard $P(\mathcal{V}_{v_i}^{k+1}|\mathcal{V}_{v_i}^{k})$ as a uniform distribution where each neighbor of $\mathcal{V}_{v_i}^{k}$ has the equal chance to be sampled. However, as our proposed Degree Penalty principle suggests, a neighbor $v_i$ of a vertex $v_j$ with high degree may not be similar to $v_j$. In other words, $v_j$ shall have less chance to be sampled as one of $v_i$'s context. Thus, we define the probability of the random walk jumping from $v_i$ to $v_j$ as \beq{ \label{} \rm{Pr}(v_j|v_i) \propto \frac{C'_{ij}}{(D_{ii} D_{jj})^\beta} } \noindent where $\mathbf{C}'$ can be found in Eq.~\ref{} and $\beta$ is the model parameter. According to Eq.~\ref{}, we find that $v_j$ will have a greater chance to be sampled when it has more common neighbors with $v_i$ and has a lower degree. After obtaining random walk generated paths, we enable skip-gram to learn effective vertex representations for $G$ by predicting each vertex's context. This results in an optimization problem \beq{ \label{} \argmin_{\mathbf{U}} -\log \rm{Pr}\left( \{ v_{i-w},\ldots, v_{i+w} \} \ v_i | \mathbf{u_i} \right) } \noindent where $2\times w$ is the path length we consider. Specifically, for each random walk $\mathcal{V}_{v_i}$, we feed it to skip-gram algorithm~\cite{mikolov2013efficient} to update the vertex representations. For implementation, we use Hierarchical Softmax~\cite{Morin05hierarchicalprobabilistic,Mnih2009hier} to estimate the concerned probability distribution. \hide{ \paragraph{Sampling Probability} Deepwalk~\cite{perozzi2014deepwalk} does not take a weighted graph as an input. Instead, our preference can be manifested during the sampling of the walking. In Deepwalk, every next destination of the walk is chosen with a uniform distribution among the neighbors. We modify the probability of walking from $x$ to $y$, on condition that the walking does not stop at $x$: \begin{equation} p(y|x) \propto \frac{C_{ij}}{(d_x d_y)^\beta} \label{} \end{equation} where $d_i$ is the degree of node $i$, and $C_{ij}$ is the corresponding element of the matrix defined by Eqn. (\ref{}). } \hide{ \subsection{Justification of our Modification} A consequence of the ubiquitous existence of big hubs in scale-free networks is the ultra-small world property \cite{cohen2003ultrasmall}. Average distance between nodes in a scale-free network is much smaller than observed in a random graph generated with the same nodes. Our modifications are specifically based on the characteristics of social networks. Eqn. (\ref{}) is in principle consistence with Eqns. (\ref{mod:1}) and (\ref{mod:2}). \subsection{Goal 1: Integrating the Influence of Degrees} However, in social network, connecting to big hubs does not imply proximity as strong as connecting to nodes with mediocre degrees. An informal example is, while a famous celebrity may receive a lot of followers, it does not imply that these followers are close to the celebrity; at the same time, two Americans, who both follow the president of the United States on twitter, could be totally different from each other. Most graph embedding algorithms, explicitly or implicitly, attempts to preserve first-order and/or second-order proximity, without given consideration to the influences of degrees on proximity. This would result in overestimation of the proximity between most low-degree nodes. Eqn. (\ref{mod:2}) tries to integrate this consideration into the training process. \subsection{Goal 2: Integrating Second-order Proximity} Still, it is not fair to part every pair of nodes whose degrees are high. When two nodes have high second-proximity, i.e. they have many common neighbors, it is reasonable that their representations are also close to each other. Eqn. (\ref{mod:1}) makes sure that connected nodes whose neighbors are similar gets higher weight. } \hide{ \para{Summary.} In this section, we introduce our proposed Degree Penalty principle and two implementations by leveraging spectral techniques and a skip-gram model respectively. By similar methods, the proposed principle can also be implemented in other network embedding frameworks. We leave it to our future work. } \hide{ \reminder{Description of algorithms. } \subsection{Algorithms for Network Embedding} In this section we introduce two network embedding algorithms: Deepwalk \cite{perozzi2014deepwalk}, and Laplacian Eigenmap \cite{belkin2003laplacian}. \paragraph{Deepwalk} Deepwalk \cite{perozzi2014deepwalk} uses random walk to define analogies of sentences in a network, which is eligible as input for Word2Vec \cite{mikolov2013word2vec,mikolov2013word2vec2}, a neural word embedding algorithm to learn vector representation of words. Specifically, random walk in Deepwalk generated by sampling from uniform distribution. Specifically, a random walk rooted at vertex $v_i$ is a stochastic process ${\mathcal{W}_{v_i}^k, k = 1,\ldots, m}$ where $\mathcal{W}_{v_i}^{k+1}$ is sampled uniformly from the neighbors of $\mathcal{W}_{v_i}^k$. By applying random walk, an analogy to language model is established on networks, where vertices are treated as words, and sequences generated by random walks are viewed as sentences. Feeding these parameters to Word2Vec, Deepwalk gains the vector representation, aka. the embedding of the original graph. The use of random walk is motivated by the following two facts: \begin{enumerate} \item random walks captures the local structure of a network, and have been used in a variety of tasks including content recommendation, community detection, link recommendation, node ranking, etc. \cite{agarwal2007noderank,Pons2005,zhang2013itemrecom,yin2010linkrecom} \item The degree of vertices in a social network follows power-law distribution, which applies also to word frequencies. \end{enumerate} } \hide{ \paragraph{Laplacian Eigenmap} Laplacian Eigenmap is a spectral based techniques for network embedding and non-linear dimensionality reduction \cite{belkin2003laplacian}. Given a graph $G=(V, E)$ and its adjacency matrix $A$, Laplacian Eigenmap chooses an embedding of the network $\mathbf{y}_1, \ldots, \mathbf{y}_n \in \mathcal{R}^m $, to minimize \begin{equation} \sum_{i,j} \Vert\mathbf{y}_i - \mathbf{y}_j\Vert^2 A_{ij} \label{} \end{equation} under the constraint \begin{equation} Y^T DY = I \label{} \end{equation} where $Y = [\mathbf{y}_1^T, \ldots, \mathbf{y}_n^T] $ is a matrix whose $i$th row is $\mathbf{y}_i^T$, the vector representation of $v_i$. The constraint (\ref{}) is provided such that the embedding does not collapse onto a subspace of dimension less than $m$. \cite{belkin2003laplacian} } \section{Experiments} \label{} \subsection{Experiment Setup} \label{} \vpara{Datasets.} We use four datasets, whose statistics are summarized in Table~\ref{} for the evaluations. \begin{itemize} \item Synthetic: We generate a synthetic dataset by the Preferential Attachment model~\cite{vazquez2003growing}, which describes the generation of scale-free networks. \item Facebook~\cite{leskovec2012learning}: This dataset is a subnetwork of Facebook\footnote{}, where vertexes indicate users of Facebook, and edges denote friendship between users. \item Twitter~\cite{leskovec2012learning}: This dataset is a subnetwork of Twitter\footnote{}, where vertexes indicate users of Twitter, and edges denote following relationships \item Coauthor~\cite{leskovec2007graph}: This network covers scientific collaborations between authors. Vertexes are authors. Vertexes are authors. An undirected edge exists between two authors if they have coauthored at least one paper. \item Citation~\cite{tang2008arnetminer}: Similar to Coauthor, this is also an academic network, where vertexes are authors. Edges indicate citations instead of coauthor relationship. \item Mobile: This is a mobile network provided by PPDai\footnote{The largest unsecured micro-credit loan platform in China.}. Vertexes are PPDai registered users. An edge between two users indicates that one of the users has called the other. Overall, it consists of over one million calling logs. \end{itemize} \begin{table}[t] \centering \resizebox{0.9\linewidth}{!}{ \begin{tabular}{ccccccc} \toprule & Synthetic & Facebook & Twitter & Coauthor & Citation & Mobile\\ \midrule $|V|$ & \small 10000 & \small 4039 & \small 81306 & \small 5242 & \small 48521 & \small 198959 \\ \midrule $|E|$ & \small 399580 & \small 88234 & \small 1768149 & \small 28980 & \small 357235 & \small 1151003 \\ \bottomrule \end{tabular} \normalsize } \caption{Statistics of datasets. \small $|V|$ indicates the number of vertexes and $|E|$ denotes the number of edges. \normalsize } \label{} \end{table} \vpara{Baseline methods.} We compare the following four network embedding algorithms in our experiments: \begin{itemize} \item Laplacian Eigenmap (LE)~\cite{belkin2003laplacian}: This represents spectral-based network embedding algorithms. It aims to learn the low-dimensional representation to expand the manifold where the data lie. \item DeepWalk~\cite{perozzi2014deepwalk}: This represents skip-gram model based network embedding algorithms. It first generates random walks on the network, and defines the context of a vertex by its co-occurred vertexes in paths. Then, it learns vertex representations by predicting each vertex's context. Specifically, we perform $10$ random walks starting from each vertex, and each random walk will have a length of $40$. \item DP-Spectral: This is a spectral technique based implementation of our degree penalty principle. \item DP-Walker: This is another implementation of our approach. It is based on a skip-gram model. \end{itemize} Unless otherwise specified, the embedding dimension for our experiments is $200$. \vpara{Tasks.} We first utilize different algorithms to learn vertex representations for a certain dataset, then apply the learned embeddings in three different tasks: \begin{itemize} \item Network reconstruction: this task aims to validate if an algorithm is able to preserve the scale-free property of networks. We evaluate the performance of different algorithms by the correlation coefficients between the reconstructed degrees and the degrees in the given network. \item Link prediction: given two vertexes $v_i$ and $v_j$, we feed their embedding vectors, $\mathbf{u_i}$ and $\mathbf{u_j}$, to a linear regression classifier and determine whether there exists an edge between $v_i$ and $v_j$. Specifically, we use $\mathbf{u}_i-\mathbf{u}_j$ as the feature, and randomly select about $1\%$ pairs of vertexes for training and evaluation. \item Vertex classification: on this task, we consider the vertex labels. For instance, in Citation, each vertex has a label to indicate the researcher's research area. Specifically, given a vertex $v_i$, we define its feature vector as $\mathbf{u_i}$, and train a linear regression classifier to determine its label. \end{itemize} \subsection{Network Reconstruction} \label{} \hide{ \vpara{Evaluation Metrics.} In the scale-free property reconstruction task, we consider three different correlation coefficients of the reconstructed degrees and original degrees: Pearson~\reminder{reference}, Spearman~\reminder{reference}, and Kendall~\reminder{reference}. Let $(d_1, \ldots, d_n)$ to indicate the degrees of the given network and $(x_1, \ldots, x_n)$ as our reconstructed degrees. The formula of Pearson correlation coefficient is: \beq{ \label{} \mathtt{Pearson} = \frac{\sum_{i=1}^n (x_i-{\bar{x}})(d_i-\bar{d})}{\sqrt{\sum_{i=1}^n (x_i - \bar{x})^2} \sqrt{\sum_{i=1}^n (d_i - \bar{d})^2}} } \noindent where $\bar{x}$ indicates the average value of $(x_1, \ldots, x_n)$. Spearman correlation coefficient is calculated as follows: \beq{ \label{} \mathtt{Spearman} = 1 - \frac{6 \sum z_i^2}{n(n^2-1)} } \noindent where $z_i$ is the difference between the ranks of corresponding values $d_i$, $x_i$. At last, the formula of Kendall correlation coefficient is: \beq{ \label{} \mathtt{Kendall} = \frac{n_c - n_d}{\frac{1}{2}n(n-1)} } \noindent where $n_c$ is the number of concordant pairs, and $n_d$ the number of discordant pairs. Concordant means for $i,j, i \neq j$, the sort order by $d$ and by $x$ agrees; i.e. if $x_i > x_j$, then $d_i > d_j$, vice versa. Discordant means the opposite. If $x_i=x_j$ or $d_i=d_j$, the pair is counted as neither concordant or discordant. } \hide{ \paragraph{Recognition of Scale-free property.} It is necessary to confirm that the experimented networks follow a power-law distribution. We use powerlaw, a Python package for fitting and analysis of power-law distribution \cite{alstott2014powerlaw}. \paragraph{Experiment 1: Correlation Coefficients.} In this experiment, we proceed with the following steps: \begin{enumerate} \item Apply an graph embedding algorithm to a certain dataset. \item Reconstruct a graph with the method mentioned in Section \ref{}, trying different choices of $\varepsilon$. \end{enumerate} We measure the goodness of the embedding by the correlation coefficients of the reconstructed degrees and the original degrees. Specifically, we roll over all values of $\varepsilon$ ranging from $0.01$ to $1$ with step $0.01$, and choose the value which maximizes the Pearson's correlation coefficient between the original and recovered degrees. \paragraph{Experiment 2: Link Prediction.} We feed the representation vector to a linear regression classifier to determine whether there exists a link between any two vertexes. The experiment is organized as follows. For each pair of nodes, we use the difference of the representation vector, $\mathbf{u}_i-\mathbf{u}_j$ as the feature. If $\mathbf{u}_i$ is connected to $\mathbf{u}_j$, we call it a positive instance, otherwise a negative instance. We sample randomly $2000$ nodes as training set, and $1000$ nodes as test set, where positive and negative instances are sampled with ratio $1:1$. } \begin{table}[t] \centering \resizebox{\linewidth}{!}{ \begin{tabular}{|l|l|c|c|c|} \hline Dataset & Method ($\varepsilon$) & Pearson & Spearman & Kendall \\ \hline \multirow{4}*{Synthetic} & LE (0.55) & 0.14 & 0.054 & 0.039 \\ \cline{2-5} & DeepWalk (0.91) & 0.47 & -0.22 & -0.18 \\ \cline{2-5} & DP-Spectral (0.52) & 0.92 & \textbf{0.79} & \textbf{0.63} \\ \cline{2-5} & DP-Walker (0.95) & \textbf{0.94} & 0.63 & 0.52 \\ \hline \multirow{4}*{Facebook} & LE (0.52) & 0.48 & 0.18 & 0.12 \\ \cline{2-5} & DeepWalk (0.81) & 0.73 & 0.65 & 0.49 \\ \cline{2-5} & DP-Spectral (0.52) & \textbf{0.87} & \textbf{0.67} & 0.51 \\ \cline{2-5} & DP-Walker (0.84) & 0.75 & 0.73 & \textbf{0.57} \\ \hline \multirow{4}*{Twitter} & LE (0.81) & 0.17 & 0.19 & 0.17 \\ \cline{2-5} & DeepWalk (0.51) & 0.08 & 0.21 & 0.26 \\ \cline{2-5} & DP-Spectral (0.93) & \textbf{0.50} & \textbf{0.34} & \textbf{0.27} \\ \cline{2-5} & DP-Walker (0.087) & 0.40 & 0.33 & \textbf{0.27} \\ \hline \multirow{4}*{Coauthor} & LE (0.50) & 0.32 & 0.04 & 0.03 \\ \cline{2-5} & DeepWalk (0.92) & 0.66 & 0.31 & 0.25 \\ \cline{2-5} & DP-Spectral (0.51) & 0.64 & \textbf{0.69} & \textbf{0.55} \\ \cline{2-5} & DP-Walk (0.93) & \textbf{0.75} & 0.44 & 0.35 \\ \hline \multirow{4}*{Citation} & LE (0.99) & 0.11 & -0.27 & -0.20 \\ \cline{2-5} & DeepWalk (0.97) & 0.51 & 0.28 & 0.20 \\ \cline{2-5} & DP-Spectral (0.50) & 0.45 & \textbf{0.72} & \textbf{0.54} \\ \cline{2-5} & DP-Walker (0.98) & \textbf{0.62} & 0.65 & 0.50 \\ \hline \multirow{4}*{Mobile} & LE (0.51) & 0.10 & 0.05 & 0.04 \\ \cline{2-5} & DeepWalk (0.71) & 0.77 & 0.22 & 0.20 \\ \cline{2-5} & DP-Spectral (0.50) & 0.40 & \textbf{0.68} & \textbf{0.60} \\ \cline{2-5} & DP-Walker (0.78) & \textbf{0.93} & 0.22 & 0.20 \\ \hline \end{tabular} } \caption{Performance of different methods on scale-free property reconstruction. For each method, $\varepsilon$ (indicated after the corresponding method) is chosen so that Pearson is maximized. \normalsize } \label{} \end{table} \begin{figure*}[htbp] \centering \subfigure[Synthetic.]{ \label{} \includegraphics{figs/dimension_syn.pdf} } \subfigure[Facebook.]{ \label{} \includegraphics{figs/dimension_fb.pdf} } \subfigure[Synthetic.]{ \label{} \includegraphics{figs/beta_syn.pdf} } \subfigure[Facebook.]{ \label{} \includegraphics{figs/beta_fb.pdf} } \caption{Model parameter analysis. (a) and (b) demonstrate the sensitivity of the embedding dimension $k$ in Synthetic and Facebook dataset respectively. (c) and (d) present the sensitivity of the degree penalty parameter $\beta$. We omit the results on other datasets due to space limitation. } \label{} \end{figure*} \vpara{Comparison results.} In this task, after obtaining vertex representations, we use the $\varepsilon$-NN algorithm introduced in Section~\ref{} to reconstruct the network. We then evaluate the performance of different methods on reconstructing power-law distributed vertex degrees by considering three different correlation coefficients of the reconstructed degrees and original degrees: Pearson's $r$~\cite{statistics}, Spearman's $\rho$~\cite{spearman}, and Kendall's $\tau$~\cite{kendalltau}. All of these statistics are used to evaluate some relationship between paired samples. Pearson's $r$ is used to detect linear relationships, while Spearman's $\rho$ and Kendall's $\tau$ are capable of finding monotonic relationships. To select $\varepsilon$, for each algorithm, we roll over all values of $\varepsilon$ ranging from $0.01$ to $1$ with step $0.01$, and choose the value which maximizes the Pearson's correlation coefficient between the original and recovered degrees. Our selection of Pearson's $r$ as evaluation metric is because of the scale-free property of degree distribution. From Table~\ref{}, we see that our algorithms outperform the baselines significantly. Especially, Pearson's $r$ of DP-Walker achieves at least $44.5\%$ improvement, and Spearman's $\rho$ of DP-Spectral achieves at least $84.3\%$ improvement. The good fitness of the vertex degree reconstructed by our algorithms suggests that we can better preserve the scale-free property of the network. Besides, we see that the best $\varepsilon$ to maximize Pearson's $r$ for DP-Spectral is more stable (i.e., around $0.51$) than other methods. Thus one can tune DP-Spectral's parameters more easily in practice. \vpara{Preserving Scale-free Property.} After reconstructing the network, for each embedding algorithm, we validate its effectiveness to preserve the scale-free property, by fitting the reconstructed degree distribution (e.g., Figure~\ref{} and (c)) to a theoretical power-law distribution~\cite{alstott2014powerlaw}. We then calculate the Kolmogorov-Smirnov distance between these two distributions and find that the proposed methods can always obtain better results compared to baselines (0.115 vs 0.225 averagely). \vpara{Parameter analysis.} We further study the sensitivity of model parameters: embedding dimension and degree penalty weight $\beta$. We only present the results on Synthetic and Facebook and omit others due to space limitation. Figure~\ref{} and \ref{} shows the Pearson correlation coefficients resulted by our algorithm with different embedding dimensions. When the embedding dimension grows, the performance increases significantly. The figures suggest that when embedding a network into Euclidean space, there is a dimension which is sufficient for preserving scale-free property, and further increase of the embedding dimension has limited effect. Correlation does not change drastically for DP-Walker as the dimension increases, as the figure suggests. The figures also show that DP-Walker is able to obtain fairly high Pearson correlation even in a lower dimension and it requires an embedding dimension lower than DP-Spectral does for a satisfactory performance. We also study how $\beta$ influences the performance. Figure~\ref{} and \ref{} shows that DP-Spectral is more sensitive to the choice of $\beta$. This is largely due to the fact that in DP-Spectral, the influence of $\beta$ is manifested in the objective function (Eq.~\ref{}), which imposes a stronger restriction than its counterpart in DP-Walker, which is embodied in the sampling process. Figure~\ref{} and \ref{} shows that the optimal choice of $\beta$ varies from graph to graph. It makes sense, since $\beta$ can be viewed as a punishment on the degrees, and the influence of $\beta$ is therefore supposedly related to the topology of the original network. \begin{table}[t] \centering \resizebox{\linewidth}{!}{ \begin{tabular}{|c|c|c|c|c|} \hline \multirow{1}*{Dataset} & \multirow{1}*{Method} & \multirow{1}*{Precision} & \multirow{1}*{\ \ Recall \ \ } & \multirow{1}*{\ \ \ \ F1 \ \ \ \ } \\ \hline \multirow{4}*{Synthetic} & LE & 0.52 & 0.53 & 0.53 \\ \cline{2-5} & Deepwalk & 0.51 & 0.51 & 0.51 \\ \cline{2-5} & DP-Spectral & \textbf{0.64} & \textbf{0.68} & \textbf{0.66} \\ \cline{2-5} & DP-Walker & 0.61 & 0.63 & 0.62 \\ \hline \multirow{4}*{\ Facebook \ } & LE & 0.75 & 0.92 & 0.83 \\ \cline{2-5} & Deepwalk & \textbf{0.84} & 0.97 & \textbf{0.90} \\ \cline{2-5} & DP-Spectral & 0.76 & \textbf{0.98} & 0.85 \\ \cline{2-5} & DP-Walker & 0.82& 0.95 & 0.89 \\ \hline \multirow{4}*{Twitter} & LE & 0.58 & 0.35 & 0.43 \\ \cline{2-5} & Deepwalk & \textbf{0.65} & 0.77 & 0.70 \\ \cline{2-5} & DP-Spectral & 0.59 & \textbf{0.98} & \textbf{0.73} \\ \cline{2-5} & DP-Walker & 0.54 & 0.58 & 0.56 \\ \hline \multirow{4}*{Coauthor} & LE & 0.61 & 0.83 & 0.70 \\ \cline{2-5} & Deepwalk & 0.55 & 0.58 & 0.56 \\ \cline{2-5} &\ \ DP-Spectral \ \ & \textbf{0.62} & \textbf{0.89} & \textbf{0.73} \\ \cline{2-5} & DP-Walker & 0.56 & 0.66 & 0.61 \\ \hline \multirow{4}*{Citation} & LE & 0.54 & 0.56 & 0.55 \\ \cline{2-5} & Deepwalk & 0.54 & 0.56 & 0.55 \\ \cline{2-5} & DP-Spectral & 0.52 &\textbf{0.99} & \textbf{0.68} \\ \cline{2-5} & DP-Walker & \textbf{0.55} & 0.57 & 0.56 \\ \hline \multirow{4}*{Mobile} & LE & \textbf{0.75} & 0.36 & 0.48 \\ \cline{2-5} & Deepwalk & 0.55 & 0.60 & 0.57 \\ \cline{2-5} & DP-Spectral & 0.63 & \textbf{0.89} & \textbf{0.74} \\ \cline{2-5} & DP-Walker & 0.54 & 0.58 & 0.56 \\ \hline \end{tabular} } \caption{Experimental results of link prediction. } \label{} \end{table} \subsection{Link Prediction} \label{} In this task, we consider the following evaluation metrics: Precision, Recall, and F1-score. Table~\ref{} demonstrates the performance of different methods on the link prediction task. For our methods, we use the model parameter $\beta$ as optimized in Table \ref{}. We see that, in most cases, DP-Spectral obtains the best performance, which suggests that with the help of the proposed principle, we can not only preserve the scale-free property of networks, but also improve the effectiveness of the embedding vectors. \subsection{Vertex Classification} \label{} Table~\ref{} lists the accuracy of vertex classification task on Citation. Our task is to determine an author's research area, which is a multi-classification problem. We define features as vertex representation obtained by the four different embedding algorithms. Generally, from the table, we see that DP-Walker and DP-Spectral beat respectively Deepwalk and Laplacian Eigenmap. In particular, DP-Spectral achieves the best result for 5 out of 7 labels. Besides, we can also observe its stability of the performance. DP-Spectral's result on all labels is more stable than other methods. In comparison, LE achieves a satisfactory result for two labels, but for others the result can be poor. Specifically, the standard deviation of DP-Spectral is 0.04, while the value is 0.26 for LE and 0.1 for DeepWalk. \section{Related Work} \label{} \begin{table}[t] \resizebox{\linewidth}{!}{ \begin{tabular}{|c|c|c|c|c|c|c|c|} \hline Method & Acrh & CN & CS & DM & THM & GRA & UNK \\ \hline LE & 0.36 & \textbf{0.75} & 0.14 & 0.37 & 0.46 & 0.13 & \textbf{0.86} \\ \hline Deepwalk & 0.54 & 0.54 & 0.52 & 0.56 & 0.56 & 0.56 & 0.85 \\ \hline DP-Walker & 0.56 & 0.57 & 0.54 & 0.58 & 0.58 & 0.55 & 0.85 \\ \hline DP-Spectral & \textbf{0.71} & 0.74 & \textbf{0.78} & \textbf{0.76} & \textbf{0.74} & \textbf{0.75} & 0.85 \\ \hline \end{tabular} } \caption{Accuracy of multi-classification task. The labels stand Architecture, Computer Network, Computer Science, Data Mining, Theory, Graphics, and Unknown, respectively. } \label{} \end{table} \paragraph{Network embedding.} Network embedding aims to learn representations for vertexes in a given network. Some researchers regard network embedding as part of dimensionality reduction techniques. For example, Laplacian Eigenmaps (LE)~\cite{belkin2003laplacian} aims to learn the vertex representation to expand the manifold where the data lie. As a variant of LE, Locality Preserving Projections (LPP)~\cite{he2005face} learns a linear projection from feature space to embedding space. Besides, there are other linear~\cite{jolliffe2002principal} and non-linear~\cite{tenenbaum2000global} network embedding algorithms for dimensionality reduction. Recent network embedding works take advancements in natural language processing, most notably models known as word2vec~\cite{mikolov2013efficient}, which learns the distributed representations of words. Building on word2vec, Perozzi et al. define a vertex's ``context'' by their co-occurrence in a random walk path~\cite{perozzi2014deepwalk}. More recently, Grover et al. propose a mixture of width-first and breadth-first search based procedure to generate paths of vertexes~ \cite{grover2016node2vec}. Dong et al. further develop the model to handle heterogeneous networks~\cite{dong2017metapath}. LINE~\cite{tang2015line} decomposes a vertex's context into first-order (neighbors) and second-order (two-degree neighbors) proximity. Wang et al. preserve community information in their vertex representations~\cite{wang2017community}. However, all of above methods focus on preserving microscopic network structure and ignore macroscopic scale-free property of networks. \vpara{Scale-free Networks.} The scale-free property has been discovered to be ubiquitous over a variety of network systems~\cite{mood1950introduction,newman2005power,clauset2009power}, such as the Internet Autonomous System graph~\cite{faloutsos1999power}, the Internet router graph~\cite{govindan2000heuristics}, the degree distributions of subsets of the world wide web~\cite{barabasi1999emergence}. \citeauthor{newman2005power} provides a comprehensive list of such work~\cite{newman2005power}. However, investigating the scale-free property in a low-dimensional vector space and establishing its cooperation with network embedding have not been fully considered. \section{Conclusion} \label{} In this paper, we study the problem of learning the scale-free property preserving network embeddings. We first analyze the feasibility of reconstructing a scale-free network based on learned vertex representations in the Euclidean space by converting our problem to the Sphere Packing problem. We then propose the \textit{degree penalty} principle as our solution and introduce two implementations by leveraging spectral techniques and a skip-gram model respectively. The proposed principle can also be implemented using other methods, which is left as our future work. We at last conduct extensive experiments on both synthetic data and five real-world datasets to verify the effectiveness of our approach. \paragraph{Acknowledgements.} The work is supported by the Fundamental Research Funds for the Central Universities, 973 Program (2015CB352302), NSFC (U1611461, 61625107, 61402403), and key program of Zhejiang Province (2015C01027). \normalsize \balance \bibliographystyle{aaai}

\section{Introduction} The conical K\"ahler-Ricci flow is the K\"ahler-Ricci flow with certain cone singularities, whose existence, regularity and convergence have been widely studied in the recent years, see e.g. \cite{CW1,CW2,Ed15,Ed17,LZ,LZ2,No,Ta,Wy,Yi,Zy} and references therein. \par In this paper, we shall discuss some properties of the conical K\"ahler-Ricci flow, focusing on the long time collapsing limits and a phenomenon on scalar curvature. \par Let's begin with the general setup as follows. Let $X$ be an $n$-dimensional compact K\"ahler manifold with a K\"ahler metric $\omega_0$ and $D=\sum_{i=1}^lD_i$ a simple normal crossing divisor on $X$ with every $D_i$ an irreducible complex hypersurface and $\beta_i\in(0,1)$. Fix a defining section $S_i$ for $D_i$ on $X$ and a Hermitian metric $h_i$ on the line bundle $L_i$ associated to $D_i$. We fix a sufficiently small positive constant $\delta$ such that $$\omega_0^*:=\omega_0+\delta\sum_{i=1}^{l}\sqrt{-1}\partial\bar\partial|S_i|^{2\beta_i}_{h_i}$$ is a conical K\"ahler metric on $X$ with cone angle $2\pi\beta_i$ along $D_i$, $i=1,...,l$. We call $\omega_0^*$ a model conical K\"ahler metric (with respect to $D$). Then consider the conical K\"ahler-Ricci flow starting from $\omega_0^*$: \begin{equation}\label{CKRF.1} \left\{ \begin{aligned} \partial_t\omega(t)&=-Ric(\omega(t))+2\pi\sum_{i=1}^l(1-\beta_i)[D_i]\\ \omega(0)&=\omega_0^*, \end{aligned} \right. \end{equation} By \cite{CW1,CW2,LZ,Sh} (which generalizes \cite{C,Ts,TZo}), the conical K\"ahler-Ricci flow \eqref{CKRF.1} has a solution up to $$T_{max}:=\{t>0|[\omega_0]+t2\pi(c_1(K_X)+\sum_{i=1}^l(1-\beta_i)[D_i])>0\}.$$ See subsection \ref{reduction} for a more precise definition of ``a solution to the conical K\"ahler-Ricci flow". \subsection{A general phenomenon on scalar curvature along the conical K\"ahler-Ricci flow}\label{general} Firstly, we would like to note that, as a smooth K\"ahler metric on the open (non-compact) manifold $X\setminus D$, the scalar curvature of $\omega_0^*$ on $X\setminus D$ may not be a bounded function (see e.g. Remark \ref{rem_unbdd}). However, we observe that, starting from a model conical K\"ahler metric $\omega_0^*$ with possibly unbounded scalar curvature, the conical K\"ahler-Ricci flow will instantly have bounded scalar curvature for $t\in(0,t_0]$ ($t_0$ is a sufficiently small number), and the bound is of the form $\frac{C}{t}$. Precisely, we have \begin{thm}\label{scal_thm} Let $\omega(t)_{t\in[0,T_{max})}$ be the solution to \eqref{CKRF.1} and $R(t):=R(\omega(t))$ the scalar curvature of $\omega(t)$ on $X\setminus D$. For any fixed $t_0\in(0,T_{max})$, there exists a constant $C\ge1$ such that for any $t\in(0,t_0]$, \begin{equation}\label{scal_bound} \sup_{X\setminus D}|R(t)|\le\frac{C}{t}. \end{equation} \end{thm} If we regard the conical K\"ahler-Ricci flow \eqref{CKRF.1} as a smooth K\"ahler-Ricci flow on the open (non-compact) K\"ahler manifold $X\setminus D$, then it seems our observation in Theorem \ref{scal_thm} fits also in the study of the K\"ahler-Ricci flow starting from a K\"ahler metric with unbounded curvature, where the phenomenon in Theorem \ref{scal_thm} has usually happened even for Riemannian curvature, i.e. once one can solve the K\"ahler-Ricci flow from a K\"ahler metric with unbounded curvature, then Riemannian curvature of solution in some cases has a bound of the form $\frac{C}{t}$ instantly, see e.g. \cite{Si,HT} and references therein for more precise results and discussions. From this viewpoint, it seems interesting to ask: \emph{can we prove the analogue result of Theorem \ref{scal_thm} for Riemannian curvature?} Even more, if this is true, then it provides an effective way to produce a (not necessarily model) conical K\"ahler metric with bounded curvature.\\ \noindent\emph{Remark} We should mention that, in the special case that $X$ is Fano, a weaker version of Theorem \ref{scal_thm} (i.e. replacing the bound $\frac{C}{t}$ in \eqref{scal_bound} by $\frac{C}{t^2}$) seems contained in arguments in \cite[Proposition 4.1, Lemma 4.6]{LZ} and \cite[Proposition 4.1, Theorem 4.4]{LZ2}. \subsection{Diameter bound and convergence}\label{diam} Next, we discuss the diameter bound and convergence of the conical K\"ahler-Ricci flow when $X$ additionally admits a holomorphic submersion over another compact K\"ahler manifold. \par Let $f:X\to Y$ be a holomorphic submersion between two compact K\"ahler manifolds with $n=dim(X)>dim(Y)=k\ge1$ and connected fibers. Fix an irreducible complex hypersurface $D'$ on $Y$, a defining section $S'$ for $D'$ on $Y$ and a Hermitian metric $h'$ on the line bundle $L'$ associated to $D'$. \par Set $D:=f^*D'$, a smooth irreducible complex hypersurface on $X$. Then $S:=f^*S'$ is a defining section for $D$ and $h:=f^*h'$ is a Hermitian metric on $L$, the line bundle associated to $D$. \par For an arbitrary K\"ahler metric $\omega_0$ on $X$ and $\beta\in(0,1)$, we fix a sufficiently small positive constant $\delta$ such that \begin{equation}\label{ref_metric.1} \omega_0^*:=\omega_0+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h \end{equation} is a model conical K\"ahler metric on $X$ with cone angle $2\pi\beta$ along $D$ and \begin{equation}\label{ref_metric} \chi^*=\chi+\delta\sqrt{-1}\partial\bar\partial|S'|^{2\beta}_{h'} \end{equation} is a model conical K\"ahler metric on $Y$ with cone angle $2\pi\beta$ along $D'$. Then, again we consider the following conical K\"ahler-Ricci flow on $X$ starting from $\omega_0^*$: \begin{equation}\label{CKRF} \left\{ \begin{aligned} \partial_t\omega(t)&=-Ric(\omega(t))+2\pi(1-\beta)[D]\\ \omega(0)&=\omega_0^*, \end{aligned} \right. \end{equation} which has a solution up to $$T_{max}:=\{t>0|[\omega_0]+t2\pi(c_1(K_X)+(1-\beta)[D])>0\}.$$ We now assume there exists a K\"ahler metric $\chi$ on $Y$ such that \begin{equation}\label{semiample} f^*\chi\in\frac{1}{T_{max}}[\omega_0]+2\pi(c_1(K_X)+(1-\beta)[D]). \end{equation} Easily, \eqref{semiample} implies \begin{equation} f^*(\chi-(1-\beta)R_{h'})\in \frac{1}{T_{max}}[\omega_0]+2\pi c_1(K_X). \end{equation} In this paper, given above setting, we study the convergence of the following twisted conical K\"ahler-Ricci flow: \begin{equation}\label{TCKRF} \left\{ \begin{aligned} \partial_t\omega(t)&=-Ric(\omega(t))-\omega(t)+\frac{1}{T_{max}}\omega_0+2\pi(1-\beta)[D]\\ \omega(0)&=\omega_0^*, \end{aligned} \right. \end{equation} In general, $0<T_{max}\le\infty$. If $T_{max}=\infty$, then \eqref{TCKRF} is just the usual (normalized) conical K\"ahler-Ricci flow and the fiber is Calabi-Yau; if $T_{max}<\infty$, then \eqref{TCKRF} is really a twisted conical K\"ahler-Ricci flow and the fiber is Fano. We will study these two different cases unifiedly. In both cases, \eqref{TCKRF} can be solved for $t\in[0,\infty)$. \par One of the most natural problems is to understand the long time behavior of the flow. Our main result can be stated as follows. \begin{thm}\label{main_thm} Assume \eqref{semiample} holds. Let $\omega(t)_{t\in[0,\infty)}$ be the solution to \eqref{TCKRF} and $(X,d_t)$ the metric completion of $(X\setminus D,\omega(t))$. We have the following conclusions: \begin{itemize} \item[(1)] There exists a positive constant $C<\infty$ such that \begin{equation} \sup_{t\in[0,\infty)}diam(X,d_t)\le C. \end{equation} \item[(2)] There exists a conical K\"ahler metric $\overline\chi$ on $Y$ with cone angle $2\pi\beta$ along $D'$ such that, as $t\to\infty$, \begin{itemize} \item[(2.1)] $\omega(t)\to f^*\overline\chi$ as currents on $X$; \item[(2.2)] there exists an $\varepsilon_0\in(0,1)$ with the following property: for any $K\subset\subset X\setminus D$ there exista a constant $C_K\ge1$ such that for any $t\in[1,\infty)$, $$|\omega(t)-f^*\overline\chi|_{C^0(K,\omega_0)}\le C_Ke^{-\varepsilon_0 t};$$ \item[(2.3)] if additionally $dim(Y)=1$, then $(X,d_t)\to(Y,\overline d)$ in Gromov-Hausdorff topology, here $(Y,\overline d)$ is the metric completion of $(Y\setminus D',\overline\chi)$. \end{itemize} \end{itemize} \end{thm} In our Theorem \ref{main_thm}, we have focused on the volume collapsing case, i.e. $dim(X)>dim(Y)$. For the volume noncollapsing case, we have convergence results on the conical K\"ahler-Ricci flow of Chen-Wang \cite{CW2} when twisted canonical line bundle $K_X+(1-\beta)L$ is trivial or ample and Shen \cite{Sh} when $K_X+(1-\beta)L$ is nef and big (also see Liu-X. Zhang \cite{LZ} on the conical K\"ahler-Ricci flow on Fano manifolds). Items (2.1) and (2.2) in Theorem \ref{main_thm} can be seen as generalizations of Chen-Wang \cite{CW2} and Shen \cite{Sh} to the volume collapsing case. Of course, Theorem \ref{main_thm} are also generalizations of results on the K\"ahler-Ricci flow \cite{ST07,ST12,FZ,TWY} to the conical setting. \par In the case $T_{max}=\infty$ in \eqref{TCKRF}, the conical K\"ahler metric $\overline\chi$ on $Y$ given in Theorem \ref{main_thm} (2) will satisfy a generalized conical K\"ahler-Einstein equation (see Remark \ref{ke}). \par In the case $\beta=1$ (i.e. we remove the conical singularity) and $T_{max}<\infty$, the equation \eqref{TCKRF} is essentially introduced (in a slightly different form) in \cite{Zo1} and that corresponding convergence result is even new for this smooth setting, solving some special case of \cite[Conjecture 3.1]{Zo1}. \par We remark that items (2.1) and (2.2) of Theorem \ref{main_thm} also hold when $D'$ (and hence $D$) is a simple normal crossing divisor as in subsection \ref{general}. However, our arguments for item (1) in Theorem \ref{main_thm} involve the upper bound for bisectional curvature of the model conical metric $\omega_0^*$, which seems only known when $D'$ is of only one irreducible component; we can extend Theorem \ref{main_thm} (1) to simple normal crossing divisor case, provided an upper bound for bisectional curvature for the simple normal crossing case is available.\\ \par Let's also take a look at Theorem \ref{main_thm} (2.3). If $dim(Y)=1$, we know the hypersurface $D'$ on $Y$ is just a single point, say $o\in Y$, and $D=f^*D'$ is a smooth fiber $X_o$ on $X$. Then Theorem \ref{main_thm} (2.3) means that the twisted conical K\"ahler-Ricci flow \eqref{TCKRF} collapses not only the gereric fiber $X_y:=f^{-1}(y)$ for $y\in Y\setminus\{o\}$, but also the cone divisor $X_o$ to a point on the base $Y$, in Gromov-Hausdorff topology. We hope that the additional condition $dim(Y)=1$ can be removed in the future work. \par We point out that the first result that the conical K\"ahler-Ricci flow will collapse the cone divisor to a point in Gromov-Hausdorff topology was recently obtained by Edwards in \cite{Ed17} on Hirzebruch surfaces (in which Edwards also proved that, in some finite-time noncollapsing case, the conical K\"ahler-Ricci flow can contract cone divisor to a point in Gromov-Hausdorff topology), under certain symmetry condition on the initial model conical metric. Our Theorem \ref{main_thm} (2.3) provides another result on such phenomenon. Note that our result doesn't need any symmetry condition. \subsection{Organization of this paper} \emph{In the remaining part of this paper, we will always be in the setup introduced in subsection \ref{diam}}. As we will pointed out later (see subsection \ref{scal.1}), a proof for Theorem \ref{scal_thm} is essentially contained in the discussions for setup introduced in subsection \ref{diam}. A more precise organization can be found as follows: \begin{itemize} \item In Section \ref{weak_conv}, by using direct arguments on the conical equations, we prove item (2.1) in Theorem \ref{main_thm}. In fact, we will show in Proposition \ref{prop_conv} that the convergence takes place exponentially fast at the level of K\"ahler potentials. \item In Section \ref{equiv}, by using direct arguments on the conical equations, we prove in Proposition \ref{prop2} that the solution is uniformly equivalent to a family of collapsing conical K\"ahler metric, which will immediately imply item (1) in Theorem \ref{main_thm}. In the proof of Proposition \ref{prop2}, a key step is carried out in Lemma \ref{C2.1}, which seems can \emph{not} be proved by passing to a smooth approximation, see Remark \ref{rem_key} for more comments. \item In Section \ref{sect_scal}, we show in Lemma \ref{scal} that the twisted scalar curvature along the twisted conical K\"ahler-Ricci flow \eqref{TCKRF} with the assumption \eqref{semiample} has a bound of the from $\frac{C}{\min\{t,1\}}$. Compare with the uniform scalar bound obtained in \cite{Ed15}, our arguments don't rely on some nontrivial properties of a smooth approximation for the initial model K\"ahler metric (see Remark \ref{rem_reasons} for more comments). We point out in subsection \ref{scal.1} that arguments for Lemma \ref{scal} can be applied to prove Theorem \ref{scal_thm}. The arguments in this Section \ref{sect_scal} need to use a smooth approximation for the twisted conical K\"ahler-Ricci flow. \item In Section \ref{sect_C0}, we use direct arguments and the strategy in \cite{TWY} to prove item (2.2) in Theorem \ref{main_thm}. \item In Section \ref{sect_gh}, we prove Gromov-Hausdorff convergence stated in item (2.3) in Theorem \ref{main_thm}, in which we will make use of Proposition \ref{prop2} crucially. \item In Section \ref{sect_rem}, we present several remarks on a twisted K\"ahler-Ricci flow and its convergence. \end{itemize} \section{The weak convergence}\label{weak_conv} In this section, we shall give a proof for item (2.1) in Theorem \ref{main_thm}. \subsection{Limiting metric $\overline\chi$}\label{subs_limit} We first need to construct the limiting metric $\overline\chi$ on $Y$. The construction is essentially the same as in \cite{ST07,ST12} and its generalization in \cite{ZyZz2}. Let's recall some details. We first fix a smooth real function $\rho$ on $X$ such that \begin{equation}\label{SRF} \overline\omega_0:=\omega_0+\sqrt{-1}\partial\bar\partial\rho \end{equation} on $X$ is a closed real $(1,1)$-form and $\overline\omega_0|_{X_y}$ is the unique K\"ahler metric on $X_y$ in class $[\omega_0|_{X_y}]$ with $Ric(\overline\omega_0|_{X_y})=\frac{1}{T_{max}}\omega_0|_{X_y}$ for every $y\in Y$. Secondly we fix a smooth positive volume form $\Omega$ on $X$ with $\sqrt{-1}\partial\bar\partial\log\Omega=f^*\chi-\frac{1}{T_{max}}\omega_0-(1-\beta)2\pi f^*R_{h'}$. Then define a positive smooth function $G:=\frac{\Omega}{C^{n}_{k}\overline\omega_0^{n-k}\wedge f^*\chi^k}$ on $X$, which descends to a smooth function on $Y$. Set $\overline\chi:=\chi+\sqrt{-1}\partial\bar\partial\psi$ be the unique solution to the following equation: \begin{equation}\label{limit} (\chi+\sqrt{-1}\partial\partial\psi)^k=\frac{e^\psi G}{|S'|_{h'}^{2(1-\beta)}}\chi^k. \end{equation} We know $\psi\in L^\infty(Y)\cap C^\infty(Y\setminus D')$ \cite{Y,Ko98} and $\overline\chi$ is a conical K\"ahler metric on $Y$ with cone angle $2\pi\beta$ along $D'$, see \cite{CGP,GP}. We will see $\overline\chi$ is the limit of the twisted conical K\"ahler-Ricci flow \eqref{TCKRF}. \begin{rem}\label{ke} If $T_{max}=\infty$, then $\overline\chi$ constructed above satisfies a generalized conical K\"ahler-Einstein equation (see \cite{ST07}): $$Ric(\overline\chi)=-\overline\chi+\omega_{WP}+2\pi(1-\beta)[D],$$ where $\omega_{WP}$ is the Weil-Petersson form on $Y$. \end{rem} \subsection{Reduction to a parabolic complex Monge-Amp\`ere equation}\label{reduction} Set $\omega_t:=e^{-t}\omega_0+(1-e^{-t})f^*\chi$ and let $\omega(t)=\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h+\sqrt{-1}\partial\bar\partial\varphi(t)$. It is well-known that \eqref{TCKRF} can be reduced to the following \begin{equation}\label{TCCMA} \left\{ \begin{aligned} \partial_t\varphi(t)&=\log\frac{e^{(n-k)t}(\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h+\sqrt{-1}\partial\bar\partial\varphi(t))^n}{|S|^{-2(1-\beta)}_h\Omega}-\varphi(t)-\delta|S|^{2\beta}_h\\ \varphi(0)&=0, \end{aligned} \right. \end{equation} With the assumption \eqref{semiample}, by \cite{CW1,CW2,LZ,Sh,Wy} we know \eqref{TCCMA} admits a unique solution $\varphi(t)$ for $t\in[0,\infty)$, where $\varphi(t)\in L^\infty(X)\cap C^\infty((X\setminus D)\times[0,\infty))$ (in fact $\varphi(t)$ is H\"older continuous on $X$ with respect to $\omega_0$ for every $t\ge0$) and $\varphi(t)$ satisfies \eqref{TCCMA} in current sense on $X\times[0,\infty)$ and in smooth sense on $(X\setminus D)\times[0,\infty)$; moreover, for any $T<\infty$ one can find a constant $C\ge1$ such that \begin{equation}\label{weak_equiv} C^{-1}\omega_0^*\le\omega(t)\le C\omega_0^* \end{equation} holds on $(X\setminus D)\times[0,T]$. \par We need some estimates along \eqref{TCKRF} and \eqref{TCCMA}. We point out that most estimates in this paper will be obtained by working with the conical equations \eqref{TCKRF} and \eqref{TCCMA} directly. Such direct arguments have been also used in \cite[Lemmas 2.1, 2.2]{LZ3}. Therefore, we will also provide direct arguments for some existing estimates (see in particular \cite{Ed15}) which have been carried out by passing to a smooth approximation. \subsection{Uniform bound for $\varphi$ and $\partial_t\varphi$} We begin with the $C^0$-estimate. In the remaining part of this paper, we assume without loss of any generality that \begin{equation}\label{ineq0} |S'|^2_{h'}\le1. \end{equation} Moreover, we fix a small positive constant $\lambda_0\in(0,1]$ such that for any $\lambda\in[-\lambda_0,\lambda_0]$ there holds on $Y\setminus D'$ that \begin{equation}\label{ineq1} \frac{1}{2}\chi^*\le\chi^*+\lambda R_{h'}\le2\chi^*, \end{equation} \begin{equation}\label{ineq1.1} \frac{1}{2}\chi\le\chi+\lambda R_{h'}\le2\chi, \end{equation} and on $X$ that \begin{equation}\label{ineq2.1} \frac{1}{2}\omega_0\le\omega_0+\lambda f^*R_{h'}\le2\omega_0. \end{equation} We also fix a sufficiently large constant $T_0\ge1$ such that for any $t\ge T_0$ there holds on $Y\setminus D'$ that \begin{equation}\label{ineq3} \frac{1}{2}\chi^*-e^{-t}\chi\ge\frac{1}{3}\chi^*. \end{equation} \begin{lem}\label{C0}\cite{Ed15} There exists a constant $C\ge1$ such that \begin{equation} \sup_{(X\setminus D)\times[0,\infty)}|\varphi|\le C. \end{equation} \end{lem} \begin{proof} This result has been obtained in \cite{Ed15} by using a smooth approximation. We now provide a direct argument without passing to a smooth approximation. Our arguments follow closely the arguments for smooth K\"ahler-Ricci flow in \cite{ST07,FZ}. We now use an auxiliary function as in \cite[Lemmas 2.1, 2.2]{LZ3}: for any given $\lambda\in(0,\lambda_0]$, set $\varphi_\lambda:=\varphi+\lambda\log|S|^{2}_h$. We have \begin{equation}\label{C0.1} \partial_t{\varphi_\lambda}=\log\frac{e^{(n-k)t}(\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h+\lambda f^*R_{h'}+\sqrt{-1}\partial\bar\partial\varphi_\lambda)^n}{|S|^{-2(1-\beta)}_h\Omega}-\varphi-\delta |S|^{2\beta}_h \end{equation} on $X\setminus D$. For any $T<\infty$, we know $\varphi$ is uniformly bounded on $[0,T]$ (see \cite{Sh,Wy}). Therefore, the maximal value of $\varphi_\lambda$ on $X\times[0,T]$ must achieved at some $(x_T,t_T)$ with $x_T\in X\setminus D$. If $t_T=0$, then $\varphi_\lambda(x_T,t_T)$ is bounded from above by a constant independent of $T$; otherwise $t_T>0$, we can apply maximum principle to \eqref{C0.1} at $(x_T,t_T)$ to see that \begin{align}\label{ineq3.5} \varphi(x_T,t_T)\le\log\frac{e^{(n-k)t}(\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h+\lambda f^*R_{h'})^n}{|S|^{-2(1-\beta)}_h\Omega}(x_T,t_T). \end{align} Using \eqref{ineq1} gives \begin{align} \omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h+\lambda f^*R_{h'}&=e^{-t}\omega_0+f^*(\chi^*+\lambda R_{h'})-e^{-t}f^*\chi\nonumber\\ &\le e^{-t}\omega_0+2f^*\chi^*. \end{align} and hence \begin{align}\label{ineq4} &(\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h+\lambda f^*R_{h'})^n\nonumber\\ &\le(e^{-t}\omega_0+2f^*\chi^*)^n\nonumber\\ &\le C'|S|^{-2(1-\beta)}_h(e^{-(n-k)t}+e^{-(n-k+1)t}+\ldots+e^{-nt}) \end{align} for some uniform constant $C'\ge1$. Plugging \eqref{ineq4} into \eqref{ineq3.5} implies $$\varphi(x_T,t_T)\le C''$$ for some uniform constant $C''$ independent on $T$ and $\lambda$. In conclusion, we can choose a constant $C\ge1$ independent on $T\in(0,\infty)$ and $\lambda\in(0,\lambda_0]$ such that \begin{equation}\label{ineq5} \sup_{(X\setminus D)\times [0,T]} \varphi_\lambda\le C. \end{equation} Let $T\to\infty$ and $\lambda\to0$ in \eqref{ineq5}, we get the desired upper bound for $\varphi$: \begin{equation} \sup_{(X\setminus D)\times [0,\infty)} \varphi\le C. \end{equation} Now let's look at the lower bound. For any $\lambda\in(0,\lambda]$, set $\varphi^\lambda:=\varphi-\lambda\log|S|^2_h$. Similarly, we have \begin{equation}\label{C0.2} \partial_t{\varphi^\lambda}=\log\frac{e^{(n-k)t}(\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h-\lambda f^*R_{h'}+\sqrt{-1}\partial\bar\partial\varphi^\lambda)^n}{|S|^{-2(1-\beta)}_h\Omega}-\varphi-\delta |S|^{2\beta}_h \end{equation} on $X\setminus D$. For any $T<\infty$, we know the minimal value of $\varphi^\lambda$ on $X\times[0,T]$ must be achieved at some $(x_T',t_T')$ with $x_T'\in X\setminus D$. If $t_T'\in[0,T_0]$, then $\varphi_\lambda(x_T',t_T')$ is bounded from below by a constant independent on $T$; otherwise $t_T'>T_0$, we can apply maximum principle to \eqref{C0.2} at $(x_T',t_T')$ to see that \begin{align}\label{ineq3.6} \varphi(x_T',t_T')\ge\log\frac{e^{(n-k)t}(\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h-\lambda f^*R_{h'})^n}{|S|^{-2(1-\beta)}_h\Omega}(x_T',t_T')-1. \end{align} Using \eqref{ineq1} and \eqref{ineq3} gives \begin{align} \omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h-\lambda f^*R_{h'}&=e^{-t}\omega_0+f^*(\chi^*-\lambda R_{h'})-e^{-t}f^*\chi\nonumber\\ &\ge e^{-t}\omega_0+\frac{1}{2}f^*\chi^*-e^{-t}f^*\chi\nonumber\\ &\ge e^{-t}\omega_0+\frac{1}{3}f^*\chi^*. \end{align} and hence \begin{align} (\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h-\lambda f^*R_{h'})^n\ge C^{-1}|S|^{-2(1-\beta)}_he^{-(n-k)t}\Omega \end{align} for some uniform constant $C\ge1$. Therefore, we find \begin{align} \varphi^\lambda(x_T',t_T')\ge-C \end{align} for some constant $C\ge1$ independent on $T\in[0,\infty)$ and $\lambda\in(0,\lambda_0]$. Then as before, by letting $T\to\infty$ and $\lambda\to0$ we conclude \begin{equation} \inf_{(X\setminus D)\times[0,\infty)}\varphi\ge-C \end{equation} foe some constant $C\ge1$. \par Lemma \ref{C0} is proved. \end{proof} \begin{lem}\label{C00}\cite{Ed15} There exists a constant $C\ge1$ such that \begin{equation} \sup_{(X\setminus D)\times[0,\infty)}|\partial_t\varphi|\le C. \end{equation} \end{lem} \begin{proof} This result has been obtained in \cite{Ed15} by using a smooth approximation. Again we present a direct proof without passing to a smooth approximation. By computation in \cite{ST07,FZ}, there holds on $(X\setminus D)\times[0,\infty)$ that \begin{equation} (\partial_t-\Delta_{\omega(t)})((e^t-1)\partial_t\varphi-\varphi-\delta|S|^{2\beta}_h)=(n-k)e^t+k-tr_{\omega(t)}\omega_0, \end{equation} and so \begin{equation} (\partial_t-\Delta_{\omega(t)})((e^t-1)\partial_t\varphi-\varphi-\delta|S|^{2\beta}_h-(n-k)e^t-kt)=-tr_{\omega(t)}\omega_0. \end{equation} As in the last lemma, we now use an auxiliary function similar to \cite[Lemmas 2.1, 2.2]{LZ3}: for any $\lambda\in(0,\lambda_0]$, set $H_\lambda:=(e^t-1)\partial_t\varphi-\varphi-\delta|S|^{2\beta}_h-(n-k)e^t-kt+\lambda\log|S|^2_h$. Then by \eqref{ineq2.1} we have, on $X\setminus D$, \begin{align} (\partial_t-\Delta_{\omega(t)})H_\lambda&=-tr_{\omega(t)}(\omega_0-\lambda f^*R_{h'})\nonumber\\ &\le-\frac{1}{2}tr_{\omega(t)}\omega_0<0. \end{align} Obviously, for any $T\in(0,\infty)$, since $\partial_t\varphi$ is bounded on $(X\setminus D)\times[0,T]$, the maximal value of $H_\lambda$ on $X\times[0,T]$ must be achieved at some point $(\tilde x_T,0)$ with $\tilde x_T\in X\setminus D$ and so $\sup_{(X\setminus D)\times[0,T]}H_\lambda$ is bounded from above by a positive constant independent on $T\in(0,\infty)$ and $\lambda\in(0,\lambda_0]$. Therefore, we have a positive constant $C\ge1$ such that, on $(X\setminus D)\times[0,\infty)$, \begin{equation} (e^t-1)\partial_t\varphi-\varphi-\delta|S|^{2\beta}_h-(n-k)e^t-kt\le C, \end{equation} which, combining with the facts that $\partial_t\varphi$ is uniformly bounded on $(X\setminus D)\times[0,1]$ and $\varphi$ is uniformly bounded by Lemma \ref{C0}, implies that $\partial_t\varphi$ is uniformly bounded from above on $(X\setminus D)\times[0,\infty)$. \par For the lower bound, by a computation similar to \cite{ST07} we have \begin{equation} (\partial_t-\Delta_{\omega(t)})(\partial_t\varphi+2\varphi+\delta|S|^{2\beta}_h)=\partial_t\varphi+tr_{\omega(t)}(\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h+f^*\chi)-n-k, \end{equation} and so, for any $\lambda\in(0,\lambda_0]$, \begin{align}\label{ineq7} &(\partial_t-\Delta_{\omega(t)})(\partial_t\varphi+2\varphi+\delta|S|^{2\beta}_h-\lambda\log|S|^2_h)\nonumber\\ &=\partial_t\varphi+tr_{\omega(t)}(\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h+f^*\chi-\lambda f^*R_{h'})-n-k, \end{align} on $(X\setminus D)\times[0,\infty)$. By \eqref{ineq1} we have \begin{align} \omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h+f^*\chi-\lambda f^*R_{h'}&=e^{-t}\omega_0+f^*\chi^*+(1-e^{-t})f^*\chi-\lambda f^*R_{h'}\nonumber\\ &\ge e^{-t}\omega_0+f^*(\chi*-\lambda R_{h'})\nonumber\\ &\ge e^{-t}\omega_0+\frac{1}{2}f^*\chi^*, \end{align} which implies \begin{align}\label{ineq8} tr_{\omega(t)}(\omega_t+\delta\sqrt{-1}\partial\bar\partial|S|^{2\beta}_h+f^*\chi-\lambda f^*R_{h'})&\ge tr_{\omega(t)}(e^{-t}\omega_0+\frac{1}{2}f^*\chi^*)\nonumber\\ &\ge n\left(\frac{(e^{-t}\omega_0+\frac{1}{2}f^*\chi^*)^n}{\omega(t)^n}\right)^{\frac{1}{n}}\nonumber\\ &\ge n\left(\frac{C^{-1}e^{-(n-k)t}|S|^{-2(1-\beta)}_h\Omega}{e^{-(n-k)t}|S|^{-2(1-\beta)}_he^{\partial_t\varphi +\varphi+\delta|S|^{2\beta}_h}\Omega}\right)^{\frac{1}{n}}\nonumber\\ &\ge C^{-1}e^{-\frac{\partial_t\varphi}{n}}, \end{align} for some uniform positive constant $C$ independent on $\lambda\in(0,\lambda_0], $where we have used $\varphi$ is uniformly bounded by Lemma \ref{C0}. Plugging \eqref{ineq8} into \eqref{ineq7} gives \begin{equation} (\partial_t-\Delta_{\omega(t)})(\partial_t\varphi+2\varphi+\delta|S|^{2\beta}_h-\lambda\log|S|^2_h)\ge\partial_t\varphi+C^{-1}e^{-\frac{\partial_t\varphi}{n}}-n-k \end{equation} on $(X\setminus D)\times[0,\infty)$. Now by maximum principle we can find a constant $C\ge1$ such that for any $\lambda\in(0,\lambda_0]$ $$\inf_{(X\setminus D)\times[0,\infty)}\partial_t\varphi+2\varphi+\delta|S|^{2\beta}_h-\lambda\log|S|^2_h\ge-C.$$ Leting $\lambda\to0$ gives the desired uniform lower bound for $\partial_t\varphi$ on $(X\setminus D)\times[0,\infty)$. \par Lemma \ref{C00} is proved. \end{proof} \subsection{Convergence of K\"ahler potentials} We come to the main result of this section. \begin{prop}\label{prop_conv} There exists a constant $C\ge1$ such that for any $t\in[0,\infty)$, \begin{equation}\label{conv_00} \sup_{X\setminus D}|\varphi(t)+\delta|S|^{2\beta}_h-f^*\psi|\le Ce^{-\frac{3}{4}t}. \end{equation} \end{prop} \begin{proof} Set $V:=\varphi+\delta|S|^{2\beta}_h-f^*\psi-e^{-t}\rho$, where $\psi$ is the unique solution to \eqref{limit} and $\rho$ is the function in \eqref{SRF}. By \eqref{limit} and \eqref{TCCMA} we have (see \cite{ST07}) \begin{equation}\label{ineq9} \partial_t V=\log\frac{e^{(n-k)t}(e^{-t}\overline\omega_0+f^*\overline\chi-e^{-t}f^*\chi+\sqrt{-1}\partial\bar\partial V)^n}{C_k^n\overline\omega_0^{n-k}\wedge f^*\overline\chi^k}-V \end{equation} on $(X\setminus D)\times[0,\infty)$. For any $\lambda\in(0,\lambda_0]$ we set $V_\lambda:=V+e^{-t}\lambda\log|S|^2_h$. By \eqref{ineq9} one easily has \begin{equation}\label{ineq10} \partial_t V_\lambda=\log\frac{e^{(n-k)t}(e^{-t}\overline\omega_0+f^*\overline\chi-e^{-t}f^*(\chi+\lambda R_{h'})+\sqrt{-1}\partial\bar\partial V_\lambda)^n}{C_k^n\overline\omega_0^{n-k}\wedge f^*\overline\chi^k}-V_\lambda \end{equation} on $(X\setminus D)\times[0,\infty)$. For any $t\in[0,\infty)$, the maximal value $V_{\lambda,max}(t)$ of $V_\lambda(t)$ must be achieved at some point in $X\setminus D$. Hence, applying the maximum principle in \eqref{ineq10} and then using \eqref{ineq1.1} give \begin{align} \partial_tV_{\lambda,max}+V_{\lambda,max}&\le\log\frac{e^{(n-k)t}(e^{-t}\overline\omega_0+f^*\overline\chi-e^{-t}f^*(\chi+\lambda R_{h'}))^n}{C_k^n\overline\omega_0^{n-k}\wedge f^*\overline\chi^k}\nonumber\\ &\le\log\frac{e^{(n-k)t}(e^{-t}\overline\omega_0+f^*\overline\chi)^n}{C_k^n\overline\omega_0^{n-k}\wedge f^*\overline\chi^k}\nonumber\\ &\le \log(1+Ce^{-t})\nonumber\\ &\le Ce^{-t}, \end{align} i.e., \begin{equation} \partial_t(e^tV_{\lambda,max}-Ct)\le0, \end{equation} and so $$V_{\lambda,max}\le C'e^{-t}(t+1)\le Ce^{-\frac{3}{4}t}$$ for some uniform constant $C\ge1$ independent on $\lambda\in(0,\lambda_0]$. Let $\lambda\to0$, we find a constant $C\ge1$ such that \begin{equation}\label{upper1} \sup_{(X\setminus D)\times[0,\infty)}V\le Ce^{-\frac{3}{4}t}. \end{equation} For the lower bound we define $V^\lambda:=V-e^{-t}\lambda\log|S|^2_h$ for any $\lambda\in(0,\lambda_0]$. We know the minimal value $V^\lambda_{min}(t)$ of $V^\lambda(t)$ must be achieved at some point in $X\setminus D$. Similarly, we have \begin{align} \partial_tV^{\lambda}_{min}+V^{\lambda}_{min}&\ge\log\frac{e^{(n-k)t}(e^{-t}\overline\omega_0+f^*\overline\chi-e^{-t}f^*(\chi-\lambda R_{h'}))^n}{C_k^n\overline\omega_0^{n-k}\wedge f^*\overline\chi^k}\nonumber\\ \end{align} We now fix a large $T_1$ such that $e^{-t}\overline\omega_0+f^*\overline\chi-2e^{-t}f^*\chi>0$ for any $t\ge T_1$ (in particular, $T_1$ does not depend on $\lambda\in(0,\lambda_0]$). Then, for any $t\ge T_1$, \begin{align} \log\frac{e^{(n-k)t}(e^{-t}\overline\omega_0+f^*\overline\chi-e^{-t}f^*(\chi-\lambda R_{h'}))^n}{C_k^n\overline\omega_0^{n-k}\wedge f^*\overline\chi^k}&\ge\log\frac{e^{(n-k)t}(e^{-t}\overline\omega_0+f^*\overline\chi-2e^{-t}f^*\chi)^n}{C_k^n\overline\omega_0^{n-k}\wedge f^*\overline\chi^k}\nonumber\\ &\ge \log(1-Ce^{-t})\nonumber\\ &\ge-Ce^{-t}, \end{align} where $C\ge1$ is a uniform positive constant independent on $t\in[0,\infty)$ and $\lambda\in(0,\lambda_0]$, and we may increase $T_1$ to make sure that $1-Ce^{-t}>0$ for any $t\ge T_1$. Therefore, we arrive at \begin{equation} \partial_t(e^tV^\lambda_{min}+Ct)\ge0, \end{equation} on $t\in[T_1,\infty)$, from which we conclude $$\inf_{(X\setminus D)\times[T_1,\infty)}V^\lambda\ge-Ce^{-\frac{3}{4}t}$$ for some constant $C\ge1$ independent on $\lambda\in(0,\lambda_0]$. By letting $\lambda\to0$, we find \begin{equation}\label{lower.1} \inf_{(X\setminus D)\times[T_1,\infty)}V\ge-Ce^{-\frac{3}{4}t}. \end{equation} Moreover, since $V$ is uniformly bounded on $(X\setminus D)\times[0,T_1]$, after possibly increasing $C$ in \eqref{lower.1} we have \begin{equation}\label{lower.2} \inf_{(X\setminus D)\times[0,\infty)}V\ge-Ce^{-\frac{3}{4}t}. \end{equation} By combining \eqref{upper1}, \eqref{lower.2} and the fact that $\rho$ is a bounded function on $X$ we conclude \eqref{conv_00} and Proposition \ref{prop_conv} is proved. \end{proof} \subsection{Proof of Theorem \ref{main_thm} (2.1)} Consequently, we have \begin{proof}[Proof of Theorem \ref{main_thm} (2.1)] By Lemma \ref{C0} and Proposition \ref{prop_conv}, we conclude that, as $t\to\infty$, $\varphi+\delta|S|^{2\beta}_h\to f^*\psi$ in $L^1(X,\Omega)$-topology and so $\omega(t)\to f^*\overline\chi$ as currents on $X$. \end{proof} \section{uniform equivalence to a family of collapsing conical K\"ahler metrics}\label{equiv} Recall \eqref{weak_equiv} that for any $T<\infty$, there exists a constant $A=A(T)\ge1$ such that \begin{equation}\label{weak_equiv.1} A^{-1}\omega_0^*\le\omega(t)\le A\omega_0^* \end{equation} holds on $(X\setminus D)\times[0,T]$. In our setting, we know the K\"ahler class $[\omega(t)]=e^{-t}[\omega_0]+(1-e^{-t})[f^*\chi]$ along the twisted conical K\"ahler-Ricci flow \eqref{TCKRF} will converge to $[f^*\chi]$, which lies on the boundary of K\"ahler cone of $X$ and is not strictly positive. Therefore, $A=A(T)$ in \eqref{weak_equiv.1} should go to $\infty$ as $T\to\infty$, i.e. $\omega(t)$ should not be uniformly equivalent to a fixed conical K\"ahler metric on $(X\setminus D)\times[0,\infty)$. \par In this section, we will show that $\omega(t)$ is uniformly equivalent to a family of conical K\"ahler metrics on $X$ whose K\"ahler classes will degenerate to $[f^*\chi]$ (see Proposition \ref{prop2}). Let's begin by the following key lemma, which will play an important role in our later discussions. Recall that, defined by \eqref{ref_metric}, $\chi^*$ is a model conical metric on $Y$ with cone angle $2\pi\beta$ along $D'$. \begin{lem}[key lemma]\label{C2.1} There exists a constant $C\ge1$ such that there holds on $(X\setminus D)\times[0,\infty)$ that \begin{equation} tr_{\omega(t)}f^*\chi^*\le C. \end{equation} \end{lem} \begin{proof} The proof makes use of Lemmas \ref{C0}, \ref{C00}, a parabolic Schwarz lemma argument \cite{ST07,Y78} and, especially, an upper bound for bisectional curvature of $\chi^*$ in \cite[Proposition A.1]{JMR}. Firstly, we note that by \eqref{weak_equiv.1} $tr_{\omega(t)}f^*\chi^*$ is uniformly bounded on $(X\setminus D)\times[0,1]$. Secondly, for any given point $x\in X\setminus D$, we choose normal coordinates for $\omega(t)$ around $x$ and $\chi^*$ around $f(x)$, in which we write $tr_{\omega(t)}f^*\chi^*=g^{\bar ji}f^\alpha_i\overline{f^\beta_j}\chi^*_{\alpha\bar\beta}$. A standard computation gives \begin{align}\label{P2.2} &(\partial_t-\Delta_{\omega(t)})tr_{\omega(t)}f^*\chi^*\nonumber\\ &=tr_{\omega(t)}f^*\chi^*-\langle\frac{1}{T_{max}}\omega_0,f^*\chi^*\rangle_{\omega(t)}+g^{\bar lk}g^{\bar j i}Rm(\chi^*)_{\alpha\bar\beta\gamma\bar\delta}f^{\alpha}_i\overline{f^\beta_j}f^\gamma_k\overline{f^\delta_l}-g^{\bar lk}g^{\bar ji}(\partial_kf^\alpha_i)(\partial_lf^\beta_j)\chi^*_{\alpha\bar\beta}\nonumber\\ &\le tr_{\omega(t)}f^*\chi^*+B(tr_{\omega(t)}f^*\chi^*)^2-g^{\bar lk}g^{\bar ji}(\partial_kf^\alpha_i)(\partial_lf^\beta_j)\chi^*_{\alpha\bar\beta}, \end{align} and so \begin{align} (\partial_t-\Delta_{\omega(t)})\log tr_{\omega(t)}f^*\chi^*\le Btr_{\omega(t)}f^*\chi^*+1 \end{align} on $(X\setminus D)\times[0,\infty)$, where $B$ is an upper bound for bisectional curvature of $\chi^*$ on $Y\setminus D'$ given by \cite[Proposition A.1]{JMR} (also see \cite[Section 7]{Br} and \cite[Lemma 2.3]{JMR} for the case $\beta\in(0,\frac{1}{2})$) and we have used \begin{equation}\label{P2.3} |\nabla tr_{\omega(t)} f^*\chi^*|^2-(tr_{\omega(t)} f^*\chi^*)g^{\bar lk}g^{\bar ji}\partial_kf^\alpha_{i}\overline{\partial_lf^\beta_{j}}\chi^*_{\alpha\bar\beta}\le0, \end{equation} On the other hand, by Lemma \ref{C00} and \eqref{ineq3} we have a uniform constant $\tilde C\ge1$ such that \begin{align} (\partial_t-\Delta_{\omega(t)})\varphi&=\partial_t\varphi-tr_{\omega(t)}(\omega(t)-e^{-t}\omega_0)+tr_{\omega(t)}f^*(\chi^*-e^{-t}\chi)\nonumber\\ &\ge tr_{\omega(t)}f^*(\chi^*-e^{-t}\chi)-\tilde C\nonumber\\ &\ge\frac{5}{6}tr_{\omega(t)}f^*\chi^*-\tilde C \end{align} on $(X\setminus D)\times[T_0,\infty)$, here $T_0$ is given in \eqref{ineq3}. Therefore, we arrive at \begin{equation} (\partial_t-\Delta_{\omega(t)})(\log tr_{\omega(t)}f^*\chi^*-\frac{6}{5}(B+1)\varphi)\le-tr_{\omega(t)}f^*\chi^*+1+\frac{6}{5}\tilde C(B+1) \end{equation} and so, for any $\lambda\in(0,\lambda_0]$, \begin{align}\label{ineq11} &(\partial_t-\Delta_{\omega(t)})(\log tr_{\omega(t)}f^*\chi^*-\frac{6}{5}(B+1)\varphi+\lambda\log|S|^2_h)\nonumber\\ &\le-tr_{\omega(t)}f^*(\chi^*+\lambda R_{h'})+1+\frac{6}{5}\tilde C(B+1)\nonumber\\ &\le-\frac{1}{2}tr_{\omega(t)}f^*\chi^*+\frac{6}{5}\tilde C(B+1). \end{align} on $(X\setminus D)\times[T_0,\infty)$. Now by applying the maximum principle in \eqref{ineq11}, Lemma \ref{C0} and \eqref{weak_equiv.1}, we can find a constant $C\ge1$ such that for any $\lambda\in(0,\lambda_0]$ we have \begin{align} \log tr_{\omega(t)}f^*\chi^*+\lambda\log|S|^2_h\le C, \end{align} which implies \begin{equation}\label{upper.2} tr_{\omega(t)}f^*\chi^*\le C \end{equation} on $(X\setminus D)\times[T_0,\infty)$. By \eqref{weak_equiv.1}, after possibly increasing $C$ in \eqref{upper.2}, we have \begin{equation}\label{upper.3} \sup_{(X\setminus D)\times[0,\infty)}tr_{\omega(t)}f^*\chi^*\le C. \end{equation} Lemma \ref{C2.1} is proved. \end{proof} We would like to make a remark on Lemma \ref{C2.1} and its proof. \begin{rem}\label{rem_key} The key Lemma \ref{C2.1} seems can \emph{not} be proved by the usual smooth approximation arguments in e.g. \cite{Ed15,LZ}. A key point is that, while $\chi^*$ has an upper bound for bisectional curvature, it seems unclear how to approximate $\chi^*$ (in a suitable sense) by a family of smooth K\"ahler metrics with bisectional curvature uniformly bounded from above. In our argument, we don't need to pass to a smooth approximation of conical equation and the upper bound for bisectional curvature of $\chi^*$ itself is enough to carry out our Lemma \ref{C2.1}. \end{rem} Next we make use of Lemma \ref{C2.1} and an argument in \cite{ST07} to show \begin{lem}\label{C2.2} There exists a constant $C\ge1$ such that there holds on $(X\setminus D)\times[0,\infty)$ that \begin{equation} tr_{\omega(t)}(e^{-t}\omega_0)\le C. \end{equation} \end{lem} \begin{proof} As before, by \eqref{weak_equiv.1} we know $tr_{\omega(t)}(e^{-t}\omega_0)$ is uniformly bounded on $(X\setminus D)\times[0,1]$. \par Firstly, we have two positive constants $B_0$ (an upper bound for bisectional curvature of $\omega_0$) and $C$ such that \begin{equation}\label{ineq12} (\partial_t-\Delta_{\omega(t)})\log tr_{\omega(t)}(e^{-t}\omega_0)\le B_0tr_{\omega(t)}\omega_{0}+C \end{equation} on $(X\setminus D)\times[0,\infty)$. \par Next, following \cite{ST07}, we define a function $\overline\varphi$ on $(Y\setminus D')\times[0,\infty)$: for any $(y,t)\in(Y\setminus D')\times[0,\infty)$, \begin{equation} \overline\varphi(y,t):=\frac{\int_{X_y}\varphi(t)(\omega_{0}|_{X_y})^{n-k}}{\int_{X_y}(\omega_{0}|_{X_y})^{n-k}}. \end{equation} Then we have \begin{align} (\omega_0|_{X_y}+\sqrt{-1}\partial\bar\partial(e^t(\varphi|_{X_y}-\overline\varphi(y,t))))^{n-k}&=(e^{t}\omega(t)|_{X_y})^{n-k}, \end{align} and so \begin{align} &\frac{(\omega_0|_{X_y}+\sqrt{-1}\partial\bar\partial(e^t(\varphi|_{X_y}-\overline\varphi)))^{n-k}}{(\omega_0|_{X_y})^{n-k}}\nonumber\\ &=\frac{(e^{t}\omega(t)|_{X_y})^{n-k}}{(\omega_0|_{X_y})^{n-k}}\nonumber\\ &=\frac{e^{(n-k)t}\omega(t)^{n-k}\wedge f^*(\chi^*)^k}{\omega_0^{n-k}\wedge f^*(\chi^*)^k}\nonumber\\ &=\frac{\omega(t)^{n-k}\wedge f^*(\chi^*)^k}{\omega(t)^n}\cdot\frac{e^{(n-k)t}\omega(t)^n}{\omega_0^{n-k}\wedge f^*(\chi^*)^k}\nonumber\\ &=\frac{\omega(t)^{n-k}\wedge f^*(\chi^*)^k}{\omega(t)^n}\cdot\frac{e^{\partial_t\varphi+\varphi+\delta|S|^{2\beta}_h}|S|^{-2(1-\beta)}_h\Omega}{\omega_0^{n-k}\wedge f^*(\chi^*)^k}\nonumber\\ &\le C' (tr_{\omega(t)}f^*\chi^*)^{k}\nonumber\\ &\le C, \end{align} where we have used Lemmas \ref{C0}, \ref{C00} and \ref{C2.1}, and $C$ in the last line is a positive constant uniform for $(x,t)\in(X\setminus D)\times[0,\infty)$ (we point out that Lemma \ref{C2.1} plays a crucial role in this step). Then by Yau \cite{Y} we get a constant $C\ge1$ such that for all $(y,t)\in(Y\setminus D')\times[0,\infty)$, \begin{equation}\label{ineq12.5} \sup_{X_y}|e^t(\varphi-\overline\varphi)|\le C. \end{equation} Note that $|S|^{2\beta}_h=f^*|S'|^{2\beta}_{h'}$ is constant along every fiber $X_y$ and hence $$(\varphi+\delta|S|^{2\beta}_h)-(\overline\varphi+\delta|S|^{2\beta}_h)=\varphi-\overline\varphi.$$ Now by using arguments in \cite{ST07} (also see \cite{FZ}), Lemma \ref{C00} and Lemma \ref{C2.1} (in fact a weaker version that $tr_{\omega(t)}f^*\chi\le C$ is enough) we have a constant $C\ge1$ such that \begin{align}\label{ineq13} (\partial_t-\Delta_{\omega(t)})(e^t(\varphi-\overline\varphi))&=(\partial_t-\Delta_{\omega(t)})\left(e^t((\varphi+\delta|S|^{2\beta}_h)-(\overline\varphi+\delta|S|^{2\beta}_h))\right)\nonumber\\ &\ge tr_{\omega(t)}\omega_0-Ce^t \end{align} holds on $(X\setminus D)\times[0,\infty)$. Combining \eqref{ineq12} and \ref{ineq13} gives \begin{equation} (\partial_t-\Delta_{\omega(t)})(\log tr_{\omega(t)}(e^{-t}\omega_0)-(B_0+1)e^t(\varphi-\overline\varphi))\le -tr_{\omega(t)}\omega_0+Ce^t \end{equation} on $(X\setminus D)\times[0,\infty)$, where $C\ge1$ is a uniform constant. Now for any $\lambda\in(0,\lambda_0]$, by \eqref{ineq2.1} there holds on $(X\setminus D)\times[0,\infty)$ that \begin{equation}\label{ineq14} (\partial_t-\Delta_{\omega(t)})(\log tr_{\omega(t)}(e^{-t}\omega_0)-(B_0+1)e^t(\varphi-\overline\varphi)+\lambda\log|S|^2_h)\le -\frac{1}{2}tr_{\omega(t)}\omega_0+Ce^t. \end{equation} By applying the maximum principle in \eqref{ineq14} and using \eqref{ineq12.5} we find a constant $C\ge1$ such that for any $\lambda\in(0,\lambda_0]$, $$\sup_{(X\setminus D)\times[0,\infty)}(\log tr_{\omega(t)}(e^{-t}\omega_0)-(B_0+1)e^t(\varphi-\overline\varphi)+\lambda\log|S|^2_h)\le C,$$ from which, by letting $\lambda\to0$ and using \eqref{ineq12.5} again, we conclude a constant $C\ge1$ such that \begin{equation} \sup_{(X\setminus D)\times[0,\infty)}tr_{\omega(t)}(e^{-t}\omega_0)\le C. \end{equation} Lemma \ref{C2.2} is proved. \end{proof} The main result of this section is the following, which says $\omega(t)$ is uniformly equivalent to a family of collapsing conical K\"ahler metrics. \begin{prop}\label{prop2} There exists a constant $C\ge1$ such that \begin{equation}\label{C2.3} C^{-1}(e^{-t}\omega_0+f^*\chi^*)\le\omega(t)\le C(e^{-t}\omega_0+f^*\chi^*) \end{equation} holds on $(X\setminus D)\times[0,\infty)$. \end{prop} \begin{proof} The left hand side follows from Lemmas \ref{C2.1} and \ref{C2.2} directly. For the right hand side, we look at \begin{align} tr_{(e^{-t}\omega_0+f^*\chi^*)}\omega(t)&\le\frac{1}{(n-1)!}(tr_{\omega(t)}(e^{-t}\omega_0+f^*\chi^*))^{n-1}\cdot\frac{\omega(t)^n}{(e^{-t}\omega_0+f^*\chi^*)^n}\nonumber\\ &\le C\frac{e^{-(n-k)t}e^{\partial_t\varphi+\varphi+\delta|S|^{2\beta}_h}|S|^{-2(1-\beta)}_h\Omega}{e^{-(n-k)t}\omega_0^{n-k}\wedge f^*(\chi^*)^k}\nonumber\\ &\le C, \end{align} on $(X\setminus D)\times[0,\infty)$, where we have used left hand side of \eqref{C2.3} and upper bound of $\partial_t\varphi,\varphi$ given by Lemmas \ref{C0} and \ref{C00}. Then we conclude that $$\omega(t)\le C(e^{-t}\omega_0+f^*\chi^*)$$ on $(X\setminus D)\times[0,\infty)$. \par Proposition \ref{prop2} is proved. \end{proof} \begin{rem} Before moving to the next step, we would like to provide a more direct argument for a special case of Lemma \ref{C2.2}, i.e. the case that $T_{max}=\infty$, $X=E\times Y$ with $E$ a smooth $(n-k)$-dimensional torus and $f:X=E\times Y\to Y$ is the projection. Denote $\tilde f:X\to E$ be the projection to the factor $E$ and $\omega_E$ a fixed flat K\"ahler metric on $E$. Then we shall show that there exists a constant $C\ge1$ such that for $(x,t)\in (X\setminus D)\times[0,\infty)$, \begin{equation}\label{e.0} tr_{\omega(t)}(e^{-t}\tilde f^*\omega_E)\le C. \end{equation} In fact, using $\omega_E$ is flat, we easily have \begin{equation} (\partial_t-\Delta_{\omega(t)})tr_{\omega(t)}(e^{-t}\tilde f^*\omega_E)\le0\nonumber \end{equation} and so \begin{equation} (\partial_t-\Delta_{\omega(t)})(tr_{\omega(t)}(e^{-t}\tilde f^*\omega_E)+\lambda\log|S|^2_h)\le \lambda tr_{\omega(t)}R_h\le \lambda \tilde Ctr_{\omega(t)}f^*\chi\le\lambda \tilde C_1\nonumber, \end{equation} namely, \begin{equation}\label{e.1} (\partial_t-\Delta_{\omega(t)})(tr_{\omega(t)}(e^{-t}\tilde f^*\omega_E)+\lambda\log|S|^2_h-\lambda\tilde C_1t)\le0 \end{equation} on $(X\setminus D)\times[0,\infty)$. For any $T<\infty$, we can apply the maximal principle in \eqref{e.1} to see $$\sup_{(X\setminus D)\times[0,T]}(tr_{\omega(t)}(e^{-t}\tilde f^*\omega_E)+\lambda\log|S|^2_h-\lambda\tilde C_1t)\le C,$$ where $C$ does not depend on $T\in(0,\infty)$ and $\lambda\in(0,\lambda_0]$. Letting $T\to\infty$ and $\lambda\to0$ gives the desired result \eqref{e.0}. Having \eqref{e.0}, we can use arguments in Proposition \ref{prop2} to see that $\omega(t)$ is uniformly equivalent to $e^{-t}\tilde f^*\omega_E+f^*\chi^*$ on $(X\setminus D)\times[0,\infty)$. \end{rem} \subsection{Proof of Theorem \ref{main_thm} (1)} We now end this section by proving diameter upper bound of $(X,\omega(t))$, i.e. item (1) of Theorem \ref{main_thm}. \begin{proof}[Proof of Theorem \ref{main_thm} (1)] By the right hand side of \eqref{C2.3} in Proposition \ref{prop2} we can find a constant $C\ge1$ such that $$\omega(t)\le C\omega_0^*$$ holds on $(X\setminus D)\times[0,\infty)$, where $\omega_0^*$ is defined in \eqref{ref_metric.1}, the reference conical K\"ahler metric on $X$ with cone angle $2\pi\beta$ along $D$. Let $(X,d_0)$ be the metric completion of $(X,\omega_0^*)$, which is of finite diameter. Therefore, $$diam(X,d_t)\le C\cdot diam(X,d_0)$$ is unifromly bounded from above. \end{proof} \begin{rem} Easily, the left hand side of \eqref{C2.3} in Proposition \ref{prop2} also implies a uniform positive lower bound for $diam(X,d_t)$. \end{rem} \section{A bound for the twisted scalar curvature}\label{sect_scal} For later discussions, we need to bound the twisted scalar curvature along the twisted conical K\"ahler-Ricci flow \eqref{TCKRF}. We define the twisted Ricci curvature by \begin{equation} \tilde{Ric}(\omega(t)):=Ric(\omega(t))-\frac{1}{T_{max}}\omega_0 \end{equation} and the twisted scalar curvature by \begin{equation} \tilde R(\omega(t)):=tr_{\omega(t)}\tilde{Ric}(\omega(t))=R(\omega(t))-\frac{1}{T_{max}}tr_{\omega(t)}\omega_0. \end{equation} Note that if $T_{max}=\infty$, the twisted term vanishes and the twisted curvatures defined above are the usual curvatures. \par We will need the following result. \begin{lem}\label{scal}\cite{Ed15} There exists a constant $C\ge1$ such that for any $t\in(0,\infty)$, \begin{equation} \sup_{X\setminus D}|\tilde R|(t)\le \frac{C}{min\{t,1\}}. \end{equation} \end{lem} In the case $T_{max}=\infty$, the twisted term $\frac{1}{T_{max}}\omega_0$ vanishes and the twisted scalar is just the scalar curvature, which is uniformly bounded by \cite[Theorem 1.1]{Ed15} using a smooth approximation argument (this is a generalization of Song and Tian's result on the K\"ahler-Ricci flow \cite{ST16}) In fact, while our Lemma \ref{scal} only provides uniform bound for $t\ge1$, \cite[Theorem 1.1]{Ed15} says the bound for scalar curvature can be uniform for all $t\in[0,\infty)$ (i.e. $\sup_{(X\setminus D)\times[0,\infty)}|\tilde R|\le C<\infty$), where some properties of a smooth approximation for $\omega_0^*$ is involved (see discussions next to \eqref{lower2}, \eqref{P2.7} and \eqref{P2.10} for more details). Though bounding scalar curvature along the conical K\"ahler-Ricci flow has been studied in some different settings, see e.g. \cite{Ed15,No}, we still would like to provide an argument for the above Lemma \ref{scal}. Let's list in the following Remark \ref{rem_reasons} the main reasons for doing this. \begin{rem}\label{rem_reasons} We provide an argument for Lemma \ref{scal}, mainly due to twofold reasons: (1) we will use slightly different quantum to apply the maximum principle (see footnotes next to \eqref{footnote1} and \eqref{P2.6} for more details); (2) we can avoid using some properties of a smooth approximation $\omega_{0,\epsilon}^*$ (given in \eqref{appro}) for $\omega_0^*$, which seems needed in \cite[Theorem 1.1, Sections 4,7,8]{Ed15} and \cite[Theorem A, Sections 6,7]{No} (see discussions next to \eqref{lower2}, \eqref{P2.7} and \eqref{P2.10} for more details). \end{rem} \begin{rem}\label{rem_unbdd} It seems in general we can not improve the bound $\frac{C}{t}$ to a uniform bound $C$ near $t=0$, as the initial model conical metric $\omega^*_0$ may have unbounded scalar curvature. For example, as pointed out in \cite[Lemma 3.14]{Ru}, if $\beta\in(\frac{1}{2},1)$, the computation in \cite[Proposition A.1]{JMR} provides not only the existence of an upper bound for bisectional curvature, but also the nonexistence of a lower bound for bisectional curvature, and hence the nonexistence of a lower bound for scalar curvature. \end{rem} \begin{proof}[Proof of Lemma \ref{scal}] We need to use a smooth approximation for the conical equations \eqref{TCKRF} and \eqref{TCCMA}, introduced in \cite{LZ,Wy} and also used in e.g. \cite{Sh,Ed15}. Following \cite{CGP} we define $$\eta_\epsilon:=\beta\int_{0}^{|S|^{2}_h}\frac{(r+\epsilon^2)^\beta-\epsilon^{2\beta}}{r}dr,$$ and \begin{equation}\label{appro} \omega^*_{0,\epsilon}:=\omega_0+\sqrt{-1}\partial\bar\partial\eta_\epsilon. \end{equation} Then one can approximate \eqref{TCCMA} by \begin{equation}\label{TCCMA.1} \left\{ \begin{aligned} \partial_t\varphi_\epsilon(t)&=\log\frac{e^{(n-k)t}(\omega_t+\delta\sqrt{-1}\partial\bar\partial\eta_\epsilon+\sqrt{-1}\partial\bar\partial\varphi_\epsilon(t))^n}{(|S|^{2}_h+\epsilon^2)^{-(1-\beta)}\Omega}-\varphi_\epsilon(t)-\delta\eta_\epsilon\\ \varphi_\epsilon(0)&=0, \end{aligned} \right. \end{equation} and, correspondingly, if we let $\omega_\epsilon(t):=\omega_t+\delta\sqrt{-1}\partial\bar\partial\eta_\epsilon+\sqrt{-1}\partial\bar\partial\varphi_\epsilon(t)$, then we have \begin{equation}\label{TCKRF.1} \left\{ \begin{aligned} \partial_t\omega_\epsilon(t)&=-Ric(\omega_\epsilon(t))-\omega_\epsilon(t)+\frac{1}{T_{max}}\omega_0+(1-\beta)(\sqrt{-1}\partial\bar\partial\log(|S|^2_h+\epsilon^2)+R_h)\\ \omega(0)&=\omega^*_{0,\epsilon}, \end{aligned} \right. \end{equation} Set $A_\epsilon:=(1-\beta)(\sqrt{-1}\partial\bar\partial\log(|S|^2_h+\epsilon^2)+R_h)$, then it suffices to get a bound for $\tilde R_\epsilon:=tr_{\omega_\epsilon(t)}(Ric(\omega_\epsilon(t))-\frac{1}{T_{max}}\omega_0-A_\epsilon)$. Let's first look at the easier part, i.e. lower bound.\\ \textbf{Claim 1}: For any $\epsilon>0$ and $t>0$, \begin{equation}\label{lower} \inf_{X}\tilde R_\epsilon\ge -\frac{3n}{\min\{t,1\}} \end{equation} \begin{proof}[Proof of Claim 1] First we have the evolution of $\tilde R_\epsilon$ along \eqref{TCKRF.1} (see e.g. \cite[Proposition 4.1]{Ed15}): \begin{equation}\label{lower1} (\partial_t-\Delta_{\omega_\epsilon(t)})\tilde R_\epsilon=|Ric(\omega_\epsilon(t))-\frac{1}{T_{max}}\omega_0-A_\epsilon|_{\omega_{\epsilon}(t)}^2+\tilde R_\epsilon, \end{equation} which, combining with $|Ric(\omega_\epsilon(t))-\frac{1}{T_{max}}\omega_0-A_\epsilon|_{\omega_{\epsilon}(t)}^2\ge\frac{1}{n}\tilde R_\epsilon^2$, implies \begin{equation}\label{lower2} (\partial_t-\Delta_{\omega_\epsilon(t)})(e^t(\tilde R_\epsilon+n))\ge \frac{1}{n}e^t(\tilde R_\epsilon+n)^2. \end{equation} To avoid involving the uniform bound for $\tilde R_\epsilon(0)$, we use a trick similar to e.g. \cite[Lemma 3.2]{ST17} and further consider \begin{equation}\label{lower3} (\partial_t-\Delta_{\omega_\epsilon(t)})(te^t(\tilde R_\epsilon+n))\ge \frac{1}{n}te^t(\tilde R_\epsilon+n)^2+e^t(\tilde R_\epsilon+n) \end{equation} We now apply the maximum principle in \eqref{lower3}. Assume $te^t(\tilde R_\epsilon+n)$ achieves the minimal value on $X\times[0,1]$ at $(x',t')\in X\times[0,1]$. If $t'=0$, $(te^t(\tilde R_\epsilon+n))(x',t')=0$ is uniformly bounded from below; otherwise $t'>0$, then by \eqref{lower3} we know, at $(x',t')$, \begin{equation}\label{lower4.1} t(\tilde R_\epsilon+n)^2\le -n(\tilde R_\epsilon+n), \end{equation} which in particular implies $(\tilde R_\epsilon+n)(x',t')\le0$. We may assume $$(\tilde R_\epsilon+n)(x',t')<0$$ (otherwise we are done), and so by \eqref{lower4.1} we find \begin{equation}\label{lower5} t(\tilde R_\epsilon+n)\ge -n, \end{equation} and so $(te^t(\tilde R_\epsilon+n))(x',t')\ge-en$. In conclusion, we have \begin{equation}\label{lower5.1} \inf_{X\times[0,1]}(te^t(\tilde R_\epsilon+n))\ge -en, \end{equation} which implies \begin{equation}\label{lower5.2} \inf_{X\times[0,1]}(t(\tilde R_\epsilon+n))\ge -en, \end{equation} and so $\inf_X(\tilde R_\epsilon+n))(1)\ge-n$. Now, back to \eqref{lower2}, we easily find that for any $t\ge1$, $$\inf_X(e^t(\tilde R_\epsilon+n))(t)\ge\inf_X(e^t(\tilde R_\epsilon+n))(1)\ge-en,$$ and \begin{equation}\label{lower5.3} \inf_{X\times[1,\infty)}(\tilde R_\epsilon+n)\ge-n. \end{equation} Combining \eqref{lower5.2} and \eqref{lower5.3}, Claim 1 is proved. \end{proof} Next we try to get an upper bound for $\tilde R_\epsilon$. Recall that \cite[Propositions 3.1,3.2,6.1]{Ed15} provides a constant $C\ge1$ such that for any small $\epsilon>0$, there holds \begin{equation} \sup_{X\times[0,\infty)}(|\varphi_\epsilon|+|\partial_t\varphi_\epsilon|+tr_{\omega_\epsilon(t)}f^*\chi)\le C. \end{equation} It suffices to remark that the twisted term $\frac{1}{T_{max}}\omega_0$ is smooth and semipositive and hence does not cause any trouble. We also recall some useful inequalities. Since $\chi$ is a K\"ahler metric on $Y$, which in particular has bounded bisectional curvature, by arguments similar to Lemma \ref{C2.1} (or see \cite[Section 6]{Ed15}) we have a constant $C\ge1$ such that for all positive $\epsilon$, \begin{align}\label{P2.2} &(\partial_t-\Delta_{\omega_\epsilon(t)})tr_{\omega_\epsilon(t)}f^*\chi\le C-C^{-1}|\nabla tr_{\omega_\epsilon(t)} f^*\chi|_{\omega_\epsilon(t)}^2, \end{align} where we have used $A_\epsilon\ge-Cf^*\chi$ and $\frac{1}{T_{max}}\omega_0\ge0$. Set $u_\epsilon:=\partial_t\varphi_\epsilon+\varphi_\epsilon+\delta\eta_\epsilon$. Then by a direct computation we have, on $X\times[0,\infty)$, \begin{equation} (\partial_t-\Delta)u_\epsilon=tr_{\omega_{\epsilon}(t)}f^*\chi-k\le C. \end{equation} To simplify the notation we set $|\cdot|:=|\cdot|_{\omega_\epsilon(t)}$. We first bound $|\nabla u_\epsilon|$. A direct computation gives \begin{align} &(\partial_t-\Delta)|\nabla u_\epsilon|^2\nonumber\\ &=|\nabla u_\epsilon|^2+2Re\langle\nabla tr_{\omega_\epsilon(t)} f^*\chi,\overline\nabla u_\epsilon\rangle_{\omega_\epsilon(t)}-|\nabla\nabla u_\epsilon|^2-|\nabla\overline\nabla u_\epsilon|^2-(\frac{1}{T_{max}}\omega_0+A_\epsilon)(\nabla u_\epsilon,\overline\nabla u_\epsilon), \end{align} which, combining with $A_\epsilon\ge-Cf^*\chi$ and $\frac{1}{T_{max}}\omega_0\ge0$, in particular implies that \begin{align}\label{P2.2.3} (\partial_t-\Delta)|\nabla u_\epsilon|^2\le-|\nabla\nabla u_\epsilon|^2-|\nabla\overline\nabla u_\epsilon|^2+C|\nabla u_\epsilon|^2+|\nabla tr_{\omega_\epsilon(t)} f^*\chi|^2. \end{align} Now we fix a sufficiently large constant $B>4$ such that \begin{equation} 2|u_\epsilon|\le B, \end{equation} and so, \begin{equation}\label{footnote1} 2<\frac{B}{2}\le B-u_\epsilon\le2B. \end{equation} Fix a constant $s\in(0,1)$ and set\footnote{We point out that in previous works on bounding scalar curvature along K\"ahler-Ricci flow (see \cite{SeT,ST16,Zo,Ed15}), people usually choose $s=1$ to define this quantity $E$. Here we choose $s<1$ to slightly simplify some arguments.} $E_\epsilon:=\frac{|\nabla u_\epsilon|^2}{(B-u_\epsilon)^s}$. Compute \begin{align}\label{P2.4} &(\partial_t-\Delta_{\omega_\epsilon(t)})E_\epsilon\nonumber\\ &=\frac{(\partial_t-\Delta_{\omega_\epsilon(t)})|\nabla u_\epsilon|^2}{(B-u_\epsilon)^s}+\frac{s|\nabla u_\epsilon|^2(\partial_t-\Delta_{\omega_\epsilon(t)}) u_\epsilon}{(B-u_\epsilon)^{s+1}}-\frac{2sRe\langle\nabla |\nabla u_\epsilon|^2,\nabla u_\epsilon\rangle_{\omega_\epsilon(t)}}{(B-u_\epsilon)^{s+1}}-\frac{s(s+1)|\nabla u_\epsilon|^4}{(B-u_\epsilon)^{s+2}}\nonumber\\ &\le\frac{-|\nabla\nabla u_\epsilon|^2-|\nabla\overline\nabla u_\epsilon|^2+2|\nabla u_\epsilon|^2+|\nabla tr_{\omega_\epsilon(t)} f^*\chi|^2}{(B-u_\epsilon)^s}+\frac{sC|\nabla u_\epsilon|^2}{(B-u_\epsilon)^{s+1}}\nonumber\\ &+\frac{2s|\langle\nabla |\nabla u_\epsilon|^2,\nabla u_\epsilon\rangle_{\omega_\epsilon(t)}|}{(B-u_\epsilon)^{s+1}}-\frac{s(s+1)|\nabla u_\epsilon|^4}{(B-u_\epsilon)^{s+2}}. \end{align} Note that \begin{align}\label{P2.5} \frac{2s|\langle\nabla |\nabla u_\epsilon|^2,\nabla u_\epsilon\rangle_{\omega_\epsilon(t)|}}{(B-u_\epsilon)^{s+1}}&\le\frac{2s|\nabla u_\epsilon|^2(|\nabla\nabla u_\epsilon|+|\nabla\bar\nabla uz-\epsilon|)}{(B-u_\epsilon)^{s+1}}\nonumber\\ &\le s\left(\frac{2(|\nabla\nabla u_\epsilon|^2+|\nabla\overline\nabla u_\epsilon|^2)}{2s(B-u_\epsilon)^s}+\frac{2s|\nabla u_\epsilon|^4}{(B-u_\epsilon)^{s+2}}\right)\nonumber\\ &=\frac{|\nabla\nabla u_\epsilon|^2+|\nabla\overline\nabla u_\epsilon|^2}{(B-u_\epsilon)^s}+\frac{2s^2|\nabla u_\epsilon|^4}{(B-u_\epsilon)^{s+2}}. \end{align} Putting \eqref{P2.5} into \eqref{P2.4} gives\footnote{Note that choosing the positive number $s$ to be in $(0,1)$ helps to make sure the factor $s-s^2$ appeared in \eqref{P2.6} is positive and avoid an additional trick used in \cite{ST16,Ed15}.} \begin{align}\label{P2.6} (\partial_t-\Delta_{\omega_\epsilon(t)})E_\epsilon&\le C|\nabla u_\epsilon|^2+|\nabla tr_{\omega_\epsilon(t)} f^*\chi|^2-(B-u_\epsilon)^{-(s+2)}(s-s^2)|\nabla u_\epsilon|^4\nonumber\\ &\le C|\nabla u_\epsilon|^2+|\nabla tr_{\omega_\epsilon(t)} f^*\chi|^2-(2B)^{-(s+2)}(s-s^2)|\nabla u_\epsilon|^4 \end{align} Now consider $F_\epsilon:=E_\epsilon+C_1tr_{\omega_\epsilon(t)} f^*\chi$. Combining \eqref{P2.2} and \eqref{P2.6} gives \begin{equation}\label{P2.7} (\partial_t-\Delta_{\omega_\epsilon(t)})F_\epsilon\le C+C|\nabla u_\epsilon|^2-(2B)^{-(s+2)}(s-s^2)|\nabla u_\epsilon|^4. \end{equation} To get an upper bound for $F_\epsilon$, one may like to apply the maximum principle in $\eqref{P2.7}$ (see last step in \cite[Section 7]{Ed15} or \cite[Section 6]{No}). Note that, arguing in such way, since the maximal value may be achieved at $t=0$, one may need a uniform bound for $|\nabla u_{\epsilon}|(0)=|\nabla\log\frac{(\omega_{0,\epsilon}^*)^n}{(|S|^2_h+\epsilon^2)^{-(1-\beta)}\Omega}|$. To avoid involving such bound, here we use a trick similar to e.g. \cite[Lemma 3.2]{ST17} and further consider \begin{align}\label{P2.8} (\partial_t-\Delta_{\omega_\epsilon(t)})(tF_\epsilon)&\le F_{\epsilon}+t(C+C|\nabla u_\epsilon|^2-(2B)^{-(s+2)}(s-s^2)|\nabla u_\epsilon|^4)\nonumber\\ &\le C(t+1)+C(t+1)|\nabla u_\epsilon|^2-t(2B)^{-(s+2)}(s-s^2)|\nabla u_\epsilon|^4. \end{align} Assume the maximal value of $tF_\epsilon$ on $X\times[0,1]$ is achieved at $(x',t')\in X\times[0,1]$. If $t'=0$, $(tF_\epsilon)(x',t')=0$ is bounded from above; otherwise $t'>0$, by \eqref{P2.8} we find a constant $C\ge1$ independent on $\epsilon$ such that, at $(x',t')$, $$t|\nabla u_\epsilon|^4\le C(t+1)|\nabla u_\epsilon|^2+C(t+1).$$ Note that $t'\in(0,1]$. We have, at $(x',t')$, \begin{align} (t|\nabla u_\epsilon|^2)^2&\le Ct(t+1)+Ct(t+1)|\nabla u_\epsilon|^2\nonumber\\ &\le C+C(t|\nabla u_\epsilon|^2) \end{align} and so $$(t|\nabla u_\epsilon|^2)(x',t')\le C.$$ Plugging into $tF_\epsilon$ we find $$(tF_\epsilon)(x',t')=\left(\frac{t|\nabla u_\epsilon|^2}{(B-u_\epsilon)^s}+tC_1tr_{\omega_\epsilon(t)}f^*\chi\right)(x',t')\le C.$$ Therefore, we have find a constant $C\ge1$ independent on $\epsilon$ such that \begin{equation}\label{upper11} \sup_{X\times[0,1]}(tF_\epsilon)\le C, \end{equation} which implies \begin{equation}\label{upper11.1} \sup_{X\times[0,1]}(t|\nabla u_\epsilon|^2)\le C. \end{equation} In particular, we have a constant $C\ge1$ independent on $\epsilon$ such that \begin{equation}\label{upper1} \sup_{X}F_\epsilon(1)\le C. \end{equation} Having \eqref{upper1}, we can now apply the maximum principle in \eqref{P2.7} to conclude that \begin{equation} \sup_{X\times[1,\infty)}F_{\epsilon}\le C\nonumber \end{equation} and so \begin{equation}\label{upper2} \sup_{X\times[1,\infty)}|\nabla u_{\epsilon}|^2\le C \end{equation} for some constant $C\ge1$ independent on $\epsilon$. Combining \eqref{P2.2}, \eqref{P2.2.3} and \eqref{upper2}, we have constants $C_1,C\ge1$ independent on $\epsilon$ such that \begin{equation}\label{P2.4.0} (\partial_t-\Delta_{\omega_\epsilon(t)})(t(|\nabla u_\epsilon|^2+C_1 tr_{\omega_\epsilon(t)} f^*\chi))\le-t|\nabla\nabla u_\epsilon|^2-t|\nabla\overline\nabla u_\epsilon|^2+C. \end{equation} on $(X\setminus D)\times[0,1]$, and \begin{equation}\label{P2.4.1} (\partial_t-\Delta_{\omega_\epsilon(t)})(|\nabla u_\epsilon|^2+C_1 tr_{\omega_\epsilon(t)} f^*\chi)\le-|\nabla\nabla u_\epsilon|^2-|\nabla\overline\nabla u_\epsilon|^2+C. \end{equation} on $(X\setminus D)\times[1,\infty)$. Next we try to bound $-\Delta_{\omega_\epsilon(t)}$ from above. Recall that \begin{equation}\label{P2.8.1} (\partial_t-\Delta_{\omega_\epsilon(t)})\Delta_{\omega_\epsilon(t)} u_\epsilon=\Delta_{\omega_\epsilon(t)} u_\epsilon+\langle Ric(\omega_\epsilon(t))-\frac{1}{T_{max}}\omega_0-A_\epsilon,\sqrt{-1}\partial\bar\partial u_\epsilon\rangle_{\omega_\epsilon}+\Delta_{\omega_\epsilon(t)} tr_{\omega_\epsilon(t)}f^*\chi. \end{equation} Using \begin{equation}\label{eq1} \sqrt{-1}\partial \bar\partial u_\epsilon=-Ric(\omega_\epsilon)-f^*\chi+\frac{1}{T_{max}}\omega_0+A_\epsilon, \end{equation} we see that the second term in \eqref{P2.8.1} \begin{align}\label{P2.9} \langle Ric(\omega_\epsilon(t))-\frac{1}{T_{max}}\omega_0-A_\epsilon,\sqrt{-1}\partial\bar\partial u_\epsilon\rangle_{\omega_\epsilon(t)}&=\langle-\sqrt{-1}\partial\bar\partial u_\epsilon-f^*\chi,\sqrt{-1}\partial\bar\partial u_\epsilon\rangle_{\omega_\epsilon(t)}\nonumber\\ &=-|\nabla\bar\nabla u_\epsilon|^2-\langle f^*\chi,\sqrt{-1}\partial\bar\partial u_\epsilon\rangle_{\omega_\epsilon(t)}\nonumber\\ &\ge-\frac{3}{2}|\nabla\bar\nabla u_\epsilon|^2-\frac{1}{2}|f^*\chi|^2\nonumber\\ &\ge-\frac{3}{2}|\nabla\bar\nabla u_\epsilon|^2-\frac{n}{2}(tr_{\omega_\epsilon(t)} f^*\chi)^2\nonumber\\ &\ge-\frac{3}{2}|\nabla\bar\nabla u_\epsilon|^2-C, \end{align} and the third term in \eqref{P2.8.1} \begin{align} \Delta_{\omega_\epsilon(t)} tr_{\omega_\epsilon(t)} f^*\chi&=-(\partial_t-\Delta_{\omega_\epsilon(t)})tr_{\omega_\epsilon(t)} f^*\chi+\partial_t tr_{\omega_\epsilon(t)} f^*\chi\nonumber\\ &\ge -C+C^{-1}_1|\nabla tr_{\omega_\epsilon(t)} f^*\chi|^2+\langle Ric(\omega_\epsilon(t))-\frac{1}{T_{max}}\omega_0-A_{\epsilon},f^*\chi\rangle_{\omega_\epsilon(t)}+tr_{\omega_\epsilon(t)} f^*\chi\nonumber\\ &\ge-C+\langle Ric(\omega_\epsilon(t))-\frac{1}{T_{max}}\omega_0-A_{\epsilon},f^*\chi\rangle_{\omega_\epsilon(t)}\nonumber\\ &=-C+\langle-\sqrt{-1}\partial\bar\partial u_\epsilon-f^*\chi,f^*\chi\rangle_{\omega_\epsilon(t)}\nonumber\\ &\ge-C-\frac{1}{2}|\nabla \overline\nabla u_\epsilon|^2-\frac{3}{2}|f^*\chi|^2\nonumber\\ &\ge-C-\frac{1}{2}|\nabla \overline\nabla u_\epsilon|^2. \end{align} Then we arrive at \begin{align} (\partial_t-\Delta_{\omega_\epsilon(t)})\Delta_{\omega_\epsilon(t)} u_\epsilon\ge\Delta_{\omega_\epsilon(t)} u_\epsilon-2|\nabla \overline\nabla u_\epsilon|^2-C. \end{align} Using \eqref{P2.4.0} and \eqref{P2.4.1}, we see \begin{align}\label{P2.10.0} &(\partial_t-\Delta_{\omega_\epsilon(t)})\left(t(\Delta_{\omega_\epsilon(t)} u_\epsilon-3(|\nabla u_\epsilon|^2+C_1 tr_{\omega_\epsilon(t)} f^*\chi))\right)\nonumber\\ &\ge (t+1)\Delta_{\omega_\epsilon(t)} u_\epsilon+t|\nabla\overline\nabla u_\epsilon|^2-C\nonumber\\ &\ge (t+1)\Delta_{\omega_\epsilon(t)} u_\epsilon+\frac{1}{n}t(\Delta_{\omega_\epsilon(t)} u_\epsilon)^2-C. \end{align} on $X\times[0,1]$, and \begin{align}\label{P2.10} &(\partial_t-\Delta_{\omega_\epsilon(t)})\left(\Delta_{\omega_\epsilon(t)} u_\epsilon-3(|\nabla u_\epsilon|^2+C_1 tr_{\omega_\epsilon(t)} f^*\chi)\right)\nonumber\\ &\ge\Delta_{\omega_\epsilon(t)} u_\epsilon+|\nabla\overline\nabla u_\epsilon|^2-C\nonumber\\ &\ge\Delta_{\omega_\epsilon(t)} u_\epsilon+\frac{1}{n}(\Delta_{\omega_\epsilon(t)} u_\epsilon)^2-C \end{align} on $X\times[1,\infty)$. To bound $\Delta_{\omega_\epsilon(t)}u_\epsilon$ from below (see \cite[Section 8]{Ed15} or \cite[Section 7]{No} for related discussions), we also try to avoid using a uniform bound for $(\Delta_{\omega_\epsilon(t)}u_\epsilon)(0)=\Delta_{\omega_\epsilon(t)}\log\frac{(\omega_{0,\epsilon}^*)^n}{(|S|_h^2+\epsilon^2)^{-(1-\beta)}\Omega}$. For convenience, we set $H_\epsilon:=t(\Delta_{\omega_\epsilon(t)} u_\epsilon-3(|\nabla u_\epsilon|^2+C_1 tr_{\omega_\epsilon(t)} f^*\chi))$, which by \eqref{P2.10.0} satisfies \begin{align}\label{P2.10.1} &(\partial_t-\Delta_{\omega_\epsilon(t)})\left(t(\Delta_{\omega_\epsilon(t)} u_\epsilon-3(|\nabla u_\epsilon|^2+C_1 tr_{\omega_\epsilon(t)} f^*\chi))\right)\nonumber\\ &\ge (t+1)\Delta_{\omega_\epsilon(t)} u_\epsilon+\frac{1}{n}t(\Delta_{\omega_\epsilon(t)} u_\epsilon)^2-C \end{align} $X\times[0,1]$. We assume the minimal value of $H_\epsilon$ on $X\times[0,1]$ is achieved at $(x',t')\in X\times[0,1]$. If $t'=0$, $H_\epsilon(x',t')=0$ is bounded from below; otherwise $t'>0$, then by maximum principle in \eqref{P2.10.1} we find, at $(x',t')$, $$t(\Delta_{\omega_\epsilon(t)}u_\epsilon)^2\le-(t+1)\Delta_{\omega_\epsilon(t)}u_\epsilon+C,$$ and so $$(t\Delta_{\omega_\epsilon(t)}u_\epsilon)^2\le-(t+1)(t\Delta_{\omega_\epsilon(t)}u_\epsilon)+Ct.$$ We may assume $(\Delta_{\omega_\epsilon(t)}u_\epsilon)(x',t')<0$ (otherwise we are done); then by noting that $t\in(0,1]$ we have $$(t\Delta_{\omega_\epsilon(t)}u_\epsilon)^2\le-2(t\Delta_{\omega_\epsilon(t)}u_\epsilon)+C,$$ from which we conclude that $$(t\Delta_{\omega_\epsilon(t)}u_\epsilon)(x',t')\ge -C.$$ Plugging into $H_\epsilon$ gives $$H_\epsilon(x',t')=\left(t(\Delta_{\omega_\epsilon(t)} u_\epsilon-3(|\nabla u_\epsilon|^2+C_1 tr_{\omega_\epsilon(t)} f^*\chi))\right)(x',t')\ge-C.$$ Therefore, we can choose a constant $C\ge1$ independent on $\epsilon$ such that $$\inf_{X\times[0,1]}H_\epsilon\ge-C,$$ which, combining with \eqref{upper11.1}, implies \begin{equation}\label{lower3.1} \inf_{X\times[0,1]}(t\Delta_{\omega_\epsilon(t)}u_\epsilon)\ge-C. \end{equation} In particular, \begin{equation}\label{lower4} \inf_X(\Delta_{\omega_\epsilon(t)} u_\epsilon-3(|\nabla u_\epsilon|^2+C_1 tr_{\omega_\epsilon(t)} f^*\chi))(1)\ge -C. \end{equation} Having \eqref{lower4}, we can easily apply the maximum principle in \eqref{P2.10} to find a constant $C\ge1$ independent on $\epsilon$ such that \begin{equation}\label{lower5} \inf_{X\times[1,\infty)}\Delta_{\omega_\epsilon(t)}u_\epsilon\ge-C. \end{equation} Combining \eqref{lower3.1} and \eqref{lower5} gives, for any $t>0$, \begin{equation}\label{lower5.1} \inf_{X}\Delta_{\omega_\epsilon(t)}u_\epsilon\ge-\frac{C}{\min\{t,1\}}. \end{equation} Finally, recall from \eqref{eq1} that \begin{equation} \tilde R_\epsilon=-tr_{\omega_\epsilon(t)}f^*\chi-\Delta_{\omega_\epsilon(t)}u_\epsilon. \end{equation} Therefore, by \eqref{lower5.1} there exists a constant $C\ge1$ independent on $\epsilon$ such that for any $t>0$ \begin{equation}\label{lower6} \sup_{X}\tilde R_\epsilon\le \frac{C}{\min\{t,1\}}. \end{equation} Combining \eqref{lower} and \eqref{lower6} gives a constant $C\ge1$ independent on $\epsilon$ such that for any $t>0$, \begin{equation}\label{scal1} \sup_{X}|\tilde R_\epsilon|\le \frac{C}{\min\{t,1\}}. \end{equation} By letting $\epsilon\to0$, Lemma \ref{scal} is proved. \end{proof} \subsection{Proof of Theorem \ref{scal_thm}}\label{scal.1} In this subsection, we explain how to apply the above arguments for Lemma \ref{scal} to prove Theorem \ref{scal_thm}. \begin{proof}[Proof of Theorem \ref{scal_thm}] It suffices to prove the conclusion in Theorem \ref{scal_thm} for the following normalized version of \eqref{CKRF}: \begin{equation}\label{NCKRF.1} \left\{ \begin{aligned} \partial_t\omega(t)&=-Ric(\omega(t))-\omega(t)+2\pi(1-\beta)[D]\\ \omega(0)&=\omega_0^*, \end{aligned} \right. \end{equation} here, without loss of any generality we assume there is only one irreducible component $D$ in the cone divisor. We have the smooth approximation \eqref{TCKRF.1} and \eqref{TCCMA.1} as before. For the fixed $t_0\in(0,T_{max})$, we fix a K\"ahler metric $\hat\omega_{t_0}\in e^{-t_0}[\omega_0]+(1-e^{-t_0})2\pi(c_1(K_X)+(1-\beta)[D])$. Also note that $\varphi_\epsilon$ and $\partial_t\varphi_\epsilon$ are uniformly bounded on $X\times[0,t_0]$ (see \cite{Sh,Wy}). In the above arguments for Lemma \ref{scal}, we replace $Y$ by $X$, $f$ by the identity map $X\to X$ and $\chi$ by $\hat\omega_{t_0}$, and then those arguments apply and Theorem \ref{scal_thm} follows. \end{proof} \section{Local $C^0$-convergence away from cone divisor}\label{sect_C0} In this section, we will prove $C^0$-convergence away from cone divisor $D$. The strategy used here is taken from Tosatti-Weinkove-Yang \cite{TWY}. \par In this section, similar to discussions in Sections \ref{weak_conv} and \ref{equiv}, we will directly work with the conical equations without passing to a smooth approximation. Let's begin with the following \begin{lem}\label{prop_conv.1} There exists a constant $C\ge1$ such that for any $t\in[1,\infty)$, \begin{equation}\label{conv} \sup_{X\setminus D}|\partial_t\varphi+\varphi(t)+\delta|S|^{2\beta}_h-f^*\psi|\le Ce^{-\frac{1}{4}t}. \end{equation} \end{lem} \begin{proof} Recall Proposition \ref{prop_conv}: there exists a contant $\tilde C$ such that for any $t\in[0,\infty)$ \begin{equation}\label{local0} \sup_{X\setminus D}|\varphi+\delta|S|^{2\beta}_h-f^*\psi|\le \tilde Ce^{-\frac{1}{4}t}. \end{equation} Also recall \begin{equation} \partial_t(\partial_t\varphi)=\partial_t\varphi-\tilde R-k. \end{equation}\label{local2} So, by Lemma \ref{scal} we can fix a constant $C_0\ge1$ such that $\sup_{(X\setminus D)\times[1,\infty)}|\partial_t(\partial_t\varphi)|\le C_0$. We set $C_1:=\sqrt{8\tilde CC_0+1}$. It suffices to show \begin{equation}\label{local2.1} \sup_{X\setminus D}|\partial_t\varphi|\le C_1e^{-\frac{1}{4}t}. \end{equation} To this end, we claim that\\ \textbf{Claim} For any $\lambda\in(0,\lambda_0]$ and $t\in[1,\infty)$, \begin{equation}\label{local3} \sup_{X\setminus D}(\partial_t\varphi+\lambda\log|S|^2_h)\le C_1e^{-\frac{1}{4}t}. \end{equation} \begin{proof}[Proof of Claim] We make use of an argument in \cite{TWY} to conclude. Suppose there exists a sequence $(x_k,t_k)\in (X\setminus D)\times[1,\infty)$ with $t_k\to\infty$ as $k\to\infty$ such that $$(\partial_t\varphi+\lambda\log|S|^2_h)(x_k,t_k)\ge C_1e^{-\frac{1}{4}t_k},$$ which in particular implies that $x_k$ in fact will be contained in some fixed $K\subset\subset X\setminus D$, since $\partial_t\varphi$ is uniformly bounded. Set $\theta_k:=\frac{k}{2C_0}e^{-\frac{1}{4}t_k}$, where $C_0$ is a constant such that $\sup_{(X\setminus D)\times[1,\infty)}|\partial_t(\partial_t\varphi)|\le C_0$ defined as before. Then, $(\partial_t\varphi+\lambda\log|S|^2_h)(x_k,t)\ge\frac{C_1}{2}e^{-\frac{1}{4}t_k}$ for all $t\in[t_k,t_k+\theta_k]$. On the one hand we have \begin{align}\label{local4} &(\varphi+\delta|S|^{2\beta}_h-f^*\psi)(x_k,t_k+\theta_k)-(\varphi+\delta|S|^{2\beta}_h-f^*\psi)(x_k,t_k)\nonumber\\ &\le\sup_{X\setminus D}|\varphi+\delta|S|^{2\beta}_h-f^*\psi|(t_k+\theta_k)+\sup_{X\setminus D}|\varphi+\delta|S|^{2\beta}_h-f^*\psi|(t_k)\nonumber\\ & \le \tilde C(e^{-\frac{3}{4}(t_k+\theta_k)}+e^{-\frac{3}{4}t_k});\nonumber\\ &\le 2\tilde Ce^{-\frac{3}{4}t_k}. \end{align} On the other hand, \begin{align}\label{local5} &(\varphi+\delta|S|^{2\beta}_h-f^*\psi)(x_k,t_k+\theta_k)-(\varphi+\delta|S|^{2\beta}_h-f^*\psi)(x_k,t_k)\nonumber\\ &\ge\int_{t_k}^{t_k+\theta_k}(\partial_t\varphi)(x_k,t)dt\nonumber\\ &\ge\int_{t_k}^{t_k+\theta_k}(\partial_t\varphi+\lambda\log|S|^2_h)(x_k,t)dt\nonumber\\ &\ge\theta_k\cdot \frac{C_1}{2}e^{-\frac{1}{4}t_k}\nonumber\\ &=\frac{C_1^2}{4C_0}e^{-\frac{1}{2}t_k}\nonumber\\ &\ge\frac{C_1^2}{4C_0}e^{-\frac{3}{4}t_k}. \end{align} By combining with \eqref{local4} and \eqref{local5} we get a contradiction: $$8\tilde CC_0+1=C^2_1\le 8\tilde CC_0.$$ Therefore, Claim is proved. \end{proof} Let $\lambda\to0$ in \eqref{local3}, we get, for all $t\in[1,\infty)$, \begin{equation}\label{local6} \sup_{X\setminus D}\partial_t\varphi\le C_1e^{-\frac{1}{4}t}. \end{equation} Similarly, we can get, for all $t\in[1,\infty)$, \begin{equation}\label{local7} \inf_{X\setminus D}\partial_t\varphi\ge -C_1e^{-\frac{1}{4}t}. \end{equation} Having \eqref{local6} and \eqref{local7}, \eqref{local2.1} follows and Lemma \ref{prop_conv.1} is proved. \end{proof} \begin{lem}\label{local5.1} There exist two positive constants $C,\gamma$ such that for any $t\in[1,\infty)$, \begin{equation}\label{local6} \sup_{X\setminus D}(|S|^{2\gamma}_h(tr_{\omega(t)}f^*\overline\chi-k))\le Ce^{-\frac{1}{8}t}. \end{equation} \end{lem} \begin{proof} Recall we have $$(\partial_t-\Delta_{\omega(t)})(\partial_t\varphi+\varphi+\delta|S|^{2\beta}_h-f^*\psi)=tr_{\omega(t)}f^*\overline\chi-k$$ and, for some $\gamma>0$, $$(\partial_t-\Delta_{\omega(t)})tr_{\omega(t)}\overline\chi\le tr_{\omega(t)}\overline\chi\le+|S|^{2\gamma}_h(tr_{\omega(t)}\overline\chi)^2\le C|S|^{2\gamma}_h,$$ where $|S|^{2\gamma}_h$ is an uniform upper bound for bisectional curvature of $\overline\chi$ on $Y\setminus D'$ and we have used $tr_{\omega(t)}f^*\overline\chi$ is uniformly bounded by Lemma \ref{C2.1} (note that $\overline\chi$ and $\chi^*$ is uniformly equivalent). Then one can apply arguments in \cite[Lemma 3.4]{TWY} to conclude \eqref{local6}. \par Lemma \ref{local5.1} is proved. \end{proof} \begin{lem} For any given $K'\subset\subset Y\setminus D'$, and $l\in \mathbb Z_{\ge1}$, there exists a constant $C=C_{K,l}\ge1$ such that for any $y\in K'$ and $t\in[1,\infty)$, \begin{equation}\label{local8} |e^t\omega(t)|_{X_y}|_{C^l(X_y,\omega_0|_{X_y})}\le C_{K,l}. \end{equation} \end{lem} \begin{proof} This lemma can be checked by the same arguments in \cite[Theorem 1.1]{ToZyg}, as we have Proposition \ref{prop2}. To see that arguments in \cite[page 2933-2934]{ToZyg} work in our twisted case, it suffices to observe that, after pulling back the twisted term $\frac{1}{T_{max}}\omega_0$ by the "stretching map" $F_k$ defined in \cite[page 2933]{ToZyg}, we get a smooth $(1,1)$-form $\frac{1}{T_{max}}F^*_k\omega_0$, whose $C^1$-norm with respect to a local Euclidean metric is uniformly bounded. Then we can get the desired conclusions. \end{proof} With all the above preparations, we now conclude the main result in this section. \begin{prop}\label{prop3} There exists a positive constant $\varepsilon_0$ such that, for any $K\subset\subset X\setminus D$ there exists a constant $C_K\ge1$ such that, for any $t\in[1,\infty)$, \begin{equation} |\omega(t)-f^*\overline\chi|_{C^0(K,\omega_0)}\le C_Ke^{-\varepsilon_0t}. \end{equation} \end{prop} \begin{proof} Having the above results, one can first prove $$|e^t\omega(t)|_{X_y}-\overline\omega_0|_{X_y}|_{C^0(X_y,\omega_0|_{X_y})}\le C_Ke^{-\varepsilon t}$$ for any $y\in K$, and then $$|\omega(t)-(e^{-t}\overline\omega_0+f^*\overline\chi)|_{C^0(K,\omega_0)}\le C_Ke^{-\varepsilon_0t}$$ by the same arguments in \cite[Section 2.5]{TWY}. We omit details here. Therefore, Proposition \ref{prop3} follows. \end{proof} \section{Gromov-Hausdorff convergence}\label{sect_gh} We are going to prove item (2.3) in Theorem \ref{main_thm}, i.e. Gromov-Hausdorff convergence. \par Throughout this section, we will always assume $dim(Y)=1$ and so $D'$ is a single point $o\in Y$. To obtain Gromov-Hausdorff convergence, the key point is to apply Proposition \ref{prop2} to bound diameter of the cone divisor $X_o$. We set $B'_\delta:=\{y\in Y|d_{\chi}(y,o)<\delta\}$ and $B_\delta:=f^{-1}(B_\delta)$. Assume without loss of generality that $\delta$ is always small enough such that $B_\delta$ is the standard disc in $\mathbb C$ and $o=0\in\mathbb C$. Fix an integer $L\ge1$ such that \begin{equation} diam(B_{\varepsilon^L},\bar d)\le \frac{\varepsilon}{4}. \end{equation} The key observation is the following \begin{lem}\label{gh0} There exists a constant $C\ge1$ such that, for any $\varepsilon>0$, there exists a positive constant $T=T_\varepsilon$ such that for all $t\ge T$, \begin{equation} diam(B_{\varepsilon^L},d_t)\le C\varepsilon \end{equation} \end{lem} \begin{proof} The proof will make use of Proposition \ref{prop2} crucially. Firstly, since $X_o$ is compact, we choose a family of local charts $\{U_i,(y,z^1,...,z^{n-1})\}_{i=1}^{I}$, such that the chart $U_i$ is centered at some point $x_i\in X_o$, $B_{\varepsilon^L}\subset \cup_{i=1}^IU_i$ and $f(y,z^1,...,z^{n-1})=y$. For any two points $x',x''\in B_{\varepsilon^L}\setminus X_o$, we may assume $x'=(y',z')\in U_1$ and $x''=(y'',z'')\in U_2$. Then when we restrict to the subset $Y_{z'}:=\{(y,z')\in U_1|y\in B'_{\varepsilon^L}\}$, we know from Proposition \ref{prop2} that $\omega(t)|_{Y_{z'}}$ is a conical metric on $Y_{z'}$ with cone angle $2\pi\beta$ at $(0,z')$. Therefore, we can easily choose a point $\bar y'\in \partial B'_{\varepsilon^L}$, and set $\bar x':=(\bar y',z')\in Y_{z'}$, such that \begin{equation}\label{gh1} d_t(x',\bar x')\le \hat C\varepsilon, \end{equation} here $\hat C$ is a positive constant only depends on the constant $C$ in Proposition \ref{prop2}, \eqref{C2.3} (in particular, $\hat C$ does not depend on $x',x''$ and $t$). Similarly, one can find a point $\bar x'':=(\bar y'',z'')$ with $\bar y''\in\partial B'_{\varepsilon^L}$ such that \begin{equation}\label{gh2} d_t(x'',\bar x'')\le \hat C\varepsilon, \end{equation} On the other hand, since $\overline\chi$ is a conical K\"ahler metric, the diameters of fibers $X_y$ away from $X_o$ uniformly collapse in the rate $e^{-\frac{1}{2}t}$ by Proposition \ref{prop2} and $\omega(t)\to f^*\overline\chi$ in $C^0(f^{-1}(Y\setminus B'_{\frac{1}{2}\varepsilon^L},\omega_0)$-topology by Proposition \ref{prop3}, one can use arguments in \cite[Section 3]{ZyZz1} to connect $\bar x',\bar x''$ by a piecewise smooth curve $\sigma\subset f^{-1}(\partial B'_{\varepsilon^L})$ such that \begin{equation}\label{gh3} L_{d_t}(\sigma)\le \varepsilon+\hat Ce^{-\frac{1}{2}t}. \end{equation} In conclusion, by combining \eqref{gh1}, \eqref{gh2} and \eqref{gh3}, we can choose a positive constant $T$ (independent on $x',x''$) such that for any $t\ge T$, \begin{equation} d_t(x',x'')\le(\hat C+2)\varepsilon. \end{equation} \par Lemma \ref{gh0} is proved. \end{proof} \begin{proof}[Proof of Theorem \ref{main_thm} (2.3)] Having Lemma \ref{gh0}, we can easily apply arguments in \cite[Section 3]{ZyZz1} (also see \cite[Section 4]{Zy17}) to conclude Theorem \ref{main_thm} (2.3). \end{proof} \section{Remarks on the twisted K\"ahler-Ricci flow}\label{sect_rem} The last section provides more remarks on the twisted K\"ahler-Ricci flow we have studied. Here we remove the conical singularity. Given a projective manifold $X$ with a rational K\"ahler metric $\omega_0\in2\pi c_1(H)$ ($H$ is some ample line bundle on $X$), we have the smooth K\"ahler-Ricci flow starting from $\omega_0$ on $X$: \begin{equation}\label{KRF} \left\{ \begin{aligned} \partial_t\omega(t)&=-Ric(\omega(t))\\ \omega(0)&=\omega_0, \end{aligned} \right. \end{equation} By \cite{C,Ts,TZo}, the K\"ahler-Ricci flow \eqref{KRF} has a smooth solution up to $$T_{max}:=\{t>0|[\omega_0]+t2\pi c_1(K_X)>0\}=\{t>0|H+tK_X>0\}.$$ Then $T_{max}=\infty$ if and only if $X$ is a smooth minimal model. \par Assume $T_{max}<\infty$. Then $T\in\mathbb Q$ and $H+T_{max}K_X$ is semi-ample (see e.g. \cite{Ma}). By semi-ample fibration theorem there exists a holomorphic map \begin{equation}\label{map} f:X\to f(X)\subset\mathbb{CP}^N, \end{equation} with the image $Y:=f(X)$ an irreducible normal projective variety of dimension $0\le k\le n$, connected fibers and \begin{equation}\label{eqn1} f^*\mathcal O_{\mathbb{CP}^N}(1)=H+T_{max}K_X. \end{equation} Let's first recall some basic ideas of Song and Tian \cite[Section 6.2]{ST17} on contracting or collapsing certain positive part of $c_1(X)$ by the K\"ahler-Ricci flow, which indicates a deep relation between finite-time singularity of the K\"ahler-Ricci flow and the Minimal Model Program in algebraic geometry. \begin{itemize} \item [(1)] If $k=0$, then $H=-T_{max}K_X$, i.e. $X$ is a Fano manifold with $\omega_0\in T_{max}c_1(X)$. The K\"ahler-Ricci flow in this case has been studied by many works, see e.g. discussions next to \cite[Conjecture 6.6]{ST17} and references therein for more comments. Here we will focus on the $k>0$ case. \item [(2)] If $1\le k\le n$, then it is conjectured in \cite{ST17} that, as $t\to T_{max}$, $(X,\omega(t))\to(Y,d_Y)$ in Gromov-Hausdorff topology, here $d_Y$ is some compact metric on $Y$ and should be induced by a smooth K\"ahler metric on $Y$ outside a subvariety. \end{itemize} Roughly speaking, \eqref{eqn1} means there are certain "positive parts" in $c_1(X)$ (and so $X$ is not a minimal model), and the picture in the above item (2) means one can use the K\"ahler-Ricci flow to contract/collapse these "positive parts" in Gromov-Hausdorff topology and arrive at a new space, which is closer to a minimal model in some sense (see \cite[Section 6.2]{ST17} for more precise discriptions). There are several progresses on above item (2), see e.g. \cite{Fo,FuZs,SSW,SW,ToZyg1}. However, it seems in general the finite time convergence of the K\"ahler-Ricci flow are still unclear. \par La Nave and Tian \cite{LT} proposed a continuity method approach to achieve the above picture, see \cite{LT,LTZ,ZyZz1,ZyZz2,Zy17} for more discussions and results on this direction.\\ \par Here we would like to discuss the possibility of deforming $X$ to $Y$ (given $f:X\to Y$ as in above \eqref{map}) by using a twisted K\"ahler-Ricci flow. Given the above setting, we consider, for an arbitrary K\"ahler metric $\omega_X$, \begin{equation}\label{TKRF} \left\{ \begin{aligned} \partial_t\omega(t)&=-Ric(\omega(t))-\omega(t)-\frac{1}{T_{max}}\omega_0\\ \omega(0)&=\omega_X, \end{aligned} \right. \end{equation} If we fix a K\"ahler metric $\chi_Y$ on $Y$ with $f^*\chi_Y\in\frac{1}{T_{max}}[\omega_0]+2\pi c_1(K_X)$, then we easily see that the K\"ahler class along \eqref{TKRF} satisfies $$[\omega(t)]=e^{-t}[\omega_X]+(1-e^{-t})[f^*\chi],$$ which stays positive for any $t\in[0,\infty)$. Therefore, \eqref{TKRF} has a smooth long time solution $\omega(t)$ on $X\times[0,\infty)$. Then one can prove $\omega(t)$ converges to $f^*\overline\chi_Y$ as currents on $X$, where $\overline\chi_Y$ is a K\"ahler current solved by a possibly singular complex Monge-Amp\`ere equation on $Y$. Precisely, we set $V$ be the singular set of $Y$ together with critical values of $f$ and fix a $(1,1)$-form $\overline\omega_X\in[\omega_X]$ on $X\setminus f^{-1}(V)$ such that $\overline\omega_X|_{X_y}$ is a K\"ahler metric on $X_y$ and $Ric(\overline\omega_X|_{X_y})=\frac{1}{T_{max}}\omega_0|_{X_y}$ for all $y\in Y\setminus V$. Also fix a smooth volume form $\Omega$ on $X$ with $\sqrt{-1}\partial\bar\partial\log\Omega=f^*\chi_Y-\frac{1}{T_{max}}\omega_0$. Then $\overline\chi_Y:=\chi_Y+\sqrt{-1}\partial\bar\partial\psi$ is solved by $$(\chi_Y+\sqrt{-1}\partial\bar\partial\psi)^k=e^{\psi}\frac{\Omega}{C^n_k(\overline\omega_X)^{n-k}\wedge f^*\chi_Y^k}\chi_Y^k.$$ Since $\chi_Y$ is rational and $0<C^{-1}\le\frac{\Omega}{C^n_k(\overline\omega_X)^{n-k}\wedge f^*\chi_Y^k}\in L^{1+\epsilon}(Y,\chi_Y^k)$ (see \cite[Proposition 3.2]{ST12}), we can find a unique $\psi\in PSH(Y,\chi_Y)\cap L^\infty(Y)\cap C^\infty(Y\setminus V)$ solving the above equation (see \cite[Section 3.2, Theorem 3.2]{ST12}). \begin{itemize} \item[(a)] Case $0<k<n$. If additionally $Y$ is smooth and $f:X\to Y$ is a submersion, then $\overline\chi_Y$ is a K\"ahler metric on $Y$ and $\omega(t)\to f^*\overline\chi_Y$ in $C^0(X,\omega_0)$-topology exponentially fast (see Theorem \ref{main_thm} (2.2)) and hence $(X,\omega(t))\to(Y,\overline\chi_Y)$ in Gromov-Hausdorff topology. This provides an alternative way (using the twisted K\"ahler-Ricci flow) to deform a Fano bundle to the base in Gromov-Hausdorff topology (compare with the works in \cite{Fo,FuZs,SSW}). \item[(b)] Case $k=n$. In this case $\overline\chi_Y$ is in fact solved by $$(f^*\chi_Y+\sqrt{-1}\partial\bar\partial\psi)^n=e^{\psi}\Omega$$ on $X$. Then $\overline\chi_Y$ is a smooth K\"ahler metric on $X\setminus f^{-1}(V)$ and $\omega(t)\to\overline\chi_Y$ smoothly on $X\setminus V$. In fact, it can be checked that $\overline\chi_Y$ coincides with the limit of the continuity method studied in \cite[Theorems 1.1, 1.2]{LTZ}, and so the metric completion of $(X\setminus f^{-1}(V),\overline\chi_Y)$ is a compact metric space homeomorphic to $Y$ by \cite[Theorem 1.2]{LTZ}; denote this limit space by $(Y,d_Y)$. On the other hand, since the twisted term $\frac{1}{T_{max}}\omega_0$ is nonnegative, it is very likely that one can extend a generalized Perelman's no-local-collapsing theorem of Wang \cite[Theorem 1.1]{W} or Q. Zhang \cite[Theorem 6.3.2]{Zq} to our twisted setting. Also note that we have a uniform bound for the twisted scalar curvature $tr_{\omega(t)}(Ric(\omega(t))-\frac{1}{T_{max}}\omega_0)$, see Section \ref{sect_scal}. Therefore, by Wang's argument in \cite[Theorem 8.2]{W}, we may obtain a uniform diameter upper bound along the twisted K\"ahler-Ricci flow \eqref{TKRF} in this volume noncollapsing case (i.e. $k=n$). The remaining question in this case is: \emph{can we prove $(X,\omega(t))\to(Y,d_Y)$ in Gromov-Hausdorff topology?} We will study this question in the future work. \end{itemize} \section*{Acknowledgements} The author thanks Prof. Huai-Dong Cao, Prof. Gang Tian, Prof. Zhenlei Zhang and Dr. Shaochuang Huang for useful discussions, Dr. Jiawei Liu for valuable comments and the referee for careful reading and valuable suggestions and comments. Part of this work was carried out while the author was visiting University of Macau and Capital Normal University, which he would like to thank for the hospitality.

\section{Introduction} A wider picture of the memory function formalism is presented. After sketching the standard derivation of the Drude formula of electrical conductivity from momentum relaxation equation, we introduce the Langevin equation for electrical conduction in metals. A derivation of the Drude formula from the Langevin equation is presented. By pointing out a fundamental problem with the Langevin equation, the problem of Drude formula is highlighted. Then generalised Langevin equation is introduced which rectifies the problem of the Langevin equation by incorporating the time dependent friction coefficient. Then generalised Drude formula is obtained from generalised Langevin equation in which constant Drude scattering rate is replaced by a frequency dependent scattering rate, which is populary known as the memory function.\\ Then an alternative route to memory function formalism or generalised Drude formalism is sketched based on the Zwanzig-Mori projection opreator method. A formal expression for the memory function in terms of the projection operator is given. We then introduce the G\"{o}tze-W\"{o}lfle perturbation expansion of the memory function and review their formalism. Special cases of electron-impurity and electron-phonon interaction are revised in detail. We come back to the question of under what conditions frequency dependent Drude scattering rate can be taken as frequency independent parameter in standard Drude formula, and discuss the relevant conditions. Our overview provides a wider picture of this field of research, and nothing new is presented here. An attempt is made to put this beautiful field of research in a nutshell. We hope that this review will be useful for pedagogical purposes. \section{The Drude Model} The traditional way to introduce AC conductivity of a metal is to use the Drude formula. In the Drude model electrons in a metal are treated as classical particles (much like gas molecules) bumping with the stationary ion cores and constitute a random thermal motion. When an external electric field is applied, electrons are accelerated and gain momentum. The momentum gained from the field is dissipated due to the collisions with ion cores and they attain an average drift speed. Due to this finite drift speed an electric current is established in the sample. When external field is removed the drift speed vanishes and electrons go back to the thermally agitated motion. Thus under the action of applied field the equation of motion of electrons can be written as: \begin{equation} m\frac{d^{2}\textbf{r}}{dt^{2}}=-e\textbf{E}(t)-m \frac{1}{\tau}\frac{d\textbf{r}}{dt}, \label{motion} \end{equation} which is nothing but Newton's second law of motion including dissipation. The momentum of an electron is being degraded due to collision processes with the ion cores at the average rate $\frac{1}{\tau}$. Considering the electric field $\textbf{E}(t)$ variation as $\textbf{E}_0\cos(\om t)$ and solving the equation of motion for momentum leads to \beq \textbf{p}(t) = \underbrace{-e\frac{\tau}{1+ \tau^2 \om^2} \textbf{E}_0 \cos(\om t)}_{In~phase~response~(dissipative)} - \underbrace{e \frac{\om \tau^2}{1+ \tau^2\om^2} \textbf{E}_0 \sin(\om t)}_{Out-of-phase~response~(reactive).} \label{mom} \eeq One can observe that the time dependent momentum consitutes two responses: (1) the in-phase response leading to dissipation that varies as $\cos(\om t)$, and (2) out-of-phase response which has $\sin(\om t)$ dependence. From the above equation of momentum, the current density is given by: \beq \textbf{j}(t) = - n e \frac{\textbf{p}(t)}{m} = \frac{n e^2 }{m}\bigg(\frac{\tau}{1+ \tau^2 \om^2} \textbf{E}_0 \cos(\om t) +\frac{\om \tau^2}{1+ \tau^2\om^2} \textbf{E}_0 \sin(\om t)\bigg). \label{mom1} \eeq The presence of current in the sample leads to dissipation of energy (Joule heating). The average heat dissipation is given as $\frac{\om}{2\pi}\int_0^{\frac{2\pi}{\om}} dt \textbf{j}(t).\textbf{E}(t)$. Clearly, this Joule heating is only due to the dissipative component, and it vanishes due to the reactive components (average of $\sin(\om t)\cos(\om t)$ over a period is zero). From equation (\ref{mom1}) the dissipative part of the conductivity ($\vec{J}=\sigma \vec{E})$ reduces to \beq \sigma_1(\om) = \sigma_{DC} \frac{1}{1+ \tau^2 \om^2}. \label{mom2} \eeq On the other hand the reactive part of the conductivity becomes \beq \sigma_2(\om) = \sigma_{DC}\frac{\tau\om}{1+ \tau^2 \om^2}. \eeq In the following paragraph we will observe that Drude expression is valid when\\ 1. Frequency ${\om}$ is very small as compared to the characterstics enegy scale in system (as discussed in section 6.2). \\ 2. Scattering with ion cores is the dominant mechanism of momentum randomisation.\\ Below we will derive the Drude formula using a different approach and we will observe that the scattering rate $\frac{1}{\tau}$ ceases to be a constant, rather it depends on frequency and temperature. \section{The Langevin equation} In the present section we introduce an important subject of non-equilibrium statistical mechanics i.e., the Langevin equation, which provides the basis for the memory function(MF) formalism. Our purpose is to write ac conductivity (in a general case) in a form resembling the Drude formula derived from the Langevin equation. Such a representation can be obtained using the MF formalism which will be our subject in the next sections. To understand memory function formalism we have to first consider a phenomenological approach. Let us start with the simple Langevin equation. The Langevin equation describe the random motion of a Brownian particle. The Brownian particle can be a pollen grain in some fluid. The origin of the random motion of Brownian particles is due to the irregular bombardment of the particle by molecules of fluid. It can also be viewed from the density fluctuations of fluid at a spatial length scale of the Brownian particle size. The molecules of fluid also provide a drag force to Brownian particle along with the continuous random force. Thus, net force on the Brownian particles can be divided into two components (1) systematic drag force, and (2) random force. With this, equation of motion can be written as \beq m \dot{u}(t) = \underbrace{- m\gamma u(t)}_{systematic ~~part} + \underbrace{R(t)}_{random~~part.} \label{lan1} \eeq This is called the Langevin equation. This equation holds great significance not only to describe the motion of a Brownian particle, but its wide regime of applicability is beautifully expressed in \cite{toda} ``Brownian motion is not merely random motion of a very fine particle; in general it is random motion of a physical quantity to be observed in a macrosystem" \vspace{5mm} Let us apply the Langevin equation to electrons in a metal which is biased by an external ac field. The presence of impurity potentials and phonons\footnote{ The Johnson-Nyquist noise is a direct manifestation of this random motion.} force electrons to move randomly in the metal. On applying external ac field ($E_0 \cos(\om_0 t)$) the Langevin equation for an electron changes to \beq m \dot{u}(t) = - m \gamma u(t) + R(t) + E_0 \cos(\om_0 t). \label{lan2} \eeq Here $m$ is the mass of an electron and $\gamma$ is the friction coefficient. Write the above equation in the Laplace domain with $u(s) = \int_0^\infty dt u(t) e^{- s t}$ \beq - m u(0) + s m u(s) = -m \gamma u(s) + R(s) + \frac{E_0}{2} \bigg(\frac{1}{ s - i \om_0} + \frac{1}{s+i\om_0}\bigg).\label{lan(3)} \eeq On simplify it and performing the ensemble average we obtain \beq \la u(s)\ra = \frac{\la u(0)\ra}{s+\gamma} + \frac{E_0}{2m (s+\gamma)}\bigg(\frac{1}{ s - i \om_0} + \frac{1}{s+i\om_0}\bigg).\label{lan(4)} \eeq The ensemble average of the random force $\la R(t) \ra = 0$. The above equation in the long time limit $t>>\frac{1}{\gamma}$ can be written in the form \beq \la u(t)\ra = Re\{ \mu(\om_0) E_0 e^{i \om_0 t}\},\label{lan(5)} \eeq where \beq \mu(\om) = \frac{1}{m}\frac{1}{i\om +\gamma} \label{lan(6)} \eeq is the expression for the electronic mobility. In the electrical conduction problem the terminal speed along the $x-$direction (say) gained by an electron due to the applied electric field is $u_t^x = -\mu (e E_x)$ (terminal speed = mobility $\times$ force i.e., Fick's Law). The induced current density can be written in terms of mobility as $J_x = -n e u_t^x = n e^2 \mu E$. Accrding to Ohm's Law ${J_x}=\sigma_{x x} { E_x},$ thus $\sigma_{x x}=\sigma=n e^2\mu $. From equation (\ref{lan(6)}) of the mobility, the conductivity takes the form. \beq \sigma(\om) = \frac{ne^2}{m} \frac{1}{i\om +\gamma}. \label{lan(7)} \eeq The dc conductivity $\sigma(0) = n e^2/m\gamma$. Here ${\gamma}$ is to be identified with the Drude scattering rate $\frac{1}{\tau}$. This is the Drude formula that we derived in the previous section. From this derivation of the Drude formula an important insight can be gained. It will be shown in the following section that Langevin equation has a serious problem in a general setting. An understanding of the problem the Langevin equation will enable us to understand the problem of the Drude formula. \subsection{Problems of the Langevin equation} The Langevin equation (\ref{lan1}) can be integrated and written as: \beq u(t) = u(t_0)e^{-\gamma(t-t_0)} + \frac{1}{m}\int_{t_0}^t dt^\p e^{-\gamma (t-t^\p)} R(t^\p). \label{e3p109} \eeq Multiply the above equation with $u(t_{0})$ and shift t to $t_{0}+t$ in the above equation. Then ensemble average becomes \beq \la u(t_0)u(t_0 + t)\ra = \la u(t_0)^2\ra e^{-\gamma t} \eeq where we have used $ \la u(t_0)\ra \la R(t^\p)\ra = 0$ . Further taking the time derivative of the above correlation function leads to \beq \frac{d}{dt} \la u(t_0) u(t_0+ t)\ra = -\gamma \la u(t_0)^2\ra e^{-\gamma t}. \eeq In the limit $t\rta 0$ we have \beq \la u(t_0) \dot{u}(t_0)\ra = -\gamma \la u(t_0)^2\ra \neq 0. \label{statio1} \eeq Therefore, the time derivative with respect to the intial time$(t_{0})$ of the velocity-velocity correlation function is not zero. It implies that according to the Langevin equation, the velocity correlation function depends on the initial time. But this is contrary to the requirement of stationarity in an equilibrium setting: \beq \frac{d}{dt_0} \la u(t_0) u(t_0+ t)\ra =0, \eeq i.e., the correlation function should not depend on the initial time $t_0$. It only dependent on the difference of time arguments. Simplifying the above equation by differentiating, and then taking the limit $t\rightarrow 0$ we get \beq \la u(t_0) \dot{u}(t_0)\ra = 0. \eeq Clearly, this is in contradiction to equation (\ref{statio1}). Thus there is serious problem with the Langevin equation. From here it can be concluded that the Drude formula derived from the Langevin equation may give wrong results. It turns out that this conclusion is correct and the Drude scattering rate which is assumed to be constant is not constant in general setting and actually becomes a frequency dependent quantity. The following section points out this problem with the Drude formula. \section{The generalized Langevin equation and the memory function (time dependent friction coefficient)} \label{s3p6} From the previous section we notice that the standard Langevin equation is inconsistent with the stationarity condition. It turns out that this problem can be resolved if the friction coefficient in the Langevin equation is taken as time dependent. In other words constant friction rate ($-m\gamma$) in the standard Langevin equation is to be replaced by time dependent friction coefficient $-m\gamma(t).$ This is equilvalent to considering the colllision history of the particle as will be made clear below. The modified Langevin equation which takes time dependent friction coefficient into account is called the generalised Langevin equation: \begin{equation} m \dot{u}(t) = \underbrace{- m \int_{-\infty}^t dt^\p M(t-t^\p) u(t^\p)}_{systematic~part~with~memory} + \underbrace{R(t)}_{random~~part} + \underbrace{E(t)}_{external~drive}. \label{glan} \end{equation} Here a time dependent friction coefficient $M(t)$ is introduced. We have phenomenologically introduced the generalized Langevin equation(GLE), however, it is well know fact in the non-equilibrium statistical mechanical literature that the GLE can be rigorously derived from equation of motion(EOM) of the brownian particle using projection operator method. We will not review this direction here, and interested readers can consult \cite{zwan}. In the following steps we will derive a modified Drude formula from GLE. Write GLE in Fourier transform $u(t) = \frac{1}{2\pi}\intmipi d\om u(\om) e^{i \om t}$ domain \beqar i\om \frac{1}{2\pi}\intmipi d\om u(\om) e^{i\om t} &=& -\int_{-\infty}^t dt^\p \left(\frac{1}{2\pi}\intmipi M[\om] e^{i\om(t-t^\p)}\right) \left(\frac{1}{2\pi}\intmipi d\om^\p u(\om^\p) e^{i\om^\p t^\p}\right)\nonumber\\ &+& \frac{1}{2\pi}\intmipi d\om (E(\om)+R(\om)) e^{i\om t}, \label{glan1} \eeqar where $M[\om]$ is defined as half-Fourier transform (or Fourier-Laplace transform) \beq M[\om] = \int_0^\infty dt M(t) e^{-i\om t}. \label{khjk} \eeq Here, $M(t') =finite$ for $t'\geq0$, and $M(t')=0$ for $t'<0$ (no memory of future). In equation (\ref{glan1}), the upper limit of the integral over $t^\p$ can be extended from $t$ to $+\infty$ as we set $M(t<0) = 0$ and then the integral over $t^\p$ gives a delta function: $\frac{1}{2\pi}\intmipi dt^\p e^{i(\om^\p-\om)t^\p} = \delta(\om^\p -\om)$. This delta function removes the integral over $\om^\p$ and we get the frequency dependent velocity $u(\om)$ in the form : \beq \la u(\om)\ra = \frac{E(\om)}{m (i\om +M[\om])}. \eeq In performing the ensemble average we set $\la R(t)\ra = 0,$ if $E(t) = E_0 \cos(\om_0 t)$, then by performing the inverse transformation\footnote{In the algebra one has to use $M[-\om] = M[\om]^\ast$ which can be easily proved from the definition of $M[\om]$.} we get $\la u(t)\ra = Re \mu(\om_0) E_0 e^{i\om_0 t}.$ Here \beq \mu(\om) =\frac{1}{m} \frac{1}{i\om +M[\om]}, \eeq is called the dynamical mobility. The dynamical conductivity is related to dynamical mobility by $\sigma(\om) = ne^2 \mu(\om)$. Thus \beq \sigma(\om) = \frac{n e^2}{m} \frac{1}{i\om +M[\om]}. \label{gdf1} \eeq {\it This is called the Generalized Drude Formula (GDF) and $M[\om]$ is called the frequency dependent friction coefficient or the memory function.} GDF can also be derived from the Zwanzig-Mori projection operators as shown in following section. The memeory dependent friction gives Generalised Drude formula. On neglecting this memory we will get back the simple Drude formula. To demonstate it let us consider the memory function to be a delta function in time i.e., $M(t) = \gamma \delta(t +i0)$ (because $M(t)$ is defined only for $t>0$). From equation (\ref{khjk}) we can show that $M[\om]=\gamma$. On substitution it into equation (\ref{gdf1}) we get \beq \sigma(\om) = \frac{n e^2}{m} \frac{1}{i\om +\gamma}. \label{gdf5} \eeq This is the same conductivity, which we have obtained in the previous section (equation $(\ref{lan(7)}))$. In a realistic situation $M[\om]$ can be a complicated frequency dependent function, and its computation can be an involved task. In the following section we derive GDF using projection operator technique and then in subsequent sections this technique will be used to derive $M(\om)$ for the important case of metals (using a pertubative procedure given by W\"{o}lfle-G\"{o}tze\cite{gotze}). Important point to remember from this section is that memoryless $M(t)$ leads to simple Drude formula. \section{The Zwanzig-Mori projection operator formalism} In the present section we are going to derive generalised Drude formula using projection operator technique. This provides a powerful computational approach as the MF can be explicitly computed in many situations of interest. In the case of electron scattering in metals we will explicitly show that memory function becomes frequency independent in the low frequency limit and can be identified with the Drude scattering rate. To derive generalised Drude formula using memory function formalism, we need to introduce Mori-Zwanzig projection operators. We start with the Kubo formula \cite{toda,Singh} \beq \smnw= V\beta\int_0^\infty dt e^{i\om t}\la J_{\nu}(0) J_{\mu}(t)\ra.\label{mori2} \eeq In the above equation we write $J_{\mu}(t)=e^{i\mcl t}J_{\mu}(0)$ where $\mcl$ is the Liouville operator\footnote{The time evolution of a function of dynamical variables is expressed as $\frac{\pr}{\pr t} f = - \mcl f,$ here $\mcl = i\{H , .\}=\frac{\pr H}{\pr \tbfp}.\frac{\pr}{\pr \tbfq} - \frac{\pr H}{\pr \tbfq}.\frac{\pr}{\pr \tbfp}$. $H$ is the total Hamiltonian of the system. The solution of the above Liouville equation reads $f(t) = e^{i\mcl t}f(0).\label{mori4}$} Thus \beq \smnw = V \beta \int_0^\infty dt e^{i\om t}\la J_\nu(0) e^{i \mcl t} J_\mu(0)\ra.\label{mori5} \eeq Time integration can be performed with the result: \beq \smnz = V\beta \la J_\nu \frac{i}{z+\mcl} J_\mu \ra.\label{mori5} \label{3.135} \eeq Let us introduce the ``Bra-Ket'' notation of quantum mechanics for convenience, and define the scalar product of two ``vector" A and B as \beq \la A | B\ra = \la A^\ast B\ra,\label{mori6} \eeq where we define $\la...\ra = tr (\rho...)$ i.e. the thermodynamical ensemble average. On implementing this notation, equation (\ref{3.135}) can be expressed as \beq \smnz = V\beta \left\la J_\nu \left| \frac{i}{z+\mcl} \right|J_\mu \right\ra. \label{3.137} \eeq Projection operator is defined as \beq \mcP \equiv \sum_{\mu^\p} \frac{1}{\la J_{\mu^\p} | J_{\mu^\p} \ra} |J_{\mu^\p} \ra\la J_{\mu^\p} |. \label{3.138} \eeq This projection operator projects an arbitrary vector $J_{\mu}$ onto $J_{\mu^\p}$. It is easy to verify that $\mcP\mcP = \mcP$ (the essential property of projection). Introduce another projection operator $\mcQ$ such that $\mcP+\mcQ = \unit$. Rewrite $\mcl$ in equation (\ref{3.137}) as $\mcl = \mcl (\mcP+\mcQ)$ \beq \smnz = V\beta \left\la J_\nu \left| \frac{i}{\underbrace{z+\mcl \mcQ}_{X} + \underbrace{\mcl \mcP}_{Y}} \right|J_\mu \right\ra. \label{3.139} \eeq With the operator identity \beq \frac{1}{X+Y} = \frac{1}{X} - \frac{1}{X} Y \frac{1}{X+Y}, \label{3.140} \eeq $\smnz$ can be divided into two terms \beq \smnz = i V\beta \underbrace{\left\la J_\nu \left| \frac{1}{z+\mcl\mcQ} \right| J_\mu \right\ra}_{1st~~term}- i V\beta \underbrace{\left\la J_\nu \left| \frac{1}{z+\mcl\mcQ} \mcl\mcP \frac{1}{z+\mcl} \right| J_\mu \right\ra}_{2nd~~term}. \label{3.141} \eeq Write the first term as $\frac{1}{z(1+\frac{1}{z} \mcl\mcQ)} = \frac{1}{z} (1-\frac{1}{z} \mcl \mcQ + \frac{1}{z^2} \mcl \mcQ \mcl\mcQ -.......)$. Except the first term in the expansion, all higher terms vanishes. It is easy to observe this: $\mcQ| J_\mu\ra = \unit |J_\mu \ra - \mcP | J_\mu\ra = \ketjm -\ketjm = 0$. For the second term, insert for the "sandwiched" operator $\mcP$ from equation (\ref{3.138}). Thus, the equation takes a new form \beq \smnz=i \beta V\la J_{\nu}\left|\frac{1}{z}\right| J_{\mu}\ra-iV\beta \chi_0^{-1}\left\la J_\nu \left| \frac{1}{z+\mcl\mcQ} \mcl \sum_{\mu^\p}\frac{\left| J_{\mu^\p}\left\ra \right\la J_{\mu^\p}\right|}{\la J_{\mu^\p} | J_{\mu^\p} \ra} \frac{1}{z+\mcl} \right| J_\mu \right\ra.\label{3.142} \eeq Replacing the magnitude form of $\la J_{\nu}|J_{\mu}\ra$ by $\frac{\chi_{0}}{\beta V}\delta_{\nu\mu}$ \cite{Singh} in both terms and manipulating the second term with projection operators, we obtain \beq \smnz=i \frac{\chi_{0}}{z}\unit-V\beta \sum_{\mu^\p}\left\la J_\nu \left| \frac{i}{z+\mcl} \right|J_\mu \right\ra V\beta\chi_{0}^{-1}\frac{1}{z}\left\la J_{\mu^\p}\left|\frac{z}{z+\mcl\mcQ}\mcl \right| J_{\mu}\right\ra. \label{3.142} \eeq We notice that the second term of the above equation can be written as a product of conductivity and the memory function \beq \sigma(z)=i \chi_0 \frac{1}{z} \unit -\frac{1}{z} \sigma(z) M(z), \label{3.143} \eeq on simplifying \beq \sigma(z)=i \frac{\chi_0}{z + M(z)} = i\frac{\om_p^2}{4\pi}\frac{1}{z+M(z)}, \label{3.144} \eeq where M(z) is the memory function \beq M(z) = V\beta \chi_0^{-1} \left\la J\left| \frac{z}{z+\mcl\mcQ}\mcl \right| J\right\ra. \label{3.145} \eeq On comparing the above equation (\ref{3.144}) with GDF (equation(\ref{gdf1})) we notice that the memory function can be expressed in terms of projection operator (equation(\ref{3.145})). In the following section we review the G\"{o}tze-W\"{o}lfle perturbation method to evaluate the memory function. \section{The G\"{o}tze-W\"{o}lfle (GW) Formalism} Expression given in the above section are formal and cannot be directly used for the computation of the memory function. In 1972, G\"{o}tze and W\"{o}lfle\cite{gotze} presented a perturbative method for the computation of the memory function. They applied their method to the computation of dynamical conductivity of metals. The purpose of this section is to present their method. Before giving the perturbative expansion for memory function we first derive a fundamental equation used by them, i.e., \beq \chi(z) = \chi_0\frac{M(z)}{z+M(z)}. \label{gw1} \eeq Here $\chi(z)$ is the Fourier-Laplace transform of the current-current correlation function \beq \chi(z) = i V\int_0^\infty dt e^{izt} \la [J(t), J(0)]\ra. \label{wg2} \eeq Dynamical conductivity $\sigma(z)$ can be written in terms of $\chi(z)$ \cite{Singh} \beq \sigma(z) = i\frac{\om_p^2}{4\pi z} - \frac{i}{z} \chi(z).\label{wg4} \eeq We have observed in the previous section that conductivity can be expressed in terms of memory function (equation(\ref{3.144})). An expression for $\chi(z)$ in terms of $M(z)$ can be ontained from equation (\ref{3.144}) and (\ref{wg4}): \beq \chi(z) = \frac{\om_p^2}{4\pi}\frac{M(z)}{z+M(z)} = \chi_0 \frac{M(z)}{z+M(z)} ~~~~~~ or,~~~~ M(z)= \frac{z \chi(z)}{\chi(0)-\chi{(z)}}. \label{gwmain} \eeq A more proof of equation (\ref{gw1}) is given in the Appendix. Memory function can be computed if we compute $\chi(z)$. The above expression is a non-perturbative (exact) expression of the memory function. In the next section we introduce the G\"{o}tze-W\"{o}lfle (GW) perturbative mehtod. \subsection{The G\"{o}tze-W\"{o}lfle (GW) approximation for the memory function} For GW metod for the computation of the memory function we need the following EOM \cite{gotze}: \beqar z\la\la J;J\ra\ra_z &=& -\la[J,J]\ra - \la\la [J,H];J\ra\ra_z \nonumber \\ &= &-\la [J,J] \ra+\la \la J;[J,H]\ra \ra_z \eeqar Proof is as follows: We have \beq z\la\la J;J\ra\ra_z = i V\int_0^\infty dt z e^{izt}\la [J(t),J]\ra = V\int_0^\infty dt \la [J(t),J]\ra \frac{d}{dt}(e^{izt}). \eeq On integrating by parts and using $\dot{J}(t) = i[H,J]$, we obtain the first EOM. The second equation can be obtained by using the cyclic property of operators under trace operation. The equal time commutators $[J,J]$ in the above equation are zero. Thus, the first equation can be written as: \beq z\chi(z) = -\la\la C; J\ra\ra_z, \label{gw2} \eeq where we define $C= [J, H]$. In the second equation of motion set $J=C$ : \beq z\la\la C;J\ra\ra_z = -\la [C, J]\ra + \la\la C;C\ra\ra_z, \label{gw3} \eeq or \beq \la\la C;J\ra\ra_z = - \frac{\la [C, J]\ra - \la\la C;C\ra\ra_z}{z}.\label{g4} \eeq Put $z=0$ in equation (\ref{gw3}) and it leads to $\la [C, J]\ra = \la\la C;C\ra\ra_0$. Using this and using equation (\ref{gw2}), the equation (\ref{g4}) gives: \beq z\chi(z) = \frac{1}{z} (\la\la C;C\ra\ra_0 - \la\la C;C\ra\ra_z). \eeq This is an exact expression for the current-current correlation function. Now comes the issue of perturbative expansion. The correlation ($\la\la C;C\ra\ra_z$) has terms proportional to the square of interaction as C is proportional to $H_{interaction}$ $ (C=[J,H]=[J,H_{0}+H_{int}])$. If $H_{int}$ is a small perturbation, then one can linearize the above expression. In equation (\ref{gwmain}) one can expand $M(z)$ as $M(z) = \frac{z\chi(z)}{\chi_0(1-\frac{\chi(z)}{\chi_0})} = \frac{z\chi(z)}{\chi_0}(1+\frac{\chi(z)}{\chi_0}-....)$. Under the condition of small perturbation $||H_{int}||<||H_{0}||$ one can keep the leading order term, and memory function can be approximately expressed as: \beq M(z) \backsimeq \frac{1}{z\chi_0} (\la\la C;C\ra\ra_0 - \la\la C;C\ra\ra_z). \label{gw100m} \eeq This is the central equation used by GW for the computation of dynamical conductivity of a metal in which electron-impurity and electron-phonon scattering is treated as perturbation. In the next two subsections we review their calculation, and then return back to the issue of Drude scattering rate. \subsection{Impurity scattering} In this subsection we review the GW formalism as applied to the impurity scattering case \cite{gotze}. The case of a simple metal (free electron gas) with impurity scattering is considered. The total Hamiltonian is written as \beqar H_{total} &=& H_{electrons} + H_{electron-impurity}\nonumber\\ H_{electrons} &=& \sum_\tbfk \ep(\tbfk) c_\tbfk^\dagger c_\tbfk\nonumber\\ H_{electron-impurity} &=& \frac{1}{N}\sum_{j,\tbfk,\tbfk^\p}\la \tbfk|U^j|\tbfk^\p\ra c_\tbfk^\dagger c_{\tbfk^\p}. \label{im1} \eeqar $N$ is the number of unit cells (we put cell volume to unity). The sum over $j$ is for all the randomly distributed impurities. $U^j$ is the electron scattering potential from $j$th impurity \cite{gotze}. Our aim is to calculate memory function (which is generalised Drude scattering rate). We need to compute $C$, and the correlator $\la\la C;C\ra\ra_z$. The operator $C$ is defined as $[J,H_{total}].$ Current density is given by : $J = \sum_\tbfk v_x(\tbfk) c_\tbfk^\dagger c_\tbfk$, where $v_x(\tbfk)$ is the $x-$component of the electron velocity. We consider a case in which the external field is applied along the x-direction and the induced current is also measured in the same direction. As the current operator commutes with the free electron part of Hamiltonian, we get \beq C = [J, H_{electron-impurity}] = \frac{1}{N}\sum_{j,\tbfk,\tbfk^\p}\la \tbfk|U^j|\tbfk^\p\ra (v_x(\tbfk)-v_x(\tbfk^\p)) c_\tbfk^\dagger c_{\tbfk^\p}. \label{im2} \eeq Define: $ \phi(z) = \la\la C;C\ra\ra_z$ then equation (\ref{gw100m}) can be written as \beq M(z) = \frac{1}{z\chi_0} (\phi(0) - \phi(z)). \label{3.180} \eeq $\phi(z)$ will have correlators of the form: written as $\la\la c_\tbfk^\dagger c_{\tbfk^\p} ; c_\tbfp^\dagger c_{\tbfp^\p}\ra\ra$. \beq \la\la c_\tbfk^\dagger c_{\tbfk^\p} ; c_\tbfp^\dagger c_{\tbfp^\p}\ra\ra = i\int_0^\infty dt e^{izt}\la [c_\tbfk^\dagger(t) c_{\tbfk^\p}(t),c_\tbfp^\dg c_{\tbfp^\p}]\ra. \label{im4} \eeq Using $c_\tbfk(t) = c_\tbfk e^{-i \ep_\tbfk t}$, the time integration can be performed with the result \beq -\frac{1}{z +\ep_\tbfk -\ep_{\tbfk^\p}}\la [c_\tbfk^\dg c_{\tbfk^\p}, c_\tbfp^\dg c_{\tbfp^\p}]\ra. \eeq By using anticommutation relations for $c$ and $c^\dg$, thermal ensemble average can be found \cite{Singh} \beq -\frac{1}{z+\ep_\tbfk - \ep_{\tbfk^\p}} (f(\tbfk) - f(\tbfk^\p)). \eeq $f(\tbfk) = \la c^\dg_\tbfk c_\tbfk\ra$ is the Fermi-Dirac distribution function. With this simplified expression of correlator, $\phi(z)$ can be written as \beq \phi(z) = \la\la C;C\ra\ra_z =- 2\frac{1}{N^2}\sum_{i,j,\tbfk,\tbfk^\p}\la \tbfk|U^i_1|\tbfk^\p\ra \la \tbfk^\p|U^j_1|\tbfk\ra (v_x(\tbfk)-v_x(\tbfk^\p))^2 \frac{f(\tbfk) -f(\tbfk^\p)}{z+\ep_\tbfk -\ep_{\tbfk^\p}}. \eeq The summations over (i,j) run over the total number of impurities ($N_{imp}$). For $i\ne j,~~\phi(z)$ is proportional to $(\frac{N_{imp}}{N})^2$. If the impurity concentration $c=\frac{N_{imp}}{N}$ is very small, then this term can be neglected as compared to the term proportional to $c$ (the diagonal elements contribution). Thus, in the leading order: \beq \phi(z) = -2\frac{c}{N}\sum_{\tbfk,\tbfk^\p} |\la \tbfk|U_1|\tbfk^\p\ra|^2 \{v_x(\tbfk)-v_x(\tbfk^\p)\}^2 \frac{f(\tbfk) -f(\tbfk^\p)}{z+\ep_\tbfk -\ep_{\tbfk^\p}}.\label{im7} \eeq Here the summation over the impurity sites $i$ is also performed because there are no correlation effects considered \cite{gotze,Singh}. And the problem reduces to one impurity problem. Further simplification can be done using isotropy in the free electron case ${\bf v}^2 = 3 v_x^2$ and writing ${\bf v} = \tbfk/m$. Substituting equation (\ref{im7}) into equation (\ref{3.180}) we get \beq M(z) = \frac{2}{3 z\chi_0}\frac{c}{m^2 N}\sum_{\tbfk,\tbfk^\p} |\la \tbfk|U_1|\tbfk^\p\ra|^2 \{\tbfk-\tbfk^\p\}^2 (f(\tbfk) -f(\tbfk^\p))\left(\frac{1}{z+\ep_\tbfk -\ep_{\tbfk^\p}} - \frac{1}{\ep_\tbfk -\ep_{\tbfk^\p}}\right). \label{im8} \eeq Substitute $z=\om+i\varepsilon$ and perform the limit $\varepsilon\rta0$, the above expression reduces to\footnote{Use the identity $\lim_{\varepsilon\rta0}\frac{1}{a\pm i\varepsilon} = \frac{1}{a}\mp i\pi \delta(a)$.} \beq Im M(\om) = \frac{2}{3} \pi (\frac{N_{imp}}{N_e}) \frac{1}{m N^2} \sum_{\tbfk,\tbfk^\p} |U(\tbfk,\tbfk^\p)|^2 (\tbfk-\tbfk^\p)^2 (\frac{f(\tbfk)-f(\tbfk^\p)}{\om})\delta(\om +\ep_\tbfk -\ep_{\tbfk^\p}). \label{im9} \eeq Using $\frac{1}{N}\sum_{\tbfk} \longrightarrow \int\frac{d^3k}{(2\pi)^3},$ equation (\ref{im9}) becomes \beqar Im M(\om) &=& \frac{2}{3}\pi (\frac{N_{imp}}{N_e})\frac{1}{m} \int\frac{d^3k}{(2\pi)^3} \int\frac{d^3k^\p}{(2\pi)^3} |U(\tbfk,\tbfk^\p,\theta)|^2 (k^2 +{k^\p}^2 - 2 k k^\p \cos(\theta))\nonumber\\ &\times& \left(\frac{f(\tbfk)-f(\tbfk^\p)}{\om}\right)\delta(\om + \ep_\tbfk -\ep_{\tbfk^\p}). \label{im10} \eeqar By changing variables $\ep = \frac{k^2}{2 m}~~(\hbar=1)$ and performing the integration over $\ep^\p=\frac{{k^\p}^2}{2 m}$ and using the properties of the delta function we obtain: \beqar Im M(\om) &=&\frac{2}{3}\pi (\frac{N_{imp}}{Ne})\frac{(2m)^3}{(2\pi)^4}\int_0^\infty d\ep \sqrt{\ep(\ep+\om)}\int_0^\pi d\theta \sin\theta |U(\ep,\ep+\om,\theta)|^2\nonumber\\ &\times&(2\ep+\om-2\sqrt{\ep(\ep+\om)}\cos\theta) \left(\frac{f(\ep)-f(\ep+\om)}{\om}\right). \label{im11} \eeqar This is a finite frequency result. In the limit $\om\rta 0$, the last factor gives a delta function and the integral over energy can be performed with the result \beq Im M(0) =\frac{4}{3}\pi (\frac{N_{imp}}{Ne})\frac{(2m)^3}{(2\pi)^4} \ep_F^2 \int_0^\pi d\theta \sin\theta |U(\ep_F,\theta)|^2 (1-\cos\theta). \label{im12} \eeq With the angle dependence of the scattering potential taken into account it is clear that the scattering rate ($\frac{1}{\tau} = Im M(0)$) has very small contribution from small angle scattering (when $\theta$ is small). Thus, small angle scattering is not efficient in the momentum degradation. The very same conclusion is obtained when one deal with the solution of the Boltzmann equation beyond Relaxation Time Approximation (RTA) in an isotropic medium \cite{Singh,ziman,madelung,wooten}. Thus the above expression from memory function formalism goes beyond the RTA \cite{Singh}. Next, consider that the scattering potential is independent of the angle between $\tbfk$ and $\tbfk^\p$. The above equation (\ref{im12}) can be further simplified: \beq Im M(0) = \frac{2}{3}\pi \frac{N_{imp}}{N_e} (U \rho_F)^2 \ep_F. \eeq Thus the real part of conductivity from equation (\ref{3.144}) is given by \beq Re\sigma(\om) = \frac{\om^2_p}{4\pi}\frac{Im M(0)}{\om^2 + (Im M(0))^2}. \label{im16} \eeq This is nothing but the Drude formula. We identify that \beq \frac{1}{\tau}=Im M(0) = \frac{1}{\taud} = \frac{2}{3}\pi \frac{N_{imp}}{N_e} (U \rho_F)^2 \ep_F. \label{memdrude} \eeq Thus, the Drude scattering rate increases linearly (in the leading order) with impurity concentration and with the square power of the electron-impurity scattering potential. {\it The most important point is that we recover the Drude formula in the low frequency limit. The low frequency here means $\om<<\ep_F$ (the Fermi energy). If Fermi energy is in $eV$s then the valid frequency regime of simple Drude formula is below say infrared frequencies. Thus} \beq \frac{1}{\tau} {\rm~~ is~frequency~independent,~when~~} \hbar\om<<\ep_F. \eeq This defines the regime of applicability of the Drude formula in metals \cite{wooten}. \subsection{Phonon scattering} Next application of the WG formalism is the computation of the Memory Funtion (MF) in the case of electron-phonon scattering in metals. The leading cause of electrical resistance at ambient temperatures is due to the electron-phonon scattering. At ambient temperature electrical resitivity is linearly proportional to tempeature i.e., $\rho \propto T.$ In a pure metalic sample at lower temperature ($T<<\Theta_{D}$, Debye Temperature), resitivity is proportional to fifth power of temperature ( $\rho \propto T^5)$. These temperature dependencies have been experimentally verified. Theoretically these can be obtained from the solution of the Bloch-Boltzmann equation \cite{Singh,ziman,madelung}. The Bloch-Boltzmann equation formulates the electron-phonon scattering in a semiclassical way. We will notice in this section that these temperature dependencies can also be obtained from memory function formalism. Before we start the calculation of the MF in the present case of electron-phonon scattering, let us note down its basic assumptions. For conduction electrons, a free electron gas is assumed. The phonon considered are long wavelength longitudinal acoustic phonons. For electron-phonon interaction the Fr\"{o}hlich Hamiltonian is used: \beqar H_0 &=& H_{el} + H_{ph},\nonumber\\ H_{el-ph} &=& \sum_{\tbfk,\tbfkp} (D(\tbfk-\tbfkp) c_\tbfk^\dg c_\tbfkp b_{\tbfk-\tbfkp} + h.c.). \eeqar Here $H_0$ is the free electron and free phonon part of the Hamiltonian with $H_{el} =\sum_\tbfk c_\tbfk^\dg c_\tbfk $ and $H_{ph} = \sum_\tbfq \om_\tbfq (b_\tbfq^\dg b_\tbfq + 1/2)$. The second equation represents the electron phonon interaction. The coefficient $D(\tbfk-\tbfkp)$ for acoustic phonon interaction is given by \beq D(\tbfq) = \frac{1}{\sqrt{2 m_{ion} N \om_\tbfq}} q C_q, \eeq $N$ is the total number of unit cells, $m_{ion}$ is the ionic mass, and $\om_\tbfq$ is the phonon frequency. $C_q$ is the electron-phonon coupling constant and it is a slowly vaying function of $q$ \cite{madelung}. As before we work with units $\hbar=k_{B}=1$. To compute the memory function we need to compute C: \beq C=[J_x, H_{el-ph}] = \sum_{\tbfk,\tbfkp,\tbfpp} v_x(\tbfk) D(\tbfp-\tbfpp)[c_\tbfk^\dg c_\tbfk, c_\tbfp^\dg c_\tbfpp b_{\tbfp-\tbfpp}]- h.c. \eeq Using the anticommutation properties of $c_\tbfk$, and $c_\tbfk^\dg$, and using the identity $[AB,C] = A[B,C] + [A,C]B$ we get \beq C = \sum_{\tbfp,\tbfpp} (v_x(\tbfp)-v_x(\tbfpp)) D(\tbfp-\tbfpp)c_\tbfp^\dg c_\tbfpp b_{\tbfp-\tbfpp} - h.c. \eeq Our next step is to calculate $\phi(z) = \la\la C;C\ra\ra_z$. Out of four terms in it only non-diagonal terms survive: \beqar \phi(z) &=&\nonumber\\ & -&\sum_{\tbfk,\tbfkp,\tbfp,\tbfpp} (v_x(\tbfk)-v_x(\tbfkp))(v_x(\tbfp)-v_x(\tbfpp)) D(\tbfk-\tbfkp) D^\ast(\tbfp-\tbfpp) \la\la c_\tbfk^\dg c_\tbfkp b_{\tbfk-\tbfkp};b^\dg_{\tbfp-\tbfpp}c^\dg_\tbfpp c_\tbfp\ra\ra\nonumber\\ &-&\sum_{\tbfk,\tbfkp,\tbfp,\tbfpp} (v_x(\tbfk)-v_x(\tbfkp))(v_x(\tbfp)-v_x(\tbfpp)) D^\ast(\tbfk-\tbfkp) D(\tbfp-\tbfpp) \la\la c_\tbfkp^\dg c_\tbfk b^\dg_{\tbfk-\tbfkp};b_{\tbfp-\tbfpp}c^\dg_\tbfp c_\tbfpp\ra\ra.\nonumber\\ \eeqar The correlation function in its expanded notation is given by: \beq \la\la c_\tbfk^\dg c_\tbfkp b_{\tbfk-\tbfkp};b^\dg_{\tbfp-\tbfpp}c^\dg_\tbfpp c_\tbfp\ra\ra = i\int_0^\infty dt e^{i z t}\la [c_\tbfk^\dg(t) c_\tbfkp(t^\p) b_{\tbfk-\tbfkp}(t), b^\dg_{\tbfp-\tbfpp}c^\dg_\tbfpp c_\tbfp]\ra. \eeq Using $c_\tbfk(t) = c_\tbfk e^{-i\ep_\tbfk t}$, the time integration and the ensemble average can be performed. The result is \beqar \phi(z) &=& \frac{2}{3}\frac{1}{m^2}\sum_{\tbfk,\tbfkp} |D(\tbfk-\tbfkp)|^2 |\tbfk-\tbfkp|^2\{f(1-f^\p)(1+n_{-})-(1-f)f^\p n_{-}\}\nonumber\\ &\times&\left(\frac{1}{\ep -\ep^\p -\om_{-} +z} +\frac{1}{\ep -\ep^\p -\om_{-} -z}\right). \eeqar Here a factor of 2 in the coefficient comes from the spin summations. We have used short hand notation: $f$ means $f(\tbfk)$; $f^\p$ means $f(\tbfkp)$; $n_{-}$ is for $n_{\om_{\tbfk-\tbfkp}}$, and $\om_{-}$ is for $\om_{\tbfk-\tbfkp}$. On inserting the above equation into (\ref{3.180}) and on simplifying we get: \beqar Im M(\om) &=& \frac{2}{3}\pi\frac{1}{m^2\chi_0} \sum_{\tbfk,\tbfk^\p} |D(\tbfk-\tbfk)|^2 |\tbfk-\tbfkp|^2 (1-f)f^\p n_{-}\nonumber\\ &\times&\left(\frac{(e^{\beta \om}-1)}{\om} \delta(\ep-\ep^\p -\om_{-} +\om) - ...({\rm terms~with~}\om \rta -\om)...\right). \eeqar Convert momentum sums into integrals. To simplify the momentum integrals, change $|\tbfk-\tbfkp|$ into q integral, by introducing $\int dq \delta(q-|\tbfk-\tbfkp|)$ and writing integration variables k and $k^\p$ as $\sqrt{2 m \ep}$ and $\sqrt{2 m \ep'}$, we obtain \beqar &&Im M(\om) =\nonumber\\ &&\frac{2}{3}\pi\frac{N m^2}{ (2\pi)^3 m_{ion} N_e}\int_0^\infty dq \frac{C_q^2}{\om_q}q^4\int_0^\infty d\ep \sqrt{\ep}\int_0^\infty d\ep^\p \sqrt{\ep^\p}\int_0^\pi d\theta \sin\theta \delta\left(q-\sqrt{2m}\sqrt{\ep+\ep^\p-2\sqrt{\ep\ep^\p}\cos\theta}\right)\nonumber\\ &\times&(1-f) f^\p n_{-}\left(\frac{(e^{\beta \om}-1)}{\om} \delta(\ep-\ep^\p -\om_{-} +\om) - ...({\rm terms~with~}\om \rta -\om)...\right). \label{ph6} \eeqar The integral over $\theta$ can be simplified in the following way. The presence of the Fermi factors $(1-f)f^\p$ cause the integrand to have a finite value only in a zone around the Fermi surface of a strip of width $2 k_B T.$ And outside of this zone integrand is vanishingly small. Thus $\ep$ and $\ep^\p$ can be approximately replaced by $\ep_F$. On implimenting all these steps, the $\theta$ integral reduces to: \beq \int_0^\pi d\theta \sin\theta \delta(q - \sqrt{2}k_F \sqrt{1-\cos\theta}). \label{ph7} \eeq This elementary integration can be performed with the result $\frac{q}{k_F^2}$ (notice that $0<q<k_F$). Insert the value $\frac{q}{k_F^2}$ of the $\theta$-integral into equation (\ref{ph6}). Perform the $\ep^\p$ integral using the property of the delta functions. Then, perform the integrations over $\ep$ using elementary methods\footnote{use $\intmipi dx \frac{e^x}{e^x+1} \frac{1}{e^{x+a}+1} = \frac{a}{e^a-1}$.}. Finally, we obtain \beqar Im M(\om) &=& \frac{1}{8}\pi^3N (m m_{ion} k_F^5)^{-1}\rho_F^2 \int_0^{q_D} dq q^5 \frac{C_q^2}{\om_q} \left(\frac{1}{e^{\beta \om_q}-1}\right)\nonumber\\ &\times& \left\{ \frac{(1-\om_q/\om)(e^{\beta\om}-1)}{e^{\beta(\om-\om_q)}-1} + ...({\rm terms~with~}\om \rta -\om)... \right\}. \label{ph8} \eeqar The above equation (\ref{ph8}) is GW's equation (\ref{im4}) in \cite{gotze}. For further simplifications we assume the linear approximation to phonon spectrum $\om_q = c_s q$ (i.e., the Debye approximation) and $C_q$ is assumed constant $C_q=\frac{1}{\rho_F}$ \cite{gotze}. In the following subsections we analyse the above expression in various special cases of interest. We will notice that Bloch-Boltzmann equation results can also be obtained by this metod. Let us first consider the D.C. case. \subsubsection{D.C. case} In the $\om\rta0$ limit, the expression in the curly brackets in the above equation (\ref{ph8}) reduces to $2\beta\frac{\om_q}{1-e^{-\beta\om_q}}.$ Thus \beq Im M(0) = \frac{1}{4}\pi^3 N (m m_{ion} k_F^5)^{-1}\int_0^{q_D} dq \frac{q^5 e^{\beta \om_q}}{(e^{\beta\om_q}-1)^2}. \eeq To simplify, write $x = \beta \om_q$, and $\beta \om_D = \Theta_D = \frac{\Theta_D}{T}$: \beq Im M(0) = \frac{1}{4}\pi^3 N q_D^6 (m m_{ion} k_F^5\Theta_D)^{-1} (T/\Theta_D)^5 J_5(\Theta_D/T),~~{\rm~where~~~}J_5(y) = \int_0^y dx \frac{x^5 e^x}{(e^x-1)^2}. \eeq The above expression is the DC scattering rate (equilvalent to the Drude scattering rate). It can be further analyzed into two special cases. In the high temperature limit, $T>>\Theta_D$, we have $J_5(y)\simeq \frac{1}{4} y^4$, thus \beq Im M(0) = \frac{1}{\tau}= \frac{1}{16} \pi^3 N q_D^6 (m m_{ion} k_F^5 \Theta_D)^{-1} (T/\Theta_D) \propto T. \eeq So, we observe that in high temperature approximation (when the temperature (T) is much greater than the Debye temperature $\Theta_{D}))$, the scattering rate or, the imaginary part of the memory function has $T-$ linear temperature dependence. This agrees with the result as obtained from the Bloch-Boltzmann equation in the high temperature limit \cite{Singh,ziman}. In the low temperature limit, i.e, $T<<\Theta_D$, we have $J_5(y)\simeq 124.4$, and the MF is \beq Im M(0) = \frac{1}{\tau}= 31.1 \pi^3 N q_D^6 (m m_{ion} k_F^5 \Theta_D)^{-1} (T/\Theta_D)^5 \propto T^5. \eeq This again agrees with the famous Bloch-Boltzmann's $T^5$ law of phonon scattering. Thus all the known results in the DC limit are reproduced (Table 1:). \begin{table}[h] \caption{Temperature dependence of memory function in DC limit.} \label{tab1} \begin{center} \begin{tabular}{|c||c|c|c|} \hline \multirow{3}{*}{Temperature limit} & \multicolumn{2}{c||}{D.C. case} \\ \cline{2-3} & $ J_{5}(y) $ & $ImM(0)$ \\ \hline $T>>\Theta_{D}$ &$ \simeq \frac{y^4}{4}$ & $\propto T$ \\ \hline $T<<\Theta_{D}$ &$ \simeq 124.4$ & $\propto T^5$ \\ \hline \end{tabular} \end{center} \end{table} \subsubsection{A.C. case} The integral in equation (\ref{ph8}) in $\om \neq 0$ limit is difficult to perform (although it can be expressed in terms of PloyLog functions). We consider a simpler and relevant case of high frequencies $\om >>\Theta_D$, and $\om>>T$. Then the expression in curly brackets in equation (\ref{ph8}) can be simplified to give $e^{\beta\om_q} +1$. On simplifying the integral over q we obtain \beq Im M(\om) \simeq \frac{1}{4}\pi^3 N q_D^6 (m m_{ion} k_F^5)^{-1} (T/\Theta_D)^5 J(\Theta_D/T),~~~J(y) =\frac{1}{2}\int_0^y dx x^4 \coth(x/2). \label{ac1} \eeq In the low temperature limit, i.e. $T<<\Theta_D$, we have $J(y) \simeq \frac{1}{10} y^5$, thus \beq Im M(\om) \simeq \frac{1}{40}\pi^3 N q_D^6 (m m_{ion} k_F^5)^{-1}. \label{ac2} \eeq We observe that at high frequency $\om>>\Theta_{D}$, but at lower temperature $T<<\Theta_{D}$, the scattering rate reduces to a constant value independent of frequency and temperature (Table 2:). And in the high temperature limit i.e., $T>>\Theta_D,$ we have $J(y)\simeq\frac{1}{4} y^4$. The temperature dependence of MF is given by \beq Im M(\om) \simeq \frac{1}{16}\pi^3 N q_D^6 (m m_{ion} k_F^5)^{-1} (T/\Theta_D) \propto T. \label{ac3} \eeq This behaviour is depicted in figure 2. An important point to be noticed is that even at zero temperature scattering rate is not zero. This non-zero scattering rate is due to an important mechanism called the Holstein mechanism. In the limit $T\rta0$, equation (\ref{ph8}) can be simplified for all values of $\om$. In this limit, the expression in the curly bracket in (\ref{ph8}) gives $(1-\om_q/\om)e^{\beta\om_q}$ if $\om>\om_q,$ and $(\om_q/\om-1)e^{\beta\om}$ if $\om<\om_q$. After performing the q integral we get: \beq Im M(\om) \simeq \frac{1}{240}\pi^3 N q_D^6 (m m_{ion} k_F^5 \Theta_D)^{-1}\times\left\{ \begin{array}{lr}(\om/\Theta_D)^5& : |\om| <\Theta_D\\ 6-5 (\om_D/\om) & : |\om| >\Theta_D. \label{ac4} \end{array} \right. \eeq In the limit $T\rta0$, the non-vanishing value of the generalised Drude scattering rate (or memory function) has a very important physical meaning as mentioned before. In a pure sample at zero temperature there are no thermally excited lattice vibrations or phonons, thus one would naviely expect that electron scattering would be prohibited and the scattering rate would vanish. But it is shown for the first time by Ted Holstein that there is an important mechanism of momentum randomization even at zero temperature \cite{ted}. The mechanism involves the simultaneous absorption of a photon and creation of an electron-hole pair along with an acoustic phonon. Thus the mechanism---now called the Holstein mechanism---involves the creation of a phonon and it occurs at finite frequencies. From equation (\ref{ac4}) we observe that G\"{o}tze-W\"{o}lfle formalism is capable of caputuring this physical effect. At low frequencies $|\om|<\Theta_{D}$, the scattering rate grows as $\om^5$ and then it saturates when $\om>>\Theta_{D}$. The transition happens around $\Theta_{D}$ (Table2). The generalised Drude scattering rate can be obtained from the experimental data of reflectance using the memory function formalism. If experimentally obtained scattering rate shows transition behaviour around $\Theta_{D}$, then it can be derived that the dominating scattering mechanism is due to phonons. This perticular issue is very important in the field of strange metals such as observed in normal states of cuprate high temperature superconductors. Absence of $\Theta_{D}$ points to other scattering mechanisms. \begin{figure}[h!] \begin{center} \includegraphics[height = 7cm, width =10cm]{fig1} \caption{A depiction of the Holstein mechanism} \label{fig1} \end{center} \end{figure} \begin{table}[h] \caption{Temperature dependence of the memory function in the AC limit} \begin{center} \begin{tabular}{|c||c|c|c|} \hline \multirow{3}{*}{($\om>>\Theta_{D}$)\newline Temperature limit} & \multicolumn{2}{c||}{A.C. case} \\ \cline{2-3} & $ J_{5}(y) $ & $ImM(0)$ \\ \hline $T>>\Theta_{D}$ &$ \simeq \frac{y^4}{4}$ & $\propto T$ \\ \hline $T<<\Theta_{D}$ &$ \simeq \frac{y^5}{10}$ & $independent ~of~ T$ \\ \hline \end{tabular} \end{center} \end{table} We also notice that the generalised Drude scattering rate (equation (\ref{ac4})) strongly deviates from the simple Drude scattering rate (which is just a constant parameter). From equation (\ref{im16}) the real part of conductivity is given by: \beq Re\sigma(\om) = \frac{\om^2_p}{4\pi}\frac{Im M(\om)}{\om^2 + (Im M(\om))^2}. \label{ac5} \eeq Where $ImM(\om)$ is given by the expression (\ref{ac4}). If the sample also has dilute impurities, then in the linear order $ImM_{total}(\om)=Im M_{imp}(\om)+ImM(\om)$, which in the $\om\rta 0$ limit goes back to Drude formula (equation (\ref{mom2})). \section{Summary of MF formalism} In this brief overview of the memory function formalism we sketched the derivations of the generalised Drude formula by two methods. In the first method it is derived from the generalised Langevin equation. An advantage of this route is that we are able to appreciate the problems of the standard Drude formula and the corresponding standard Langevin equation. Then in the second route we reviewed the Zwanzig-Mori projection operator method, and showed that dynamical conductivity from the Kubo formula can be written in a form resembling the generalised Drude formula. In doing that the memory function is expressed through projection operators. We then reviewed a very useful method for the computation of the memory function put forward by G\"{o}tze-W\"{o}lfle (GW). Two applications of this method in the computation of dynamical conductivity of metals are reviewed. In the first application of the electron-impurity scattering we notice that the Drude scattering rate is frequency dependent. It can be taken as frequency independent only when the system is probed with frequencies much less than $\frac{\ep_{F}}{\hbar}.$ We also noticed that the DC scattering rate includes $1-\cos\theta$ factor which stresses the importance of large angle scattering in momentum degradation. This factor appears in the solution of the Bloch-Boltzmann equation when vortex corrections are taken into account. In the elecron-phonon scattering case we notice that in the DC limit the memory function formalism reproduces all the known results of the Bloch-Boltzmann equation ($\rho\propto T~when ~T>>\Theta_{D}, ~and~ \rho\propto T^5 ~when~T<<\Theta_{D}).$ In the AC limit it also reproduces the Holstein mechanism of electromagnetic energy absorption at zero temperature. \section{Appendix: The G\"{o}tze-W\"{o}lfle (GW) Formalism} We need to show that \beq \chi(z) = \chi_0\frac{M(z)}{z+M(z)}, \label{wg1} \eeq where $\chi(z)$ is \beq \chi(z) = i V\int_0^\infty dt e^{izt} \la [J(t), J(0)]\ra. \label{wg2} \eeq Let us define \beq \zeta(t) = i\la [J(t),J(0)]\ra. \eeq Using the cyclic property of trace $\zeta(t)$ can be written as \beq i tr(\rho J(t) J) - i tr(\rho J J(t)) = i tr(J(t) [J,\rho]). \label{wg3} \eeq Use Kubo's Identity $e^{\beta H} [J, e^{-\beta H}] = \int_0^\beta e^{\lambda H} [H, J]e^{-\lambda H},$ and write $[J,\rho] $ it in the following way \beq [J,\rho] = \int_0^\beta d\lambda \rho [H, J(-i\lambda)]. \eeq Using $\dot{A} = i [H, A]$, the above commutator takes the form \beq [J, \rho] = -i \int_0^\beta d\lambda \rho \dot{J}(-i\lambda). \eeq Inserting this expression in equation (\ref{wg3}), $\zeta(t)$ takes the form \beq \zeta(t) = \beta \la\dot{J}(-i\lambda) J(t)\ra^\diamond. \eeq Here we have introduced a notation: $\la...\ra^\diamond = \frac{1}{\beta}\int_0^\beta d\lambda tr (\rho...)$. With this, the equation (\ref{wg2}) can be written as \beq \chi(z) = V \int_0^\infty dt e^{izt}\zeta(t) = V\beta \int_0^\infty dt e^{izt} \la \dot{J}(-i\lambda) J(t)\ra^\diamond. \label{wg10} \eeq For the next part of the proof, we will use an alternative representation of conductivity from Kubo's formula \cite{kubo}: \beq \sigma(z) = -\frac{\om_{p}^2}{4 \pi iz} - \frac{\beta V}{iz} \int_0^\infty dt e^{izt} \la J(-i\lambda)\dot{J}(t)\ra^\diamond. \label{wg11} \eeq We want to find a connection between the last two equations ( (\ref{wg10}) and (\ref{wg11})). Only problem is that the integrands differ w.r.t. time differentiation. To sort this out, we use the stationarity property of the current-current correlation function, i.e., $\la J(t_0)J(t_0+t)\ra$ must be independent from the initial time $t_0$ or \beq \frac{d}{dt_0}\la J(t_0) J(t_0 +t + i\lambda)\ra^\diamond = 0 = \la \dot{J}(t_0) J(t_0 +t + i\lambda)\ra^\diamond + \la J(t_0) \dot{J}(t_0 +t + i\lambda)\ra^\diamond. \label{wg12} \eeq Thus at $t_0=0$ \beq \la \dot{J}(0) J(t + i\lambda)\ra^\diamond = - \la J(0) \dot{J}(t + i\lambda)\ra^\diamond. \eeq Using this, from equation (\ref{wg10}) and (\ref{wg11}) we obtain the required result.

\section{Introduction} Relativistic dissipative hydrodynamics has become an important tool to study bulk properties of the near-equilibrium system formed in relativistic heavy-ion collisions \cite{}. The large densities and short mean-free times in the system allow for a coarse graining in hydrodynamics which integrates out the microscopic length and time scales. The effective degrees of freedom are then the average conserved quantities, namely the energy, momentum, electric charge and baryon number, which are dynamically evolved according to the hydrodynamic equations \cite{}. When low enough densities are reached and the interaction times become longer, the system falls out of equilibrium, which ultimately leads to a kinetic freeze-out. In spite of this inherent coarse graining \cite{}, hydrodynamics has been remarkably successful in explaining several experimental observables pertaining to relativistic heavy-ion collisions, for example, the anisotropic flow $v_n$ \cite{}, that characterizes the final-state momentum anisotropy in the plane transverse to the beam direction. The flow has been well understood as the collective hydrodynamic response to the initial collision geometry fluctuating event by event \cite{}. The long-range rapidity structures observed in multiparticle correlation measurements in heavy-ion collisions \cite{} as well as in high-multiplicity collision events involving small projectiles ($p/d/^3$He) \cite{} can also be related to the hydrodynamic behavior. It is then instructive to investigate whether the thermal noise or the fast microscopic degrees of freedom that survive coarse graining, have any measurable effect on the experimental observables. The fluctuation-dissipation theorem already forces one to consider fluctuations in systems that are in thermal equilibrium. Further, as the size of the fireball formed is just about 10 fm, and there are only a finite number of particles in each coarse-grained fluid cell, fluctuations may play a crucial role. Thermal noise could be even more important for proper interpretation of observables near the critical point for confinement-deconfinement transition where all fluctuations are large in general. In contrast to the perturbation created in the medium due to energy deposition by a propagating jet \cite{}, thermal fluctuations are produced at all space-time points in the fluid cells. These local fluctuations are propagated/diffused via the fluid dynamic evolution equations. Nevertheless, the thermal fluctuations in heavy-ion collisions may not be quite large, other than near the critical point, as the strongly coupled quark-gluon plasma (QGP) is formed with a small shear viscosity to entropy density of $\eta_v/s \simeq 0.08-0.20$ \cite{}. The theory of hydrodynamic fluctuations or noise in a nonrelativistic fluid \cite{Landau} was extended into the relativistic regime for Navier-Stokes (first-order) viscous fluid by Kapusta {\it et al.} \cite{}. The thermal fluctuation, $\Xi^{\mu\nu}$, of the energy-momentum tensor was shown to have a nontrivial autocorrelation $\langle \Xi^{\mu\nu}(x) \Xi^{\alpha\beta}(x') \rangle \sim T \eta_v \: \delta^4(x-x')$, where $T$ is the temperature. \cite{}. Due to the occurrence of the Dirac delta function, the energy and momentum density averaged value of this white noise becomes $\sim 1/\sqrt{\Delta V \Delta t}$. Thus, even for small shear viscosity $\eta_v$, the white noise sets a lower limit on the system cell size $\Delta V$ that is essentially comparable to the correlation length. Consequently, white noise could lead to large gradients which makes the basic hydrodynamic formulation (based on gradient expansion) questionable. The divergence problem can be circumvented, by treating the white noise as a perturbation (in a linearized hydrodynamic framework) on top of a baseline nonfluctuating hydrodynamic evolution \cite{}. Analytic solutions for hydrodynamic fluctuations were obtained in the case of boost-invariant longitudinal expansion without transverse dynamics (Bjorken flow) \cite{} and with transverse dynamics (Gubser flow) \cite{}. However, both these calculations were performed in the relativistic Navier-Stokes theory for a conformal fluid. It is important to recall that the first-order dissipative fluid dynamics or the Navier-Stokes theory, displays acausal behavior that may lead to unphysical effects. On the other hand, the second-order dissipative fluid dynamics, based on the M\"uller-Israel-Stewart (MIS) framework \cite{}, gives hyperbolic equations and restores causality. The commonly used MIS formulation has been quite successful in explaining the spectra and azimuthal anisotropy of particles produced in heavy-ion collisions. Recently, formally new dissipative equations have been derived from Chapman-Enskog-like iterative expansion of the Boltzmann equation in the relaxation-time approximation \cite{} and from the modified 14-moment method which was developed by Denicol {\it et al.} \cite{}. In this work, we shall present the formulation of thermal fluctuations for these forms of hydrodynamic dissipative evolution equations in the case of a boost-invariant longitudinal expansion. The fluctuation equations so obtained are rather general and will be used along with an equation of state (EOS) corresponding to a conformal fluid and then with the lattice QCD EOS. Since analytical solutions for hydrodynamic fluctuations cannot be obtained for the baseline second-order hydrodynamic approaches, we shall perform numerical simulations of thermal noise and its evolution as a perturbation on top of boost-invariant longitudinal expansion of the viscous medium. The paper is organized as follows. In Sec. II, we formulate hydrodynamic fluctuations in the linearized limit as a perturbation on top of second-order dissipative equations for boost-invariant expansion. In Sec. III, we calculate the phenomenology of particle freeze-out and the effect of fluctuation on two-particle rapidity correlations. In Sec. IV, we present results from numerical simulations for the rapidity correlations with ideal gas and lattice QCD equations of state. Finally in Sec. V, we summarize our results and conclude. \section{Thermal fluctuations in boost invariant hydrodynamics} In this section we formulate thermal fluctuations in the boost-invariant longitudinal expansion of matter within second-order viscous hydrodynamics. In the presence of a thermal noise tensor $\Xi^{\mu\nu}$, the total energy-momentum tensor becomes \begin{align}\label{FTt:eq} T^{\mu\nu} = \epsilon u^\mu u^\nu - p\Delta^{\mu\nu} + \pi^{\mu\nu} + \Xi^{\mu\nu}. \end{align} We shall work in the Landau-Lifshitz frame and disregard particle diffusion current, which is a reasonable approximation due to very small values of the net baryon number formed at RHIC and LHC; we further ignore bulk viscosity in our calculation. In the above equation, $\epsilon$ and $p$ are respectively the energy density and pressure in the fluid's local rest frame (LRF), and $\pi^{\mu\nu}$ is the shear pressure tensor. $\Delta^{\mu\nu}=g^{\mu\nu}-u^\mu u^\nu$ is the projection operator on the three-space orthogonal to the hydrodynamic four-velocity $u^\mu$ which is defined by the Landau matching condition $T^{\mu\nu}u_{\nu} = \epsilon u^\mu$. The total energy-momentum tensor $T^{\mu\nu}$ consists of an average part $T_0^{\mu\nu}$ (represented by subscript ``0") and a thermally fluctuating part $\delta T^{\mu\nu}$ (represented by $\delta$). In the presence of fluctuations, the energy density (or temperature), flow velocity, and shear pressure tensor can be written as \cite{} \begin{align}\label{nonlin:eq} \epsilon &= \epsilon_0 + \delta\epsilon, \nonumber\\ u^\mu &= u_0^\mu + \delta u^\mu, \nonumber\\ \pi^{\mu\nu} &= \pi_0^{\mu\nu} + \delta\pi^{\mu\nu}. \end{align} In the linearized limit (keeping terms up to first order in fluctuations), the total energy-momentum tensor becomes: \begin{align}\label{FT:eq} T^{\mu\nu} &= \epsilon u^\mu u^\nu - p\Delta^{\mu\nu} + \pi^{\mu\nu} + \Xi^{\mu\nu}, \nonumber\\ &= T_0^{\mu\nu} + \delta T_{\rm id}^{\mu\nu} + \delta\pi^{\mu\nu} + \Xi^{\mu\nu} \equiv T_0^{\mu\nu} + \delta T^{\mu\nu}. \end{align} The total fluctuating part $\delta T^{\mu\nu}$ has contributions from the viscous term $\delta\pi^{\mu\nu}$, the noise term $\Xi^{\mu\nu}$, and the ideal energy-momentum tensor term \begin{align}\label{DTid:eq} \delta T_{\rm id}^{\mu\nu} =& \delta\epsilon \: u_0^\mu u_0^\nu - \delta p \: \Delta^{\mu\nu}_0 \nonumber\\ & + \left(\epsilon_0 + p_0 \right)\left( u_0^\mu \delta u^\nu + \delta u^\mu u_0^\nu \right) + {\cal O}(\delta^2). \end{align} All of these can be determined by the fluctuating variables ($\delta\epsilon, \delta u^\mu, \delta \pi^{\mu\nu}$). The conservation equations for the total energy-momentum tensor, $\partial_\mu T^{\mu\nu}=0$, along with the usual conservation for the average part, $\partial_\mu T_0^{\mu\nu}=0$, lead to \begin{align}\label{FCT:eq} \partial_\mu (\delta T_{\rm id}^{\mu\nu} + \delta\pi^{\mu\nu} + \Xi^{\mu\nu}) \equiv \partial_\mu (\delta T^{\mu\nu}) = 0. \end{align} Though in a single event thermal noise causes $\delta T^{\mu\nu} \neq 0$, the average over many events results in $\langle \delta T^{\mu\nu} \rangle = 0$. However, noise induces a nonvanishing correlator $\langle \delta T^{\mu\nu} \delta T^{\alpha\beta} \rangle$, that contributes to event-by-event distribution of an observable, e.g., two-particle rapidity correlations \cite{}. In order to solve (numerically) Eq. (\ref{FCT:eq}), which involves the evolution of viscous fluctuation and noise, one requires the averaged quantities ($\epsilon_0, u_0^\mu, \pi_0^{\mu\nu}$). We first deal with the background viscous evolution equations and then formulate the evolution of fluctuation on top this background. For Bjorken longitudinal expansion, we work in the Milne coordinates ($\tau,x,y,\eta$) where proper time $\tau = \sqrt{t^2-z^2}$, space-time rapidity $\eta = \ln[(t+z)/(t-z)]/2$, and four-velocity $u_0^\mu = (1,0,0,0)$. The conservation equation for the average part of the energy-momentum tensor, $\partial_\mu T_{0\: id}^{\mu\nu} = -\partial_\mu \pi_0^{\mu\nu}$, gives the evolution equation of noiseless $\epsilon_0$ as \begin{align}\label{energy:eq} \frac{d\epsilon_0}{d\tau} = -\frac{1}{\tau} \left(\epsilon_0 + p_0 - \pi_0 \right), \end{align} where $\tau^2\pi_0^{\eta\eta} \equiv -\pi_0$ is taken as the independent component of the shear pressure tensor. For the three independent variables, we need two more equations, namely, the viscous evolution equation and the equation of state. The simplest choice for the dissipative equation would be the relativistic Navier-Stokes theory, where the instantaneous constituent equation for the shear pressure is $\pi^{\mu\nu} = 2\eta_v \nabla^{\langle \mu}u^{\nu\rangle} \equiv 2\eta_v \sigma^{\mu\nu}$. Using Eq. (\ref{nonlin:eq}), the average (noiseless) shear part in the Bjorken case becomes \begin{align}\label{NS:eq} \pi_0 = \frac{4\eta_v}{3}\theta_0 , \end{align} where $\eta_v \geq 0$ is the shear viscosity coefficient, and $\nabla^{\langle \mu}u^{\nu\rangle} = (\nabla^\mu u^\nu + \nabla^\nu u^\mu)/2 - (\nabla \cdot u) \Delta^{\mu\nu}/3$ and $\nabla^\mu = \Delta^{\mu\nu}\partial_\nu$. For boost-invariant case, the local expansion rate and the time derivative in the LRF are $\theta_0 = 1/\tau$. In the following we shall use the standard notation $A^{\langle\mu\nu\rangle}\equiv \Delta^{\mu\nu}_{\alpha\beta}A^{\alpha\beta}$ for traceless symmetric projection orthogonal to $u^{\mu}$, where $\Delta^{\mu\nu}_{\alpha\beta} \equiv (\Delta^\mu_\alpha \Delta^\nu_\beta + \Delta^\mu_\beta \Delta^\nu_\alpha)/2 - (1/3)\Delta^{\mu\nu}\Delta_{\alpha\beta}$. The most commonly used second-order dissipative hydrodynamic equation derived from positivity of the divergence of entropy four-current is based on the works of M\"uller-Israel-Stewart (MIS) \cite{}: \begin{align}\label{IS:eq} \dot\pi^{\langle \mu\nu \rangle} =& - \frac{1}{\tau_\pi} \left( \pi^{\mu\nu} - 2\eta_v \nabla^{\langle \mu}u^{\nu\rangle} \right) \nonumber \\ & -\frac{1}{2} \pi^{\mu\nu} \frac{\eta_v T}{\tau_\pi} \partial_{\lambda} \left( \frac{\tau_\pi}{\eta_v T} u^\lambda \right) , \end{align} where the above equation involves the full hydrodynamic variables, and $T = T_0 + \delta T$ is the total temperature corresponding to $\epsilon$. In contrast to the first-order equation, the above equation restores causality by enforcing the shear pressure to relax to its first-order value via the relaxation time $\tau_\pi = 2\eta_v \beta_2$, where $\beta_2$ is the second-order transport coefficient. In the boost-invariant scaling expansion, the noiseless part of the dissipative Eq. (\ref{IS:eq}) reduces to \begin{align}\label{MIS:eq} \frac{d\pi_0}{d\tau} + \frac{\pi_0}{\tau_\pi} = \frac{4\eta_v}{3\tau_\pi} \theta_0 - \lambda_\pi\theta_0\pi_0, \end{align} where terms up to second-order in the velocity gradients are kept in the expansion of the last term in Eq. (\ref{IS:eq}). The resulting expansion coefficient is \begin{align}\label{lpi:eq} \lambda_\pi = \frac{1}{2} \left[ 1 + \frac{\epsilon_0 + p_0}{T_0} \frac{dT_0}{d\epsilon_0} \left(1 - \frac{T_0}{\beta_2} \frac{d\beta_2}{dT_0} \right) \right], \end{align} which for an ultrarelativistic EOS reduces to $\lambda_\pi = 4/3$. The relaxation time $\tau_\pi$ depends on the underlying microscopic theory namely, weakly coupled QCD, lattice QCD, and ${\cal N}=4$ SYM \cite{}. For all these theories, one can express $\tau_\pi = \chi \eta_v/(sT_0)$, where the coefficient $2 \lesssim \chi \lesssim 6$. In the present study we consider $\tau_\pi = 2\eta_v \beta_2 = 5\eta_v/(sT_0)$. Hereafter $\eta_v/s$ is kept fixed, where $s = s_0 + \delta s$ is the total entropy density in the linearized limit with $s_0$ being the average entropy density. Alternatively, dissipative evolution equations can be derived from Chapman-Enskog-like (CE) gradient expansion of the nonequilibrium distribution function about the local value, using Knudsen number as a small expansion parameter \cite{}. The relativistic Boltzmann equation, in the relaxation-time approximation for the collision term, can be solved iteratively to yield \begin{equation}\label{SOSHEAR:eq} \dot{\pi}^{\langle\mu\nu\rangle} \!+ \frac{\pi^{\mu\nu}}{\tau_\pi}\!= \frac{\sigma^{\mu\nu}}{\beta_2} \!+2\pi_\gamma^{\langle\mu}\omega^{\nu\rangle\gamma} \!-\frac{10}{7}\pi_\gamma^{\langle\mu}\sigma^{\nu\rangle\gamma} \!-\frac{4}{3}\pi^{\mu\nu}\theta, \end{equation} where $\omega^{\mu\nu}\equiv(\nabla^\mu u^\nu-\nabla^\nu u^\mu)/2$ is the vorticity tensor. In the boost-invariant case, the noiseless part of Eq. (\ref{SOSHEAR:eq}) gives \begin{align}\label{CE:eq} \frac{d\pi_0}{d\tau} + \frac{\pi_0}{\tau_\pi} = \frac{4\eta_v}{3\tau_\pi} \theta_0 - \lambda_\pi\theta_0\pi_0. \end{align} In the Chapman-Enskog-like approach, the relaxation time naturally comes out to be $\tau_\pi = 2\eta_v\beta_2 = 5\eta_v/(s T_0)$ and $\lambda_\pi = 38/21$ \cite{}. In this limit the CE equation is equivalent to that obtained by Denicol et. al. \cite{} where the expansion is controlled by the Knudsen number and the inverse Reynolds number. We shall explore the effects of the above viscous equations on the thermal noise correlators and the two-particle rapidity correlations. For the equation of state (EOS), we have employed the conformal QGP fluid with the thermodynamic pressure $p = \epsilon/3$, and the s95p-PCE EOS \cite{} which is obtained from fits to lattice data for crossover transition and matches a realistic hadron resonance gas model at low temperatures $T$, with partial chemical equilibrium (PCE) of the hadrons at temperatures below $T_{\rm PCE} \approx 165$ MeV. The EOS influences the longitudinal expansion of the fluid and the two-particle correlations. In order to obtain the evolution equations for fluctuations, we use the normalization $u^\mu u_\mu = (u_0^\mu + \delta u^\mu)(u_{0\mu} + \delta u_{\mu}) = 1$, orthogonality $\pi^{\mu\nu} u_\nu =0$, and tracelessness $\pi^\mu_\mu =0$. It is important to note that noise breaks the boost invariance, as a result of which the fluctuating quantities depend explicitly on both space-time rapidity and proper time. Thus the three independent variables are $\delta\epsilon \equiv \delta\epsilon(\tau,\eta)$, $\delta u^\eta \equiv \delta u^\eta(\tau,\eta)$, $\delta \pi^{\eta\eta} \equiv \delta \pi^{\eta\eta}(\tau,\eta) = -(\delta\pi^{xx} + \delta\pi^{yy})/\tau^2$. Further, since $\Xi^{\mu\nu}$ satisfies the same constraints as $\pi^{\mu\nu}$, viz. $u_\mu \Xi^{\mu\nu} =0$ and $\Xi^\mu_\mu =0$, this results in one independent component, which we take as $\Xi^{\eta\eta}$. The fluctuating part of the energy-momentum conservation equation (\ref{FCT:eq}) can then be written as \begin{align} \label{FTevol:eq} &\frac{\partial}{\partial\tau} (\tau \delta \epsilon) + \frac{\partial}{\partial\eta} (\tau {\cal U}_0 \delta u^\eta) = -\delta {\cal V}, \\ \label{FT1evol:eq} &\frac{\partial}{\partial\tau}(\tau {\cal U}_0 \delta u^\eta) + \frac{\partial}{\partial\eta} \left(\frac{\cal \delta V}{\tau} \right) = - 2 {\cal U}_0 \delta u^\eta , \end{align} where ${\cal U}_0(\tau) \equiv \epsilon_0 + p_0 - \pi_0 = w_0 - \pi_0$ depends on the background variables that are functions of proper time only; $w_0$ is the enthalpy of the fluid. On the other hand, ${\cal \delta V}(\eta,\tau) \equiv \delta p + \tau^2 \delta\pi'^{\eta\eta}$ consists of the fluctuating variables which depend on the space-time rapidity as well. We have introduced a stochastic variable \begin{align}\label{pip:eq} \delta \pi'^{\eta\eta} = \delta\pi^{\eta\eta} + \Xi^{\eta\eta} \equiv - \delta \pi'/\tau^2, \end{align} whose evolution will be derived below. The stochastic part of the dissipative equation corresponding to MIS or CE, can be obtained from Eq. (\ref{IS:eq}) or (\ref{SOSHEAR:eq}) by using the linearization Eq. (\ref{nonlin:eq}). For Bjorken expansion, the evolution equation for the independent fluctuating component, $\delta\pi'$, reads \begin{align}\label{Fpievol:eq} \frac{\partial\delta\pi'}{\partial\tau} + \frac{\delta\pi'}{\tau_\pi} =& \frac{1}{\tau_\pi} \left[ \tau^2 \xi^{\eta\eta} + \frac{4\eta_v}{3s} \left(s_0\delta\theta + \delta s\theta_0 \right) \right] \nonumber\\ & -\lambda_\pi \left( \theta_0 \delta \pi' + \delta\theta \pi_0 \right) \nonumber \\ & -\frac{\delta\tau_\pi}{\tau_\pi} \left( \lambda_\pi\theta_0\pi_0 + \frac{d\pi_0}{d\tau} \right). \end{align} The new noise term $\xi^{\eta\eta}$ defines the equation of motion of $\Xi^{\eta\eta}$, which for the MIS equation is \begin{align}\label{ximis:eq} \dot \Xi^{\langle\eta\eta\rangle} = - \frac{1}{\tau_\pi} \left( \Xi^{\eta\eta} - \xi^{\eta\eta} \right) - \lambda_\pi \Xi^{\eta\eta}\theta , \end{align} and for the CE equation is \begin{align}\label{xice:eq} \dot \Xi^{\langle\eta\eta\rangle} = - \frac{1}{\tau_\pi} \left( \Xi^{\eta\eta} - \xi^{\eta\eta} \right) - \frac{10}{7} \Xi^{\langle\eta}_\gamma \sigma^{\eta \rangle\gamma} - \lambda_\pi \Xi^{\eta\eta}\theta . \end{align} We recall that in the derivation of Eq. (\ref{Fpievol:eq}), the ratio of the shear viscosity and the total entropy density $\eta_v/s$ is kept fixed during the entire evolution. The local expansion rate of the fluid due to velocity fluctuation is of the form $\delta \theta \equiv \partial \cdot \delta u^\eta = \partial_\eta \delta u^\eta$. The variation of the relaxation time $\tau_\pi$ due to thermal fluctuation is \begin{align}\label{vartpi:eq} \delta\tau_\pi = \delta \left( 2\eta_v\beta_2 \right) = - \tau_\pi \frac{\delta T}{T_0}. \end{align} Equation (\ref{Fpievol:eq}) involves the noise term $\xi^{\eta\eta}$ that generates the fluctuations, which in turn, evolve via the fluctuating hydrodynamic equations. The equations are closed once the noise $\xi^{\eta\eta}$ (or equivalently $\Xi^{\eta\eta}$) is specified. The autocorrelation of $\Xi^{\mu\nu}$ can be obtained using Eq. (\ref{FCT:eq}): \begin{align}\label{auotocor:eq} \langle \partial_\mu \Xi^{\mu\nu}(x) \partial_\alpha \Xi^{\alpha\beta}(x') \rangle =& \langle \partial_\mu \left(-\delta T_{\rm id}^{\mu\nu} - \delta\pi^{\mu\nu}\right)(x) \nonumber\\ &\times \partial_\alpha \left(-\delta T_{\rm id}^{\alpha\beta} - \delta\pi^{\alpha\beta}\right)(x') \rangle, \end{align} along with the use of modes in the dissipative hydrodynamic equations and also employing the fluctuation-dissipation theorem \cite{}. Alternatively, the autocorrelations can also be derived from the non-equilibrium entropy four-current and using the fluctuation-dissipation theorem \cite{}. In the theory of quasi-stationary fluctuations \cite{Landau}, the rate of change of total entropy can be expressed as $dS/dt = - \sum_a \dot{x_a} X_a$, where the generalized forces $X_a = - \partial S/\partial x_a$ are conjugate to the set of variables $x_a$. For a system close to equilibrium, the evolution of $x_a$ may be approximated as \begin{equation}\label{Onsager:eq} \dot{x}_a = - \sum_b \gamma_{ab}X_b + y_a , \end{equation} where $\gamma_{ab}$ are the Onsager coefficients. The random fluctuations $y_a$ then satisfy the autocorrelations $\langle y_a(t) y_b(t')\rangle = (\gamma_{ab}+\gamma_{ba})\delta(t-t')$. In terms of the nonequilibrium part of total energy-momentum tensor, $\pi'^{\mu\nu} = \pi^{\mu\nu} + \Xi^{\mu\nu}$, the entropy four-current (up to second-order) in the MIS and CE theories can be written as \cite{} \begin{equation}\label{S_MIS:eq} S^{\mu} = s u^{\mu} - \frac{\beta_2}{2T}u^{\mu}\pi'^{\alpha\beta}\pi'_{\alpha\beta}, \end{equation} where the equilibrium entropy density $s = (\epsilon + p)/T$. From the total (average plus noise) conservation $\partial_{\mu}T^{\mu\nu} = 0$, one obtains from Eq. (\ref{S_MIS:eq}) \begin{equation} \frac{dS}{dt} = \int d^3 \! x \frac{\pi'^{\mu\nu}}{T} \left[ \nabla_{\mu} u_{\nu} - \beta_2 \dot{\pi}'_{\mu\nu} - \beta_2 \lambda_\pi \theta \pi'_{\mu\nu} \right]. \end{equation} Identifying \begin{align}\label{ide} \dot x_a \to & \pi'^{\mu\nu} \nonumber \\ X_b \to & - \frac{1}{T} \left[\nabla_\mu u_\nu - \beta_2\dot{\pi}'_{\mu\nu} - \beta_2 \lambda_\pi \theta \pi'_{\mu\nu} \right] \equiv X_{\mu\nu}, \nonumber \end{align} and in analogy with Eq. (\ref{Onsager:eq}), one can write \begin{equation}\label{Landau:eq} \pi'^{\mu\nu} = - \gamma^{\mu\nu\alpha\beta} X_{\alpha\beta} + \xi^{\mu\nu}. \end{equation} Owing to symmetries of $\pi'^{\mu\nu}$, one gets $\gamma^{\mu\nu\alpha\beta} = \gamma^{\nu\mu\alpha\beta}$, $\gamma^{\mu\alpha\beta}_{\mu} = 0$, and $\gamma^{\mu\nu\alpha\beta} u_\mu = 0$. Note that the identification of $X_{\mu\nu}$ is not unique as the transformation $X_{\mu\nu} \to X_{\mu\nu} + H_{\mu\nu}$, keeps $dS/dt$ invariant, where $H_{\mu\nu}$ is any tensor orthogonal to $\pi'^{\mu\nu}$. To obtain an autocorrelation which is insensitive to such transformations, the above constraints for $\gamma^{\mu\nu\alpha\beta}$ with respect to $\mu,\nu$ indices should also follow for the $\alpha, \beta$ indices. In the MIS theory, we obtain $\gamma^{\mu\nu\alpha\beta}$ by comparing Eq. (\ref{IS:eq}) (using $\pi^{\mu\nu} = \pi'^{\mu\nu} - \Xi^{\mu\nu}$) and Eq. (\ref{Landau:eq}): \begin{align} \gamma^{\mu\nu\alpha\beta} = 2 \eta_v T \Delta^{\mu\nu\alpha\beta}. \end{align} From Eqs. (\ref{Onsager:eq}) and (\ref{Landau:eq}) and using the above expression of $\gamma^{\mu\nu\alpha\beta}$, the autocorrelation in the MIS theory can be written as \begin{align}\label{MInois:eq} \langle \xi^{\mu\nu}(x) \xi^{\alpha\beta}(x') \rangle = 4\eta_v T \Delta^{\mu\nu\alpha\beta} \; \delta^4(x-x'). \end{align} Similarly, to obtain the autocorrelation in the Chapman-Enskog (CE) theory we compare Eq. (\ref{Landau:eq}) with Eq. (\ref{SOSHEAR:eq}) for $\pi'^{\mu\nu}$. The $\gamma^{\mu\nu\alpha\beta}$, that is consistent with the constraints as stated above, is found to be \begin{align} \gamma^{\mu\nu\alpha\beta} =& 2 \eta_v T \Big( \Delta^{\mu\nu\alpha\beta} - \frac{10}{7}\beta_2 \Delta^{\mu\nu}_{\zeta\kappa} \pi'^\zeta_\gamma \Delta^{\kappa\gamma\alpha\beta} \nonumber \\ & + 2 \tau_\pi \Delta^{\mu\nu}_{\zeta\kappa} \omega^\zeta_\gamma \Delta^{\kappa\gamma\alpha\beta} \Big). \end{align} Note that the second and third terms in the above equation reproduce $\pi'^{\langle\mu}_\gamma \sigma^{\nu\rangle\gamma}$ and $\omega^{\langle\mu}_\gamma \sigma^{\nu\rangle\gamma}$, respectively, in the CE theory. Moreover, these terms also give (via contraction with $X_{\alpha\beta}$) additional higher order terms which can be neglected in our second-order formalism. Correspondingly one obtains the noise autocorrelation: \begin{align}\label{CEnois:eq} \langle \xi^{\mu\nu}(x) \xi^{\alpha\beta}(x') \rangle = & 4\eta_v T \Big( \Delta^{\mu\nu\alpha\beta} - \frac{5}{7}\beta_2 \Delta^{\mu\nu}_{\zeta\kappa}\pi'^\zeta_\gamma \Delta^{\kappa\gamma\alpha\beta} \nonumber \\ & - \frac{5}{7}\beta_2 \Delta^{\alpha\beta}_{\zeta\kappa}\pi'^\zeta_\gamma \Delta^{\kappa\gamma\mu\nu} \nonumber \\ & + \omega-{\rm terms} \Big) \delta^4(x-x'). \end{align} In the boost-invariant case, the autocorrelation for the independent component $\xi^{\eta\eta}$ in the MIS (Eq. (\ref{MInois:eq})) and CE (Eq. (\ref{CEnois:eq})) dissipative formalisms reduce to \begin{align}\label{Bjnoise:eq} \langle \xi^{\eta\eta}(\eta,\tau) \xi^{\eta\eta}(\eta',\tau') \rangle = & \frac{8 \eta_v(\tau) T_0(\tau)}{3 A_\perp \tau^5} \left[1 - {\cal A}\beta_2\pi_0 \right] \nonumber \\ & \times \delta(\tau-\tau')\delta(\eta-\eta'). \end{align} Note that the autocorrelation depends only on the background quantities as we have treated the noise as a perturbation on top of background evolution. The coefficient ${\cal A} = 0$ in the MIS theory, and ${\cal A} = 5/7$ in the CE formalism. The delta function in the transverse direction $\delta({\bf x} - {\bf x'})\delta({\bf y} - {\bf y'}) = 1/A_\perp$ is represented by the inverse of the transverse area $A_\perp$ of the colliding nuclei. The random variable $\xi^{\eta\eta}(\eta,\tau)$ is the stochastic source that obeys the energy-momentum conservation, and propagates each fluctuation $\delta T^{\mu\nu}$ up to later times to their thermal expectation values. In the Navier-Stokes limit, one can show \cite{} that the autocorrelation for $\Xi^{\eta\eta}$ has an identical form of Eq. (\ref{Bjnoise:eq}) with ${\cal A} = 0$. It may be mentioned that the autocorrelation has nonvanishing values in the transverse directions, $\langle \Xi^{ii}(\eta,\tau) \Xi^{ii}(\eta',\tau') \rangle ~(i\equiv x,y)$. Consequently, for boost-invariant longitudinal expansion with transverse symmetry, a perturbation generated at any space-time point in the fluid will propagate also in the transverse direction with the sound velocity of the medium. In the present study, we have ignored such transverse motion of these ``ripples" \cite{}. The hydrodynamic fluctuation Eqs. (\ref{FTevol:eq})-(\ref{Fpievol:eq}) are solved perturbatively in the $\tau$-$\eta$ coordinates using the MacCormack (a predictor-corrector type) method. \section{Freeze-out and two-particle rapidity correlations} The freeze-out of a near-thermalized fluid to a free-streaming (noninteracting) particles can be obtained via the standard Cooper-Frye prescription \cite{}. For a boost-invariant scenario without fluctuations, freeze-out on a hypersurface of constant temperature would be equivalent to freeze-out at a constant proper time. Inclusion of fluctuation, breaks the boost invariance of the system. In an event, the total temperature would be the sum of constant background temperature and the fluctuating temperature which varies for different cells. We shall consider freeze-out at a constant background temperature $T_f$ so that for any hydrodynamic variable, $X(\tau_f,\eta)= X_0(\tau_f) + \delta X (\eta,\tau_f)$, the fluctuating field $\delta X(\eta,\tau_f)$ varies on the hypersurface; $\tau_f$ is the freeze-out time corresponding to $T_f$. For such isothermal (and isochronous) freeze-out at a constant background $T_f$, the particle spectrum can be obtained from \begin{align}\label{CF:eq} E\frac{dN}{d^3p} = \frac{g}{(2\pi)^3} \int_\Sigma d\Sigma_\mu p^\mu f(x,p), \end{align} where $p^\mu$ is the four-momentum of the particle with degeneracy $g$ and $d\Sigma^\mu$ is the outward-directed normal vector on an infinitesimal element of the hypersurface $\Sigma(x)$. In the present ($\tau,x,y,\eta$) coordinate system, the three-dimensional volume element at freeze-out is \begin{align}\label{volel} d\Sigma_\mu =& \tau_f \left(\cosh\eta, {\bf 0}, - \sinh\eta \right) d\eta d{\bf x}_\perp. \end{align} The four-momentum of the particles is \begin{align}\label{4mom} p^\mu = (m_T\cosh y, {\bf p}_T, m_T\sinh y), \end{align} where $m_T = \sqrt{p_T^2 + m^2}$ is the transverse mass of the particle with transverse momentum $p_T$ and kinetic rapidity $y =\tanh^{-1}(p_z/p_0)$. The integration measure at the constant temperature freeze-out hypersurface $\Sigma(x)$ is then $p^\mu d\Sigma_\mu = m_T \: \cosh(y-\eta) \tau_f d\eta d{\bf x}_\perp$. The phase-space distribution function at freeze-out, $f(x,p) = f_{\rm eq}(x,p) + f_{\rm vis}(x,p)$ contains the equilibrium contribution \begin{align}\label{feq:eq} f_{\rm eq} = {\rm exp}[p \cdot u/T \pm 1]^{-1} \approx {\rm exp}(-p\cdot u)/T. \end{align} The nonequilibrium viscous correction in the MIS theory has the form based on Grad's 14-moment approximation \cite{Grad}: \begin{align}\label{fvis_MIS:eq} f_{\rm vis} = f_{\rm eq}(1 \mp f_{\rm eq}) \frac{p^\mu p^\nu \pi'_{\mu\nu}}{2(\epsilon+p)T^2} \approx f_{\rm eq} \frac{p^\mu p^\nu \pi'_{\mu\nu}}{2(\epsilon+p)T^2}. \end{align} The last expression in Eq. (\ref{fvis_MIS:eq}) is valid in the large temperature limit, and the total values (noiseless plus noise) for the hydrodynamic variables $X \equiv X(\tau_f,\eta)$ are evaluated at the freeze-out hypersurface coordinates. In the linearized limit, the total distribution function $f(x,p)$ can be expanded as \begin{align}\label{f:eq} f(x,p) = f_0(x,p) + \delta f(x,p). \end{align} The noiseless part of the distribution function $f_0(x,p)$ has contributions from ideal and viscous fluctuations: \begin{align}\label{f0:eq} f_0 =& (f_{\rm eq})_0 \left( 1 + {K_0}_{\mu \nu} \pi_0^{\mu\nu} \right), \end{align} where $K_0^{\mu\nu} = p^\mu p^\nu[2(\epsilon_0+p_0)T_0^2)]^{-1}$, and the total temperature $T=T_0 + \delta T$. We recall that $T_0 \equiv T_f$ for freeze-out at a constant background temperature. Similarly, the noise part $\delta f(x,p)$ can be written as the sum of ideal and viscous fluctuations. Using the linearization Eq. (\ref{nonlin:eq}), this becomes \begin{align}\label{f1:eq} \delta f =& \delta f_{\rm eq} + {K_0}_{\mu\nu} \Big[ \delta f_{\rm eq} \: \pi_0^{\mu\nu} + (f_{\rm eq})_0 \: \delta \pi'^{\mu\nu} \nonumber \\ & - (f_{\rm eq})_0 \: \pi_0^{\mu\nu} \left( 2\frac{\delta T}{T_0} + \frac{\delta \epsilon+\delta p}{\epsilon_0+p_0} \right) \Big] , \end{align} where $(f_{\rm eq})_0 = {\rm exp}(-u_0^\mu p_\mu/T_0)$ and $\delta f_{\rm eq} = (f_{\rm eq})_0 (\delta T u_0^\mu p_\mu/T_0^2 - \delta u^\mu p_\mu/T_0)$ are, respectively, the noiseless and the noise parts of the ideal distribution function, and the terms inside the square brackets in Eq. (\ref{f1:eq}) stem from viscous fluctuations. The rapidity distribution of the particle, corresponding to Eq. (\ref{CF:eq}), then reduces to \begin{align}\label{rapidity:eq} \frac{dN}{dy} =& \frac{g\tau_f A_\perp}{(2\pi)^3} \int d\eta \: \cosh(y-\eta) \nonumber\\ & \times \int dp_x dp_y \: m_T [f_0(x,p) + \delta f(x,p)] \nonumber\\ & \equiv (dN/dy)_0 + \delta (dN/dy) . \end{align} Here $A_\perp = \int d{\bf x}_\perp$ is the usual transverse area of Eq. (\ref{Bjnoise:eq}). For the boost-invariant longitudinal flow, the particle rapidity distribution of the average part can be written as \begin{align}\label{rap0:eq} \left(\frac{dN}{dy}\right)_{\!\! 0} =& ~\frac{g\tau_f T_0^3 A_\perp}{(2\pi)^2} \int \frac{d\eta}{\cosh^2(y-\eta)} \: \Big[ \Gamma_3(y-\eta) \nonumber \nonumber \\ & + \frac{\pi_0}{4 w_0} \left({\cal C}(y-\eta) \Gamma_5(y-\eta) - \frac{m^2}{T_0^2}\Gamma_3(y-\eta) \right) \Big]. \end{align} Here $\Gamma_k(x) \equiv \Gamma(k, m\cosh x /T_0)$ denotes the incomplete Gamma function of the $k$th kind \cite{Stegun} and ${\cal C}(x) = 3~{\rm sech}^2x -2$. The second term within the brackets corresponds to viscous corrections. The fluctuating parts can be expressed as \begin{align}\label{rapf:eq} \delta\frac{dN}{dy} &= ~\frac{g\tau_f T_0^3 A_\perp}{(2\pi)^2} \int d\eta \: \Big[ {\cal F}_T (y-\eta) \frac{\delta T(\eta)}{T_0} \nonumber\\ &+ {\cal F}_u (y-\eta) \tau_f \delta u^\eta(\eta) + {\cal F}_\pi (y-\eta) \frac{\delta\pi'(\eta)}{w_0} \Big]. \end{align} Here ${\cal F}_{T,u,\pi}$ are the coefficients of the fluctuations, ($\delta T, \delta u^\eta, \delta \pi'$), that are obtained by performing the momentum integrals. In the MIS theory these are given by \begin{align}\label{Coef_Tup:eq} {\cal F}_T \cosh^2 x =& \Gamma_4(x) - \frac{\pi_0}{4w_0} \Big[ \frac{m^2}{T_0^2}\left(\Gamma_4(x) - \kappa\Gamma_3(x) \right) \nonumber\\ & - {\cal C}(x) \left(\Gamma_6(x) - \kappa\Gamma_5(x) \right) \Big], \\ {\cal F}_u \cosh^2 x =& \tanh x ~\Gamma_4(x) - \frac{\pi_0}{4w_0} \tanh x ~\Big[ \frac{m^2}{T_0^2}\Gamma_4(x) \nonumber \\ & - {\cal C}(x)\Gamma_6(x) - 4\Gamma_5(x) \Big], \\ {\cal F}_\pi \cosh^2 x =& \frac{1}{4} \Big[{\cal C}(x) \Gamma_5(x) - \frac{m^2}{T_0^2}\Gamma_3(x) \Big], \end{align} where $\kappa = 2 + (T_0/w_0)\partial w_0 /\partial T_0$. The two-particle rapidity correlator due to fluctuations can then be written as \begin{align}\label{FRapCor:eq} \left\langle \delta\frac{dN}{dy_1} \ \delta\frac{dN}{dy_2} \right\rangle =& \left[ \frac{g\tau_f T_0^3 A_\perp}{(2\pi)^2} \right]^2 \int \! d\eta_1 \int \! d\eta_2 \nonumber \\ & \times \sum_{X,Y} {\cal F}_X (y_1-\eta_1){\cal F}_Y (y_2-\eta_2) \nonumber\\ & \times \langle X(\eta_1) Y(\eta_2) \rangle . \end{align} Here $(X,Y) \equiv (\delta T, \delta u^\eta, \delta \pi')$ and $\langle X(\eta_1) Y(\eta_2) \rangle$ are the two-point correlators between the fluctuating variables calculated at the freeze-out hypersurface. In the Chapman-Enskog-like approach of iteratively solving Boltzmann equation, the viscous correction in the nonequilibrium distribution function has the form \cite{} \begin{align}\label{fvis_CE:eq} f_{\rm vis} \approx f_{\rm eq} \frac{5 p^\mu p^\nu \pi'_{\mu\nu}}{8pT(u.p)}, \end{align} with a total flow velocity $u^\mu \equiv u^\mu(\tau_f,\eta)$. Following similar procedure as done in the MIS theory, the two-particle rapidity correlations in the CE formalism give the same form as in Eq. (\ref{FRapCor:eq}) but with modified coefficients ${\cal F}_{T,u,\pi}$. \section{Results and discussions} We shall explore two-particle rapidity correlations induced by thermal fluctuations in the Bjorken expansion. A clear understanding of this can be achieved by calculating the time evolution of the correlations among the hydrodynamic variables. The fluctuations employed in our study generates singularities in the correlators at zero separation in rapidity and at the sound horizons that corresponds to maximum distance propagated by the sound wave along rapidity. In the Navier-Stokes theory, the singular and regular parts of the correlators can be obtained analytically \cite{}; see Appendix A. Figure 1 shows the rapidity dependence of these equal-time correlators for inviscid fluid $\langle X(\tau,\eta_1) Y(\tau,\eta_2) \rangle$ with $(X,Y) \equiv (\delta T, \delta u^\eta)$. Shear viscosity is neglected in the evolution but accounted for in the noise correlator $\Xi^{\eta\eta}$ of Eq. (\ref{Bjnoise:eq}). These correlators represent a wake of the medium behind the shock front associated with the noise propagation. The regular part of the temperature-temperature correlator, $\langle \delta T \delta T \rangle$, peaks at zero separation due to short-range correlation which builds up rapidly with increasing proper time. At later times, the peak value decreases and the correlator spreads farther in rapidity due to expansion of the fluid. The time evolution of these structures in the Bjorken expansion is to be contrasted with that of a uniform static system; see Appendix B. The equal-time long-range correlations in the static fluid vanish and correlations at later times are due to propagation of sound waves in opposite directions. This clearly underscores the importance of underlying background flow that influences the propagation of fluctuations. \begin{figure}[t] \includegraphics[width=\linewidth,height=4.0in]{Corr_eta_NS.eps} \caption{Regular and singular parts of the equal time temperature-temperature, velocity-velocity and temperature-velocity correlators computed as a function of space-time rapidity difference $\Delta\eta$ at various proper times in the Navier-Stokes theory. The results are for ultrarelativistic gas EOS ($p=\epsilon/3$) with initial temperature of $T_0 = 550$ MeV, proper time $\tau_0 = 0.2$ fm/c. The shear viscosity to entropy density ratio of $\eta_v/s = 1/4\pi$ is accounted only in the noise correlator.} \label{} \end{figure} In Fig. \ref{}, we also show the regular part of the velocity-velocity correlator, $\langle \delta u^\eta \delta u^\eta \rangle$. It exhibits a similar rapidity dependence as seen in $\langle \delta T \delta T \rangle$, however with a much smaller magnitude. In contrast, the regular part of the temperature-velocity correlator, $\langle \delta T \delta u^\eta \rangle$ is an odd function in $\Delta\eta$. As a consequence this correlator vanishes at $\Delta\eta =0$ and turns negative (positive) for positive (negative) values of rapidity separation. Note that the ``cross" correlators follow $\langle \delta T \delta u^\eta \rangle =-\langle \delta u^\eta \delta T\rangle$. Analytic results for the singular part of the correlator can be obtained in the Navier-Stokes theory for an ultra-relativistic gas EOS, see Appendix A. Figure 1 shows the singular part of the equal-time correlators wherein the theta function and its higher-order derivatives have been smeared using a normalized Gaussian distribution of width $\sigma_\eta =0.2$. In all these correlators, the singularities at $\Delta\eta = 0$ arise from self-correlations, and those at large rapidity separations are induced by sound horizons at $\Delta\eta = \pm 2c_s\log(\tau/\tau_0)$. We note that inclusion of viscosity would dampen the singularities and thereby smear the structures in the longitudinal correlations. As analytic solutions for thermal fluctuations in second-order dissipative hydrodynamics do not exist, the singular and regular parts of the correlators cannot be separated. However, in the numerical simulation of noise in second-order hydrodynamics, the smearing functions ${\cal F}_X$ in Eq. (\ref{FRapCor:eq}) smoothen out all the singularities in the total correlation function $\langle X(\tau,\eta_1) Y(\tau,\eta_2) \rangle$. Hence the computed two-particle rapidity correlation at freeze-out would show a clear structure. \begin{figure}[t] \includegraphics[width=\linewidth]{FFCor_MIS_eta.eps} \caption{Two-particle rapidity correlations from various fluctuations calculated at freeze-out for charged pions in the M\"uller-Israel-Stewart (MIS) dissipative hydrodynamics. The results are for ideal gas EOS ($p=\epsilon/3$) with initial temperature $T_0 = 550$ MeV, proper time $\tau_0 = 0.2$ fm/c, freeze-out temperature $T_f = 150$ MeV and shear viscosity to entropy density ratio $\eta_v/s = 1/4\pi$.} \label{} \end{figure} In the second-order M\"uller-Israel-Stewart viscous hydrodynamics, we present in Fig. \ref{} the various components, $X,Y \equiv \delta T, \delta u^\eta, \delta \pi$, of the rapidity correlators $\langle (\delta dN/dy_1) (\delta dN/dy_2) \rangle_{X,Y}$ (of Eq. (\ref{FRapCor:eq})) for charged pions as a function of kinematic rapidity separation $\Delta y = y_1-y_2$. This has been obtained by convoluting the two-point correlators $\langle X(\tau,y_1) Y(\tau,y_2) \rangle$, at freeze-out with the respective smearing functions ${\cal F}_{X,Y}$. The calculations are for initial values of temperature $T_0 = 550$ MeV, proper time $\tau_0 = 0.2$ fm/c and the freeze-out temperature is taken as $T_f = 150$ MeV. A constant $\eta_v/s = 0.08$ is used in both the average and noise parts of the evolution equations. The two-particle correlation functions are essentially manifestations of the sum of their regular and singular parts; see Fig. \ref{} for the correlators in the Navier-Stokes case. The smearing functions, namely, $F_{\delta T}$ (which is Gaussian about $\delta\eta=0$) and $F_{\delta u^\eta}$ (which peaks at $\Delta\eta \simeq \pm 1.5$ and vanishes at $\Delta\eta =0$) broadens these correlators when convoluted. While the peak at $\delta\eta=0$ is dominated by the temperature-temperature correlation function, the structures seen at $\Delta\eta \simeq 2-4$ for the $\delta T \delta T$, $\delta u^\eta \delta u^\eta$ and $\delta T \delta u^\eta$ correlations are similar in magnitude but have distinct rapidity dependence. The contributions to the correlation functions involving the viscous stress tensor $\delta\pi$ are found to be much smaller. \begin{figure}[t] \includegraphics[width=\linewidth]{N12_eta_Bj.eps} \caption{Correlation function for charged pions normalized with the single-particle rapidity distribution in the Navier-Stokes (NS), M\"uller-Israel-Stewart (MIS), and Chapman-Enskog (CE) formalisms for the evolution of thermal noise and compared with the ideal background (noiseless) hydrodynamic evolution. The initial and freeze-out conditions are same as in Fig. \ref{}.} \label{} \end{figure} In Fig. \ref{} we compare the total two-particle rapidity correlation for charged pions for the Navier-Stokes, M\"uller-Israel-Stewart and Chapman-Enskog viscous evolutions for an ultra-relativistic gas EOS ($p=\epsilon/3$). The initial and freeze-out conditions are the same as in Fig. \ref{}. We shall first consider the case of ideal hydrodynamics for the background evolution and explore various dissipative equations for the evolution of thermal fluctuations. The fluctuation in the Navier-Stokes theory gives rise to a larger peak at small rapidity separations as compared to that in the second-order viscous evolutions. This is mainly due to faster build-up of all the correlators and in particular the dominant temperature-temperature correlations in the first-order viscous evolution. In the Chapman-Enskog case, the correlation strength at $\Delta y \approx 0$ is smallest due to the larger coefficient $\lambda_\pi =38/21$, that results in a slower approach of the viscous fluctuations towards the Navier-Stokes limit. The inclusion of viscosity in the background evolution damps the correlation peak for all the cases studied. As expected, the maximum reduction in correlation strength occurs in the first-order theory. It may be mentioned that previous studies of rapidity correlations ignored the variation of the relaxation time $\delta \tau_\pi$ of Eq. (\ref{vartpi:eq}) \cite{}. We have found that such an assumption is justified as the rapidity correlation remains practically unaltered when thermal fluctuation of $\tau_\pi$ was not considered. \begin{figure}[t] \includegraphics[width=\linewidth]{N12_eta_Bj_lat.eps} \caption{Similar to Fig. \ref{} but with lattice QCD EOS; see text for details.} \label{} \end{figure} Figure \ref{} shows numerical results for two-particle rapidity correlation of charged pions in the NS, MIS, and CE formalisms for the lattice QCD EOS that incorporates the transition to a hadron resonance gas at $T_{\rm PCE} \approx 165$ MeV \cite{}. It may be mentioned that analytical results for the correlations cannot be obtained for the lattice EOS even in the Navier-Stokes limit. While the initial time and freeze-out temperature are considered the same as used for ideal gas EOS, the initial temperature is set at $T_0 = 378$ MeV. This choice stems from the consideration that the event-averaged single particle rapidity distribution for direct charged pions, $\langle dN/dy \rangle$, in this case is practically identical to that in the ideal gas EOS. Moreover, the freeze-out times for the lattice and conformal equation of states are found similar for each of the dissipative theories. We find that the magnitudes of the correlation between the particles are enhanced for all the cases in the lattice EOS as compared to that for ideal gas EOS. This can be understood as due to smaller sound velocity of the fluid near the critical temperature $T_c$ which slows the fluid expansion. Consequently, the correlation is solely from the short-range temperature-temperature correlator and the structures associated with the velocity and shear pressure correlators are largely damped and do not spread in rapidity. \section{Summary and Conclusions} We have studied the evolution of thermal fluctuations within relativistic second-order dissipative hydrodynamics. The fluctuations were treated in the linearized hydrodynamic framework as a perturbation on top of boost-invariant longitudinal expansion of matter. The analytic form of the autocorrelation function was found to be identical for the acausal Navier-Stokes and the causal M\"uller-Israel-Stewart theories. However, for the Chapman-Enskog-like dissipative equations, the correlation has an explicit dependence on the shear stress tensor. Within the analytically solvable Navier-Stokes limit in the Bjorken scenario, we demonstrated that the two-particle rapidity correlation at small rapidity separation, $\Delta y \lesssim 2$, is mostly due to temperature-temperature correlations and structures seen in the correlations at $\Delta y \approx 2-4$, are caused by varying contributions involving fluid velocity and shear pressure tensor correlations. In general, the two-particle rapidity correlations produced from thermal fluctuations were found to spread to large distance in rapidity separation with magnitude (and pattern) that can be well measured in relativistic heavy-ion collisions. While viscous damping of the correlation is at most $\sim 20\%$, there is further significant damping at small $\Delta y$, if one goes from the first-order Navier-Stokes theory to a second-order dissipative hydrodynamic formulation. As compared to the conformal equation of state, the softer lattice QCD EOS, with smaller sound velocity, causes reduced propagation of the fluctuations but leads to a pronounced peak in the rapidity correlations. \begin{acknowledgments} We thank Ananta Mishra for useful discussions during the initial stages of the work. We also thank Jean-Yves Ollitrault for illuminating discussions. \end{acknowledgments}

\section{Introduction} The Ruderman-Kittel-Kasuya-Yosida (RKKY) interaction is an indirect exchange interaction of external spins mediated by the virtual excitations of an electron gas\cite{kl14,kl15,kl16}. While the RKKY interaction is strongly suppressed in insulators due to the bulk band gap, this changes dramatically in topological insulators (TI) that feature inherent metallic surface states. These surface states are spin polarized and reflect the nontrivial topology of the bulk electronic structure, physics that will also be encoded in the RKKY interaction. However, for most topological insulators, which are rather complex materials, it is practically impossible to provide in analytical form the connection between the bulk and the surface topological states and hence to understand the surface RKKY coupling in terms of the bulk electronic structure. As a consequence, the RKKY interaction in topological insulators has, to date, been addressed on the basis of \emph{generic} Dirac-Weyl (or, equivalently, Rashba) Hamiltonians\cite{wang15,zyu14,qin14,lit14,efi14,chei12,li12,aba11,zhu11,liu09}. These investigations have uncovered a rich RKKY physics in the topological surface state which features a wide variety of coupling interactions - Ising, Heisenberg, as well as Dzyaloshinskii-Moriya (DM). Such an approach, however, excludes the possibility of analyzing the relation between the bulk electronic structure in the surface RKKY interaction. In particular, interesting questions such as how the rich RKKY behaviour of topological insulators might be tuned by manipulating the bulk physics, e.g. by alloying, cannot be answered. In this paper we remedy this situation by establishing an explicit link between bulk band gap states and the topological surface wavefunctions for the IV-VI crystalline topological insulators\cite{vol78,vol83,vol85,pan87,saf13,tan12,tan13,tan13a,bar13,dru14,shi14,pol14}, and on this basis analyze the impact of bulk physics on the RKKY interaction. We find that despite the universal form of the Dirac-Weyl topological surface state the spin structure, as probed by the RKKY interaction, is strikingly non-universal, and depends qualitatively on the crystal-field and spin-orbit parameters of the semiconductor bulk. As a specific example we show that for the well known SnTe topological insulator the form of the RKKY interaction can be strongly influenced by alloying with Pb in the bulk. \section{The bulk band structure model for the IV-VI compounds} We first briefly summarize the band structure model that we will use; for further details we refer the reader to Ref.~\citeonline{vol78,vol83}. The IV-VI materials crystallize in a cubic NaCl-type lattice with 6 $p$-bands around the Fermi energy (3 conduction and 3 valence bands) that originate from the atomic $p$-shells of the group IV and group VI species. \begin{figure}[ht] \begin{center} \includegraphics[width=0.6\linewidth]{./L-point-bandstructure.pdf} \caption{Schematic illustration of the band ordering at the $L$ points (a) showing the band inversion that occurs in SnTe due to crystal field and spin orbit coupling effects. (b) low energy band structure in vicinity of the L-points exhibiting the two linear surface bands (gray) that limit at the energy \(\epsilon_{max}\) tangentially to the upper bulk band \(\epsilon_1^+\) (black).} \label{pic_BZ_BS} \end{center} \end{figure} As shown in Ref.~\citeonline{vol78,vol83} the wave function parity at the high symmetry $L$ points provides a good quantum number and thus the Bloch function is a sum of atomic orbitals from either the group IV or the group VI species (our coordinate system well be chosen such that the group IV Bloch states have odd parity and the group VI Bloch states even parity). While the ad-mixture of the atomic $s$-states can be considerable, it does not change the \emph{form} of the effective Hamiltonian, and thus can be included by tuning the value of the model parameters to fit band structure calculated \emph{ab-initio}. In the absence of interaction between these Bloch states, two atomic p-states would give rise to two degenerate triplets of cubically symmetric bands at the $L$ points, ordered such that the triplet of odd parity (-) group IV states is higher in energy than the triplet of even parity (+) group VI states. The lifting of the degeneracy by the crystal field matrix elements $w^\pm$ (which mix $p$-orbitals) and spin-orbit coupling $\lambda^\pm$ (which mix spin channels) reduces the band gap dramatically, as shown in the schematic illustration presented in Fig.~\ref{pic_BZ_BS}(a). In the case of SnTe or cubic SnSe this is sufficient to invert the ordering of the $\epsilon_1^+$ and $\epsilon_2^-$ states as compared to the vacuum, as illustrated in Fig.~\ref{pic_BZ_BS}(a), resulting in a topologically protected surface state. The Bloch states at the $L$-points are given by \begin{align} \Phi^+_{x,y,z}(\mathbf{r})=&i\sqrt{\frac{2}{N}}\sum_{\mathbf{R}} \sin\left(\mathbf{k}_L\mathbf{R}\right)p_{x,y,z}(\mathbf{r}-\mathbf{R})\label{Phi_n^+},\\ \Phi^-_{x,y,z}(\mathbf{r})=&\sqrt{\frac{2}{N}}\sum_{\mathbf{R}} \cos\left(\mathbf{k}_L\mathbf{R}\right)p_{x,y,z}(\mathbf{r}-\mathbf{R})\label{Phi_n^-}. \end{align} where ${\bf k}_L = \pi/a (\pm1,\pm1,\pm1)$ is one of the eight $L$-point wave vectors and ${\bf R}$ runs over all lattice vectors. As a result of spin and orbital angular momentum mixing, the band the band edge states at the $L$ point are given by the Kramers conjugate pairs \begin{align} \Phi_2^- &=-\sin\frac{\theta_-}{2}\Phi_+^{-\downarrow}+\cos\frac{\theta_-}{2}\Phi_0^{-\uparrow}\\ K\Phi_2^- &=-\sin\frac{\theta_-}{2}\Phi_-^{-\uparrow}+\cos\frac{\theta_-}{2}\Phi_0^{-\downarrow}\\ \Phi_1^+ &=\ \ \cos\frac{\theta_+}{2}\Phi_+^{+\downarrow}+\sin\frac{\theta_+}{2}\Phi_0^{+\uparrow}\\ K\Phi_1^+ &=\ \ \cos\frac{\theta_+}{2}\Phi_-^{+\uparrow}+\sin\frac{\theta_+}{2}\Phi_0^{+\downarrow} \end{align} where $\Phi_0^{\pm\uparrow\downarrow}=\Phi_{z'}^{\pm\uparrow\downarrow}\), \(\Phi_\pm^{\pm\uparrow\downarrow}=1/\sqrt{2}\left(\Phi_{x'}^{\pm\uparrow\downarrow}\pm i \Phi_{y'}^{\pm\uparrow\downarrow}\right)$ and where the primed coordinate system has the $z'$-axis aligned along the $(111)$ symmetry axis which forms a natural coordinate system. The \emph{spin mixing parameters} \(\theta_\pm\) are connected to the crystal field and the spin orbit interaction as follows\cite{vol83}: \begin{align} &\tan\theta_\pm=-\frac{2\sqrt{2}\hbar \lambda^\pm}{\hbar\lambda^\pm+3w^\pm}. \label{SM} \end{align} This parameter will be subsequently shown to dramatically influence the RKKY interaction on the (111) surface of these materials. As shown in Ref.~\citeonline{vol83}, the \(C_{3v}\) symmetry of the $L$-point dictates the following Hamiltonian \begin{align} H^\pm=\mp\Delta_{ion}+2w^\pm\cos\left(\frac{2\pi}{3}L_{z^\prime}\right)+\lambda^\pm\left(\mathbf{L}.\boldsymbol{\sigma}\right)^\pm \label{eqH} \end{align} where the \(\pm\) index refers to the conduction band edge (group VI derived even parity state) or the valence band edge (group IV derived odd parity state) and where ${\bf L}$ is the angular momentum operator. The first term in Eq.~\ref{eqH} represents the difference in ionization energy between the group IV and group VI species; the second term encodes the mixing of the $p$-orbitals which, in the Hilbert space of the three $p$-states, is equivalent to the actions of the rotation operator by angles of \(2\pi /3\) around the $(111)$ symmetry axis; the third term is the spin-orbit coupling. The parameters \(w^\pm\) and \(\lambda^\pm\) are tabulated in Refs.~\citeonline{vol78,vol83} for all IV-VI compounds. Finally a \(\mathbf{k}.\mathbf{p}\) expansion around the $L$-point leads to a twelve-band \(\mathbf{k}.\mathbf{p}\) Hamiltonian, which reduces at low energies to a four-band Dirac-type Hamiltonian\cite{vol83} \begin{align} H=\begin{pmatrix} \Delta&\hbar\left[v_\perp \boldsymbol{\tau}_\perp^\prime.\mathbf{k}'_\perp+v_\parallel \tau_z^\prime k_z^\prime\right]\\ \hbar\left[v_\perp \boldsymbol{\tau}_\perp^\prime.\mathbf{k}'_\perp+v_\parallel \tau_z^\prime k_z^\prime\right]&-\Delta \end{pmatrix} \label{H_D}, \end{align} where \(\boldsymbol{\tau}_\perp^\prime\) stands for the vector of Pauli matrices $(\tau_x^\prime,\tau_y^\prime)$. In addition to the bulk band gap $\Delta$, the Hamiltonian contains \(v_\parallel\) (\(v_\perp\)), the velocity parallel (perpendicular) to the $(111)$ symmetry axis, and which are responsible for anisotropic effective masses parallel or perpendicular to this axis. \section{The topological surface states} To describe the low energy electron excitations on the surface we take Eq.~\ref{H_D} and replace the band gap $\Delta$ by a \emph{band gap function} \(\Delta(z)=\Delta_0f(z)\) with \(\Delta_0<0\) and $f(z)$ ($f(+\infty) = 1$ and $f(-\infty) \to -\infty$) such that $\Delta(-\infty)\Delta(\infty) < 0$, as required for a topologically non-trivial band structure\cite{vol85}. Furthermore, we introduce an energy shift function \(\varphi(z)=\varphi_0f(z)\) that models the expected overall shift in energy of the states at the surface\cite{pan87} (i.e., band bending). This yields the bulk-boundary eigenvalue problem \begin{align} \begin{pmatrix} \Delta(z)& v_\parallel \tau_z p_z+v_\perp \boldsymbol{\tau}_\perp.\mathbf{p}_\perp\\ v_\parallel \tau_z p_z+v_\perp \boldsymbol{\tau}_\perp.\mathbf{p}_\perp&-\Delta(z) \end{pmatrix} \psi=\left(\epsilon-\varphi(z)\right)\psi, \label{EVP_shift} \end{align} Note that for the experimentally observed downward band bending\cite{tas14} we require \(\varphi_0>0\). There are four inequivalent $L$-points in the bulk with, for the (111) surface, ${\bf k}_L=\pi/a(1,1,1)$ projecting to the $\Gamma$ point of the (111) surface Brillouin zone, and ${\bf k}_L=\pi/a(1,1,1)$, ${\bf k}_L=\pi/a(1,1,1)$, and ${\bf k}_L=\pi/a(1,1,1)$ to the three inequivalent $M$-points. For reasons that will be explained below we focus on the $\Gamma$-point of the (111) surface. This choice is, in fact, already implied by Eq.~\ref{EVP_shift}: setting up the bulk-boundary problem for another crystal facet -- or a different $L$-point -- would require rotation of $k_z$ in Eq.~\ref{H_D} such that the principle axis of the ellipsoidal band manifold at the $L$-point coincides with the surface normal. After a basis transformation Eq.~\ref{EVP_shift} takes the form of a supersymmetric Dirac equation: \begin{align} &h_\pm^\dagger h_\pm\zeta_\pm=\left[\epsilon^2-\hbar^2v_\perp^2k_\perp^2+\frac{\epsilon^2\varphi_0^2}{\Delta_0^2-\varphi_0^2}\right]\zeta_\pm \label{SSE_W} \end{align} where $h_\pm = \left[i\sigma_zv_\parallel p_z\pm W(z)\right]$ with \(W(z)=\sqrt{\Delta_0^2-\varphi_0^2}\left[f(z)+\epsilon\varphi_0/\left(\Delta_0^2-\varphi_0^2\right)\right]\) the superpotential and where we have defined the components of two (spinor valued) parts of the transformed wavefunction \(\zeta=\begin{pmatrix}\zeta_s, & \zeta_{-s}\end{pmatrix}\). The only normalizable solution for the ground state of the supersymmetric equation are the pure spinors $\zeta_s \propto \begin{pmatrix} 1, & 0 \end{pmatrix} g(z)$ and $\zeta_{-s} \propto \begin{pmatrix} 0, & 1 \end{pmatrix} g(z)$. Inserting \(\zeta\) into Eq. \ref{EVP_shift} leads to a differential equation in $z$ (the direction normal to the surface) that is solved by the envelope function \(g(z)\) (decaying exponentially into the vacuum and bulk). This leads to an effective Dirac-Weyl equation on the surface \begin{align} H_{eff}(\mathbf{k}_\perp)\chi=\epsilon \chi \end{align} where the effective surface Hamiltonian is given by \(H_{eff}(\mathbf{k}_\perp)=\hbar \gamma v_\perp\boldsymbol{\eta}_\perp.\mathbf{k}_\perp\) with \(\boldsymbol{\eta}\) the two-dimensional vector of Pauli matrices and \(\gamma=\sqrt{1-\varphi_0^2/\Delta_0^2}\). This equation, as is well known, describes massless Dirac fermions, and the surface spectrum consists of two linear dispersing energy bands \(\epsilon_{k_\perp l}=l\hbar \tilde{v}_\perp k_\perp\) with the velocity \(\gamma v_\perp\). The corresponding eigenstates are given by \begin{align} \chi_{\mathbf{k}_\perp l}= \frac{1}{\sqrt{2}} \begin{pmatrix} l e^{-i\frac{\phi}{2}}\\ e^{i\frac{\phi}{2}} \end{pmatrix} \label{chi_SnTe} \end{align} This solution is, however, not the physical wavefunction of the surface state. The effective surface Hamiltonian and its eigenstates are defined on a pseudospin space that is spanned by the Kramers conjugate pair $X$ and $K X$ with \begin{equation} X=\frac{s}{\sqrt{2}}\left[e^{-i\frac{\pi}{4}}\sqrt{1-\frac{\varphi_0}{\Delta_0}}\Phi_2^-+e^{i\frac{\pi}{4}}\sqrt{1+\frac{\varphi_0}{\Delta_0}}\Phi_1^+\right] \end{equation} leading to the physical wavefunction of the surface state given by \begin{equation} \Psi_{\mathbf{k}_\perp l} = (c_+ X + c_- KX) e^{i\mathbf{k}_\perp.\mathbf{r}_\perp}e^{i\kappa z}g(z) \label{yolo} \end{equation} The surface states are characterized by two quantum numbers, the chirality \(l\in\{\pm\}\) and the wave vector \(\mathbf{k}_\perp\) that in polar coordinates is given by \((k_\perp,\phi)\). The function $g(z)=N \exp\left(-\frac{s}{\hbar v_\parallel}\int_0^z dz\ W(z)\right)$ ensures an exponential decay away from the surface \emph{provided that the superpotential changes its sign asymptotically}. From the definition of the superpotential, it is seen that this behaviour is ensured by the function \(f(z)\) provided \(|\varphi_0|<|\Delta_0|\) and $\epsilon_{-s} < \epsilon <\epsilon_{s}$ with $\epsilon_{s}=-\varphi(-s\infty)\left(\Delta_0^2/\varphi_0^2-1\right)$. The energy shift function (\(\varphi_0(z)\)) guarantees that the two surface bands limit tangentially to the bulk band \(\epsilon_1^+(k_\perp)\) for $s = -1$, or \(\epsilon_2^-(k_\perp)\) for $s = +1$. This relation between bulk bands and the surface state is illustrated in Fig.~\ref{pic_BZ_BS}(b). As our theory distinguishes in a clear way the physical electron spin from the pseudospin degree of freedom we can calculate expectation values of both the pseudo- as well as electron spin operators for a given surface state \(\chi_{\mathbf{k}_\perp l}\). For the pseudospin we find the usual \emph{irrotational} pseudospin density vector field $\langle\chi_{\mathbf{k}_\perp l} \mid\boldsymbol{\eta}\mid \chi_{\mathbf{k}_\perp l}\rangle=l \begin{pmatrix} \cos\phi, & \sin\phi, & 0 \end{pmatrix}$ with unit magnitude while, in contrast, the physical spin density is the well known \emph{solenoidal} field $\langle\Psi_{\mathbf{k}_\perp l} \mid \boldsymbol{\sigma}\mid\Psi_{\mathbf{k}_\perp l}\rangle=l\sigma \begin{pmatrix}-\sin\phi, & \cos\phi, & 0 \end{pmatrix}$, with a magnitude that depends explicitly on the bulk band structure parameters \(\sigma=1/2\cos\left[\left(\theta_-+\theta_+\right)/2\right]-\varphi_0/(2\Delta_0) \cos\left[\left(\theta_--\theta_+\right)/2\right]\). Note that as we have derived our surface state from a microscopic theory of the bulk crystal, the value of $\sigma$ incorporates the bulk spin-orbit coupling that is known to reduce the spin polarization of the surface state\cite{yaz10}. \section{The RKKY Interaction} \begin{figure}[ht] \centering \includegraphics[width=\linewidth]{./abc_EF} \caption{The evolution of the strength of the RKKY coefficients \(a_\alpha\) (in-plane coupling), \(b_\alpha\) (out-of-plane coupling) and \(c_\alpha\) (Dzyaloshinskii-Moriya coupling) as a function of the spin mixing parameters \(\theta_\alpha\) of a topological insulator of the SnTe class. The parameters $\theta_\pm$ reflect the relative strengths of spin orbit coupling and crystal field in the bulk for the Sn (-) and Te (+) species (with $\theta_\pm=0$ corresponding to no spin-orbit coupling, see Eq.~\ref{SM}). The changing coupling strength of these interaction types results in three qualitatively distinct regions of the RKKY interaction, highlighted by the different background color: in the brown and in the turquoise region both spins cant with respect to a ferromagnetic or anti-ferromagnetic reference state parallel to the \(x\)- or \(z\)-axis respectively, while in the yellow region a collinear coupling parallel to the \(y\)-axis is preferred (the choice of coordinate system is such that the $x$-axis is aligned with the connection vector of the spin impurities and $z$ points out-of-plane). The values of $\theta_\pm$ that correspond to the Pb$_x$Sn$_{1-x}$Te system are indicated by the shaded area.} \label{abc} \end{figure} We now come to the RKKY interaction between two magnetic moments on the surface of a IV-VI topological insulator. We will consider the $\Gamma$-point Dirac cone on the (111) surface and ignore possible contribution from the M-point cones. This is likely to be a rather good approximation as the energy separation of the $\Gamma$ and $L$-point cones ($\approx 170\,$meV), along with the limited energy range within which the Dirac cone exists ($\approx 200\,$meV), implies that for energies at which the RKKY signal from the $\Gamma$-cone is strong, that from the $M$-cone will be weak. We consider the indirect exchange interaction to be mediated by the topological surface states, Eq.~\ref{chi_SnTe}, and assume that each impurity spin \(\mathbf{S}_{1,2}\) couples via a contact interaction to the electron spin density \(\mathbf{s}_{1,2}\) \begin{align} V=-\beta\left(\mathbf{S}_1.\mathbf{s}_1+\mathbf{S}_1.\mathbf{s}_2\right), \label{eqV} \end{align} where the coupling strength is given by \(\beta\). As a further simplification of the model we set the energy shift function $\varphi_0 = 0$ (i.e. no band bending at the surface), an approximation expected to be good provided the Fermi energy does not approach the tangent point of the Dirac cone at which the surface state merges with the bulk spectrum. \begin{figure*}[t!] \centering \includegraphics[width=0.94\linewidth]{./RKKY-finite-doping} \caption{The RKKY interaction on the tin terminated (111) surface of SnTe (panels a and c) and \(\mathrm{Pb}_{0.25}\mathrm{Sn}_{0.75}\mathrm{Te}\) (panels b and d) at a Fermi energy of \(0.1\)~eV. The SnTe crystal features two distinct spin states: canting with respect to the ferromagnetic coupling parallel to the \(x\)-axis (spin state 1) and canting with respect to the antiferromagnetic reference state parallel to the \(x\)-axis (spin state 2). The canting angle is given in panel c). By alloying the SnTe crystal with lead it is possible to open a small window where ferromagnetic coupling parallel to the \(y\)-axis is preferred (spin state 3).} \label{cantr} \end{figure*} To calculate the RKKY interaction we require an understanding of the impurity scattering mechanism due to Eq.~\ref{eqV}. This is encoded in the scattering matrix $\boldsymbol{\sigma}.\mathbf{S}_1$. For scattering at a group IV site impurity the \(\boldsymbol{\sigma}.\mathbf{S}_1\)-matrix in the pseudospin basis takes the form \begin{align} \left(\boldsymbol{\sigma}.\mathbf{S}_1\right)_-= \frac{1}{2}\begin{pmatrix} S_1^z\cos\theta_-&iS_1^-\cos^2\frac{\theta_-}{2}\\ -iS_1^+\cos^2\frac{\theta_-}{2}&-S_1^z \cos\theta_- \end{pmatrix} \label{sigmaS-} \end{align} while for scattering at a group VI site impurity it is given by \begin{align} \left(\boldsymbol{\sigma}.\mathbf{S}_1\right)_+= \frac{1}{2}\begin{pmatrix} -S_1^z\cos\theta_+&-iS_1^-\sin^2\frac{\theta_+}{2}\\ iS_1^+\sin^2\frac{\theta_+}{2}&S_1^z \cos\theta_+ \end{pmatrix} \label{sigmaS+} \end{align} with \(S_1^\pm=S_1^x\pm i S_1^y\). Employing these scattering matrices with a standard RKKY calculation leads to the following interaction energy \begin{align} \notag E_\alpha^{int}(R_\perp)=-\frac{2\pi^2\hbar \lambda^2}{\Omega_{SBZ}^2 v_\perp}&\Bigl(\left[A_{k_F}(R_\perp)-B_{k_F}(R_\perp)\right]\left[a_\alpha S_1^xS_2^x+b_\alpha S_1^zS_2^z\right]\Bigr. +\left[A_{k_F}(R_\perp)+B_{k_F}(R_\perp)\right]a_\alpha S_1^yS_2^y\\ &\Bigr.+\left[C_{k_F}(R_\perp)+D_{k_F}(R_\perp)\right]c_\alpha\left[S_1^xS_2^z- S_1^zS_2^x\right]\Bigl) \label{E_int_TI} \end{align} where the material dependent coefficients \(a_\alpha\), \(b_\alpha\) and \(c_\alpha\) are given in Table \ref{sigma-rho} and where \begin{align} A_{k_F}(R_\perp)=\frac{\pi}{2}\lim_{s\to 0}\int_{k_F}^\infty dk_\perp\ k_\perp^ 2 J_0\left(k_\perp R_\perp\right)Y_0\left(k_\perp R_\perp\right)e^{-sk_\perp}\label{A1}\\ B_{k_F}(R_\perp)=\frac{\pi}{2}\lim_{s\to 0}\int_{k_F}^\infty dk_\perp\ k_\perp^ 2 J_1\left(k_\perp R_\perp\right)Y_1\left(k_\perp R_\perp\right)e^{-sk_\perp}\\ C_{k_F}(R_\perp)=\frac{\pi}{2}\lim_{s\to 0}\int_{k_F}^\infty dk_\perp\ k_\perp^ 2 J_1\left(k_\perp R_\perp\right)Y_0\left(k_\perp R_\perp\right)e^{-sk_\perp}\\ D_{k_F}(R_\perp)=\frac{\pi}{2}\lim_{s\to 0}\int_{k_F}^\infty dk_\perp\ k_\perp^ 2 J_0\left(k_\perp R_\perp\right)Y_1\left(k_\perp R_\perp\right)e^{-sk_\perp}\label{D1} \end{align} with \(k_\perp=E/(\hbar v_\perp)\). Following an asymptotic expansion Eq.~\ref{E_int_TI} can be brought to the form \begin{align} \notag E_\alpha^{int}(R_\perp)=-\frac{\pi^2k_F\hbar\lambda^2}{\Omega_{SBZ}^2 v_\perp R_\perp^2}&\Biggl[\sin\Bigl(2k_FR_\perp\Bigl) \Bigl(a_\alpha S_1^xS_2^x+b_\alpha S_1^zS_2^z\Bigr)\Biggr. \Biggl.-c_\alpha \cos\left(2k_FR_\perp\right)\Bigl(S_1^xS_2^z-S_1^zS_2^x\Bigr)\Biggr]\\ \notag-\frac{\pi^2\hbar\lambda^2}{4\Omega_{SBZ}^2v_\perp R_\perp^3}&\Biggl[\cos\left(2k_F R_\perp\right)\Bigl(3a_\alpha S_1^xS_2^x-2a_\alpha S_1^yS_2^y+3b_\alpha S_1^zS_2^z\Bigr)\Biggr. \Biggl.+3c_\alpha \sin\left(2k_F R_\perp\right)\Bigl(S_1^xS_2^z-S_1^zS_2^x\Bigr)\Biggr] \end{align} Note that fact that in this equation (as well as in the full interaction energy Eq.~\ref{E_int_TI}) the dependence is solely through the impurity separation \(R_\perp\), and not the polar angle \(\theta\), arises from the choice of the coordinate system in which we have rotated \(x\)-axis such that it is aligned with the impurity separation vector. \begin{table}[h!] \begin{align*} \arraycolsep=12pt\def1.8{1.8} \begin{array}{c|c|c} \hline\hline a_\alpha&b_\alpha&c_\alpha\\ \hline \frac{1}{4}\cos^4\left(\frac{2\theta_\alpha+\pi(1+\alpha)}{4}\right)& \frac{1}{4}\cos^2\theta_\alpha& \frac{1}{4}\cos\theta_\alpha\cos^2\left(\frac{2\theta_\alpha+\pi(1+\alpha)}{4}\right)\\[0.2cm] \hline\hline \end{array} \end{align*} \caption{The coefficients for the RKKY interaction on the surface of a topological insulator of the SnTe class. The spin mixing parameter \(\theta_\alpha\) is for a range of IV-VI semiconductors given in Ref.~\citeonline{vol83}. The parameter \(\alpha\) takes the value +1 on the tellurium terminated surface and -1 on the tin terminated surface.} \label{sigma-rho} \end{table} In the lowest order ($1/R_\perp^2$) there are two Ising-type terms ($S_1^xS_2^x+b_\alpha S_1^zS_2^z$) that favour a collinear ferromagnetic (FM) or antiferromagnetic (AFM) coupling, as well as a Dzyaloshinskii-Moriya (DM) term ($S_1^xS_2^z-S_1^zS_2^x$) that favours the spins to be (i) in the plane formed by the connection vector and the surface normal and (ii) have a relative angle of 90$^\circ$. Curiously, the collinear terms are $\pi/2$ out of phase with the DM term (a fact also true at order $1/R_\perp^3$). However, the most interesting aspect of the interaction energy is that each coupling term is endowed with its own material constant and, as may be seen in Fig.~\ref{abc}, these are oscillatory functions of the spin mixing angles and, hence, of the materials crystal field and spin orbit coupling parameters. The \emph{qualitative form} of the RKKY interaction in the IV-VI topological insulators is, therefore, profoundly sensitive to the details of the bulk electronic structure: there is no ``universal'' form of the RKKY interaction corresponding to, or resulting from, the universality of the Dirac-Weyl operator that describes the topological surface state. From Fig.~\ref{abc} it can be noticed that (i) the materials parameters for the group IV terminated surface ($-$) and group VI terminated surface ($+$) are, with a sign change of the DM constant, related to each other by a phase shift of $\pi$ and (ii) for $\lambda^- = 0$, i.e. zero spin orbit coupling, the non-collinear DM coupling on the group VI terminated surface is at \emph{maximum strength} (provided, of course, that the crystal field by itself is enough to invert the band order). A strong DM interaction is usually associated with substantial spin orbit coupling; this example highlights the fact that it is the non-trivial topology of the bulk band structure, not the spin orbit coupling, that is responsible for the DM interaction. To investigate impact of the bulk physics on the impurity spin coupling for a specific example we now consider spin impurities on the tin surface of SnTe and its alloy \(\mathrm{Pb}_{0.25}\mathrm{Sn}_{0.75}\mathrm{Te}\), illustrated in panels (a) and (b) of Fig.~\ref{cantr} respectively. On the tin terminated surface of SnTe and at an impurity separation of 20\AA~both impurity spins couple at FM and parallel to the connection vector. At higher impurity separation the two spins cant away from the connection vector leading to an increasing out-of-plane component. At an impurity separation of \(\approx80\)\AA~the canting angle reaches its maximum and the configuration flips into an AFM coupling, following which the canting angle then begins to decrease. In the further course of the evolution with increasing separation the two spins oscillate between these two types of spin states with successively increasing and decreasing canting angle, as shown in \ref{cantr}(c). The period of oscillation is given by the Fermi energy which is here set to \(0.1\)~eV. The corresponding behaviour on the tin terminated surface of \(\mathrm{Pb}_{0.25}\mathrm{Sn}_{0.75}\mathrm{Te}\) is shown in Fig. \ref{cantr}(b): in large parts the interaction resembles that on the surface of SnTe but in small window from \(59\)\AA~to \(67\)\AA~a coupling in the direction parallel to the \(y\)-axis can be observed. As shown in Fig.~\ref{cantr}(d) the canting angle is in that region rather small and thus the coupling can considered to be, as a good approximation, ferromagnetic. \section{Discussion and conclusions} We have employed the RKKY interaction to probe the coupling between the spin structure of the surface state and that of the bulk insulator wavefunctions in the IV-VI semiconductor crystalline topological insulators. For the (111) surface we have derived the Dirac-Weyl surface state from a fully microscopic description of the bulk insulator, finding in this way an explicit connection between the surface and bulk wavefunctions. This, in turn, provides a ``fully electronic'' route from the atomic degrees of freedom of the bulk to the emergent pseudospin structure of the surface state and allows one to probe the impact of bulk physics on the surface RKKY interaction. We find that the nature of the topological surface state, as probed by the RKKY interaction, is coupled very strongly to the bulk physics: the balance of crystal field and spin-orbit coupling in the bulk insulator determines even the qualitative form of the RKKY interaction. For the group IV terminated surface, when the crystal field dominates spin-orbit in the bulk the RKKY interaction favours an out-of-plane configuration, while when the the reverse is true an in-plane configuration is favoured. On the group VI terminated surface the situation is reversed. The equilibrium spin configuration is generally non-collinear, with collinear spins found only when the spin-orbit coupling in the bulk is switched off. In short, while the Dirac-Weyl surface state is ``universal'' in the sense that the surface state spectrum is always, at low energies, a conical intersection, the surface wavefunction is not and depends strongly on details of the bulk electronic structure, a fact reflected in the rich dependence of the type of RKKY interaction on the microscopic parameters of the bulk band structure. \bibliographystyle{apsrev}

\section{Introduction} Deep learning really took off in 2012 when Krizhevsky et al. showed record breaking results on the Imagenet dataset \cite{krizhevsky2012imagenet}. They demonstrated that deep convolutional neural networks trained end to end on large labelled datasets can beat most other techniques for image recognition. Deep learning quickly became the default algorithm for image classification and now even achieves super-human level performance \cite{}. Deep learning has also revolutionized other fields like speech recognition and natural language processing \cite{lecun2015deep}. \\ \newline The two key ingredients needed to successfully apply deep neural networks are large amounts of labelled training data and powerful computing systems such as GPUs. Mobile devices including smartphones, Internet-of-Things (IoT) devices or smart home assistants have very limited processing power because of their intrinsic limitations on size and energy consumption. One possible solution is to offload all computations to the cloud but this introduces a latency and potentially even a privacy risk when sensitive data is processed remotely. \\ \newline There is a considerable amount of active research on techniques to reduce the computational cost of deep learning models. One approach is to prune the network by removing redundant weights. The idea of pruning already goes back to the eighties when LeCun et al. used second order derivatives to calculate the impact of each weight on the loss of the network \cite{NIPS1989_250}. Weights with a small impact are then removed from the network. More recently pruning was used on modern deep neural networks. Song et al. proposed a pruning pipeline where first weights with a small magnitude where removed after which the network was fine-tuned to recover the lost accuracy \cite{han2015learning}. Network pruning is especially effective when a network is used for transfer learning. In transfer learning the model is first trained on a large dataset such as Imagenet and is then fine-tuned on a small domain specific dataset.\\ Because the network was pretrained on a general dataset it will contain convolutional kernels that are not useful for the domain specific dataset. Molchanov et al. introduced a criterion based on a first-order Taylor expansion to decide which feature maps to remove and demonstrated impressive results when used together with transfer learning \cite{molchanov2016pruning}. \\ \newline Instead of compressing a trained model it is also possible to train efficient models from scratch. The recently introduced MobileNets \cite{howard2017mobilenets} use depthwise separable convolutions to reduce the computational cost. Depthwise separable convolutions factorize a standard convolution into a depthwise convolution and a 1x1 pointwise convolution. The depthwise convolution applies a single convolution to each input channel while the pointwise convolution combines the information in the different channels. Factorizing a traditional convolution into these two convolutions dramatically reduces the computational cost and size of the network with only a minimal reduction in accuracy. \\ \newline Most implementations of deep neural networks use 32 bit floating point numbers for weights and activations. Various works have shown that this is not necessary and that it is possible to use 16 bit \cite{micikevicius2017mixed} or 8 bit \cite{vanhoucke2011improving} numbers. In the extreme case it is even possible to use binary weights and activations. Courbariaux et al. successfully trained convolutional neural networks for image recognition with binary weights and activations \cite{courbariaux2016binarized}. This works surprisingly well for small scale datasets such as CIFAR10 but there is still a large drop in accuracy for larger datasets such as Imagenet. Neural networks with binary weights and activations are attractive because they replace the costly floating point multiplications and additions with bitwise XNORs and left and right bit shifts. These operations are very efficient to implement in hardware. \\ \newline Another problem with deep learning is the need for large labelled datasets. Training a new model from scratch requires a large amount of training data. A well known technique is to use transfer learning where a model is first trained on a large dataset like Imagenet and afterwards the last layer is removed and retrained using a small amount of new domain specific training data. Transfer learning works because the first layers in the network learn to detect features such as color transitions and basic shapes that are present in images from different domains \cite{yosinski2014transferable}. \section{A hybrid hardware-software architecture} We propose a hybrid hardware-software architecture based on this idea of transfer learning. We train a neural network with binary weights and activations on the Imagenet dataset and use this network as a fixed feature extractor that could be optimised on hardware level. The last layer is implemented in software and is evaluated on the CPU (or GPU) of the device. Most of the computations are offloaded to the custom circuit. Since we only need to retrain the last layer of the network it even becomes feasible to train on the device itself instead of offloading the training to the cloud. This is very attractive from a privacy point of view because the training data never leaves the mobile device. Our architecture also allows app developers to embed custom neural networks into their apps. Right now this is often impossible because a typical neural network quickly requires hundreds of megabytes of storage just for the weights. If the device however is equipped with the fixed feature extractor we only need to ship the last (domain specific) layer of the network with the app. A schematic overview of our architecture is shown in Figure \ref{}. \begin{figure}[h] \centering \includegraphics[scale=0.43]{tekening2.pdf} \caption{Schematic overview of our proposed architecture. Most layers of a binary neural network trained on Imagenet are embedded in a specialised circuit in hardware. This circuit is used as a fixed feature extractor. The last layer is implemented in software and retrained for different applications.} \label{} \end{figure} \section{Experiments} In this section we show that it is possible to use a binary neural network trained on Imagenet as a feature extractor for other datasets. We trained a binary version of the Alexnet architecture \cite{krizhevsky2012imagenet} on the ILSVRC2012 training set following the binarization technique of Courbariaux et al. \cite{courbariaux2016binarized}. Our binary network obtains a top 5 accuracy of 67\% (42\% top 1) while the floating point Alexnet obtains a top 5 accuracy of 80\% (57\% top 1). We then retrained the last layer on the domain specific datasets without changing the weights of the other layers. We report results for the three fine-grained image datasets summarized in Table \ref{}. For all our models we resized the input images such that the longest side was of length 256. During training we took random 224 by 224 pixel crops. For the test set we used center crops. The floating point networks where trained with stochastic gradient descent with momentum. For the binary networks we found that Adam \cite{kingma2014adam} gives better results which is consistent with Courbariaux et al. \cite{courbariaux2016binarized}. \begin{table}[h] \caption{The different datasets used in our experiments. We pretrained our network on the Imagenet dataset and then retrained the last layer on the three smaller domain specific datasets. } \label{} \centering \begin{tabular}{llll} \toprule Dataset & Number of classes & Training images & Testing images \\ \midrule ILSVRC2012 & 1000 & 1,200,000 & 50,000\\ \midrule Flowers \cite{Nilsback08} & 102 & 6,149 & 1,020\\ UCSD Birds \cite{WelinderEtal2010} & 200 & 5,994 & 5,794\\ MIT Indoor scenes \cite{quattoni2009recognizing} & 67 & 5.360 & 1,340\\ \bottomrule \end{tabular} \end{table} The results are summarized in Table \ref{sample-table}. The first part of the table shows the accuracies for the baseline floating point networks, either trained from scratch on the domain specific datasets or fine-tuned from the Imagenet model. As expected the fine-tuned models consistently outperform the models trained from scratch. The second part of the table shows the results for our binary models. The fine-tuned models again outperform the models trained from scratch which shows that transfer learning also works when the network uses binary weights and activations. For the last layer we can use binary weights and activations (a) but because this layer is evaluated in software in our proposed architecture we also experimented with floating point weights and activations (b). This consistently increases the accuracy on all datasets and pushes the accuracy closer to the accuracy of the floating point models. \begin{table}[h] \caption{The top 1 and top 5 accuracies for each domain specific dataset. We report results for the binary network trained from scratch and for the fine-tuned networks, both with binary weights in the last layer (a) and floating point weights (b). We compare to a floating point baseline.} \label{sample-table} \centering \begin{tabular}{llllllll} \toprule & & \multicolumn{2}{c}{Flowers} & \multicolumn{2}{c}{Birds} & \multicolumn{2}{c}{Indoor scenes}\\ & & Top 1 & Top 5 & Top 1 & Top 5 & Top 1 & Top 5 \\ \midrule \multirow{2}{*}{FP} & trained from scratch & 71.7\% & 90.5\% & 32.4\% & 57.2\% & 29.7\% & 56.4\%\\ & fine-tuned & 85.4\% & 97.2\% & 50.4\% & 79.6\% & 57.4\% & 86.8\%\\ \midrule \multirow{3}{*}{Binary}& trained from scratch & 63.2\% & 85.1\% & 13.8\% & 32.4\% & 29.0\% & 54.9\% \\ & (a) fine-tuned binary last layer & 80.6\% & 94.5\% & 33.9\% & 62.9\% & 46.7\% & 77.4\%\\ & (b) fine-tuned FP last layer & 84.0\% & 95.5\% & 36.7\% & 63.4\% & 48.3\% & 78.6\% \\ \bottomrule \end{tabular} \end{table} \section{Conclusion and future work} We introduced a hybrid hardware-software architecture where a binary neural network trained on Imagenet can be embedded in a dedicated circuit. The last layer is implemented in software and is retrained for each specific task. We showed that transfer learning works very well for binary neural networks and experimented with a hybrid binary-floating point network where only the last layer uses floating point operations. This is a good trade-off between accuracy and computational cost since most of the computations can be offloaded to the fixed hardware accelerator. In future work we will explore other hybrid architectures. We now only considered finetuning the last layer of the network but we can also retrain more than one layer which should give a higher accuracy but will also incur a higher computational cost since a larger part of the network is evaluated in software. \section*{Acknowledgments} We gratefully acknowledge the support of NVIDIA Corporation with the donation of GPU hardware used for this research. Steven Bohez is funded by a Ph.D. grant of the Agency for Innovation by Science and Technology in Flanders (IWT). \bibliographystyle{unsrt}

\section{Introduction} Li-Yorke sensitivity and Li-Yorke chaos are well-known properties of dynamical systems, where by a dynamical system we mean a phase space $X$ endowed with an evolution map $T$. We require that the phase space $(X,d)$ is a compact metric space and the evolution map is a continuous surjective mapping $T:X \rightarrow X$. \\ The definition of Li-Yorke sensitivity is a combination of sensitivity and Li-Yorke chaos. The Li-Yorke chaos was introduced in 1975 by Li and Yorke in \cite{LY}. A dynamical system is \emph{Li-Yorke chaotic} if there is an uncountable scrambled set. A set $S$ is scrambled if any two distinct points $x,y\in S$ are proximal (i.e. trajectories of $x$ and $y$ are arbitrarily close for some times) but not asymptotic, that means $$\liminf _{n\to\infty} d(T^n(x),T^n(y))=0\qquad\mbox{ and }\qquad \limsup _{n\to\infty} d(T^n(x),T^n(y))>0.$$ The initial idea of sensitivity goes back to Lorenz \cite{Lor}, but it was firstly used in topological dynamics by Auslander and Yorke in \cite{AY} and popularized later by Devaney in \cite{D}. A map $T$ is \emph{sensitive} if there is $\epsilon>0$ such that that for each $x\in X$ and each $\delta>0$ there is $y\in X$ with $d(x,y)<\delta$ and $n\in\mathbb{N}$ such that $d(T^n(x),T^n(y))>\epsilon$. By Huang and Ye in \cite{HY}, $T$ is sensitive if and only if there is $\epsilon>0$ with the property that any neighbourhood of any $x\in X$ contains a point $y$ such that trajectories of $x$ and $y$ are separated by $\epsilon$ for infinitely many times, that is $$\limsup _{n\to\infty} d(T^n(x),T^n(y))>\epsilon.$$ Inspired by the above results, Akin and Kolyada introduced Li-Yorke sensitivity in \cite{AK}. A map $T$ is \emph{Li-Yorke sensitive} if there is $\epsilon>0$ with the property that any neighbourhood of any $x\in X$ contains a point $y$ proximal to $x$, such that trajectories of $x$ and $y$ are separated by $\epsilon$ for infinitely many times. Thus, $$\liminf _{n\to\infty} d(T^n(x),T^n(y))=0\qquad\mbox{ and }\qquad \limsup _{n\to\infty} d(T^n(x),T^n(y))>\epsilon.$$ Authors in \cite{AK} proved, among others, that weak mixing systems are Li-Yorke sensitive and stated five conjectures concerning Li-Yorke sensitivity. Three of them were disproved in \cite{M1} and \cite{M2}, one was confirmed recently in \cite{Y}. Only one problem remained open until today:\\ {\bf Question 1.} Are all Li-Yorke sensitive systems Li-Yorke chaotic?\\ This question was also included in the list of important open problems in the contemporary chaos theory in topological dynamics in \cite{YeL}.\\ We show that the answer is negative. We construct an infinite-dimensional compact metric space $X$, which is a closed subset of $\mathbb{S}\times\mathbb{H}$, where $\mathbb{S}$ is the unit circle and $\mathbb{H}$ is the Hilbert cube, and a skew-product map $F$, which is a combination of a rotation on $\mathbb{S}$ and a contraction on $\mathbb{H}$, such that $(X,F)$ is Li-Yorke sensitive but possesses at most countable scrambled sets. The mapping $F$ can be continuously extended to get a connected dynamical system with the same properties, see Remark 1. \\ We recall here some notations used throughout the paper. A pair of points $(x,y)$ in $X^2$ is \emph{asymptotic} if $\lim_{n\to\infty} d(T^n(x),T^n(y))=0$. A pair of points $(x,y)$ in $X^2$ is \emph{proximal} if $\liminf_{n\to\infty} d(T^n(x),T^n(y))=0$, if $(x,y)$ is not proximal then it is called \emph{distal}. A pair of points $(x,y)$ in $X^2$ is \emph{scrambled} if it is proximal but not asymptotic. A pair of points $(x,y)$ in $X^2$ is \emph{scrambled with modulus $\epsilon$} if it is proximal and $\limsup_{n\to\infty} d(T^n(x),T^n(y))\geq \epsilon$. A system $(X,T)$ is \emph{minimal} if every point $x\in X$ has a dense orbit $\{T^n(x)\}_{n=0}^{\infty}$. A system is \emph{transitive} if, for every pair of open, nonempty subsets $U,V\subset X$, there is a positive integer $n\in\mathbb{N}$ such that $U\cap T^{n}(V)\neq \emptyset.$ A system $(X,T)$ is \emph{weakly mixing} if the product system $(X\times X,T\times T)$ is transitive.\\ \section{Main result} Here we state the main result and outline of its proof. Technical details of the proof can be found in a form of lemmas and claims in the last section. \begin{theorem} There is a Li-Yorke sensitive dynamical system which is not chaotic in the sense of Li-Yorke. \end{theorem} \begin{proof} Let $X_0$ be the unit circle $\mathbb{S}=\mathbb{R}/\mathbb{Z}$ equipped with the metric $d_0(x,y)=\min\{|x-y|,1-|x-y|\}$ and, for $i\geq 1$, $X_i=\mathbb{N}\cup\{\infty\}$ equipped with the metric $d_i(x,y)=|\frac{1}{x}-\frac{1}{y}|$, where $\frac{1}{\infty}=0$. Then $\Pi_{i=0}^{\infty}X_i$ with the product topology is a compact space. The product topology is equivalent to the metric topology induced by the metric $D(x,y)=\sum_{i=0}^{\infty}\frac{d_i(x_i,y_i)}{2^i}.$ Let $Y=\{x\in\Pi_{i=0}^{\infty}X_i: \{x_i\}_{i=1}^{\infty}\mbox{ is nondecreasing}\}.$ $Y$ is a closed subset of $\Pi_{i=0}^{\infty}X_i$ and therefore it is a compact metric space. Notice that, for $i\geq 1$, $X_i$ can be embed into the unit interval $[0,1]$ equipped with the natural topology, so $Y$ can be identify with a closed subset of $\mathbb{S}\times\mathbb{H}$, where $\mathbb{H}$ is the Hilbert cube.\\ Let $F:Y\rightarrow Y$ be a mapping defined for a point $x=(x_0,x_1,x_2,\ldots)$ in Y by $F(x)=(f_0(x),f_1(x),f_2(x),\ldots)$, where \begin{align}f_0(x)&=(x_0+\sum_{i=1}^{\infty}\frac{1}{2^i}\cdot \frac{1}{x_i})\mod 1,\\ f_i(x)&=x_i+1,\qquad\mbox{ for }i\geq 1, \end{align} where $\infty+1=\infty.$ $F$ is a continuous mapping, since $f_i$ is continuous, for every $i\geq 0$. First, we will show that $(F,Y)$ is Li-Yorke sensitive. It is enough to show that, for a given $x=(x_0,x_1,x_2,\ldots)\in Y$ and $U\in Nb_x$, there is $y=(y_0,y_1,y_2,\ldots)\in U$ such that $$\liminf _{n\to\infty} D(F^n(x),F^n(y))=0\qquad\mbox{ and }\qquad \limsup _{n\to\infty} D(F^n(x),F^n(y))\geq \frac 1 2.$$ \emph{ I. $\{x_i\}_{i=1}^{\infty}$ is a nondecreasing sequence containing at least one $\infty$}\\ Since $\{x_i\}_{i=1}^{\infty}$ is nondecreasing, there is $M\in\mathbb{N}$ such that $x_i$ is finite, for $i<M$, and $x_i=\infty$, for $i\geq M$. The neigbourhood $U$ is defined by $U=V\cap Y$, where $V$ is a neigbourhood of $x$ in $\Pi_{i=0}^{\infty}X_i$. Let $V=V_0\times V_1\times V_2\times\ldots$, where $V_i$ is a neigbourhood of $x_i$ such that $V_i=X_i$ for all but finitely many $i\geq 0$. Let $K\in \mathbb{N}$, sufficiently large to satisfy $K\in V_i$, for $i\geq M$, and simultaneously $K\geq x_{M-1}$. We define the point $y$ as follows: \begin{align}\label{y1} y_i=&x_i, \mbox{ for }0\leq i<M,\\ \label{y11}y_i=&K, \mbox{ for }i\geq M. \end{align} It is easy to see that $y$ belongs to $U$. By Claim 1, $(x,y)$ is scrambled with modulus $\frac 1 2$.\\ \emph{II. $\{x_i\}_{i=1}^{\infty}$ is a nondecreasing sequence of finite numbers}\\ The neigbourhood $U$ is defined by $U=V\cap Y$, where $V$ is a neigbourhood of $x$ in $\Pi_{i=0}^{\infty}X_i$. Without loss of generality, suppose $V=V_0\times V_1\times V_2\times\ldots$, where $V_0=(x_0-\delta,x_0+\delta)$, for some $\delta>0$, and, for $i>0$, $V_i$ is a neigbourhood of $x_i$ such that $V_i=X_i$ for all but finitely many $i$. Let $M\in \mathbb{N}$ such that $2^{-M}<\delta$ and simultaneously $V_i=X_i$, for $i\geq M$. We define the point $y$ as follows: \begin{align} \label{y2}y_0=&\big(x_0+1-\sum_{i=1}^{\infty}\frac{1}{2^{i+M}}\delta_i\big)\mod 1,\\ \label{y22}y_i=&x_i, \mbox{ for }0<i<M,\\ \label{y222}y_i=&\infty, \mbox{ for }i\geq M, \end{align} where $$\delta_i = \left\{ \begin{array}{ll} 0, \quad&\mbox{if } x_{i+M}=x_M, \\ \big(\sum_{j=0}^{x_{i+M}-x_M-1}\frac{1}{x_M+j}\big)\mod 1,\quad & \mbox{otherwise.} \end{array} \right.$$ It is easy to see that $\sum_{i=1}^{\infty}\frac{1}{2^{i+M}}\delta_i\leq 2^{-M}$ and $y$ belongs to $U$. By Claim 2, $(x,y)$ is scrambled with modulus $\frac 1 2$.\\ Notice that, in both cases, one point of the pair $(x,y)$ has $\infty$ coordinates while the other has all coordinates finite. By Claim 3, if $x_i$ and $y_i$ are finite, for all $i\geq 1$, then $\lim_{n\to\infty}D(F^n(x),F^n(y))$ exists and $(x,y)$ is not a scrambled pair. Therefore in each scrambled set $S\subset Y$, there is at most one $z\in S$ such that $z_i$ is finite, for $i\geq 1$. We finish our proof by finding an injection between $S\setminus\{z\}$ and $\mathbb{N}$.\\ Let $l_x=\min\{i: x_i=\infty\}$. Then the mapping $\ \mathbb{N}$ defined by $\iota(x)=\l_x$ is injective. We proceed by assuming the opposite. Let $x\neq y$ in $S\setminus\{z\}$ such that $l=l_x=l_y$. Since $\{x_i\}_{i=1}^{\infty}$ and $\{y_i\}_{i=1}^{\infty}$ are nondecreasing, \begin{equation}\label{y3} x_i<\infty \wedge y_i<\infty, \quad\mbox{ for }0<i<l,\qquad x_i=y_i=\infty, \quad\mbox{ for }i\geq l. \end{equation} By Claim 4, $\lim_{n\to\infty}D(F^n(x),F^n(y))$ exists which is in contradiction with $(x,y)$ being a scrambled pair. \end{proof} \begin{remark} The mapping $F$ can be continuously extended to get a \emph{connected} dynamical system with the same properties. Let $X_0$ be the unit circle $\mathbb{S}=\mathbb{R}/\mathbb{Z}$ equipped with the metric $d_0(x,y)=\min\{|x-y|,1-|x-y|\}$ and, for $i\geq 1$, $X_i$ be the unit interval $[0,1]$ equipped with the natural topology. Then $\Pi_{i=0}^{\infty}X_i$ equipped with the product topology is $\mathbb{S}\times \mathbb{H}$, where $\mathbb{H}$ is the Hilbert cube. The product topology is equivalent to the metric topology induced by the metric $D(x,y)=\sum_{i=0}^{\infty}\frac{d_i(x_i,y_i)}{2^i}.$ Let $Z=\{x\in\mathbb{S}\times \mathbb{H}: \{x_i\}_{i=1}^{\infty}\mbox{ is nonincreasing}\}.$ $Z$ is closed and pathwise connected subset of $\mathbb{S}\times \mathbb{H}$.\\ Let $x=(x_0,x_1,x_2,\ldots)$ be a point in $Z$. We will express every $x_i\in X_i\setminus\{0\}=(0,1]$, for $i\geq1$, as $x_i=\frac{1}{k_i}+t_i\cdot|\frac{1}{k_i-1}-\frac{1}{k_i}|,$ where $t_i\in(0,1]$ and $k_i\in\mathbb{N}\setminus\{1\}$. Let $G:Z\rightarrow Z$ be a mapping defined by $G(x)=(g_0(x)g_1(x),g_2(x),\ldots)$, where $$g_0(x)=(x_0+\sum_{i=1}^{\infty}\frac{1}{2^i}\cdot {x_i})\mod 1,$$ and for $i\geq 1,$ $$g_i(x)= \left\{ \begin{array}{ll} 0, \quad&\mbox{if } x_{i}=0,\\ \frac{1}{k_i+1}+t_i\cdot|\frac{1}{k_i}-\frac{1}{k_i+1}|,\quad &\mbox{otherwise.} \end{array} \right.$$ Then $G$ is a continuous extension of $F$ and $(G,Z)$ is a Li-Yorke sensitive but not Li-Yorke chaotic system. \end{remark} \begin{remark} The mapping $F$ is not minimal (it is even not transitive). In case of minimal maps, we have still an open question:\\ {\bf Question 2.} Are all Li-Yorke sensitive minimal systems Li-Yorke chaotic? \end{remark} \begin{remark} $Y$ is an infinite-dimensional space. We can examine the relation between Li-Yorke sensitivity and Li-Yorke chaos for low-dimensional dynamical systems. It is known that in case of graph mappings (in particular, interval mappings) Li-Yorke sensitivity implies Li-Yorke chaos, since, for graph mappings, the existence of a single scrambled pair implies the existence of an uncountable scrambled set. But this is not true for other classes of dynamical systems - shifts, maps on dendrites, triangular maps of the square.\\ {\bf Question 3.} Are all Li-Yorke sensitive shifts/maps on dendrites/triangular maps of the square Li-Yorke chaotic?\\ \end{remark} \section{Proofs} \begin{lemma} Let $p\in\mathbb{N}$,$m\in\mathbb{N}$ and $\epsilon>0.$ There are sequences $\{v_n\}_{n=1}^{\infty}$ and $\{u_n\}_{n=1} ^{\infty}$ such that \begin{equation}\label{v}\mbox{for every }n\in\mathbb{N}, \qquad \Big(\frac{1}{2^p}\sum_{j=0}^{v_n-1}\frac{1}{m+j}\Big) \mod 1<\epsilon, \end{equation} and \begin{equation}\label{u}\mbox{for every }n\in\mathbb{N},\qquad \Big|\Big(\frac{1}{2^p}\sum_{j=0}^{u_n-1}\frac{1}{m+j}\Big) \mod 1-\frac 1 2\Big|<\epsilon. \end{equation} \end{lemma} Lemma 1 follows by the simple fact that the harmonic series is divergent while its increment tends to 0. Therefore the $n$th partial sum of harmonic series modulo 1 is $\epsilon$-close to any number from $[0,1)$ for infinitely many $n$. \begin{lemma} For any $i\in\mathbb{N}$, let $\{\delta^n_i\}_{n=1}^{\infty}$ be a sequence of positive numbers not greater than 1, such that $\lim_{n\to\infty}\delta_i^n=0$. Then $$\lim_{n\to\infty}\sum_{i=1}^{\infty}\frac{1}{2^i}\delta_i^n=0.$$ \end{lemma} \begin{proof} For every $\epsilon>0$, there is $k\in\mathbb{N}$ such that $\epsilon>2^{-k+1}$. Since $\lim_{n\to\infty}\delta_i^n=0$, for every $i\in\mathbb{N}$, there is $N\in\mathbb{N}$ such that, for $n\geq N$ and $i\leq k$, $\delta_i^n<2^{-k}$. We can estimate, for $n\geq N$, $$\sum_{i=1}^{\infty}\frac{1}{2^i}\delta_i^n<\sum_{i=1}^{k}\frac{1}{2^i}\delta_i^n+\sum_{i=k+1}^{\infty}\frac{1}{2^i}<(1-2^{-k})\cdot 2^{-k}+2^{-k}<\epsilon.$$ \end{proof} \begin{lemma} Let $k\in\mathbb{N}$, $r\in\mathbb{N}$. Then $$\lim_{n\to\infty}\sum_{j=0}^{n-1}\Big(\frac{1}{k+j}-\frac{1}{k+r+j}\Big)=\sum_{j=0}^{r-1}\frac{1}{k+j}.$$ \end{lemma} \begin{proof} For sufficiently large $n$, $$\sum_{j=0}^{n-1}\Big(\frac{1}{k+j}-\frac{1}{k+r+j}\Big)=\sum_{j=0}^{r-1}\frac{1}{k+j}-\sum_{j=n-r}^{n-1}\frac{1}{k+r+j},$$ where the second term on the right side tends to 0 for $n\to\infty.$ \end{proof} \begin{claim} $x$ and $y$ defined in (\ref{y1}) and (\ref{y11}) is a scrambled pair with modulus $\frac 1 2$. \end{claim} \begin{proof} Denote the $i$th coordinate of $F^n(x)$ by $x_i^n$. First members of the sequences $\{x^n_0\}_{n=1}^{\infty}$ and $\{y^n_0\}_{n=1}^{\infty}$ are: $$x_0\mapsto \big(x_0+\sum_{i=1}^{M-1}\frac{1}{2^i}\frac{1}{x_i}\big)\mod 1\mapsto \big(x_0+\sum_{i=1}^{M-1}\frac{1}{2^i}\frac{1}{x_i}+\sum_{i=1}^{M-1}\frac{1}{2^i}\frac{1}{x_i+1}\big)\mod 1 \ldots,$$ $$y_0=x_0\mapsto \big(x_0+\sum_{i=1}^{M-1}\frac{1}{2^i}\frac{1}{x_i}+\sum_{i=M}^{\infty}\frac{1}{2^i}\frac{1}{K}\big)\mod 1\mapsto \big(x_0+\sum_{i=1}^{M-1}\frac{1}{2^i}\frac{1}{x_i}\sum_{i=M}^{\infty}\frac{1}{2^i}\frac{1}{K}+\sum_{i=1}^{M-1}\frac{1}{2^i}\frac{1}{x_i+1}+\sum_{i=M}^{\infty}\frac{1}{2^i}\frac{1}{K+1}\big)\mod 1 \ldots.$$ The following equations are with modulus 1 whenever necessary. Since $d_0(x_0^n,y_0^n)\leq |x^n_0-y^n_0|$, where \begin{equation}\label{abs}|x^n_0-y^n_0|=\sum_{i=M}^{\infty}\sum_{j=0}^{n-1}\frac{1}{2^i}\frac{1}{K+j}=2^{-M+1}\cdot\sum_{j=0}^{n-1}\frac{1}{K+j},\end{equation} and $$ d_i(x_i^n,y_i^n)=0, \quad\mbox{ for }0<i<M,\qquad d_i(x_i^n,y_i^n)=\frac{1}{K+n}, \quad\mbox{ for }i\geq M,$$ we can estimate $$D(F^n(x),F^n(y))\leq2^{-M+1}\cdot\sum_{j=0}^{n-1}\frac{1}{K+j}+\sum_{i=M}^{\infty}\frac{1}{2^{i}}\frac{1}{K+n}.$$ Let $\epsilon>0.$ By (\ref{v}) in Lemma 1, there is $\{v_n\}_{n=1}^{\infty}$ such that $\big(2^{-M+1}\cdot\sum_{j=0}^{v_n-1}\frac{1}{K+j}\big)\mod 1<\epsilon$, for $n\geq 1$. By Lemma 2 and since $\lim_{n\to\infty}\frac{1}{K+n}=0$, it holds, for sufficiently large $v$, $\sum_{i=M}^{\infty}\frac{1}{2^{i}}\frac{1}{K+{v_n}}<\epsilon.$ Therefore $\lim_{v\to\infty}D(F^{v_n}(x),F^{v_n}(y))<2\epsilon$ and $\liminf_{n\to \infty}D(F^n(x),F^n(y))=0.$ Similarly by (\ref{u}) in Lemma 1, there is $\{u_n\}_{n=1}^{\infty}$ such that $\Big|\Big(2^{-M+1}\sum_{j=0}^{u_n-1}\frac{1}{K+j}\Big) \mod 1-\frac 1 2\Big|<\epsilon,$ for sufficiently large $n$. Therefore, by (\ref{abs}), $d_0(x_0^{u_n},y_0^{u_n})>\frac 1 2 -\epsilon$ and $\limsup_{n\to \infty}D(F^n(x),F^n(y))\geq\frac 1 2.$ \end{proof} \begin{claim} $x$ and $y$ defined in (\ref{y2}), (\ref{y22}) and (\ref{y222}) is a scrambled pair with modulus $\frac 1 2$. \end{claim} \begin{proof} Denote the $i$th coordinate of $F^n(x)$ by $x_i^n$. First members of the sequences $\{x^n_0\}_{n=1}^{\infty}$ and $\{y^n_0\}_{n=1}^{\infty}$ are: $$x_0\mapsto \big(x_0+\sum_{i=1}^{\infty}\frac{1}{2^i}\frac{1}{x_i}\big)\mod 1\mapsto \big(x_0+\sum_{i=1}^{\infty}\frac{1}{2^i}\frac{1}{x_i}+\sum_{i=1}^{\infty}\frac{1}{2^i}\frac{1}{x_i+1}\big)\mod 1 \ldots,$$ \begin{multline*}y_0=\big(x_0+1-\sum_{i=1}^{\infty}\frac{1}{2^{i+M}}\delta_i\big)\mod 1\mapsto \big(x_0+1-\sum_{i=1}^{\infty}\frac{1}{2^{i+M}}\delta_i+\sum_{i=1}^{M-1}\frac{1}{2^i}\frac{1}{x_i}\big)\mod 1\mapsto \\\mapsto\big(x_0+1-\sum_{i=1}^{\infty}\frac{1}{2^{i+M}}\delta_i+\sum_{i=1}^{M-1}\frac{1}{2^i}\frac{1}{x_i}+\sum_{i=1}^{M-1}\frac{1}{2^i}\frac{1}{x_i+1}\big)\mod 1 \ldots . \end{multline*} Notice that for sufficiently large $n$, \begin{equation}\label{gamma}\delta_i+\sum_{j=0}^{n-1}\frac{1}{x_{i+M}+j}=\sum_{j=0}^{n-1}\frac{1}{x_M+j}+\gamma_i^n,\end{equation} where $$\gamma^n_i = \left\{ \begin{array}{ll} 0, \quad&\mbox{if } x_{i+M}=x_M, \\ \big(\sum_{j=x_M}^{x_{i+M}-1}\frac{1}{n+j}\big)\mod 1,\quad & \mbox{otherwise.} \end{array} \right.$$ The following equations are with modulus 1 whenever necessary. Since \begin{multline*} d_0(x_0^n,y_0^n)\leq |x^n_0-y^n_0|=\sum_{i=1}^{\infty}\frac{1}{2^{i+M}}\delta_i+\sum_{i=M}^{\infty}\sum_{j=0}^{n-1}\frac{1}{2^i}\frac{1}{x_i+j}=\sum_{i=1}^{\infty}\frac{1}{2^{i+M}}\big(\delta_i+\sum_{j=0}^{n-1}\frac{1}{x_{i+M}+j}\big)+\sum_{j=0}^{n-1}\frac{1}{2^M}\frac{1}{x_M+j}\\ \overset{(\ref{gamma})}{=}\sum_{i=0}^{\infty}\frac{1}{2^{i+M}}\big(\sum_{j=0}^{n-1}\frac{1}{x_{M}+j}+\gamma_i^n\big)+\sum_{j=0}^{n-1}\frac{1}{2^M}\frac{1}{x_M+j}=2^{-M+1}\cdot\sum_{j=0}^{n-1}\frac{1}{x_M+j}+\sum_{i=0}^{\infty}\frac{1}{2^{i+M}}\gamma_i^n, \end{multline*} and $$ d_i(x_i^n,y_i^n)=0, \quad\mbox{ for }0<i<M,\qquad d_i(x_i^n,y_i^n)=\frac{1}{x_i+n}, \quad\mbox{ for }i\geq M,$$ we can estimate $$D(F^n(x),F^n(y))\leq2^{-M+1}\cdot\sum_{j=0}^{n-1}\frac{1}{x_M+j}+\sum_{i=0}^{\infty}\frac{1}{2^{i+M}}\gamma_i^n+\sum_{i=M}^{\infty}\frac{1}{2^{i}}\frac{1}{x_i+n}.$$ Let $\epsilon>0.$ By (\ref{v}) in Lemma 1, there is $\{v_n\}_{n=1}^{\infty}$ such that $\big(2^{-M+1}\cdot\sum_{j=0}^{v_n-1}\frac{1}{x_M+j}\big)\mod 1<\epsilon$, for $n\geq 1$. By Lemma 2 and since $\lim_{n\to\infty}\gamma_i^n=0$ and $\lim_{n\to\infty}\frac{1}{x_i+n}=0$, for $i\geq 1$, it holds, for sufficiently large $n$, $$\sum_{i=0}^{\infty}\frac{1}{2^{i+M}}\gamma_i^{v_n}<\epsilon\qquad\mbox{ and }\qquad\sum_{i=M}^{\infty}\frac{1}{2^{i}}\frac{1}{x_i+{v_n}}<\epsilon.$$ Therefore $\lim_{v\to\infty}D(F^{v_n}(x),F^{v_n}(y))<3\epsilon$ and $\liminf_{n\to \infty}D(F^n(x),F^n(y))=0.$ Similarly by (\ref{u}) in Lemma 1, there is $\{u_n\}_{n=1}^{\infty}$ such that $$\Big|\Big(2^{-M+1}\sum_{j=0}^{u_n-1}\frac{1}{x_M+j}\Big) \mod 1+\sum_{i=0}^{\infty}\frac{1}{2^{i+M}}\gamma_i^{u_n}-\frac 1 2\Big|<2\epsilon,$$ for sufficiently large $n$. Therefore $d_0(x_0^{u_n},y_0^{u_n})>\frac 1 2 -2\epsilon$ and $\limsup_{n\to \infty}D(F^n(x),F^n(y))\geq\frac 1 2.$\\ \end{proof} \begin{claim} If $x_i$ and $y_i$ are finite, for all $i\geq 1$, then $\lim_{n\to\infty}D(F^n(x),F^n(y))$ exists and $(x,y)$ is not a scrambled pair.\end{claim} \begin{proof} Let $r_i=|x_i-y_i|.$ The following equations are with modulus 1 whenever necessary. First, observe by Lemma 3, \begin{equation}\label{x}\lim_{n\to\infty}\sum_{\substack{i=1 \\ x_i\leq y_i}}^{\infty}\frac{1}{2^i}\sum_{j=0}^{n-1}\big(\frac{1}{x_i+j}-\frac{1}{y_i+j}\big)=\lim_{n\to\infty}\sum_{\substack{i=1 \\ x_i\leq y_i}}^{\infty}\frac{1}{2^i}\sum_{j=0}^{n-1}\big(\frac{1}{x_i+j}-\frac{1}{x_i+r_i+j}\big)=\sum_{\substack{i=1 \\ x_i\leq y_i}}^{\infty}\frac{1}{2^i}\sum_{j=0}^{r_i-1}\frac{1}{x_i+j}\end{equation} and similarly, \begin{equation}\label{y}\lim_{n\to\infty}\sum_{\substack{i=1 \\ x_i>y_i}}^{\infty}\frac{1}{2^i}\sum_{j=0}^{n-1}\big(\frac{1}{y_i+j}-\frac{1}{x_i+j}\big)=\sum_{\substack{i=1 \\ x_i>y_i}}^{\infty}\frac{1}{2^i}\sum_{j=0}^{r_i-1}\frac{1}{y_i+j}.\end{equation} Since \begin{multline}\label{cele}|x^n_0-y^n_0|=\big|x_0-y_0+\sum_{i=1}^{\infty}\frac{1}{2^i}\sum_{j=0}^{n-1}\big(\frac{1}{x_i+j}-\frac{1}{y_i+j}\big)\big|=\\\big|x_0-y_0+\sum_{\substack{i=1 \\ x_i\leq y_i}}^{\infty}\frac{1}{2^i}\sum_{j=0}^{n-1}\big(\frac{1}{x_i+j}-\frac{1}{y_i+j}\big)-\sum_{\substack{i=1 \\ x_i>y_i}}^{\infty}\frac{1}{2^i}\sum_{j=0}^{n-1}\big(\frac{1}{y_i+j}-\frac{1}{x_i+j}\big)\big|,\end{multline} and $$d_i(x_i^n,y_i^n)=\big|\frac{1}{x_i+n}-\frac{1}{y_i+n}\big|, \quad\mbox{ for }i\geq 1,$$ it follows by Lemma 2 and by $\lim_{n\to\infty}\big|\frac{1}{x_i+n}-\frac{1}{y_i+n}\big|=0$, $$\lim_{n\to\infty}D(F^n(x),F^n(y))=\lim_{n\to\infty}d_0(x_0^n,y_0^n)+\lim_{n\to\infty}\sum_{i=1}^{\infty}\frac{1}{2^i}\big|\frac{1}{x_i+n}-\frac{1}{y_i+n}\big|=\min\{\lim_{n\to\infty}|x_0^n-y_0^n|,1-\lim_{n\to\infty}|x_0^n-y_0^n|\},$$ where by equations (\ref{x}), (\ref{y}), (\ref{cele}), $$\lim_{n\to\infty}|x_0^n-y_0^n|=\big|x_0-y_0+\sum_{\substack{i=1 \\ x_i\leq y_i}}^{\infty}\frac{1}{2^i}\sum_{j=0}^{r_i-1}\frac{1}{x_i+j}-\sum_{\substack{i=1 \\ x_i>y_i}}^{\infty}\frac{1}{2^i}\sum_{j=0}^{r_i-1}\frac{1}{y_i+j}\big |.$$ Therefore $\lim_{n\to\infty}D(F^n(x),F^n(y))$ exists which is in contradiction with $(x,y)$ being a scrambled pair. \end{proof} \begin{claim} For $x$ and $y$ defined in (\ref{y3}), $\lim_{n\to\infty}D(F^n(x),F^n(y))$ exists and $(x,y)$ is not a scrambled pair.. \end{claim} \begin{proof} Let $r_i=|x_i-y_i|$, for $0<i<l$. By similar calculation as in Claim 3, $$\lim_{n\to\infty}D(F^n(x),F^n(y))=\lim_{n\to\infty}d_0(x_0^n,y_0^n)+\lim_{n\to\infty}\sum_{i=1}^{l-1}\frac{1}{2^i}\big|\frac{1}{x_i+n}-\frac{1}{y_i+n}\big|=\min\{\lim_{n\to\infty}|x_0^n-y_0^n|,1-\lim_{n\to\infty}|x_0^n-y_0^n|\},$$ where $$\lim_{n\to\infty}|x_0^n-y_0^n|=\big|x_0-y_0+\sum_{\substack{i=1 \\ x_i\leq y_i}}^{l-1}\frac{1}{2^i}\sum_{j=0}^{r_i-1}\frac{1}{x_i+j}-\sum_{\substack{i=1 \\ x_i>y_i}}^{l-1}\frac{1}{2^i}\sum_{j=0}^{r_i-1}\frac{1}{y_i+j}\big |.$$ Therefore $\lim_{n\to\infty}D(F^n(x),F^n(y))$ exists which is in contradiction with $(x,y)$ being a scrambled pair. \end{proof} \paragraph{\bf Acknowledgment} The author thanks to professor Jaroslav Sm\' ital for stimulating discussions and useful suggestions. Research was funded by institutional support for the development of research organization (I\v C 47813059).

\section{Introduction} Let $k\ge 2$ be an integer. A {\it $k$-automatic sequence} $(u_n)$ over an alphabet ${\cal A}$ is the output sequence of a finite automaton, where the input is the $k$-ary digital expansion of $n$. Automatic sequences have gained much attention during the last decades. For monographs and surveys about automatic sequences we refer to \cite{al,alsh,dr,RecurenceSequences}. For a prime power $k=q$, $q$-automatic sequences $(u_n)$ over the finite field ${\cal A}=\mathbb{F}_q$ of $q$ elements can be characterized by a result of Christol, see \cite{christol} for prime $q$ and \cite{chka} for prime power $q$ as well as \cite[Theorem~12.2.5]{alsh}: Let $$G(t)=\sum_{n=0}^{\infty}u_nt^n$$ be the {\em generating function} of the sequence $(u_n)$ over $\mathbb{F}_q$. Then $(u_n)$ is $q$-automatic over $\mathbb{F}_q$ if and only if $G(t)$ is algebraic over $\mathbb{F}_q[t]$, that is, there is a polynomial {\color{black} $h(s,t)\in \mathbb{F}_q[s,t]\setminus\{0\}$} such that $h(G(t),t)=0$. (Note that for all $m=1,2,\ldots$ a sequence is $k$-automatic if and only if it is $k^m$-automatic by \cite[Theorem~6.6.4]{alsh} and even a slightly more general version of Christol's result holds: {\color{black} For a prime $p$ and positive integers $m$ and $r$, $(u_n)$ is $p^m$-automatic over $\mathbb{F}_{p^r}$ if and only if $G(t)$ is algebraic over $\mathbb{F}_{p^r}$.}) Diem \cite{di} defined the {\em $N$th expansion complexity} $E_N(u_n)$ of $(u_n)$ as the least total degree of a nonzero polynomial $h(s,t)\in \mathbb{F}_q[s,t]$ with $$h(G(t),t)\equiv 0 \bmod t^N$$ if the first $N$ sequence elements are not all $0$ and $E_N(u_n)=0$ otherwise. Hence, the $q$-automatic sequences over $\mathbb{F}_q$ are exactly the sequences over $\mathbb{F}_q$ with $$E(u_n)=\sup_{N\ge 1} E_N(u_n)<\infty.$$ Sequences $(u_n)$ with small $E(u_n)$ are predictable and not suitable in cryptography. For example, the {\em Thue--Morse sequence} over $\mathbb{F}_2$ is defined by $$t_n=\left\{ \begin{array}{cl} t_{n/2} & \mbox{if $n$ is even},\\ t_{(n-1)/2}+1 & \mbox{if $n$ is odd}, \end{array}\right.\quad n=1,2,\ldots $$ with initial value $t_0=0$. Taking $$h(s,t)=s(t+1)^2+s^2(t+1)^3+t$$ its generating function $G(t)$ satisfies $h(G(t),t)=0$ and thus $E(t_n)\le 5$. More precisely, in the proof of Corollary~\ref{} below we will see that $G(t)$ is not rational and thus $h(s,t)=(t+1)^3(s+G(t))(s+G(t)+\frac{1}{t+1})$ is irreducible over $\mathbb{F}_2(t)[s]$ and we get $$E(t_n)=5.$$ The \textit{$N$th linear complexity} $L(u_n, N)$ of a sequence $(u_n)$ over $\mathbb{F}_q$ is the length~$L$ of a shortest linear recurrence relation satisfied by the first $N$ elements of $(u_n)$: $ u_{n+L}=c_{L-1}u_{n+L-1}+\dots +c_1u_{n+1}+c_0u_n, \quad 0\leq n\leq N-L-1, $ for some $c_0,\ldots,c_{L-1}\in \mathbb{F}_q$. We use the convention that $L(u_n,N)=0$ if the first $N$ elements of $(u_n)$ are all zero and $L(u_n,N)=N$ if $u_0=\dots=u_{N-2}=0\ne u_{N-1}$. The sequence $(L(u_n,N))$ is called {\em linear complexity profile} of $(u_n)$ and $$L(u_n)=\sup_{N\ge 1} L(u_n,N)$$ is the {\em linear complexity} of $(u_n)$. It is well-known (\cite[Lemma~1]{Niederreiter88-b}) that $L(u_n)<\infty$ if and only if $(u_n)$ is ultimately periodic, that is, its generating function is rational: $G(t)=g(t)/f(t)$ with polynomials $g(t),f(t)\in \mathbb{F}_q[t]$. The $N$th linear complexity is a measure for the unpredictability of a sequence as well. A large $N$th linear complexity (up to sufficiently large $N$) is necessary (but not sufficient) for cryptographic applications. Sequences of small linear complexity are also weak in view of Monte-Carlo methods, see \cite{do,domewi,dowi,DorferWinterhof}. For more background on linear complexity and related measures of pseudorandomness we refer to \cite{mewi,Niederreiter2003,towi,Winterhof2010}. In particular, for ultimately periodic sequences the relation between linear complexity and expansion complexity was studied in \cite{meniwi}. In this paper, we show that any $q$-automatic sequence over $\mathbb{F}_q$ which is not ultimately periodic has $N$th linear complexity of (best possible) order of magnitude $N$. Hence, we provide many examples of sequences with high $N$th linear complexity which are still predictable (since $E(u_n)$ is small). For example, for the Thue--Morse sequence over $\mathbb{F}_2$ we prove, see Theorem~\ref{} below, \begin{equation}\label{tmlin} L(t_n,N)=2\left\lfloor\frac{N+2}{4}\right\rfloor,\quad N=1,2,\ldots \end{equation} In Section~\ref{sec2} we prove a bound on the $N$th linear complexity of any $q$-automatic sequence over $\mathbb{F}_q$ which is not ultimately periodic. We apply this result to several famous automatic sequences including pattern sequences and sum-of-digits sequences. For example, for the Thue--Morse sequence our result implies the bound $$\left\lceil\frac{N-1}{2}\right\rceil \le L(t_n,N)\le \left\lfloor \frac{N}{2}\right\rfloor+1.$$ By $(\ref{tmlin})$ the lower bound is attained if $N\equiv 0,1\bmod 4$ and the upper bound if $N\equiv 2,3\bmod 4$. In Section~\ref{sec3} we determine the exact value of the $N$th linear complexity in the special case of binary pattern sequences with the all one pattern. The Thue--Morse sequence and the Rudin--Shapiro sequence are the simplest examples of such sequences. Besides a small expansion complexity, the Thue--Morse and Rudin--Shapiro sequences have another deficiency, a very large correlation measure of order~$2$, see \cite{mauduitSarkozy-TM}. By a result of \cite{BrandstatterWinterhof}, small correlation measures of order $k$ (up to a sufficiently large $k$) imply large $N$th linear complexities. The converse is not true. For example, the correlation measure of order $4$ of the {\em Jacobi-sequence} $(j_n)$ of length $pq$ with two distinct odd primes $p$ and $q$ is of order of magnitude $pq$, see \cite{risa}, which is based on the relation $j_n+j_{n+p}+j_{n+q}+j_{n+p+q}=0$ for all {\color{black}$n$ with $1\le n<pq$ and} $\gcd(n,pq)=1$. However, its linear complexity profile is quite large, see \cite{brwi}. This lower bound on the linear complexity profile can be obtained using an analog of the result of \cite{BrandstatterWinterhof}, see \cite{hepawawi}, for a modified correlation measure with {\em bounded lags}. Examples of sequences with large correlation measure of small order with bounded lags but large linear complexity profile were not known before. However, the results of this paper for the Thue--Morse and Rudin--Shapiro sequence and the results of \cite{mauduitSarkozy-TM} (using lags $(0,1)$) show that both sequences are such examples. \section{Arbitrary $q$-automatic sequences over $\mathbb{F}_q$} \label{sec2} {\color{black} Ultimately periodic sequences} (that {\color{black} is linear recurrence sequences) are automatic and} correspond to rational generating functions \cite[Lemma 1]{Niederreiter88-b}. If $(u_n)$ is an {\color{black} ultimately periodic sequence}, then it has finite linear complexity. However, we show now that if $(u_n)$ is {\color{black} automatic but not} ultimately periodic, then the $N$th linear complexity of $(u_n)$ is of order of magnitude $N$ for all $N$. \begin{thm}\label{} Let $q$ be a prime power and $(u_n)$ be a $q$-automatic sequence over $\mathbb{F}_q$ which is not ultimately periodic. Let $h(s,t)=h_0(t)+h_1(t)s+\dots + h_d(t)s^d\in\mathbb{F}_q[s,t]$ be a non-zero polynomial with $h(G(t),t)=0$ with no rational zero. Put $$M=\max_{0\leq i\leq d}\{\deg h_i-i\}.$$ Then we have \[ \frac{\displaystyle N-M}{d}\leq L(u_n,N)\leq \frac{\displaystyle (d-1)N+M+1}{d}. \] \end{thm} \begin{proof} Since $(u_n)$ is not ultimately periodic, $G(t)\not\in \mathbb{F}_q(t)$ is not rational by \cite[Lemma~1]{Niederreiter88-b}. Let $g(t)/f(t)\in\mathbb{F}_q(t)$ be a rational zero of $h(s,t)$ modulo $t^N$ with $\deg(f)\le L(u_n,N)$ and $\deg(g)<L(u_n,N)$. More precisely, put $L=L(u_n,N)$. Then we have \begin{equation}\label{lincomp} \sum_{\ell=0}^Lc_{\ell}u_{n+\ell}=0 \quad \mbox{for }0\le n\le N-L-1 \end{equation} for some $c_0,\ldots,c_{L}\in \mathbb{F}_q$ with $c_L\neq 0$. Take \[ f(t)=\sum_{\ell=0}^L c_{\ell}t^{L-\ell} \] and \[ g(t)=\sum_{m=0}^{L-1}\left(\sum_{\ell=L-m}^Lc_\ell u_{m+\ell-L}\right)t^m \] and verify $$f(t)G(t)\equiv g(t)\bmod t^N:$$ $$f(t)G(t)\equiv \sum_{\ell=0}^L c_\ell t^{L-\ell}\sum_{k=0}^{N-1}u_k t^k \equiv \sum_{m=0}^{N-1} \sum_{\ell=0}^L c_\ell u_{m+\ell-L} t^m\equiv g(t)\bmod t^N,$$ where we put $m=L-\ell+k$ and used $(\ref{lincomp})$ with $n=m-L$ in the last step. Then \[ h_0(t)f^d(t)+h_1(t)g(t)f^{d-1}(t)+\dots + h_d(t)g(t)^d=K(t) t^N. \] Here $K(t)\neq 0$ since $h(s,t)$ has no rational zero. Comparing the degrees of both sides we get \[ dL+M\geq N \] which gives the lower bound. The upper bound for $N=1$ is trivial. For $N\geq 2$ the result follows from the well-known bound (see for example \cite[Lemma 3]{DorferWinterhof}) \[ L(u_n,N)\leq \max\left\{L(u_n,N-1), N-L(u_n,N-1) \right\} \] by induction \begin{eqnarray*} L(u_n,N)&\le& \max\left\{\frac{(d-1)(N-1)+M+1}{d},N-\frac{N-1-M}{d}\right\}\\ &\le& \frac{(d-1)N+M+1}{d}, \end{eqnarray*} where we also used the lower bound on $L(u_n,N-1)$ to estimate $N-L(u_n,N-1)$. \end{proof} \subsection*{Examples} Now we state bounds on the $N$th linear complexity of some famous automatic sequences as corollaries of Theorem \ref{}. In the following let $p$ be a prime. \subsubsection*{Pattern sequences} Let $P\in \mathbb{F}_p^k\setminus\{00\ldots 0\}$ be a \emph{pattern of length $k$}. Let $e_P(n)$ be the number of occurrences of $P$ in the $p$-ary representation of $n$. For example if $p=2$, then $e_{11}(7)=2$, $e_1(9)=2$, $e_{11}(9)=0$ and $e_{101}(21)=2$. For a pattern $P$ of length $k$ define the sequence $(r_n)$ by \begin{equation*} r_n\equiv e_P(n) \mod p, \quad r_n\in\F_p,\quad n=0,1,\dots \end{equation*} The sequence $(r_n)$ over $\mathbb{F}_p$ satisfies the following recurrence relation \begin{equation}\label{} r_n= \left\{ \begin{array}{cl} r_{\lfloor n/p\rfloor}+1 & \text{if } n\equiv a \mod p^k,\\ r_{\lfloor n/p\rfloor} & \text{otherwise,} \end{array} \right. n=1,2,\dots \end{equation} with initial value $r_0=0$, where $a$ is the {\color{black} integer in the range} $0< a <p^k$ such that its $p$-ary expansion {\color{black} is} the pattern $P$. Classical examples for binary pattern sequences are the Thue--Morse sequence ($p=2$, $k=1$ and $P=1$ ($a=1$)) and the {\em Rudin--Shapiro sequence} ($p=2$, $k=2$ and $P=11$ ($a=3$)). \begin{cor}\label{} Let $p$ be a prime number and $a,k$ integers with $1\leq a<p^k$. If $(r_n)$ is the pattern sequence defined by \eqref{}, then \[ \frac{N-p^k+1}{p}\leq L(r_n,N) \leq \frac{(p-1)N+p^k}{p}. \] \end{cor} \begin{proof} By the recurrence relation \eqref{} we have \begin{equation*} G(t)=(1+t+\dots + t^{p-1})G(t)^p+\sum_{i=0}^\infty t^{p^k i+a}. \end{equation*} Multiplying with $t^{p^k}-1=(t-1)^{p^k}$ we have \begin{equation}\label{} (t-1)^{p^k}G(t)=(t-1)^{p^k+p-1}G(t)^p-t^a, \end{equation} {\color{black} see also \cite[Th\'eor\`em 3]{chka}.} Put $h(s,t)=(t-1)^{p^k+p-1}s^p-(t-1)^{p^k}s-t^a$. Then $h(s,t)$ has no rational zero in $\F_p(t)$. Indeed, if $g(t)/f(t)$ is a zero (with $\gcd(g,f)=1$), that is \[ (t-1)^{p^k+p-1}g(t)^p-(t-1)^{p^k}g(t)f(t)^{p-1}-t^af(t)^p=0, \] then $(t-1)\mid f(t)$ and so $(t-1)^{p^k+p-1}\mid f(t)^p$ thus $(t-1)^{p^{k-1}+1}\mid f(t)$ which means that $t-1\mid g(t)$, a contradiction. Finally, the result follows from Theorem \ref{}. \end{proof} \subsubsection*{The sum-of-digits sequence} Let $k>1$ be an integer and $\sigma_m(n)$ be the sum of digits of $n$ in the $k$-ary representation. Then define $s_n=\sigma_m(n)\bmod k$. Clearly $(s_n)$ satisfies the following recurrence relation \begin{equation}\label{} s_n\equiv s_{\lfloor \frac{n}{k}\rfloor}+n \mod k,\quad n=1,2,\ldots \end{equation} with initial value $s_0=0$. \begin{cor}\label{} Let $p$ be a prime number and $(s_n)$ be the sum-of-digit sequence modulo $p$ defined by \eqref{} with $k=p$. Then \[ \frac{N-1}{p} \leq L(s_n,N) \leq \frac{(p-1)N+2}{p}. \] \end{cor} For $p=2$ $(s_n)$ is the Thue--Morse sequence again and in this case Corollary \ref{} coincides with Corollary \ref{}. \begin{proof} As before, by the recurrence relation \eqref{} we have \begin{align}\label{} G(t)&=(1+t+\dots + t^{p-1})G(t)^p+\sum_{i=0}^\infty t^{ip }\cdot \sum_{a=1}^{p-1}at^a \notag \\ &=\frac{1-t^p}{1-t}G(t)^p+\frac{1}{1-t^{p}}\cdot t(1-t)^{p-2}.\notag \end{align} Multiplying by $(1-t)^{2}$ we get that $G(t)$ is a root of \[ h(s,t)=(1-t)^{p+1}s^p- (1-t)^{2}s+ t. \] As in the previous proof, it can be shown that $h(s,t)$ has no rational zero in $\F_p(t)$, thus the result follows from Theorem \ref{}. \end{proof} \subsubsection*{Baum--Sweet sequence} The Baum--Sweet sequence $(b_n)$ is a $2$-automatic sequence defined by the rule $b_0=1$ and for $n\ge 1$ $$ b_n= \left\{ \begin{array}{cl} 1& \text{if the binary representation of $n$ contains no block of consecutive} \\ & \text{$0$'s of odd length;}\\ 0& \text{otherwise.} \end{array} \right. $$ Equivalently we have for $n\ge 1$ of the form $n=4^km$ with $m$ not divisible by $4$ \begin{equation*} b_n=\left\{\begin{array}{ll} 0 & \mbox{if $m$ is even},\\ b_{(m-1)/2} & \mbox{if $m$ is odd}.\end{array}\right. \end{equation*} \begin{cor} Let $(b_n)$ be the Baum--Sweet sequence. Then \[ \frac{N}{3}\leq L(b_n,N)\leq \frac{2N+1}{3}. \] \end{cor} \begin{proof} It can be easily checked that the generating function $G(t)$ of $(b_n)$ is the unique root of $h(s,t)=s^3+t\cdot s+1$, {\color{black} see \cite[p.\ 403]{chka}}. Then the bounds follow from Theorem~\ref{}. \end{proof} \subsubsection*{Regular paperfolding sequence} The value of any given term $v_n\in \mathbb{F}_2$ in the regular paperfolding sequence can be defined as follows. If $n = m\cdot2^k$ where $m$ is odd, then \begin{equation*} v_n= \left\{ \begin{array}{cl} 1& \text{if } m\equiv 1 \mod 4,\\ 0& \text{if } m\equiv 3 \mod 4, \end{array} \right. n=1,2,\ldots \end{equation*} and any $v_0\in \mathbb{F}_2$. \begin{cor} Let $(v_n)$ be the regular paperfolding sequence. Then \[ \frac{N-3}{2}\leq L(v_n,N)\leq \frac{N}{2}+2. \] \end{cor} \begin{proof} It can be checked that the generating function $G(t)$ of $(v_n)$ is a zero of $h(s,t)=(t^4+1)\cdot s^2+(t^4+1)\cdot s+t=(t^4+1)(s+G)(s+G+1)$ which has no rational zero. Then the bounds follow from Theorem \ref{}. \end{proof} \subsubsection*{A sequence with perfect lattice profile and perfect linear complexity profile} The generating function $G(t)$ of the sequence $(w_n)$ over $\mathbb{F}_2$ defined by $$w_{2n}=1\quad \mbox{and} \quad w_{2n+1}=w_n+1,\quad n=0,1,\ldots$$ satisfies the functional equation $$t(t+1)G(t)^2+(t+1)G(t)+1=0.$$ {\color{black} (This sequence or its complement are sometimes called {\it Toeplitz sequence} or {\it period doubling sequence}, see .)} Hence, $L(w_n,N)=\lfloor \frac{N+1}{2}\rfloor$. Such a linear complexity profile is called perfect. This is the only sequence with both a perfect linear complexity profile and a perfect 'lattice profile', see \cite{domewi} for more details. Sequences with the first are characterized by $w_0=1$ and $w_{2n+2}=w_{2n+1}+w_n$ but the choice of $w_{2n+1}$ is free for $n\ge 1$, see \cite{Niederreiter88-b}. Sequences with the latter are characterized by $w_{2n+1}=w_n+1$ but the choice of any $w_{2n}$ is free, see \cite{domewi}. \section{Pattern sequences with the all one pattern} \label{sec3} In this section we slightly improve Theorem \ref{} for the binary pattern sequences when the pattern $P$ is the all one string $11\dots 1$ of length $k$ (that is $a=2^k-1$). \begin{thm}\label{} The $N$th linear complexity $L(r_n,N)$ of the sequence $(r_n)$ defined by \eqref{} with $a=2^k-1$ (and $p=2$) satisfies \[ L(r_n,N)= \left\{ \begin{array}{cl} 2(2^k-1)\left\lfloor \frac{N}{4(2^k-1)}\right\rfloor+2^k & \text{if }2^k \leq N \bmod 4(2^k-1) \leq 3(2^k-1),\\[5pt] 2(2^k-1)\left\lfloor \frac{N+2^k-2}{4(2^k-1)}\right\rfloor & \text{otherwise.} \end{array} \right. \] \end{thm} The proof of the theorem is based on the theory of continued fractions. Thus first we summarize some basic facts about them (Section \ref{}), then we prove the result (Section \ref{}). \subsection{Linear complexity and continued fractions}\label{} Now we describe the connection between linear complexity profile of a binary sequence and the continued fraction expansion of its the generating function (see for example \cite{Niederreiter88-b}). Let $\mathbb{F}_2((x^{-1}))$ be the ring of formal Laurent series \[ R=\sum_{i=-m}^\infty r_ix^{-i}, \quad r_{-m}, r_{-m+1}, \dots \in \mathbb{F}_2 \] over $\mathbb{F}_2$ in $x^{-1}$. In fact $\mathbb{F}_2((x^{-1}))$ is a field and the coefficients of the inverse of a non-zero $R$ can be computed recursively, namely, if $r_{-m}=1$, then \[ R^{-1}=\sum_{i=m}^\infty z_ix^{-i} \] with coefficients \begin{equation}\label{S-inverse} z_m=1 \quad \text{and } z_{i+m}=r_{i-m}+\sum_{j=-m+1}^{i-m-1}r_{j}z_{i-j}, \quad i=1,2,\dots \end{equation} Every formal non-rational Laurent series $R\in\mathbb{F}_2((x^{-1}))$ has a unique continued fraction expansion \[ R=A_0+ \cfrac{1}{A_1+ \cfrac{1}{A_2+ \cfrac{1}{\ddots} }}, \] where $A_j\in\mathbb{F}_2[x]$ are polynomials for $j\geq 0$ and $\deg A_j\geq 1$ for $j\geq 1$. For \[ R=\sum_{i=-m}^\infty r_ix^{-i}\in\mathbb{F}_2((x^{-1})) \] we define its polynomial part by \[ \mathop{\mathrm{Pol}}(R)= \sum_{i=-m}^0 r_ix^{-i}. \] The polynomials $A_j$ ($j\geq 0$) are obtained recursively by \begin{equation*}\renewcommand{\arraystretch}{1.5} \begin{array}{ll} A_0=\mathop{\mathrm{Pol}}(R), & B_0=R-\mathop{\mathrm{Pol}}(R),\\ A_{j+1}=\mathop{\mathrm{Pol}}(B_j^{-1}), & B_{j+1}=B_{j}^{-1}-\mathop{\mathrm{Pol}}(B_j^{-1}), \quad j\geq 0. \end{array} \end{equation*} If the continued fraction expansion is broken off after the term $A_j$ ($j\geq 0$), we get the rational convergent $P_j/Q_j$. The polynomials $P_j, Q_j$ can be calculated recursively by \begin{equation}\label{} \begin{array}{llll} P_{-1}=1, & P_0=A_0, & P_j=A_jP_{j-1}+P_{j-2}, & j\geq 1, \\[5pt] Q_{-1}=0, & Q_0=1, & Q_j=A_jQ_{j-1}+Q_{j-2}, & j\geq 1. \end{array} \end{equation} The following formulas are shown by straightforward induction on $j$: \begin{align} & \deg (Q_j)=\sum_{h=1}^j \deg A_h \quad \text{for } j\geq 1, \label{} \\[10pt] & P_{j-1}Q_j+P_jQ_{j-1}=1 \quad \text{for } j\geq 1,\label{} \\[12pt] & R=\frac{P_j+B_jP_{j-1}}{Q_j+B_jQ_{j-1}} \quad \text{for } j\geq 0. \label{} \end{align} With $x=t^{-1}\in\mathbb{F}_2(t)$ we obtain that the generating function of the sequence $(r_n)$ over $\mathbb{F}_2$ is \[ G=\sum_{n=0}^\infty r_nx^{-n}. \] The following lemma \cite[Theorem 1]{Niederreiter88-b} gives an explicit description of the linear complexity profile of the sequence $(r_n)$ in terms of the polynomials $Q_j$ that are obtained from the continued fraction expansion of \begin{equation}\label{} R=x^{-1}G=\sum_{i=1}^{\infty}r_{i-1}x^{-i}. \end{equation} \begin{lemma}\label{} For any $N\geq 1$ the $N$th linear complexity $L(r_n,N)$ is given by \[ L(r_n,N)=\deg Q_j \] where $j\geq 0$ is uniquely determined by the condition \[ \deg Q_{j-1}+\deg Q_{j}\leq N < \deg Q_{j}+\deg Q_{j+1}. \] \end{lemma} We define the (exponential) valuation $\v$ on $\mathbb{F}_2((x^{-1}))$ as \[ \v(R)=m, \quad \text{if } R=\sum_{i=-m}^\infty r_ix^{-i}\in\mathbb{F}_2((x^{-1})) \text{ and } r_{-m}\neq 0. \] For $R=0$ we put $\v(R)=-\infty$. We have the following properties of $\v$. For $R,S\in \mathbb{F}_2((x^{-1})) $ we have \begin{enumerate}[(i)] \item $\v(RS)=\v(R)+\v(S)$, \item $\v(R+S)\leq \max\{\v(R),\v(S)\}$, \item $\v(R+S)= \max\{\v(R),\v(S)\}$ if $\v(R)\neq \v(S)$. \end{enumerate} The valuation $\v$ extends the degree function on $\mathbb{F}_2[x]$: for $f\in\mathbb{F}_2[x]$ we have $\v(f)=\deg f$ and \[ \v\left( \frac{f}{g}\right)=\deg f -\deg g \quad \text{for } f,g\in\mathbb{F}_2[x], g\neq 0. \] \subsection{Proof of Theorem \ref{} It follows from the functional equation \eqref{} for the generating function of the sequence $(r_n)$ that the function $R$ defined in \eqref{} satisfies \begin{equation}\label{} (1+x)R^2+R+U^{2^k}x^{-2^k}=0, \end{equation} where $U=\sum_{i=0}^\infty x^{-i}$. \begin{lemma}\label{} If for a polynomial $Q\in\mathbb{F}_2[x]$ we have \[ QU^{2^{k}}=\sum_{i=-\deg Q}^{\infty} b_ix^{-i}, \] then $b_i=b_j$ for $i,j\geq 0$ with $i\equiv j \mod 2^{k}$ and \[ Q\equiv b_1 x^{2^{k}-1} + b_2 x^{2^{k}-2}+ \dots + b_{2^{k}} \mod x^{2^{k}}+1. \] \end{lemma} \begin{proof} Write $Q=x^d+q_{d-1}x^{d-1}+\dots +q_1x+q_0$. Then for $i \geq 0$ the coefficient of $QU^{2^k}$ of $x^{-i}$ is \begin{equation*} b_{i}=\sum_{j: \ i+j\equiv 0 \bmod 2^{k}} q_j, \end{equation*} and the result follows. \end{proof} \begin{proof}[Proof of Theorem \ref{}] We prove the following assertions by induction. \begin{enumerate}[(i)] \item \label{} $A_1=x^2+x+1$ if $k=1$ and $A_1=x^{2^k}+x$ if $k\geq 2$. \item \label{ $j\geq 2$, $A_j=x^2+1$ if $k=1$ and for $j\geq 1$, $A_{2j}=x^{2^k-2}+x^{2^k-4}+\dots +x^2+1$, $A_{2j+1}=x^{2^k}+1$ if $k\geq 2$. \item \label{ $j\geq 0$, $Q_{j}\equiv 1 \mod x+1$ if $k=1$ and $Q_{2j}\equiv 1 \mod x^{2^{k-1}}+1$, $Q_{2j+1}\equiv x+1 \mod x^{2^{k-1}}+1$ for $k\geq 2$. \end{enumerate} Then the result follows from Lemma \ref{}, \eqref{}, \eqref{} and (\ref{}). The first part follows from straightforward computation. Namely, observe that the first $2^{k+1}$ elements of the sequence $(r_n)$ are zeros, except the elements $r_n$ with $n=2^k-1$ and $n=2^{k+1}-2$. Thus \[ R=\frac{1}{x^{2^k}}+\frac{1}{x^{2^{k+1}-1}}+ \sum_{i=2^{k+1}+1}^\infty r_{i-1}x^{-i}, \] whence by \eqref{S-inverse} we have \[ R^{-1}=x^{2^k}+x+x^{2-2^k}+\dots \] so \[ A_1=\mathop{\mathrm{Pol}} (R^{-1})= \left\{ \begin{array}{ll} x^{2}+x+1 & \text{if $k=1$},\\ x^{2^k}+x & \text{if $k\geq 2$} \end{array} \right. \] which proves \eqref{}. By Corollary \ref{}, $R$ is irrational and thus $\deg A_j \geq 1$ for all $j\geq 1$. Now by \eqref{}, \eqref{}, (\ref{}) and the properties of $\v$ it follows for $l\geq 1$ that \begin{equation}\label{} \v(Q_{l-1}R-P_{l-1})=-\v(Q_{l}) \end{equation} whence \begin{equation}\label{} \v(Q_{l-1}^2R^2-P^2_{l-1})=-2\v(Q_{l}). \end{equation} Consider \begin{align} T&=(x+1)P_{l-1}^2+P_{l-1}Q_{l-1}+U^{2^k}x^{-2^k}Q_{l-1}^2 \label{} \\ &=(x+1)\left(Q^2_{l-1}R^2-P_{l-1}^2 \right)+Q_{l-1}(P_{l-1}+Q_{l-1}R),\label{} \end{align} where the second equality follows from \eqref{}. Now we have \begin{equation}\label{} \v(T)=-\v(A_{l}), \quad l\geq 1, \end{equation} by \eqref{}, \eqref{}, \eqref{}, \eqref{} and the properties of $\v$. It follows from \eqref{}, \eqref{} and \eqref{}, that \begin{equation}\label{eq-A_l-inverse} Q_{l-1}(P_{l-1}+Q_{l-1}R)=Q_{l-1}\frac{1}{A_{l}Q_{l-1}+Q_{l-2}+B_{l}Q_{l-1}}=\frac{1}{A_{l}+\frac{Q_{l-2}}{Q_{l-1}}+B_{l}} \end{equation} where we have \begin{equation}\label{}} \v(B_{l}), \v\left(\frac{Q_{l-2}}{Q_{l-1}}\right)<0. \end{equation} On the other hand \begin{align}\label{} \v\left((x+1)\left(Q^2_{l-1}R^2-P_{l-1}^2 \right)\right)&=1-2\v(Q_l)=-2\v(A_l)+1-2\v(Q_{l-1}) \notag \\ &\leq -2\v(A_l)+1-2\v(A_1)<-2\v(A_l) \end{align} for $l\ge 2$. We now prove \eqref{} and \eqref{} by induction. We remark that the assertion \eqref{} for $Q_{0}$ and $Q_1$ follows from \eqref{} and \eqref{}. Take $l\geq 2$. If $k=1$ or $k\geq 2$ and $l=2j+1$ is odd, then $Q_{l-1}\equiv 1 \mod x^{2^{k-1}}+1$, thus $Q_{l-1}^2\equiv 1 \mod x^{2^{k}}+1$. By Lemma \ref{}, the coefficients of $x^{-i}$ ($i=1,\dots, 2^k-1$) in $Q_{l-1}^2U^{2^k}$ are all zero but the coefficient of $x^{-2^k}$ is one. Since $R$ is irrational, the degree $\deg A_l$ and so $\v(T)$ cannot be zero. Thus by \eqref{} $x^{-2^k}$ is the leading term of $T$ so $\deg (A_l)=2^k$ by \eqref{}. By Lemma \ref{}, the coefficient of $x^{-i}$ ($i=2^k+1,\dots, 2^{k+1}-1$) in $Q_{l-1}^2U^{2^k}$ is zero. Thus by \eqref{} \[ Q_{l-1}(P_{l-1}+Q_{l-1}R)= x^{-2^k}+x^{2^{k+1}}+\sum_{i=2^{k+1}}^{\infty}a_ix^{-i}. \] It follows from \eqref{}} that \[ A^{-1}_{l}= x^{-2^k}+x^{2^{k+1}}+\sum_{i=2^{k+1}}^{\infty}a'_ix^{-i}, \] whence $A_l=x^{2^k}+1$ by \eqref{S-inverse}. If $k\geq 2$ and $l=2j$ is even, the proof of \eqref{} is similar. Since $Q_{l-1}\equiv x+1 \mod x^{2^{k-1}}+1$, thus $Q_{l-1}^2\equiv x^2+1 \mod x^{2^{k}}+1$. By Lemma \ref{}, the coefficients of $x^{-i}$ ($i=1,\dots, 2^k-3$) in $Q_{l-1}^2U^{2^k}$ are all zero but the coefficient of $x^{-2^k+2}$ is one. Thus by \eqref{} $x^{-2^k+2}$ is the leading term of $T$ so $\deg (A_l)=2^k-2$ by \eqref{}. By Lemma \ref{}, the coefficient of $x^{-i}$ ($i=2^k-1,\dots, 2^{k+1}-1$) in $Q_{l-1}^2U^{2^k}$ are all zeros except for $i=2^{k+1}-2$. Thus By \eqref{} \[ Q_{l-1}(P_{l-1}+Q_{l-1}R)= x^{-2^k+2}+x^{2^{k}} +x^{-2^{k+1}-2}+\sum_{i=2^{k+1}-1}^{\infty}a_ix^{-i}. \] It follows from \eqref{}} that \[ A^{-1}_{l}= x^{-2^k+2}+x^{2^{k}} +x^{-2^{k+1}-2}+\sum_{i=2^{k+1}-1}^{\infty}a'_ix^{-i}, \] whence $A_{l}=x^{2^k-2}+x^{2^k-4}+\dots +x^2+1$ by \eqref{S-inverse}, which proves \eqref{}. Since $x^{2^k}+1\mid A_l$ for $l\geq 2$, we have by \eqref{} that \[ Q_l=A_lQ_{l-1}+Q_{l-2}\equiv Q_{l-2} \mod x+1, \] which proves \eqref{}. \end{proof} \subsection*{Acknowledgement} The authors are partially supported by the Austrian Science Fund FWF Project 5511-N26 which is part of the Special Research Program {\color{black}``}Quasi-Monte Carlo Methods: Theory and Applications''.

\section{Introduction} Consider the nonlinear state-space model \begin{subequations}\label{} \begin{align} x_{t}\mid x_{t-1} &\sim f_\vartheta(x_{t}\mid x_{t-1}),\\ y_{t}\mid x_t &\sim g_\vartheta(y_{t}\mid x_t), \end{align} \end{subequations} where the hidden states $x_t \in \mathsf{X} \subset \mathbb{R}^{N}$ and observations $y_t \in \mathsf{Y} \subset \mathbb{R}^{n_y}$ evolves over time $t = 0, 1, \dots$. We assume that the probability densities $f_\vartheta$ and $g_\vartheta$ are parameterized by unknown parameters $\vartheta$, which we are interested to learn from data. More specifically we are seeking the maximum likelihood estimate of $\vartheta$ from given observations $\{y_1, \dots, y_T\} \triangleq y_{1:T}$, i.e., to solve \begin{equation} \smash{\widehat{\param}} = \argmax_\vartheta p_\param(y_{1:T}).\label{eq:ml} \end{equation} We shall refer to $p_\param(y_{1:T})$ as the \emph{likelihood function} when considered a function of $\vartheta$. It follows from~\eqref{} and the initial state density $p(x_0)$ that \begin{equation} p_\param(y_{1:T}) = \int p(x_0)\prod_{t=1}^{T}f_\vartheta(x_{t}\mid x_{t-1}) g_\vartheta(y_t\mid x_t)dx_{0:T}.\label{} \end{equation} This can be evaluated analytically only for a few special cases. Approximations, such as the particle filter, are needed for the nonlinear case. For any given $\vartheta$, the particle filter can be run to give a Monte Carlo based estimate $\smash{\widehat{z}_\param}$ of~\eqref{}. It is well-known \citep{} that $\smash{\widehat{z}_\param}$ is unbiased, i.e., $\Exp{\smash{\widehat{z}_\param}} = p_\param(y_{1:T})$, where the expectation is over the stochasticity in the particle filter algorithm itself. Albeit unbiased, $\smash{\widehat{z}_\param}$ is also stochastic (i.e., a different value is obtained every time the particle filter algorithm is run, since it is a Monte Carlo solution), often with a considerable variance and non-Gaussian characteristics, and standard optimization routines can therefore not be used to solve~\eqref{eq:ml}. To this end, schemes using, optimization for noisy function evaluations have been applied (e.g., \citealt{}). We propose in this paper to iteratively run a particle filter with some current parameter estimate $\vartheta_{k-1}$ and scrutinize the particle filter algorithm to construct a \emph{deterministic}\\local approximation of the likelihood function around $\vartheta_{k-1}$. In a sense we hereby manage to circumvent the stochasticity from the Monte Carlo procedure in a consistent way, and can allow for standard optimization routines to search for a new parameter value $\vartheta_{k}$, for which we can run a new particle filter, etc. The idea is outlined as Algorithm~\ref{}. Related ideas have been proposed earlier \citep{}, but they have to the best of the authors knowledge not yet been fully explored in the system identification context. \begin{algorithm}[h!] \For{$k = 1, \dots$}{ Run a particle filter with $\vartheta_{k-1}$ and save all the generated random numbers $\pss{x^{n}_{0:T},a^{n}_{1:T}}$.\; Re-write the likelihood estimator $\smash{\widehat{z}_\param}$ as a deterministic function of the particle system $\pss{x^{n}_{0:T},a^{n}_{1:T}}$, and use conventional optimization to find its maximizing argument $\vartheta_k$.\; } \caption{Identification idea} \label{} \end{algorithm} \section{Background on particle filtering} The particle filter was originally proposed as a Monte Carlo solution to the state filtering problem \citep{}, i.e., to compute $p_\vartheta(x_t\mid y_{1:t})$. It was soon \citep{} realized that the particle filter could also be used to \emph{estimate the likelihood function} for given parameter values, essentially by inserting the particles (Monte Carlo samples) into the integral in \eqref{} and thereby obtain an estimate $\smash{\widehat{z}_\param}$ of $p_\param(y_{1:T})$. Thanks to this likelihood estimate, the particle filter can be used for system identification purposes.\footnote{There are several alternative ways in which the particle filter can be used for system identification, for example approaches based on the EM algorithm or Gibbs sampling.} As mentioned, $\smash{\widehat{z}_\param}$ is unbiased, but it often has a heavy-tailed and asymmetric distribution with a non-negligible variance. Its exact properties depends on the particle filter settings and the model. A textbook introduction to particle filters is given in, e.g., \citet{, }. We summarize a rather general formulation of the particle filter in Algorithm~\ref{alg:pf}, a version of the auxiliary article filter \citep{}. We use $q(x_t\mid x_{t-1},y_t)$ to denote an almost arbitrary\footnote{The support of $q$ has to cover the support of $f_\vartheta$.} proposal distribution. Furthermore, $\nu_t^{_{n}}$ are the resampling weights, $w_t^{_{n}}$ the importance weights, and $\mathcal{C}(\pss{\nu_t^{_{n}}})$ denotes the categorical distribution on the set $\{1, \dots, N\}$ with (possibly unnormalized) weights $\pss{\nu_t^{_{n}}}$, and $N$ is the number of particles. Let us be explicit about the stochastic elements of Algorithm~\ref{alg:pf}: Particles are initially drawn from the initial distribution $p(x_0)$ on line \ref{} (which we assume to be independent of $\vartheta$). Furthermore, the ancestor indices $a_t^{_{n}}$ on line~\ref{} are drawn with respect to the resampling weights $\nu_t^{_{n}}$, and finally the propagation of particles $x_t^{_{n}}$ on line~\ref{}, where the new particles are drawn from the proposal $q(x_t\mid x_{t-1},y_t)$. Whereas $f_\vartheta(x_t\mid x_{t-1})$, $g_\vartheta(y_t \mid x_t)$ and $p(x_0)$ in Algorithm~\ref{alg:pf} are given by the model specification and $w_t^{_{n}}$ follows from the algorithm, the choices for $q(x_t\mid x_{t-1}, y_t)$, $\nu_t^{_{n}}$ and $N$ are left to the user. The number of particles $N$ is usually taken as large as the computational budget permits, and two common choices for weights and proposals are \begin{itemize} \item the \emph{bootstrap particle filter} \citep{} with \[q(x_t\mid x_{t-1}, y_t) = f_\vartheta(x_t\mid y_t)\] and \[\nu_t^{n} = w_t^{n},\] and consequently $w_t^{_{n}} = g_\vartheta(y_t\mid x_t^{_{n}})$. This choice is generic and requires very little from the user, but has inferior performance compared to \item the \emph{fully adapted particle filter} \citep{}, with \[q(x_t\mid x_{t-1}, y_t) = p_\vartheta(x_t\mid x_{t-1}, y_t)\] and \[\nu_t^{n} = p_\vartheta(y_t\mid x_{t-1}).\] This choice is superior to the bootstrap choice in terms of variance of the obtained approximation, but it is only available for a quite limited set of models. The literature on approximations to this choice is therefore rich (e.g., \citealt{}). \end{itemize} We will in this paper exploit the relatively large freedom that is available when it comes to choosing the proposal density and the resampling weights. By making a choice that only depends on the \emph{current} parameter value~$\vartheta_{k-1}$ it is possible to evaluate the likelihood estimator (which we will denote by $\likestn{\vartheta_{k-1}}$) for different values of $\vartheta$, while at the same time making use of the \emph{same} realization of $\pss{x^{_{n}}_{0:T},a^{_{n}}_{1:T}}$. \begin{algorithm}[t] Draw ${x}_0^{n} \sim p({x}_0)$ and set $w_0^{n} = 1, \nu_0^{n} = 1$.\label{; \For{$t = 1$ \KwTo $T$}{ Draw ${a_t^{n}} \sim \mathcal{C}(\{\nu_{t-1}^{j}\}_{j=1}^N).$\label{; Propagate ${x^{{n}}_{t}} \sim q({x_{t}} \mid {x_{t-1}^{a_t^{n}}}, y_t)$.\label{; Set $w_t^{n} \leftarrow \frac{w_{t-1}^{a_t^{n}}/\sum_{j=1}^{N}w^{j}_{t-1}}{\nu_{t-1}^{a_t^{n}}/\sum_{j=1}^N \nu_{t-1}^{a_t^{j}}}~\frac{f_{\theta}(x_t^{n} \mid x_{t-1}^{a_t^{n}})}{ q(x_t^{n} \mid x_{t-1}^{a_t^{n}}, y_t)}g_{\theta}(y_t \mid x_t^{n})$.\; Set $z_t \leftarrow \frac{1}{N} \sum_{n=1}^{N} w_t^{n}$.\; } Compute $\smash{\widehat{z}_\param} = \prod_{t=1}^{T} z_t$.\; \BlankLine \mycomment{All statements with $n$ are for $n = 1, \dots, N$.} \caption{The auxiliary particle filter} \label{alg:pf} \end{algorithm} \section{Related work} The use of the likelihood estimator $\smash{\widehat{z}_\param}$ as an objective function in optimization, in the search for a maximum likelihood estimate $\smash{\widehat{\param}}$, has been subject to several studies: \citet{} differentiate $\smash{\widehat{z}_\param}$ and use it in a stochastic gradient descent scheme, whereas \citet{} use an optimization scheme based on Gaussian processes. \citet{} use a fixed random seed and run the particle filter for different $\vartheta$. For a fixed random seed, $\smash{\widehat{z}_\param}$ is indeed deterministic, however with discontinuous `jumps' due to different resampling decisions being made for different $\vartheta$. To this end, \citeauthor{} proposes an approximative smoothing to obtain a continuous function. The idea used in this paper, i.e., to make the ancestor indices $a^{_{n}}_{1:T}$ and particles $x^{_{n}}_{0:T}$ depending only on the current, or reference, parameter value $\vartheta_{k-1}$ (instead of $\vartheta$) has been analyzed by \citet{}, but has (to the best of the authors knowledge) not yet been applied to the system identification problem. The work by \citeauthor{} concerns central limit theorems for the likelihood estimator $\likestn{\vartheta_{k-1}}$. The application of the results from \citet{} to this work is subject for further work. The work presented in this paper shares similarities with the authors recent work \citet{}, which however is concerned with the different topic of Bayesian identification for state-space models with highly informative observations. Other approaches to maximum likelihood estimation in nonlinear state-space models include the combination of the popular Expectation Maximization (EM) algorithm and particle filters \citep{}. \section{The solution} The key idea in this paper is to choose $q(x_t\mid x_{t-1}, y_t)$ and $\nu_t^{_{n}}$ such that they are independent of $\vartheta$. By such a choice, we note that all the random elements in Algorithm~\ref{alg:pf}, $\pss{x^{_{n}}_{0:T},a^{_{n}}_{1:T}}$, become independent of $\vartheta$. If we then condition on a certain realization of $\pss{x^{_{n}}_{0:T},a^{_{n}}_{1:T}}$, the estimate $\smash{\widehat{z}_\param}$ becomes a deterministic function in $\vartheta$, and any standard optimization routine can subsequently be applied to solve~\eqref{} and find $\smash{\widehat{\param}}$. The strength of the particle filter, however, lies in the sequential build-up of the samples on the high-dimensional space $\mathsf{X}^{{T+1}}$, where the resampling operation provides important `feedback' on which parts of the state space to be explored further. With an arbitrary choice of $\vartheta$-independent resampling weights $\nu_t^{_{n}}$, this feature will be lost, and we may expect an extremely high variance in the obtained estimate. In fact, a particle filter with $\vartheta$-independent resampling weights $\nu_t^{_{n}}$ van be understood as importance sampling on the space $\mathsf{X}^{{T+1}}$, and we can in general not expect such an approach to be successful. In order to obtain a deterministic function in $\vartheta$, but avoid the bad consequences of a $\vartheta$-independent resampling, we propose to let the resampling weights $\nu_t^{_{n}}$ and proposal $q(x_t\mid x_{t-1}, y_t)$ depend on some current parameter estimate $\theta_{k-1}$, as, e.g., \begin{subequations}\label{} \begin{equation} q(x_t\mid x_{t-1}, y_t) = f_{\vartheta_{k-1}}(x_t\mid y_t) \end{equation} and \begin{equation} \nu_t^{n} = g_{\vartheta_{k-1}}(y_t\mid x_t^{n}), \end{equation} \end{subequations} i.e., the bootstrap choice for $\vartheta_{k-1}$ (instead of $\vartheta$). If then $\vartheta$ is somewhat close to $\vartheta_{k-1}$, we can expect the variance of the corresponding estimate of the likelihood function, which we denote by $\likestn{\vartheta_{k-1}}$, not to be forbiddingly large. However, if the current $\vartheta_{k-1}$ is far from $\smash{\widehat{\param}}$, we cannot expect $\likestn{\vartheta_{k-1}}$ to be a particularly good estimator at the value $\smash{\widehat{\param}}$, and in particular, not expect the maximum of $\likestn{\vartheta_{k-1}}$ to lie at $\smash{\widehat{\param}}$. For this reason, we have to iterate the parameter values over $k$ until we arrive in the vicinity of $\smash{\widehat{\param}}$. By inserting~\eqref{} into Algorithm~\ref{alg:pf}, combined with an outer optimization loop as discussed, and re-arranging, we arrive at Algorithm~\ref{}. For numerical reasons, we work with the logarithm of the likelihood function. The conceptual idea is illustrated in Figure~\ref{}. \begin{algorithm}[t] \SetKwFunction{prar}{particle\_filter} \SetKwFunction{likf}{log\_likelihood} Set $\vartheta_{0}$\; \For{$k = 1, \dots$}{ Call $\pss{x^{n}_{0:T},a^{n}_{1:T}}\leftarrow$ \prar{$\vartheta_{k-1}$}\; Solve \mbox{$\vartheta_k \leftarrow \argmax_\vartheta$ \likf{$\vartheta,\vartheta_{k-1},\pss{x^{n}_{0:T},a^{n}_{1:T}}$}} using an off-the-shelf optimization routine.\label{; } \BlankLine \setcounter{AlgoLine}{0} \SetKwProg{myprocp}{Function}{}{} \myprocp{\prar{$\vartheta_{k-1}$}}{ Draw ${x}_0^{n} \sim p({x}_0)$ and set $w_0^{n} = 1$.\; \For{$t = 1$ \KwTo $T$}{ Draw ${a_t^{n}} \sim \mathcal{C}(\{w_{t-1}^{j}\}_{j=1}^N)$.\; Propagate ${x^{i}_{t}} \sim f_{\vartheta_{k-1}}({x_{t}} \mid {x_{t-1}^{a_t^{n}}}, y_t)$.\; Set $w_t^{n} \leftarrow g_{\theta_{k-1}}(y_t \mid x_t^{n})$.\; } \KwRet $\pss{x^{n}_{0:T},a^{n}_{1:T}}$ } \BlankLine \setcounter{AlgoLine}{0} \SetKwProg{myprocl}{Function}{}{} \myprocl{\likf{$\vartheta,\vartheta_{k-1},\pss{x^{n}_{0:T},a^{n}_{1:T}}$}}{ \For{$t = 1$ \KwTo $T$}{ Set $w_t^i \leftarrow \frac{w_{t-1}^{a_{t}^{n}}/\sum_j w_{t-1}^{a_{t}^{j}}}{(\star)}~\frac{f_{\vartheta}(x_{t}|{x}_{t-1}^{a_{t}^{n}})}{f_{\vartheta_{k-1}}(x_{t}|{x}_{t-1}^{a_{t}^{n}})} ~g_{\vartheta}({y}_t|{x}_t^{n})$.\; Set $z_t \leftarrow \frac{1}{N}\sum_{n=1}^{N}w_t^{n}$.\; } \KwRet $\log \likestn{\vartheta_{k-1}} \leftarrow \sum_{t=1}^{T}\log z_t$.\; } \mycomment{$(\star) = g_{\vartheta_{k-1}}(x_{t-1}^{a_{t}^{n}}\mid y_{t-1})/\sum_j g_{\vartheta_{k-1}}(x_{t-1}^{a_{t}^{j}}\mid y_{t-1})$} \caption{Proposed method} \label{} \end{algorithm} \begin{figure}[h] \centering\footnotesize \includegraphics{arxiv-figure0} \caption{ The gray dots are log likelihood estimates obtained by running individual bootstrap particle filters with $N=100$ particles and parameter value $\theta_{k-1}$. Conditioned on the particle system underlying each particle filter (gray dots), the likelihood function is approximated also for other $\vartheta$ values (red lines), which we can expect to be practically useful in the vicinity of $\vartheta_{k-1}$. The idea of Algorithm~\ref{} is as follows: Start with some $\theta_{k-1}$ and sample a corresponding gray dot (\texttt{particle\_filter}), and then apply a standard optimization scheme to find the maximum of the corresponding red line (\texttt{log\_likelihood}). We save the result as $\vartheta_k$, and start over again with a new particle filter for~$\vartheta_k$, etc.} \label{} \end{figure} \subsection{Solving the maximization problem}\label{} On line~\ref{} in Algorithm~\ref{}, we have formulated a step where $\argmax_\vartheta \likestn{\vartheta_{k-1}}$ is to be solved. Importantly, this is now a completely deterministic problem and it can be handled by any standard numerical optimization tool. We will in the experiments demonstrate this by simply applying the general-purpose optimization tool \texttt{fminunc} in Matlab\footnote{Similar functions in other languages are \texttt{fminunc} (Octave), \texttt{scipy.optimize} (Python), \texttt{optim} (R) and \texttt{optimize} (Julia).}. The particular structure of $\likestn{\vartheta_{k-1}}$ (defined implicitly in the function \texttt{likelihood} in Algorithm~\ref{}) could, however, possibly be utilized by a more tailored optimization scheme. Its structure can be written as \begin{align}\small \likestn{\vartheta_{k-1}} = \frac{1}{N}\prod_{t=1}^T\sum_{n=1}^{N}c_t^{n}~\omega_t^{n}(\vartheta)~f_{\vartheta}(x_{t}|{x}_{t-1}^{a_{t}^{n}})g_{\vartheta}({y}_t|{x}_t^{n}), \end{align} where $c_t^{n}$ is a constant that is independent of $\vartheta$, $\omega_t^{_{n}}(\vartheta)$ depends on $\vartheta$ but always fulfil $\sum_{n=1}^{N} \omega_t^{_{n}}(\vartheta) = 1$, and $f_{\vartheta}$ and $g_{\vartheta}$ depends on the model. Whether this function exhibits any particular properties that could be exploited in optimization is subject to further research. \section{Analysis} We will now present a brief analysis of the proposed Algorithm~\ref{}. First, we conclude in Section~\ref{} that the proposed scheme has desirable asymptotic properties as $N\to\infty$. Second, we make an attempt in Section~\ref{} to understand the behavior also for finite $N$, and third in Section~\ref{} we discuss an alternative solution that would place the proposed method in the framework of stochastic gradient descent methods. \subsection{Convergence as $N\to\infty$}\label{} Asymptotically as $N\to \infty$ in the particle filter, the proposed method becomes exact and converges (in principle) in one iteration. This can be realized as follows: The log-likelihood is estimated by \begin{align} \log p_\param(y_{1:T}) \approx \sum_{t=1}^T \log\left( \sum_{i=1}^{N} w_t^{n} p(y_t\mid x_t^{n})\right).\label{} \end{align} Assuming that the proposal distribution used to generate the particles $\{x_t^{n}\}_{t=1}^T$ is everywhere non-zero, this log-likelihood approximation is consistent under weak conditions and converges point-wise in $\theta$ as $N\rightarrow\infty$ \citep{}. Thus, as long as the global solution to $\argmax_\vartheta$ can be found, it is indeed the likelihood that has been maximized and $\smash{\widehat{\param}}$ found, which happens in a single iteration $k = 1$. \subsection{Convergence as $k\to\infty$ and finite $N$}\label{} It is indeed reassuring that our proposed scheme is consistent as the number of particles $N\to\infty$, as discussed in the previous section. For practical purposes, however, the behavior for a finite $N$ (which always is the case in an implementation) is probably more interesting. We start by noting that for $N<\infty$, it holds that $\Exp{\likestn{\vartheta_{k-1}}} = p_\param(y_{1:T})$ \citep{}. Note, however, that this does \emph{not} imply \begin{align} \Exp{\argmax_\vartheta \likestn{\vartheta_{k-1}}} = \argmax_\vartheta p_\vartheta(y), \end{align} so we do not have a theoretical justification to simply average the obtained sequence $\{\vartheta_k\}$ to obtain $\smash{\widehat{\param}}$ To obtain some guidance on how to extract a final estimate from the obtained sequence $\{\vartheta_k\}$, we can make the simplifying assumption that the error $\log\likestn{\vartheta_{k-1}}-\log p_\param(y_{1:T})$, viewed as a stochastic process with index variable $\theta$, is stationary. In such a case, we can (under some technical assumptions) expect that $\smash{\widehat{\param}}$ is equal to the maximum mode of the distribution for $\vartheta_k = \argmax_\vartheta \log \likestn{\vartheta_{k-1}}$. A proof sketch for this claim is found in Appendix~\ref{}. This suggests that we should look at the maximum mode in the histogram for $\{\vartheta_k\}$ to find a good estimate of $\smash{\widehat{\param}}$ when we are using the method in practice (i.e., with a finite $N$). This will be illustrated in Example~2 and Figure~\ref{}. \subsection{Stability} Stability, i.e., that the sequence $\vartheta_0, \vartheta_1, \dots$ does not diverge, is another important property. We have not experienced any such issues with the proposed scheme. The key to a stable algorithm is that the solution to the maximization problem on line~\ref{} in Algorithm~\ref{} is not too far from $\theta_{k-1}$. To motivate that this is likely to be the case, we note that while $\likestn{\theta}$ is unbiased for all $\vartheta$, its variance $\sigma^2(\vartheta)$ tends to increase as the distance between $\theta$ and $\theta_{k-1}$ increases. It is also known that $\log \likestn{\theta}$ (rather than $\likestn{\theta}$) has approximately a Gaussian distribution, which implies that $\log \likestn{\theta}$ has a bias in the order of $-\sigma^2(\vartheta)$. This motivates that for $\vartheta$ far from $\vartheta_{k-1}$, the value of $\log \likestn{\theta}$ is likely to be small, and hence cause the algorithm not to deviate to much from the current iterate. \subsection{Stochastic gradient descent}\label{} An alternative approach would be not to solve the $\argmax_\vartheta$ problem, but only use $\likestn{\vartheta_{k-1}}$ to estimate a gradient around $\vartheta_{k-1}$ and take an (inevitable stochastic) gradient step. Indeed, this has already been proposed by \citet{}. Designing step lengths based on stochastic approximation ideas \citep{} yields the well-studied stochastic gradient descent method. Our practical experience, however, is that (stochastic) gradient steps have inferior performance compared to the proposed $\argmax_\vartheta$ scheme for our problem, including slower convergence and severe stability issues. \section{Numerical experiments} We will in this section apply our proposed method to two simulated examples, in order to illustrate and evaluate it. First a common example form the literature will be considered, and comparisons to alternative methods made. Thereafter a more difficult example will be studied. The code for the examples is available via the first author's homepage. \subsection{Example 1}\label{ex:1} In this first example, we consider $T=100$ measurements generated by the model \begin{subequations} \begin{align} x_{t+1} &= 0.5x_t + b\tfrac{x_t}{1+x_t^2} + 8\cos(1.2t) + qw_t,\\ y_t &= 0.05x_t^2+e_t, \end{align} \end{subequations} where $w_t\sim\mathcal{N}(0,1)$, $e_t\sim\mathcal{N}(0,1)$, and $\vartheta=\{b,q\}$. The true values of $\vartheta$ are $\{25,\sqrt{0.1}\}$, and this example (with $q = \sqrt{0.1}$ and $\theta = b$) was also used to generate Figure~\ref{}. The proposed Algorithm~\ref{} is implemented with $N=100$, and employing the generic optimization routine \texttt{fminunc} in Matlab to solve the optimization problem on line~\ref{} in Algorithm~\ref{}. The initial~$\vartheta_0$ is chosen randomly on the intervals $[10,40]$ and $(0,4]$, respectively, and the entire example is repeated 100 times. Each example took approximately 6 seconds on a standard laptop, and the results are found in Figure~\ref{}. We compare with two alternative methods for maximum likelihood estimation in nonlinear state-space models: The results for particle stochastic approximation EM (PSAEM, \citealt{}) applied to the very same problem are reported in Figure~\ref{}. The results for the same problem with a stochastic optimization approach using the particle filter to estimate the likelihood and the Gaussian process (GP) to model the likelihood surface and its derivatives \citep{} are found in Figure~\ref{}. With the number of iterations chosen as in the figures, the computational load are of the same order of magnitude for all three methods. \begin{figure}[t] \centering \begin{subfigure}{.49\linewidth} \includegraphics{arxiv-figure1} \caption{The proposed method.} \label{} \end{subfigure} \begin{subfigure}{.49\linewidth} \centering\footnotesize \includegraphics{arxiv-figure2} \vspace*{-1em} \caption{PSAEM \citep{}.} \label{} \end{subfigure} ~ ~ \begin{subfigure}{.49\linewidth} \centering\footnotesize \includegraphics{arxiv-figure3} \vspace*{-1em} \caption{GP-based optimization \citep{}.} \label{} \end{subfigure} \caption{ Example 1. The results from each method is shown by red lines, and the true parameter values are shown in black. The practical convergence properties of the proposed method appears to be promising.} \end{figure} From this, we conclude that our proposed method tend to converge faster than the alternatives (counting the number of iterations needed), but that each iteration is computationally more involved. The computational load of our algorithm is partly due to the use of a generic optimization routine (\texttt{fminunc} in Matlab), which makes no use of the particular structure of the objective function (\texttt{likelihood} in Algorithm~\ref{}), as discussed in Section~\ref{}. \subsection{Example 2} Now, consider the following state-space model \begin{subequations}\label{} \begin{alignat}{3} x_{t+1} &= \frac{x}{a + x^2} + b u + w_t,& ~~ w_t&\sim \mathcal{N}(0,1),\label{ y_t &= x + e_t,& ~~ e_t&\sim \mathcal{N}(0,1). \end{alignat} \end{subequations} with $T = 1\,000$ and $\vartheta = \{a, b\}$. One data set is generated with $\vartheta = \{0.5, -2\}$, and our method is applied, with different initializations, 100 times to find $\smash{\widehat{\param}}$. This problem is significantly harder than Example~1 due to the location of $a$ in the denominator (and not the numerator) in~\eqref{}. As an illustration, independent particles filter were run to estimate the log-likelihood for different values of $a$ in Figure~\ref{}, from which we conclude that the likelihood estimate is rather noisy. This can be compared to Example~1 and the gray dots in Figure~\ref{}, where the likelihood estimation is clearly less challenging. Again, we use the proposed Algorithm~\ref{} with \texttt{fminunc} in Matlab to solve the optimization problem on line~\ref{}. The results are shown in Figure~\ref{} and \ref{}. Despite the challenging likelihood estimation, our proposed method manages to eventually converge towards meaningful values, and following the guidance discussed in Section~\ref{}, we take the final estimates as the maximum of the histograms in Figure~\ref{}, ${0.59, -1.995}$, which corresponds well to the true parameters. \begin{figure}[t] \centering \begin{subfigure}{.6\linewidth} \centering\footnotesize \includegraphics{arxiv-figure4} \caption{Log-likelihood estimates (vertical axis) for the model~\eqref{} in Example~2 for different $a$ (horizontal axis, true $a=0.5$) and $b = -2$, obtained with independent particle filters with $N = 100$. As can be seen, the variance in $\smash{\widehat{z}_\param}$ is rather high in this example, which is to be compared with the gray dots in Figure~\ref{}, the corresponding plot for Example~1. We thus expect maximum likelihood estimation to be significantly more challenging in this example.} \label{} \end{subfigure} ~ ~ \begin{subfigure}{.49\linewidth} \centering\footnotesize \includegraphics{arxiv-figure5} \caption{ The results for our proposed method on Example~2. Our proposed method manages, despite the poor quality likelihood estimates (Figure~\ref{}), to eventually converge towards sensible parameter values for a wide range of initializations. These traces are shown as histograms in the figure below.} \label{} \end{subfigure}\hfill \begin{subfigure}{.49\linewidth} \centering\footnotesize \includegraphics{arxiv-figure6} \caption{ The traces from Figure~\ref{} above as a histograms, after discarding the transient phase up to $k=50$. Using the principle suggested in Section~\ref{}, the final estimate~$\smash{\widehat{\param}}$ should be taken as the maximum of the histograms, i.e., ${0.59, -1.995}$ , which corresponds well to the true parameter values $\{0.5,-2\}$.} \label{} \end{subfigure} \caption{ Results for Example 2.}\label{} \end{figure} \section{Conclusions} We have proposed a novel method, Algorithm~\ref{}, to find maximum likelihood estimates of unknown parameters in nonlinear state-space models. The method builds on the particle filter, and allows for parameter inference in any model where also a particle filter can be applied. In fact, the method can be understood as a particular choice of $\vartheta$-independent proposal and resampling weights in the auxiliary particle filter. One key reason for the promising results is probably that we propose to solve an optimization problem at each iteration, instead of only taking a gradient step or similar: heuristically this seems to lead to fast convergence, and has not caused any problems with instability. The theoretical analysis, however, becomes more involved. We have presented an attempt to such an analysis in Section~\ref{}, but all questions have not been answered. As mentioned, the work by \citet{} could potentially be useful in a more thorough analysis of the proposed method. A tailored choice of optimization routine would be interesting for further work. Furthermore, the applicability of the proposed method for the particle marginal Metropolis-Hastings sampler \citep{} would be another interesting question.

\section{Introduction} Electron-positron pair production in strong electric fields, the Sauter-Schwinger effect, is a long-standing theoretical prediction \cite{} which still awaits experimental verification. Multiphoton pair production, on the other hand, has already been observed in a laboratory \cite{}. The dynamically assisted Sauter-Schwinger effect \cite{PhysRevLett.101.130404,Linder} exploits the idea that a combination of a low with a high frequency (``multi-photon'') laser pulse will lead to pair production rates which are significantly larger than the sum of the rates for the two separate pulses. Together with the arising capabilities of high-intensity laser technology, see, e.g., refs.~\cite{}, such a combination of laser pulses will make experimental tests in this regime of non-linear QED in the near future possible. Besides on the technological progress future experimental tests and their interpretations will depend substantially on more reliable calculations which include the inhomogeneities of the electromagnetic fields as they typically occur in the focus of crossing laser beams. Computations for inhomogeneous and time-dependent fields (i.e., beyond the previously well-studied case of time-dependent but homogeneous electric fields) have been started a few years ago and have seen steady progress \cite{}. These studies are, however, still far from providing a satisfactory understanding of the effects which originate from the finite spatial extension of the considered laser pulses. As the particle creation rate depends on the laser intensity, see, e.g., ref.~\cite{} for a discussion in the context of the planning of the XFEL, one would na\"ively expect that better focusing and therefore higher local field intensities lead always to an increase of the particle yield. However, the dynamics in high-intensity laser fields is much more complex and results not only in strong but also non-monotonic dependencies of the yield on the characteristics of the laser field \cite{Linder,}. Other quantities like, e.g., distribution functions and particle spectra, display a rich structure with sometimes surprisingly large sensitivities to small changes of the field parameters. Clearly such a situation calls for the search of concepts which are able to allow an, at least qualitative, understanding. One possibility to quantify the effects associated with a spatially inhomogeneous field is based on the notion of an effective mass which an electron acquires in a background electromagnetic field \cite{}. This parameter, as every effective mass, reflects the ``integrated`` collective interactions of a particle with its surroundings. It thus provides the possibility of a drastic simplification but nevertheless allows the coarse-grained description of highly intricate effects. Although this might be an oversimplification with respect to details of the resulting spectra, the concept of an effective mass works astonishingly well, a fact which can be attributed to the unique conditions in high-intensity laser experiments \cite{}. Correspondingly, the idea of an effective mass has been applied recently to multiphoton pair production \cite{}. In these studies, however, the employed electric fields were homogeneous thereby greatly simplifying the process under investigation. In the following we will discuss particle creation in inhomogeneous fields and introduce to this end a more general concept for the effective mass and relate it via a semi-classical analysis to ponderomotive forces. (NB: A ponderomotive force is a nonlinear force that a ``classical'' charged particle experiences in an inhomogeneous oscillating electromagnetic field, {\it cf.} ref.~\cite{Bauer} and references therein.) We will concentrate in a first step hereby on multiphoton pair production in the nonperturbative threshold regime. As is evident from the discussion above, in the parameter regions of interest the effects of a laser pulse's finite spatial extension cannot be neglected. (For a more detailed discussion of this issue, see, e.g., ref.\ \cite{Diss}.) Hereby one should distinguish two aspects: First, how does the finiteness of the fields' extension influence the pair production process, and second, how does a spatially inhomogeneous field alter the subsequent electron/positron dynamics. To keep the calculational complexity in a manageable range we will assume cylinder symmetry of the electric field. The spatial dependence can be inferred, e.g., from a gauge potential which is directed along a direction and is inhomogeneous only w.r.t.\ the same direction. This has the advantage that the corresponding magnetic field vanishes. However, one then clearly has no propagating waves. Nevertheless, such a configuration provides some, although simplistic, model of the focus of two counter-propagating lasers. This paper is organized as follows: In Sect.\ \ref{Multiphoton} several aspects of multiphoton pair production in the nonperturbative threshold regime underlying our work will be briefly summarized: In subsection \ref{DHW} a short description of the Dirac-Heisenberg-Wigner (DHW) formalism for the case of an electric field with cylinder symmetry is provided; subsection \ref{NumTreat} deals with the numerical treatment via Fourier transform; and in subsection \ref{Fields} the form of the electric field employed later on is given. In sect.\ \ref{Interpretational} two interpretational concepts are introduced and discussed: the effective mass and a trajectory-based semi-classical analysis. In sect.\ \ref{Results} numerical results, ordered according to the longitudinal and transverse momentum distributions, are presented and discussed. Sect.\ \ref{Conclusions} contains the conclusions of the presented study. Furthermore, we use $\hbar = c = 1$ throughout this paper. \section{Theoretical description of multiphoton pair production} \label{Multiphoton} \subsection{DHW Formalism} \label{DHW} The present study is based on the DHW formalism, a relativistic phase-space approach, which has been developed for the case of pair production in refs.~\cite{}. Hereby the electron is treated as a quantum field but the laser pulse is approximated by its mean-field. Given the magnitudes of the electric field needed in pair production this is a justified approximation. A convenient starting point is provided by the gauge-invariant density operator of the system, \begin{equation} \hat {\mathcal C}_{\alpha \beta} \left( r , s \right) = \mathcal U \left(A,r,s \right) \ \left[ \bar \psi_\beta \left( r - s/2 \right), \psi_\alpha \left( r + s/2 \right) \right], \end{equation} in terms of the electron's spinor-valued Dirac field $\psi_\alpha (x)$. Hereby $r$ denotes the center-of-mass and $s$ the relative coordinate. The Wilson line factor \begin{equation} \mathcal U \left(A,r,s \right) = \exp \left( \mathrm{i} \ e \ s \int_{-1/2}^{1/2} d \xi \ A \left(r+ \xi s \right) \right) \end{equation} is introduced to render the density operator gauge-invariant. It depends on the elementary charge $e$ and the background gauge field $A$. Treating the background field in Hartree approximation, i.e., \begin{equation} F^{\mu \nu} \left( {x} \right) \approx \langle \hat F^{\mu \nu} \left( {x} \right) \rangle , \end{equation} no path ordering is needed, and in a given Lorentz frame and gauge the background gauge field $A \left(\mathbf{x}, t \right)$ is a fixed $c$-number valued function. The covariant Wigner operator, \begin{equation} \hat{\mathcal W}_{\alpha \beta} \left( r , p \right) = \frac{1}{2} \int d^4 s \ \mathrm{e}^{\mathrm{i} ps} \ \hat{\mathcal C}_{\alpha \beta} \left( r , s \right), \end{equation} reflects thus the quantum fluctuations of the electron but not the one of the laser field. The main implication of the Hartree approximation for the electromagnetic field becomes evident when taking the vacuum expectation value of the covariant Wigner operator to obtain the covariant Wigner function \begin{equation} \mathbbm{W} \left( r,p \right) = \langle \Phi \vert \hat{\mathcal W} \left( r,p \right) \vert \Phi \rangle. \end{equation} In its equation of motion the electromagnetic field can be factored out, e.g., \begin{equation} \langle \Phi \vert F_{\mu \nu} \ \hat{\mathcal{C}} \vert \Phi \rangle = F_{\mu \nu} \langle \Phi \vert \hat{\mathcal{C}} \vert \Phi \rangle \end{equation} which allows to resolve the otherwise infinite BBGKY hierarchy of correlation functions. Being a Dirac-matrix valued quantity the Wigner function is best decomposed into 16 covariant Wigner coefficients \begin{equation} \mathbbm{W} = \frac{1}{4} \left( \mathbbm{1} \mathbbm{S} + \textrm{i} \gamma_5 \mathbbm{P} + \gamma^{\mu} \mathbbm{V}_{\mu} + \gamma^{\mu} \gamma_5 \mathbbm{A}_{\mu} + \sigma^{\mu \nu} \mathbbm{T}_{\mu \nu} \right) \label{decomp} \end{equation} where the corresponding transformation properties are made evident by the notation. Working in a definite frame it proves advantageous to project on equal times which yields the equal-time Wigner function \begin{align} \mathbbm{w} \left( \mathbf{x}, \mathbf{p}, t \right) = \int \frac{d p_0}{2 \pi} \ \mathbbm{W} \left( r,p \right) \end{align} and by an analogous decomposition to eq. (\ref{decomp}) the corresponding equal-time Wigner coefficients $\mathbbm{s},\mathbbm{p}, \mathbbm{v}_{0,x,y,z}, \ldots$. Exploiting cylindrical symmetry, which will be assumed throughout the following, and keeping the electric field homogeneous in transversal direction results in a reduced system of differential equations\footnote{The coefficients given in the following are obtained by linear combinations of the equal-time Wigner coefficients. The quantity $\overline{\mathbbm{v}}$, for example, is defined as a linear superposition of $\mathbbm{v}_z$ and tensor components $\mathbbm{t}_{xz}$ and $\mathbbm{t}_{yz}$. The details of the derivation can be found in ref. \cite{Diss}.}: \begin{alignat}{5} & D_t \overline{\mathbbm{s}} && && -2 p_z \overline{\mathbbm{p}} &&+ 2 p_{\rho} \overline{\mathbbm{v}} &&= 0, \label{eq_Cy1} \\ & D_t \overline{\mathbbm{v}} &&+\partial_{x} \overline{\mathbbm{v}}_0 && &&- 2 p_{\rho} \overline{\mathbbm{s}} &&= -2 \overline{\mathbbm{p}}, \\ & D_t \overline{\mathbbm{p}} && && +2 p_z \overline{\mathbbm{s}} && &&= 2 \overline{\mathbbm{v}}, \\ & D_t \overline{\mathbbm{v}}_0 &&+ \partial_{z} \overline{\mathbbm{v}} && && &&= 0, \label{eq_Cy4} \end{alignat} where the pseudo-differential operator $D_t$ reads \begin{align} D_t = \partial_{t} + e \int \ d\xi \ E \left(z+\textrm{i} \xi \partial_{p_z},t \right) \ \partial_{p_z}. \label{Pseudo} \end{align} We refrained from putting a spatial index on the electric field $E$ as by construction it is oriented in the $z$-direction. The advantage of a representation with Wigner coefficients lies in the fact, that it allows to identify $\overline{\mathbbm s}$ as mass, $\overline{\mathbbm v}_0$ as charge and $\overline{\mathbbm v}$ as current density \cite{}. In order to perform calculations within the DHW formalism we employ vacuum initial conditions: \begin{align} \overline{\mathbbm{s}}_{in} = -\frac{2}{\omega},\ \overline{\mathbbm{v}}_{in} = -\frac{2 p_z}{\omega} ,\ \overline{\mathbbm{p}}_{in} = -\frac{2 p_{\rho} }{\omega}, \end{align} where $\omega = \omega \left(p_z,p_{\rho} \right) = \sqrt{m^2 + p_z^2 + p_{\rho}^2}$ is the one-particle energy. It is convenient to subtract these initial conditions, and therefore we define \begin{equation} \overline{\mathbbm{w}}^v := \overline{\mathbbm{w}} - \overline{\mathbbm{w}}_{in}, \end{equation} where $\overline{\mathbbm{w}}$ is a placeholder for any Wigner component. Additionally, we define the particle number density per unit volume in momentum space \begin{align} N \left(p_z, p_{\rho} \right) = & \ \int dz \ \frac{ \overline{\mathbbm{s}}^v + p_z \overline{\mathbbm{v}}^v + p_{\rho} \overline{\mathbbm{p}}^v } {\omega \left(p_z,p_{\rho} \right)} \, .\label{NN} \end{align} Consequently, the total particle yield per unit volume is defined via \begin{align} N = & \int dp_z \ \int dp_{\rho} \ N \left(p_z, p_{\rho} \right). \label{NNfull} \end{align} \subsection{Fourier transform and numerical treatment} \label{NumTreat} Eqs.\ \eqref{eq_Cy1}-\eqref{eq_Cy4} can be solved numerically without any further approximations or truncations \cite{}. The challenging part is the non-locality of the pseudo-differential operator \eqref{Pseudo} which can be nevertheless treated if the electric field $E \left( z, t \right)$ can be Taylor expanded and integrated sufficiently efficient. The differential operator \eqref{Pseudo} splits naturally in two parts: \begin{equation} D_t = \partial_{t} + e \int d\xi \ E \left( z+\textrm{i} \xi \partial_{p_z},t \right) \partial_{p_z} =: \partial_{t} + \Delta. \label{Pseudo2} \end{equation} To apply the operator $\Delta$ on the (subtracted) Wigner components we Fourier transform and inverse Fourier transform $\Delta \overline{\mathbbm{w}}^v$ w.r.t.\ to the variable $p_z$, i.e., we employ \begin{equation} f \left( p_z \right) = {\cal F}_{p_z}^{-1} \left[ {\cal F}_{p_z}[ f \left( p_z \right)] \right] = {\cal F}_{p_z}^{-1} \left[ \tilde f \left( {\rm k}_{p_z} \right)\right] , \end{equation} Taylor expand the electric field, use \begin{equation} {\cal F}_{p_z}[\frac{d^n}{dp_z^n}f(p_z)]= \left( \textrm{i} {\rm k}_{p_z} \right)^n \tilde f \left( {\rm k}_{p_z} \right) , \end{equation} and resum then to obtain the generic form (for more details see ref.~\cite{Diss}): \begin{align} &\Delta \ \overline{\mathbbm{w}}^v \left( z, p_z, t \right) \notag \\ &= \mathcal {F}_{p_z}^{-1} [ \textrm{ie} {\rm k}_{p_z} \int d\xi \ E \left( z-\xi {\rm k}_{p_z},t \right) \ \tilde{\overline{\mathbbm{w}}} \left( z, {\rm k}_{p_z}, t \right) ]. \end{align} Due to the fact, that a Fourier transform takes into account all points in a domain, the introduction of global basis functions turned out to be favorable compared to the finite-difference-method used in ref.~\cite{}. Hence, we have equidistantly discretized spatial and momentum directions, respectively, turning the system of partial differential equations \eqref{eq_Cy1}-\eqref{eq_Cy4} into a high-dimensional ($N_z \times N_{p_\rho} \times N_{p_z}$) system of ordinary differential equations with $t$ as the only continuous parameter. Furthermore, we choose periodic boundary conditions in $z$ and $p_z$, \begin{equation} \overline{\mathbbm{w}}^v \left( z_0,p_z,t \right) = \overline{\mathbbm{w}}^v \left( z_{N_z},p_z,t \right) \end{equation} and \begin{equation} \overline{\mathbbm{w}}^v \left( z, p_{z,0},t \right) = \overline{\mathbbm{w}}^v \left(z,p_{z,N_{p_z}},t \right), \end{equation} respectively. Additionally, we set \begin{align} \overline{\mathbbm{w}}^v \left( z_{k_i}, p_{z,{k_j}},t \right) = 0 \quad \textrm{if} \quad k_i = 0 \ \textrm{or} \ k_j = 0. \end{align} These choices do not influence the numerical results as long as the chosen discretized domain is sufficiently large, and the number of grid points in every direction is sufficiently high. After discretizing eqs.\ \eqref{eq_Cy1}-\eqref{eq_Cy4} in this way they can be solved using pseudospectral methods \cite{Boyd}. The time integration is done using a Dormand-Prince Runge-Kutta integrator of order 8(5,3) \cite{NR}. \subsection{Model for the fields} \label{Fields} As stated above, if the spatial dependence of the electric field is inferred from a gauge potential which is directed along a direction and is inhomogeneous only w.r.t.\ the same direction the corresponding magnetic field vanishes. In addition, two time scales are needed to tune a pulse of finite duration to the multiphoton regime. These requirements are fulfilled by the ansatz \begin{align} E \left( z,t \right) &= \varepsilon E_0 \ {\cal E} \left( z \right) \ F \left( t \right) \notag \\ &= \varepsilon E_0 \ {\cal E} \left( z \right) \ \cos^4 \left( \frac{t}{\tau} \right) \ \cos \left( \omega t \right), \label{EE0} \end{align} for $t \in \left[ -\pi \tau/2,~ \pi \tau/2 \right]$, and $E=0$ otherwise. Hereby the critical field strength $E_0=m^2/e$ has been factorized out for convenience. Non-perturbative multiphoton pair production is probed if one choses the product $\omega \tau > 1$ and a Keldysh parameter of $\gamma = \omega / m \varepsilon > 1$ \cite{Diss}. As we want to investigate how focusing influences the particle distribution rate we chose a well-localized electric field with a Gaussian shape of width $\lambda$: \begin{equation} {\cal E} \left( z \right) = \exp \left( -\frac{z^2}{2 \lambda^2} \right). \label{EE1} \end{equation} \section{Semi-classical analysis} \label{Interpretational} In Sect.\ \ref{Results} it will become obvious that the dependencies of observables on the field parameters are of an astonishing complexity. In order to obtain an interpretation and such an understanding of the results we will analyze them referring to the concepts developed in this section. Introducing a generalized effective mass concept and relating it to arising ponderomotive forces is a central aspect of this paper. We discuss the improvements compared to previous definitions of the effective mass in the context of pair production in section \ref{Ch_Con1}. Moreover, based upon refs. \cite{Bauer,Pond}, we draw a connection between a spatially dependent effective mass and ponderomotive forces. However, an interpretation of the final particle distribution on the basis of an effective mass becomes more involved for spatially inhomogeneous fields due to the position-dependence of the gradient. Hence, we rely on a semi-classical trajectory-based model, which allows us to determine the overall scheme by simple means. \subsection{Effective mass and ponderomotive forces} \label{Ch_Con1} Various studies have used effective masses to simplify intermediate calculations \cite{}. It was suggested only recently to employ the concept of an effective mass to determine directly observable quantities regarding particle creation, see ref. \cite{}. In case of a monochromatic plane wave, the effective mass takes the form \cite{}: \begin{equation} M_{\ast} = m \sqrt{1 + \xi^2} ,~ \textrm{with} ~ \xi = \frac{e}{m} \sqrt{- \langle A^{\mu} A_{\mu} \rangle }. \end{equation} More general definitions have been proposed in ref. \cite{Kibble}. We, however, adopt essentially but modify slightly the definition above and parameterize the effective mass as follows \begin{align} m_{\ast} \left( \mathbf{x} \right) &= m \sqrt{1+\tilde{\xi} \left( \mathbf{x} \right)^2}, \\ &\tilde{\xi} \left( \mathbf{x} \right) = \frac{e}{m} \sqrt{- \langle A_{\mu} \left( \mathbf{x},t \right) A^{\mu} \left( \mathbf{x},t \right) \rangle }. \label{meff} \end{align} Similarly to refs. \cite{}, we cope with the temporal finiteness of the pulse by averaging over one field oscillation around $t=0$ only; this approximation is well justified for long, flat-topped multicycle pulses due to the minor influence of the envelope function. The relativistic ponderomotive force then yields, see ref. \cite{Bauer}, \begin{equation} F_{p} = -\left( \mathbf{v}_0 \cdot \boldsymbol{\nabla}_x m_{\ast},~ \boldsymbol{\nabla}_x m_{\ast} + \frac{\gamma_0 -1}{v_0^2} \left( \mathbf{v}_0 \cdot \boldsymbol{\nabla}_x m_{\ast} \right) \mathbf{v}_0 \right), \label{Pond_Rel} \end{equation} where $\gamma_0$ is the Lorentz factor and $\mathbf{v}_0$ denotes the velocity of the quasi-particle. We do not use equation \eqref{Pond_Rel} directly to calculate reference values. However, as equation \eqref{Pond_Rel} describes an effective force, where all short scale contributions are ''integrated out``, it primarily serves as a tool helping to interpret the results obtained from solving the system \eqref{eq_Cy1}-\eqref{eq_Cy4}. Analyzing equation \eqref{Pond_Rel} analytically, we can deduce that the term $\boldsymbol{\nabla}_x m_{\ast}$ is the decisive factor in order to understand the effective force on the particles. In turn, we expect that all particles are forced from strong-field towards weak-field regions. Furthermore, in the case under consideration we primarily expect a boost in the parallel momentum. If particles are created with vanishing transversal momentum, ponderomotive forces only act upon them in direction parallel to the applied field. \subsection{Trajectory-based semi-classical model} \label{Ch_Con2} The virtue of the DHW method, namely to take into account various effects in one common approach, obstructs the analysis of its results. Hence, we introduce a trajectory-based model in order to overcome these difficulties. Contrary to Schwinger pair production the formation time in multiphoton pair production is quite long. For the sake of simplicity, we still want to assume that particles are created at points in space-time, where the electric field takes on its maximal values.\footnote{We are well aware of the fact that this assumption leads to an oversimplification of the description of multiphoton processes as particles can be created at all times. However, evaluating particle trajectories that start at turning points of the applied fields yields reference values for the final particle momenta which suffices, in the context of the current investigation, to validate the method.} Moreover, for the case of $n$-photon pair production we expect that the initial particle momenta $p$ can be derived via energy conservation laws \cite{}: \begin{equation} p_n^2 = \left( \frac{n \omega}{2} \right)^2 - m_{\ast}^2. \label{En} \end{equation} These assumptions are sufficient to determine the trajectory of a particle in an external field as they provide all needed initial conditions. Due to the form of the electric field and especially due to the absence of a magnetic field, spin-effects will be ignored at this point. Hence, we employ the relativistic Lorentz equation \begin{equation} \frac{d P^{\mu}}{d \tau} = e F^{\mu \nu} U_{\nu} \label{Lorentz} \end{equation} to analyze the particle's trajectory in the external field. Here, $F^{\mu \nu}$ is the electromagnetic field strength tensor, $U_{\nu}$ the four-velocity and $P^{\mu}$ the four-momentum. This method is a convenient yet powerful tool to analyse the dynamics of pair production. However, it should be understood as a simple approximation, which clearly cannot replace a full quantum field theoretical treatment of the process. In ref. \cite{Diss} its usefulness is demonstrated for a variety of field configurations. \section{Numerical results for multiphoton pair production in the threshold regime} \label{Results} In the following we discuss the solutions of the PDEs \eqref{eq_Cy1}-\eqref{eq_Cy4} and compare the outcome with the results obtained from the trajectory-based approach. We analyze the distribution function for parallel particle momenta and subsequently the total production rate. For the pulse length we have chosen a value of $\tau = 100$ $m^{-1}$ for all calculations. Such a pulse length is sufficient to capture all essential features of multiphoton pair production. \subsection{Parallel momentum distribution} \label{Ch_Res1} In this subsection we only analyze particle spectra where $p_\rho=0$. In Fig. \ref{Fig1} a typical spectrum for multiphoton-dominated particle creation is displayed. Especially for quasi-homogeneous fields, $\lambda=1000$ $m^{-1}$, the characteristic multiphoton peaks are easily distinguishable. Here, the peak in the middle stems from a \mbox{$3$-} photon process and the side maxima are related to \mbox{$4$-}, and $5$-photon pair production. \begin{figure}[th] \begin{center} \includegraphics[width=0.48\textwidth]{distrC.eps} \end{center} \caption{Reduced particle distribution function obtained from a DHW calculation for a field of strength $\varepsilon=0.5$, length $\tau=100$ $m^{-1}$ and frequency $\omega=0.7$ $m$. The peaks can be related to $n$-photon pair production. The stronger the focus the stronger the ponderomotive forces, thus the higher the final particle momentum.} \label{Fig1} \end{figure} However, comparing the relative peak sizes with results obtained from calculations within homogeneous fields one finds that the strong peak around $p_z=0$ is now more pronounced. Nevertheless, we recover the quantum kinetic limit \cite{} when evaluating the results for $\lambda \to \infty$ at $z=0$. Going beyond this limit reveals that a decrease of the pulse's spot size leads to dramatic changes in the distribution function. In case of, e.g. $\lambda=10$ $m^{-1}$, the dominant peak in the particle momentum spectrum takes on a much wider form which can be related to quantum interferences. It is known from atomic ionization \cite{Corkum} that $n$-photon peaks can be interpreted as a result of particle trajectories adding up to the interference pattern observable. Around $\lambda = 10$ $m^{-1}$ the finite size of the laser pulse seems to prevent a coherent superposition. In turn, the corresponding interference pattern become disturbed resulting in the broadened distribution. A quantitative comparison between the trajectory-based approach and the DHW formalism is summarized in Tab. \ref{Tab1}. At an even smaller focus size, $\lambda = 2.5$ $m^{-1}$, we observe so-called peak splitting. This effect can be understood in the context of ponderomotive forces. Concentrating on the $3-$photon peak, particles are created with close to vanishing momentum $p_z$ and subsequently follow the oscillations of the background field. Due to the spatial inhomogeneity these particles then drift to low-intensity regions in space. As the applied electric field is symmetric in $z$ there are, however, two equally likely options: either the particles are accelerated in $z$ or in $-z$ direction. Hence, the two peaks at nonzero final momentum. \setlength{\extrarowheight}{3pt} \setlength\tabcolsep{15pt} \begin{table}[bh] \caption{Trajectory analysis of particles within an external field. The results are obtained by evaluating the relativistic Lorentz force equation for particles seeded at $t_0=0,~ z_0=0$ in a field of strength $\varepsilon=0.5$, length $\tau=100$ $m^{-1}$, frequency $\omega=0.7$ $m$ and spatial extent $\lambda$. The different initial momenta $p_{z,0}$ correspond to different $n$-photon processes. The final momenta $p_{z,f}$ are obtained at asymptotic times. For comparison we provide the results from a DHW calculation $p_{\rm DHW}$.} \label{Tab1} \begin{ruledtabular} \begin{tabular}{cccc} $\lambda$ $[m^{-1}]$ & $p_{z,0}$ $[m]$ & $p_{z,f}$ $[m]$ & $p_{\rm DHW}$ $[m]$ \\ \colrule 1000 & 0 & $10^{-7}$ & 0 \\ 10 & 0 & 0.162 & 0 - 0.38 \\ 2.5 & 0 & 0.444 & 0.33 \\ \colrule 1000 & 0.92 & 0.92 & 0.92 \\ 10 & 0.92 & 0.99 & 0.98 \\ 2.5 & 0.92 & 1.12 & 1.03 \\ \colrule 1000 & 1.4 & 1.4 & 1.40 \\ 10 & 1.4 & 1.43 & 1.43 \\ 2.5 & 1.4 & 1.65 & 1.45 \\ \end{tabular} \end{ruledtabular} \end{table} \setlength{\extrarowheight}{0pt} \setlength\tabcolsep{0pt} Particles with non-zero initial momenta acquire only an additional push due to ponderomotive forces. The strength of this boost depends on the initial conditions, but seems to be non-monotonic following the results in Tab. \ref{Tab1}. We interpret the increase in momentum as a consequence of particle acceleration out of the strong field region within one half-cycle. If the spatial extent of the background field is sufficiently small these particles are basically unaffected by the following field oscillations and keep their respective momentum. {A finite spatial extent also affects the particle production rate. Fig. \ref{Fig4} shows the reduced particle yield for different photon frequencies and various spot sizes. In the case of $3$-photon pair production (blue line in Fig. \ref{Fig4}) the introduction of a spatial extent only lowers the particle production rate. However, in case of photon frequencies of $\omega = 0.7$ $m$ and a field strength of $\varepsilon=0.5$ a $3$-photon pair production process is not possible due to energy conservation laws, see equation \eqref{En} and ref. \cite{}. The fact, that we still observe a peak at $p_z \sim 0$ in the particle spectrum, see Fig. \ref{Fig1}, is a hint towards a dynamically assisted tunneling process, where absorbing $3$ photons lowers the energy barrier and, in turn, increases the likelihood of a tunneling process to happen. Additionally, the $4$-photon creation process contributes towards the yield in Fig. \ref{Fig4} (dashed red line). All in all, it seems as if especially particles created via the assistance mechanism \cite{Linder} benefit from a small spot size and the resulting strong ponderomotive forces leading to an overall increase in the reduced yield. Concerning the drop-off for small values of the spatial extent $\lambda$: If the spatial extension of the field is of the order of the Compton wavelength the total electric field energy becomes too small to produce a sizable amount of particles \cite{}.} \begin{figure}[th] \begin{center} \includegraphics[width=0.48\textwidth]{N2.eps} \end{center} \caption{Log-log plot of the reduced particle yield $N/\lambda$ calculated from the DHW equations for various field frequencies $\omega$ as a function of the spatial extent $\lambda$. The sharp decrease for $\lambda \sim \mathcal{O} \left( 1 \right)$ is directly related to the lack of sufficient field energy to produce particles. Parameters: $\varepsilon = 0.5$, $\tau=100$ $m^{-1}$.} \label{Fig4} \end{figure} \subsection{Transversal momentum distribution} \label{Ch_Res2} Following the discussion on particle acceleration parallel to the applied field we now turn our attention towards electrons/positrons, which are created with non-vanishing transversal momentum. For demonstration purpose we have chosen a slightly higher field frequency. In this way, the implications of ponderomotive forces in the particle momentum spectrum become more evident. The Keldysh parameter for the field used is $\gamma = 1.68$ indicating a process in the crossover regime with multiphoton dominance \cite{Diss}. \begin{figure}[bh] \includegraphics[width=0.49\textwidth]{716_Mom5.eps} \includegraphics[width=0.49\textwidth]{727_Mom5.eps} \caption{Momentum maps $N/\lambda$ obtained from evaluating the DHW equations. The electric field features a field strength of $\varepsilon=0.5$, a pulse length of $\tau=100$ $m^{-1}$, a field frequency of $\omega=0.84$ $m$ and spatial extents $\lambda=100$ $m^{-1}$(top) or $\lambda=2$ $m^{-1}$(bottom). The additional acceleration in parallel direction (in the second plot) can be attributed to strong ponderomotive forces. The result is line broadening and peak splitting.} \label{Fig2} \end{figure} The comparison between a flat ($\lambda = 100$ $m^{-1}$) and a sharp peak ($\lambda = 2$ $m^{-1}$) is illustrated in Fig. \ref{Fig2} with focus on a $3$-photon process. The figure at the top shows a situation close to a result obtained via a purely time-dependent electric field \cite{Paulus,Li}. Particles created via photon absorption obtain a well-defined total momentum \eqref{En}. Additionally, the resulting ring-like structure is superimposed by quantum interferences \cite{}. \setlength{\extrarowheight}{3pt} \setlength\tabcolsep{15pt} \begin{table \caption{ The relativistic Lorentz force equation is evaluated for particles seeded at local maxima of the field, $t_0$ and $z_0=0$, in a field of strength $\varepsilon=0.5$, length $\tau=100$ $m^{-1}$, frequency $\omega=0.84$ $m$ and spatial extent $\lambda$. The initial momenta were chosen to be $p_{z,0} = 0.665$ $m$ and $p_{\rho,0}=0$ resembling the photon peak positions in Fig. \ref{Fig2}.} \label{Tab4} \begin{ruledtabular} \begin{tabular}{rcc} $t_0$ $[m^{-1}]$ & $p_{z,f} \left( \lambda=100 \right)$ $[m]$ & $p_{z,f} \left( \lambda=2 \right)$ $[m]$ \\ \colrule $-2 \pi / \omega$ & 0.70 & 0.79 \\ $-\pi / \omega$ & 0.70 & 0.85 \\ $0$ & 0.70 & 0.80 \\ $\pi / \omega$ & 0.70 & 0.84 \\ $2 \pi / \omega$ & 0.70 & 0.80 \\ \end{tabular} \end{ruledtabular} \end{table} \setlength{\extrarowheight}{0pt} \setlength\tabcolsep{0pt} A smaller spatial extension of the laser focus leads to an increase in strength of ponderomotive forces, clearly visible in Fig. \ref{Fig2} in the form of an acceleration in $p_z$-direction. In similarity to the previous case, the bunch of particles at $p_z = 0$, $p_\rho = 0.7$ $m$ is boosted in either $p_z$-direction leading to peak splitting. Moreover, the interference pattern diminishes and the sharp peaks signalizing $n$-photon processes are washed out. Once again, we analyze the particle trajectories to understand this line broadening. Constraining ourselves to the center of the laser pulse we seed the particles at local maxima of the applied electric field and calculate the particles final momenta in dependence of the spatial extent, see Tab. \ref{Tab4}. In the case of a broad spatial extent all particles acquire the same final momenta. In case of a small extent, however, the finiteness of the pulse plays a crucial role because then different particle creation times lead to different behaviour. \setlength{\extrarowheight}{3pt} \setlength\tabcolsep{15pt} \begin{table}[bh] \caption{Evaluating the relativistic Lorentz force equation for particles seeded at $t_0=0,~ z_0=0$ with parallel momentum $p_{z,0}=0$ and transversal momentum $p_{\rho,0}$. The external field shows a strength of $\varepsilon=0.5$, a pulse length of $\tau=100$ $m^{-1}$ and a frequency of $\omega=0.7$ $m$. The spatial extent $\lambda$ is varied. The final parallel momenta $p_{z,f}$ are obtained at asymptotic times. Note, that the transversal momenta are nearly unaffected by the external field.} \label{Tab2} \begin{ruledtabular} \begin{tabular}{ccc} $p_{\rho,0}$ $[m]$ $\sim$ $p_{\rho,f}$ $[m]$ & $\lambda$ $[m^{-1}]$ & $p_{z,f}$ $[m]$ \\ \colrule 0 & 1000 & $8 \cdot 10^{-6}$ \\ 0.5 & 1000 & $7 \cdot 10^{-6}$ \\ 1 & 1000 & $4 \cdot 10^{-6}$ \\ \colrule 0 & 20 & 0.026 \\ 0.5 & 20 & 0.021 \\ 1 & 20 & 0.013 \\ \colrule 0 & 10 & 0.16 \\ 0.5 & 10 & 0.12 \\ 1 & 10 & 0.07 \\ \end{tabular} \end{ruledtabular} \end{table} \setlength{\extrarowheight}{0pt} \setlength\tabcolsep{0pt} Due to the fact, that we seed the particles at maxima of the field only, we obtain an upper and a lower limit for the particles' final momenta. Assuming that electrons and positrons are created also at intermediate times, we expect that their respective final momenta are distributed in between these two limits. It is therefore not surprising that the peaks in the momentum spectrum appear much wider compared to the quasi-homogeneous case. The smaller boost for higher transversal momenta is a consequence of relativistic mechanics as can be seen by comparison with the effective ponderomotive forces \eqref{Pond_Rel}. For the sake of completeness, we provide the data obtained through a trajectory analysis in Tab. \ref{Tab2}. \section{Conclusions} \label{Conclusions} We have presented numerical solutions describing multiphoton pair production for oscillating, spatially inhomogeneous electric fields in the DHW formalism. Spatial inhomogeneities introduce effective ponderomotive forces, which directly affect the particle momentum distribution and subsequently the total yield. Moreover, we have shown that these forces can be understood via a generalized effective mass concept. With the aid of a semi-classical trajectory-based model we produced reference values to analyse our findings regarding new phenomena connected with a finite spatial pulse size: Peak splitting and line broadening. As for the first effect, we note that the peaks split due to strong ponderomotive forces altering the particle momentum spectrum. Line broadening happens because the particle spectra properties which are characteristic for multiphoton pair production erode and instead of sharp lines one obtains broad bunches. In summary, we presented here further evidence how important it is to take spatial inhomogeneities of the fields underlying pair production processes into account. Therefore, further investigations aiming at an understanding of matter creation from fields will have, at least, to include spatial variations of the fields. Given sufficient computational resources the DHW formalism can be readily extended such that inhomogeneous magnetic fields can be included. This would imply a major step towards understanding multiphoton pair production in realistic scenarios further closing the gap between theory and experiment. \section{Acknowledgements} We thank T. Heinzl for helpful and interesting discussions. C. K. acknowledges financial support by the FWF (Austrian Science Fund) Doctoral Program DK W1203-N16, by the BMBF under grant No.\ 05P15SJFAA (FAIR-APPA-SPARC), and by the Helmholtz Association through the Helmholtz Postdoc Programme (PD-316). Computations were performed on the ``Supermicro Server 1028TR-TF'' in Jena, which was funded by the Helmholtz Postdoc Programme (PD-316).

\section{Introduction} Low Density Parity Check (LDPC) codes have been widely applied nowadays\cite{LDPC}. Although various decoding methods have been developed for LDPC, e.g. Belief Propagation (BP) and Min-Sum (MS) \cite{EFAF, LDPC}, nearly all of the existing solutions are approximation-based (not truly Maximum Likelihood solution) and consequently suffer an early error floor, especially in the case of high Signal-to-Noise Ratio (SNR). Therefore, it is important yet challenging to develop an error-floor-free decoding scheme for LDPC with more accurate decoding performance. For binary LDPC codes over symmetric channels, Feldman introduced a relaxed version \cite{DECC, ULPT} of Maximum Likelihood (ML) decoding problem which can be interpreted as a Linear Program (LP) decoding, whose certificate of correctness (ML certificate) and performance are reported in \cite{OLCL}. Compared with other decoding methods, such as BP and MS, LP decoding can reduce the error floor\cite{MPAA}. Many algorithms have been proposed for LP decoding. The most related work is the Alternating Direction Method of Multipliers (ADMM) \cite{ADMM} introduced in \cite{DMFL}, which can be easily implemented in a distributed way. In order to obtain the ideal decoding performance, Liu et al. \cite{SPBP} developed a penalized-LP decoding model by adding different types of penalty terms in the objective to drive the solution away from 0.5, and reported that the $\ell_2$ penalty term generally achieved better error performance. However, one remaining hurdle is the difficulty in properly choosing the penalty parameter. The effect of the penalty parameter is two-fold. First, the penalty parameter closely affects the performance of the algorithm and an improper value can make the algorithm extremely inefficient. Most importantly, the optimal solution of the penalized-LP decoding problem may vary for different penalty parameter values. As a result, the decoder often finds solutions with low accuracy. It is worth noting that the penalized-LP formulation in fact approximates the original problem via penalizing the violation of the binary constraint. Hence its optimal solution is rarely the optimal solution of the original LP decoding problem. In this paper, we apply a parameter-free continuous optimization model \cite{AVFF}, which is an \underline{\textbf{\emph{exact}}} reformulation of the binary LP decoding. The binary constraint in the LP decoding is replaced by the intersection of a box and an $\ell_p$ sphere. We also design an efficient ADMM solver, which can be parallelized to reduce the computational cost. Simulation results demonstrate that the proposed model along with the ADMM solver can further bring down the error floor for large SNRs. \section{Background}\label{sec.back} We consider a binary linear LDPC code $C$ of length $N$, which can be defined by an $M \times N$ parity check matrix $\bm{H}$. Each column of the parity check matrix corresponds to a codeword symbol, indexed by $\mathcal{I}:=\{1,2,...,N\} $. Each row of the parity check matrix corresponds to a check, which specifies a subset of codeword symbols that add to 0 (modulo 2), indexed by $\mathcal{J}:=\{1,2,...,M\} $. The neighborhood of check $j$, denoted as $\mathcal{N} (j)$, is the set of variables that check $j$ constrains to add to 0. That is \cite{GECC}, $\mathcal{N} (j) = \{ i \in \mathcal{I}:\bm{H}_{j,i} = 1 \}$. For a binary linear code transmitted over a symmetric memoryless channel, let $\mathcal{X} = \{0, 1\}$ be the input space, $\mathcal{Y}$ be the output space and $P(\bm{y}|\bm{x})$ denotes the probability that $\bm{y} \in \mathcal{Y}$ is received if the codeword $\bm{x} \in \mathcal{X}$ is sent over the channel. Let $\mathcal{C}$ be the set of possible codewords. By Bayes' rule, the decoding problem can be modeled as $\arg \max_{\bm{x} \in \mathcal{C}} P(\bm{y}|\bm{x})$. In particular, the Maximum Log-likelihood (ML) decoding problem \cite{DECC,ULPT} takes the form $ \hat{\bm{x}}= \arg \max_{\bm{x} \in \mathcal{C}} P(\bm{y}|\bm{x}) = \arg \min_{\bm{x} \in \mathcal{C}} \bm{\gamma}^T \bm{x}, $ where $\bm{\gamma}$ is a length-$N$ vector of Log-likelihood Ratios (LLRs) with $\bm{\gamma}_i = \log (P(y_i|0)/P(y_i|1) )$. Some work \cite{DMFL,ANLP} has been done in the past decades focusing on the properties of the convex hull of all possible codewords. ML decoding can be formulated as the problem of minimizing $\bm{\gamma}^T \bm{x}$ over the convex hull---an Integer Programming (IP) decoding problem that is NP-hard in general\cite{OTII}. Denote the subset of coordinates of $\bm{x}$ participating in the $j$th check as the matrix $\bm{P}_j$, so $\bm{P}_j$ is a binary $d \times N$ matrix consisting of $d$ components participating in the $j$th check. From \cite{DMFL}, each local codeword constraint can be relaxed to satisfy $\bm{P}_j \bm{x} \in \mathbb{PP}_d$, where $\mathbb{PP}_d$ is the specific expression of the convex hull of all possible codewords. The LP decoding problem is then derived in \cite{DMFL}, by relaxing the binary constraints \begin{equation}\label{LP} \min_{\bm{x}}\quad \bm{\gamma}^T \bm{x} \quad \text{s.t.} \quad \bm{x} \in [0,1]^N,\quad \bm{P}_j \bm{x} \in \mathbb{PP}_d , \forall j \in \mathcal{J}, \end{equation} where $[0, 1]^N$ is the $N$-dimensional box. The binary constraint $\bm{x} \in \{0,1\}^N$ in ML decoding is relaxed to $\bm{x} \in [0,1]^N$ now. The simulation results in \cite{DMFL} suggest that an ideal decoder should perform like BP at low SNRs. To achieve this, a penalized-LP decoding is proposed in \cite{SPBP} by penalizing the fractional solutions, that is, the objective of \eqref{LP} is replaced with $\bm{\gamma}^T \bm{x} + \sum_{i=1}^N g(x_i)$, where the penalty function $g : [0, 1]\rightarrow \mathbb{R} \cup \{\pm \infty \} $ has three options, namely, $\ell_1$, $\ell_2$ and $\log$ penalty. That is, $g_1(x) = −\alpha|x - 0.5|$, $g_2(x) = −\alpha(x - 0.5)^2$, $g_3(x) = −\alpha\log(|x - 0.5|)$, with penalty parameter $\alpha>0$. It is reported that the $\ell_2$ penalty has the best error performance and the fastest convergence among the options. Other improved penalty functions are proposed and experimented in \cite{IPFO}. \section{ $\ell_p$-box decoding} In this section, we propose a new formulation for LP decoding. Compared to the existing formulations suitable to decentralized processing, where mixed integer formulation are usually involved, our proposed formulation is to design a continuous and exact reformulation of the LP decoding to circumvent using the traditional method to solve the IP problem, such as branch and bound and Lagrangian relaxation. The major technique used in our model is the $\ell_p$-box recently proposed in \cite{AVFF} for solving IP problems. The main idea of this technique is to replace the binary constraint by the intersection of a box and an $\ell_p$ sphere with $p \in (0, \infty)$: \begin{equation}\label{lpbox} \bm{x} \in \{0,1\}^N \Leftrightarrow \bm{x} \in [0,1]^N \cap \{ \bm{x}: \|\bm{x} - \tfrac{1}{2} \mathbf{1}_N\|_p^p = \tfrac{N}{2^p} \}, \end{equation} where $\mathbf{1}_N$ is the $N$-dimensional vector filled with all $1$s. Now consider this technique in the LP decoding problem \eqref{LP}. Besides the existing box constraint $[0,1]^N$, we only need an additional $\ell_p$-sphere constraint to enforce the solution to be binary, yielding the $\ell_p$-box decoding problem: \begin{equation}\label{lpbox_ADMM} \begin{aligned} \min_{\bm{x}}\ &\bm{\gamma}^T \bm{x}\\ \text{s.t.}\ & \bm{x} \in [0,1]^N, \|\bm{x} -\tfrac{1}{2}\mathbf{1}_N\|_p^p = \tfrac{N}{2^p}, \bm{P}_j \bm{x} \in \mathbb{PP}_d , \forall j \in \mathcal{J}. \end{aligned} \end{equation} For the penalized-LP decoding, the choice of parameters is critical to the solution performance and, unfortunately, an appropriate choice can be very intricate to obtain. If the penalty parameter is set too large, the algorithm may quickly converge to a binary point far from the global solution. On the other hand, for a too small penalty parameter, the algorithm may converge to a fractional point. Compared to the parameterized penalized-LP formulation, there is no parameter involved in \eqref{lpbox_ADMM}, rescuing users from adjusting the penalty parameters according to SNR. It should be noticed that \eqref{lpbox_ADMM} is equivalent to the binary decoding problem. Therefore, the optimal solution of the binary decoding problem is the global optimal solution of \eqref{lpbox_ADMM}. Furthermore, \eqref{lpbox_ADMM} is a continuous problem, and thus can be attacked by various continuous optimization algorithms. There are many options for the selection of $p$, and \eqref{lpbox_ADMM} becomes nonsmooth with $p\in(0,1]$. In this paper, we use $p=2$ to keep the projection of a point onto the sphere simple. \section{$\ell_2$-box ADMM}\label{sec.admm} In this section, we design an efficient algorithm for solving the proposed $\ell_2$-box decoding problem by incorporating the $\ell_2$-box technique in the ADMM framework. As discussed in previous section, we set $p=2$. In order to apply the ADMM, we introduce two sets of auxiliary variables $\bm{y}$ and $\bm{z}$ to decouple the box and the $\ell_2$ sphere, resulting in an $\ell_2$-box decoding problem: \begin{equation}\label{l2box_ADMM} \begin{aligned} \min_{\bm{x, y, z}}\quad &\bm{\gamma}^T \bm{x}\\ \text{s.t.}\quad &\bm{y}= \bm{x},\quad \|\bm{y}-\tfrac{1}{2}\mathbf{1}_N\|_2^2 = \tfrac{N}{4},\quad \bm{x} \in [0,1]^N\\ &\bm{z}_j= \bm{P}_j \bm{x} ,\quad \bm{z}_j\in \mathbb{PP}_d ,\quad \forall j \in \mathcal{J}, \end{aligned} \end{equation} where $\bm{y}$ is a $N$-dimensional vector and $\bm{z}_j \in \mathbb{PP}_d $ for all $j \in \mathcal{J}$. The ADMM works with the augmented Lagrangian $ L_{\mu_1,\mu_2}(\bm{x},\bm{y},\bm{z},\bm{\lambda}_1,\bm{\lambda}_2) = \bm{\gamma}^T \bm{x} + \sum\limits_{j\in \mathcal{J}} \bm{\lambda}_{1,j}^T(\bm{P}_j \bm{x} -\bm{z}_j) +\frac{\mu_1}{2}\sum\limits_{j\in \mathcal{J}}\|\bm{P}_j \bm{x} -\bm{z}_j\|_2^2 + \bm{\lambda}_2^T(\bm{x} -\bm{y}) +\frac{\mu_2}{2}\|\bm{x} -\bm{y}\|_2^2 $ with constants $\mu_1$ and $\mu_2$ all positive. Here $\bm{\lambda}_1$ is the dual variable associated with the constraint $\bm{z}_j= \bm{P}_j x$ and its dimension is the cardinality of $|\mathcal{J}|$, and the $N$-dimensional $\bm{\lambda}_2$ is the dual variable associated with the coupling constraint $\bm{y}= \bm{x}$. Let $\mathcal{X}$, $\mathcal{Y}$ and $\mathcal{Z}$ denote the feasible regions for variables $\bm{x}$, $\bm{y}$ and $\bm{z}$ respectively, i.e., $\mathcal{X} =\mathcal{Y}= [0,1]^N$ and \begin{equation*} \begin{array}{rll} \mathcal{Z} = & \underbrace{\mathbb{PP}_{|\mathcal{N}(j)|} \times ...\times \mathbb{PP}_{|\mathcal{N}(j)|}}_{number\ of \ checks\ \mathcal{J}} \end{array} . \end {equation*} The ADMM iteration then can be elaborated as follows {\small \begin{equation*} \bm{x}\text{-update}: \ \bm{x}^{k+1} = \arg\min_{\bm{x} \in \mathcal{X}} L_{\mu_1,\mu_2}(\bm{x},\bm{y}^k,\bm{z}^k,\bm{\lambda}_1^k,\bm{\lambda}_2^k), \end{equation*} \begin{equation*} (\bm{y},\bm{x})\text{-update}: \left\{\begin{split} &\bm{y}^{k+1} = \arg\min_{\bm{y} \in \mathcal{Y}} L_{\mu_1,\mu_2}(\bm{x}^{k+1},\bm{y},\bm{z}^k,\bm{\lambda}_1^k,\bm{\lambda}_2^k) \\ \ & \text{s.t.} \ \|\bm{y}-\tfrac{1}{2}\mathbf{1}_N\|_2^2 = \tfrac{N}{4} \\ &\bm{z}^{k+1} = \arg\min_{z \in \mathcal{Z}} L_{\mu_1,\mu_2}(\bm{x}^{k+1},\bm{y}^k,\bm{z},\bm{\lambda}_1^k,\bm{\lambda}_2^k)\\ \ & \text{s.t.} \ \bm{z}_j\in \mathbb{PP}_d ,\quad \forall j \in \mathcal{J}, \end{split}\right. \end{equation*} \begin{equation*} (\bm{\lambda}_1,\bm{\lambda}_2)\text{-update}: \left\{\begin{split} &\bm{\lambda}_{1,j}^{k+1} = \bm{\lambda}_{1,j}^{k} + \mu_1(\bm{P}_j\bm{x}^{k+1} - \bm{z}_j^{k+1}) ,\ \forall j \in \mathcal{J}\\ &\bm{\lambda}_2^{k+1} = \bm{\lambda}_2^{k} + \mu_2(\bm{x}^{k+1} - \bm{y}^{k+1}). \end{split}\right. \end{equation*} } Next, we discuss the solution of the ADMM subproblems. (a) $\bm{x}$-update. The $\bm{x}$-update is to minimize $L_{\mu_1,\mu_2}(\bm{x},\bm{y},\bm{z},\bm{\lambda}_1,\bm{\lambda}_2)$ subject to $\bm{x}\in[0,1]^N$ with fixed $(\bm{y},\bm{z},\bm{\lambda}_1,\bm{\lambda}_2)$, The $i$th component of the solution of the $\bm{x}$-update subproblem can be given explicitly by $ \bm{x}_i^{k+1} = \Pi_{[0,1]^N}( \tfrac{ d^k- \gamma_i - \bm{\lambda}_{2,i}^k +\mu_2 \bm{y}_i^k }{\mu_1|\mathcal{N}(i)|+\mu_2} ). $ Here $d^k = \sum_{j \in \mathcal{N}(i)} \mu_1 (\bm{P_j}^T\bm{z}_j^k)^{(i)} - (\bm{P_j}^T\bm{\lambda}_{1,j}^k)^{(i)} $ where $(\bm{P_j}^T\bm{z}_j^k)^{(i)}$ denotes the $i$th component of $\bm{P_j}^T\bm{z}_j^k$, and so forth for $(\bm{P_j}^T\bm{\lambda}_{1,j}^k)^{(i)}$, $\bm{x}_i$, $\bm{\lambda}_{2,i}$ and $\bm{y}_i$. Define $\bm{P} = \sum_{j\in \mathcal{J}} \bm{P}_j^T\bm{P}_j$ and $ \Pi_{[0,1]^N} $ as the projection operator onto the $N$-dimensional box $[0, 1]^N$. Note that for any $j$, $\bm{P} = \sum_{j\in \mathcal{J}} \bm{P}_j^T\bm{P}_j$ is an $N\times N$ diagonal binary matrix with non-zero entries at $(i, i)$ if and only if $i$ participates in the $j$th parity check, i.e., $i \in \mathcal{N}(j)$. This implies that $\bm{P} = \sum_{j\in \mathcal{J}} \bm{P}_j^T\bm{P}_j$ is a diagonal matrix with the $(i, i)$th entry equal to $|\mathcal{N}(i)|$. Hence $\bm{P}^{-1}$ is diagonal with $1/|\mathcal{N}(i)|$ as the $i$th diagonal entry. See details of this calculation in \cite{DMFL}. (b) $(\bm{y},\bm{z})$-update. Notice that the two variables $\bm{y}$ and $\bm{z}$ are independent to each other, thus they can be updated separately. By fixing $(\bm{x},\bm{z},\bm{\lambda}_1,\bm{\lambda}_2)$, the $\bm{y}$-update subproblem is given by \begin{equation}\label{prob.y} \begin{aligned} &\bm{y}^{k+1} = \arg\min_{\bm{y} \in \mathcal{Y}} L_{\mu_1,\mu_2}(\bm{x}^{k+1},\bm{y},\bm{z}^k,\bm{\lambda}_1^k,\bm{\lambda}_2^k),\\ \ &\text{s.t.} \ \|\bm{y}-\tfrac{1}{2}\mathbf{1}_N\|_2^2 = \tfrac{N}{4}. \end{aligned} \end{equation} Thus, by eliminating the constants, \eqref{prob.y} can be written as \begin{equation*} \min_{\bm{y}}\ - \left( \mu_2(\bm{x}^{k+1}-\tfrac{1}{2}\mathbf{1}_N) + \bm{\lambda}_2^{k} \right)^T \bm{y} \quad \text{s.t.} \ \|\bm{y}-\tfrac{1}{2}\mathbf{1}_N\|_2^2 = \tfrac{N}{4}, \end{equation*} which has an explicit solution: $ \bm{y}^{k+1} = \tfrac{\mu_2(\bm{x}^{k+1}-\tfrac{1}{2}\mathbf{1}_N) + \bm{\lambda}_2^{k}}{\|\mu_2(\bm{x}^{k+1}-\tfrac{1}{2}\mathbf{1}_N) + \bm{\lambda}_2^{k}\|} \times \tfrac{\sqrt{N}}{2} + \tfrac{1}{2}. $ For updating $\bm{z}$, the subproblem is to solve \begin{equation} \min_{\bm{z}_j} \ \tfrac{\mu_1}{2}\|\bm{P}_j \bm{x}^{k+1} -\bm{z}_j\|_2^2 - \bm{\lambda}^k_{1,j} \bm{z}_j, \ \text{s.t.} \ \bm{z}_j \in \mathbb{PP}_d, \end{equation} for each $j \in \mathcal{J}$. The solution of this subproblem can be given explicitly \cite{SPBP}: $ \bm{z}^{k+1}_j =\Pi_{\mathbb{PP}_d} (\bm{P}_j\bm{x}^{k+1} + \tfrac{1}{\mu_1} \bm{\lambda}^k_{1,j} ), \ \forall j \in\mathcal{J}, $ where $ \Pi_{\mathbb{PP}_d} $ is the projection operator onto the codeword polytope. (c) $(\bm{\lambda}_1,\bm{\lambda}_2)$-update. Note that $\bm{\lambda}_1$ and $\bm{\lambda}_2$ are also independent to each other, and thus can be updated separately: \begin{equation} \begin{aligned} &\bm{\lambda}_{1,j}^{k+1} = \bm{\lambda}_{1,j}^{k} + \mu_1(\bm{P}_j\bm{x}^{k+1} - \bm{z}_j^{k+1}), \ \forall j\in\mathcal{J}\\ &\bm{\lambda}_2^{k+1} = \bm{\lambda}_2^{k} + \mu_2(\bm{x}^{k+1} - \bm{y}^{k+1}). \end{aligned} \end{equation} The entire $\ell_2$-box ADMM decoding algorithm is stated in Algorithm~\ref{alg.admm}. \begin{algorithm}[htbp] \caption{$\ell_2$-box ADMM decoding} \label{alg.admm} \begin{algorithmic}[1] \STATE Given a $M \times N$ parity check matrix $\bm{H}$, and parameters $\mu_1>0$, $\mu_2>0$, tolerance $\epsilon>0$. \STATE Construct the log-likelihood vector $\bm{\gamma}$ and the $d \times N$ matrix $\bm{P}_j $ for all $j \in \mathcal{J}$. \STATE Initialize $\bm{y}$, $\bm{\lambda}_2$, $\bm{z}_j$ and $\bm{\lambda}_{1,j}$ for all $j \in \mathcal{J}$. \STATE Set $k$ = 0. \REPEAT \STATE \textbf{Update $\bm{x}$}:\ $\bm{x}_i^{k+1} = \Pi_{[0,1]^N}( \frac{ d^k- \gamma_i - \bm{\lambda}_{2,i}^k +\mu_2 \bm{y}_i^k }{\mu_1|\mathcal{N}(i)|+\mu_2} ). $ \STATE \textbf{Update $\bm{y}$}:\ $ \bm{y}^{k+1} = \frac{\mu_2(\bm{x}^k -\tfrac{1}{2}\mathbf{1}_N) + \bm{\lambda}_2^k}{\|\mu_2(\bm{x}^k -\tfrac{1}{2}\mathbf{1}_N) + \bm{\lambda}_2^k\|}\times \frac{\sqrt{N}}{2} + \tfrac{1}{2}$ \STATE \textbf{Update $\bm{z}$}:\\ \ for all $j\in \mathcal{J}$ do $\bm{z}^{k+1}_j =\Pi_{\mathbb{PP}_d} (\bm{P}_j\bm{x}^{k+1} + \frac{1}{\mu_1} \bm{\lambda}^k_{1,j} )$ \STATE \textbf{Update $\bm{\lambda}_1$}:\\ \ for all $j\in \mathcal{J}$ do $ \bm{\lambda}_{1,j}^{k+1} = \bm{\lambda}_{1,j}^{k} + \mu_1(\bm{P}_j\bm{x}^{k+1} - \bm{z}_j^{k+1})$ \STATE \textbf{Update $\bm{\lambda}_2$}:\\ \ $ \bm{\lambda}_2^{k+1} = \bm{\lambda}_2^{k} + \mu_2(\bm{x}^{k+1} - \bm{y}^{k+1}) $ \STATE Set $k = k+1$ \UNTIL{{\small$\max\limits_j \|\bm{P}_j\bm{x}^{k+1} - \bm{z}_j^{k+1} \|_{\infty} < \epsilon $ and $\|\bm{x}^{k+1} - \bm{y}^{k+1}\|_{\infty}<\epsilon$}} \STATE \textbf{return} $\bm{x}$ \end{algorithmic} \end{algorithm} \section{Numerical Experiments}\label{sec.test} In this section, we design numerical experiments to test the proposed decoding model and algorithm. We use the experimental framework of Liu\cite{SPBP} in our test, and design a simulator that the $[2640,1320]$ ``Margulis" code \cite{CCCM}, which is known for its strong error-floor, with randomly added noise transmitting over an Additive White Gaussian Noise (AWGN) channel. Many acceleration techniques exist for ADMM \cite{FADO}, and the $\ell_2$-penalized ADMM LP decoder in our comparison is in fact an accelerated version \cite{FADO}. Our algorithm, however, is simply the basic ADMM without using acceleration, since our primary focus is on the effectiveness of our proposed decoding model. The ADMM parameters in our implementation are roughly tuned to ensure the algorithm achieve an acceptable performance. \subsection{Effectiveness of the parameter-free model} To emphasize the advantage of our new formulation, we first show for the $\ell_2$-penalized ADMM LP decoder how the penalty parameter value could affect the decoding accuracy and running time. We consider the $\ell_2$-penalized ADMM LP decoder \cite{SPBP} in our comparison, which is reported to be the most effective and efficient among three penalty terms in \cite{SPBP}. In this experiment, we set $\text{SNR}=1.6$dB, and the maximum number of iterations to be 1000. The decoder is terminated whenever 200 errors have been encountered. We record the Word error rate (WER) and the running time of the $\ell_2$-penalized ADMM for penalty parameter value varying from 0.24 to 5.0, and depict the results in Fig.~\ref{}. \begin{figure}[htbp] \centering \subfigure[WER versus $\alpha$]{ \label{} \includegraphics[width=1.65in]{alpha_WER.eps}} \hspace{0in} \subfigure[Running time versus $\alpha$]{ \label{} \includegraphics[width=1.65in]{alpha_time.eps}} \caption{WER and Running time versus $\alpha$} \label{} \end{figure} From Fig.~\ref{} and Fig.~\ref{}, we can see that the penalty parameter for the $\ell_2$-penalized ADMM LP decoder plays an important role in the decoding accuracy and efficiency. The algorithm becomes extremely slowly for too small or too large penalty parameter, and the WER could also be extremely large for improper parameter values. Thus users need to select an appropriate value to achieve acceptable performance. As for our proposed parameter-free decoding model, no parameter tuning task is needed for the users. It should be noticed that the WER and running time of algorithm is competitive with the best performance of the $\ell_2$-penalized ADMM LP decoder. Therefor, the performance of our decoder is stable in different environment. In real situations, generally SNR cannot be estimated accurately, and it is not realistic to adjust penalty parameter in real time for broadcast transmission. These disadvantages may limit the practicability of the $\ell_2$-penalized ADMM LP decoder, while our proposed decoder does not suffer from such issue and thus is more practical in real applications. \subsection{Performance} We test our proposed decoder against existing standard algorithms including the penalized ADMM decoder in \cite{SPBP} and the BP decoder in \cite{DMFL}. The decoding model parameters for each decoder are tuned to achieve the best performance, and the algorithms are terminated once 200 errors are encountered. The WER of each decoder is depicted in Fig.~\ref{SNR-WER}. To show the practicality on other type of codes, we also add a comparison on $[2304, 768]$ WiMAX~\cite{wimax} in Fig.~\ref{wimax_WER}. The MS decoder and corrected MS decoder can be referred to~\cite{ryan}. The max iterations for BP decoder, MS decoder and corrected MS decoder are set as 60. \begin{figure}[htbp] \centering \subfigure[Margulis]{ \label{SNR-WER} \includegraphics[width=1.65in]{WER-SNR_all4.eps}} \hspace{0in} \subfigure[WiMAX]{ \label{wimax_WER} \includegraphics[width=1.65in]{wimax_WER_4.eps}} \caption{WER versus different SNRs} \label{WER} \end{figure} We can see that from Fig.~\ref{SNR-WER}, the overall performance of our proposed $\ell_2$-box ADMM decoder achieves lower WER than all other contemporary decoders. As the SNR turns large, the difference between decoders become increasingly significant, and this is especially the case when the SNR is between 1.8 and 2.4, where the error floor for BP starts from 2. From Fig.~\ref{wimax_WER}, we can also see a significant difference when SNR is between 4 and 4.4, where the error floor for MS occurs. Generally the error-floor effect can be caused by weak distance spectrum of the code or the decoder's poor ability of escaping from local minimizers. Our experiments show that the proposed decoder is more capable of avoiding early convergence to local solutions, thus can achieve better error performance. \begin{table}[htbp] \renewcommand{\arraystretch}{1.3} \caption{Comparison of complexity of each decoder} \label{complexity} \centering \begin{tabular}{|c|c|c|c|c|} \hline \multicolumn{2}{|c|}{Decoders} & BP & $\ell_2$-penalized & $\ell_2$-box\\ \hline \multirow{2}*{Time (s)} &SNR=2.0dB & 0.027 & 0.010 & 0.018 \\ \cline{2-5} &SNR=2.5dB & 0.019 & 0.007 & 0.012 \\ \hline \end{tabular} \end{table} We also recorded CPU time needed for BP, $\ell_2$-penalized ADMM and $\ell_2$-box ADMM in Table~\ref{complexity}. The simulation results confirm that LP decoding outperforms BP decoding as reported in~\cite{MPAA}. It can also be observed that the greatly improved performance on WER by our proposed decoder only cost slightly additional amount of time (less than 0.008 second). \subsection{Choice of algorithmic parameters} Our proposed decoder involves the ADMM parameters $\mu_1$ and $\mu_2$, which may affect the efficiency of the decoder. Fig.~\ref{2mu1-200} depicts the simulations when the decoder encounter 200 errors for $(\mu_1,\mu_2) \in [0,200]\times[0,200]$ with $\text{SNR}=1.6$dB.. We can see that for $(\mu_1,\mu_2) \in [0,200]\times[10,200]$ our proposed decoder is relatively insensitive to the choice of $\mu_1$ and $\mu_2$. \begin{figure} \centering \includegraphics[width=2in]{2mu1-200.eps} \caption{The iterations of choice of the two parameters} \label{2mu1-200} \end{figure} \section{Conclusion}\label{sec.conclusion} In this paper, we have proposed a parameter-free $\ell_p$-box formulation for LDPC decoding and implemented the new formulation with an ADMM solver. We have also developed an efficient ADMM-based algorithm to solve this newly formulated problem in a distributive manner. It should be emphasized that the proposed decoder can be easily applied to various situations with different SNR and channels since the decoding model is parameter-free. \bibliographystyle{IEEEtran}

\section{Introduction} \label{} Detecting opinion leaders and popular users within an online community is a desirable task for a number of practical applications such as social media activity monitoring, content placing, and trend detection. All these tasks, if effectively carried out, can substantially lift the value of a community, however the detection of users who consistently generate content appreciated by the community is not always straightforward. We assume that a community's reception of user-generated content will be somewhat close to a Gaussian one, like is shown in Figure \ref{}, with a relatively small number of items attracting the most negative feedback, another small fraction attracting the most positive feedback, and the vast majority receiving a mild reception or just little attention. While most literature effort is directed towards describing \cite{buckels2014trolls,correa2010interacts} or detecting \cite{arnt2003learning} the ``bad'' users or content that occupies the left hand end of the distribution, in this work we focus on detecting the ``good'' ones, standing on the right hand end. This task is of particular interest since recent studies show evidence of how positive feedback tends to generate more herding effect in Social media than negative one \cite{muchnik2013social} and therefore can be regarded as more interesting from a community management point of view as well. \begin{figure}[ht] \centering \includegraphics[width=0.7\textwidth]{comments_distribution.png} \caption{How we assume the posts to be distributed with respect to the community's reception} \label{} \end{figure} Some platforms provide their users with the possibility of expressing explicit feedback on user-generated content by means of upvoting and downvoting. The proportion between upvotes and downvotes is a simple and straightforward indicator of how the community received the published content: a strong majority of positive feedback clearly indicates a well received content and the identification of users publishing appreciated content in regular basis is rather straightforward. However not all social media provide such tools or allow external applications to access them. As a matter of fact most platforms, like Facebook and Twitter, allow only positive feedback to be explicitly expressed by means of voting rather than commenting. Sentiment Analysis is a valuable tool to understand the reactions of a community to a given content \cite{pang2008opinion}, however Sentiment Analysis systems, being trained on the textual contents of a set of labelled texts, are highly conditioned by writing style nuances and the considered domain of application. Moreover the usage of negation \cite{wiegand2010survey} and irony \cite{bosco2013developing} are notoriously hard problems to face when building a Sentiment Analysis application. We claim that the very structure of the conversations hosted on Social Media platforms embeds information about the sympathy and overall appreciation aroused by a post, and hence the reputation of its author. By leveraging such information, it is possible to abstract over the content of the conversation and therefore avoid non-trivial problems such as the aforementioned negation and irony managing. To support our claim, we present in this work a novel approach exploiting Dung's well known Abstract Argumentation Framework in the context of social media. While other studies focus on understanding the topic of the argumentation, identifying the coalitions of arguments or users, or determining which are the strongest arguments, our work is content agnostic and aims at modelling the social interaction of users involved in an online conversation to extract from its flow insights about users' social influence. The rest of the paper is organised as follows: in Section~\ref{} we introduce some related work, in Section~\ref{} we provide a brief overview of Abstract Argumentation and describe the procedure we used to build a AAF from a social media conversation, in Section~\ref{} we describe the features extracted from the AAF and provide an overview of the used classifiers, in Section~\ref{} we present an evaluation of our technique performed over Reddit conversations. We discuss the results in Section~\ref{}, and we finally draw conclusions and describe some future works directions in Section~\ref{}. \section{Related Work} \label{} Since their birth, social network sites have attracted a growing interest in the research community and have been analysed by several authors in the literature. Extensive data mining activities have been performed to address a broad span of problems including text categorisation, user profiling, communities detection, and content recommendation \cite{abel11}. Noticeable examples of this kind research can be found in~\cite{social-enrich}, where the authors exploit contextual enrichment to improve the topic extraction process from texts posted on the Twitter platform; in~\cite{twitter-sim} a new user model approach has been proposed to compute user similarity based on a network representing the semantic relationships between the words occurring in the same tweet and the related topics; in~\cite{tao12} is presented a user model that features topic detection and entity extraction for tweets and links the latter to news articles to describe the context of the tweets; finally, a recent survey~\cite{riquelme15} illustrates how there is a growing interest in measuring user influence on Twitter. Understanding the behavior of users on various social media platforms, their interaction paradigms \cite{kirschner2015facebook}, and predicting which contents are potentially harmful for the community are still open and debated problems. In \cite{arnt2003learning} a seminal methodology for automated moderation leveraging relatedness between posts and a Naive Bayes classifier is presented. In \cite{arora2015good} a content-based technique to predict which Stack Overflow answers will get negative scores is presented. Most of the proposed approaches, however, rely on content-based features to assess the potential reception of the community to a given post, giving little importance or ignoring the discourse structure of social media interaction. The problem of analysing discourse and in particular argumentation has recently become prominent in the Artificial Intelligence research community with several studies focusing on externalising implicit argumentation structures hidden in user generated content. The authors of \cite{rahwan2009argumentation} provide a comprehensive introduction to the study of argumentation in Artificial Intelligence, including theoretical and computational aspects of the subject. Several formal frameworks have been proposed to describe user argumentation, mostly focused on claims and justifications. Ontological descriptions of argumentation like the SIOC argumentation ontology module \cite{lange2008expressing}, DILIGENT \cite{tempich2005argumentation}, and several others \cite{schneider2013review} have been developed, but found little application due to their intrinsic complexity. Machine learning based classification of social media argumentation has been explored as well, however, as pointed out by the authors of \cite{llewellyn2014re} the scarce availability of human annotated trained corpora and the domain dependency of such approaches are severe limiting factors. Abstract Argumentation Frameworks (herein AAF), on the other hand, are a powerful, yet simple graph based representation of argumentation, introduced in \cite{dung1993acceptability,dung1995acceptability}; the main advantage of Dung's AAF is the abstraction over the actual content of single arguments and the focus on the relationships among considered arguments, which is a substantial difference from more traditional approaches that leverage the breakdown of an argument into claim, premises, and evidence \cite{toulmin2003uses}. AAF allow the evaluation of sets of acceptable arguments, may it be a binary labelling (accepted or rejected) or a ranking with several levels of acceptability. This latter approach can come in many flavours and a detailed comparative survey of ranking-based semantics for AAF is presented in \cite{bonzon2016comparative}. Over the years AAF found application in different domains, each of them requires specific semantics \cite{charwat2015methods} and specific extensions of the original attack-based model. The authors of \cite{leite2011social} investigate the usage of AAF in the context of social media, highlighting how due to the large number of arguments to be considered in such a scenario, ranking based approaches are to be preferred, moreover introduce a Social extension of AAF considering also the votes (such as likes, thumbs up, and so on) expressed by the community. The authors of \cite{grosse2015integrating} propose an extension of the AAF to mine opinions from Twitter. Another substantial extension of AAF are Bipolar AAF \cite{cayrol2009bipolar,cayrol2010coalitions} that include the notion of support as a primitive. Such a representation can be further extended by including other relations among arguments, such as noncommittal, in an arbitrary high number \cite{brewka2013abstract}. Recent developments in Natural Language Processing in particular allow the extraction of claims \cite{}, supporting evidence \cite{rinott2015show}, and other key components of argumentation with satisfactory precision. The authors of \cite{lippi2016argumentation} provide a detailed survey of the state of the art in argument mining to which we address the curious reader. However in this work we are not mining actual arguments out of our conversation, but rather exploiting AAF as a modelling framework to describe the flow of an online conversation in which several users interact, thus argumentation mining is out of our scope. \section{Conversation Modelling} \label{} In this section we present the modelling framework we designed to describe the interactions between users occurring within a conversation and the methodology we use to build such a model. In the first part of this section we will briefly introduce the basic concepts of AAF, then we will pinpoint the assumption we introduced to fit AAF modelling to our problem, and finally we will illustrate the graph construction rules we used to model Social Media conversation. \subsection{Abstract Argumentation Basics} As introduced in Section \ref{}, Abstract Argumentation Frameworks \cite{dung1993acceptability,dung1995acceptability} provide a simple, yet powerful, network representation of the structure of argumentation. Formally speaking, an argumentation framework is a pair $(AR, attacks)$ where $AR$ is the set of considered arguments, and $attacks$ is a binary relation on $AR$. Since the attack relationship is binary, a discourse can be represented as a directed graph. The notion of \emph{attack} is the key concept of an abstract argumentation framework: arguments are meant to attack each other and the structure of the network of attacks between arguments embeds the semantic information we aim to extract. The word attack in the context of AAF must not be misinterpreted since it does not imply aggressive behaviour, but rather represents an argument trying to weaken another one. To further expand the representation of the interaction between users, supporting arguments can be modelled as well. Intuitively, an argument can be considered supporting if it strengthens another one, either by bringing actual supporting evidence or by attacking its attacking arguments. The adoption of support edges as well allows to model a discourse as a multigraph. One of the main advantages of adopting AAF is their ability to introduce an abstraction layer over the contents of the discourse and the nature of the attacking and supporting arguments, hence, once built, AAF can be considered content agnostic which is coherent to our goal. \subsection{Considerations and Assumptions on Social Media Discourse} Abstract argumentation frameworks are typically used to detect \emph{conflict free} sets of arguments, i.e. arguments that do not attack each other and therefore are most likely to be accepted by the participants to the discussion\footnote{Being conflict-free however is a necessary but not sufficient condition for a set of arguments to be accepted: an argument set should also be admissible in that it defends all its members against attack. Moreover, there can be multiple and mutually conflicting admissible sets in the same discussion.}. However, our point is not finding which arguments are considered valid by the community, but rather which posts and authors attract the most sympathy and appreciation. To fit this goal, some assumptions must be relaxed and some others must be introduced. First and foremost, reasoning over AAF implies assuming that the involved agents are \emph{rational}, which in the context of social media platforms is a strong assumption since emotional factors are prominent and greatly affect the interaction between users \cite{kramer2014experimental}. Moreover, sympathy is not rational by definition, hence we will not consider our agents as rational and we will not perform actual reasoning on the constructed argument graphs. On the other hand, we assume that the discourse structure provided by AAF modelling embeds behavioural intentions and therefore patterns implying sympathy and appreciation can be learnt from it. The discourses hosted on Social Media platforms can arguably differ significantly from the ones conducted on more traditional media due to their peculiar characteristics and constraints. We can pinpoint four main characteristics which are relevant to extents of our modelling activity. \begin{itemize} \item \emph{Brevity}: posts on a social media platform are typically short\footnote{Some platforms, like Twitter, force users to produce short texts.}; this fact makes unlikely to attack and to defend the same user within a single post's space. \item \emph{No need of writing to express support}: typically Social Media platforms offer their users tools to express positive feedback (such as like, re-tweet, or similar options) that do not imply writing a reply. On the other hand, on most platforms, writing a comment is the only way to express criticism. \item \emph{Multiple posting}: it is not unusual for social media users to post sequences of consecutive messages and those self-replies are highly unlikely to be attacks. This behaviour can be seen as a consequence of the aforementioned typical brevity of social media posts. \item \emph{Referencing}: it is generally easy to identify to whom a post is replying since most social platforms provide reply indentation\footnote{Plaforms such as Reddit and Youtube indent posts according to whom they reply to.} or an explicit annotation\footnote{Platforms such as Twitter and Facebook provide a referral to the replied post or to its author}, hence there is no need to infer it from the textual content of a post. \end{itemize} Attack and support relationships are defined in our approach, which exploits AAF concepts, leveraging the above considerations. We define as an attack to a given post any replying message authored by a distinct user from the one that wrote the original message. We consider messages including explicit referrals to a post or an user to be direct replies to that post or the last post authored by the mentioned user. Intuitively, since Social Media provide dedicated tools to express positive feedback on a post, a reply is likely to include some degree of criticism. This definition of attack is clearly a simplification of reality, however we claim that this rough approximation allows to extract meaningful information to the extents of recognising positive reception of content. To enrich our modelling we are also considering supporting arguments. For the sake of simplicity, we choose to consider as supporting arguments to a given post the messages that attack the replies to the considered post. Since this notion of support is a simplified one and differs from its more established meaning, we will herein refer to it as \emph{defence}. \subsection{Graph Construction Rules} Building upon the considerations and assumptions illustrated in the previous paragraphs, we define two rules to construct a graph model from a conversation extracted from a social media platform: \begin{itemize} \item \emph{Attack rule}: a post $a$ is attacking another post $b$ whenever $a \neq b$, the author of $a$ is not the author of $b$, and $a$ is either a reply to $b$ or it contains an explicit referral to a user $u$ and $b$ is the last post authored by such a user at the time $a$ is posted. If $a$ is not a reply or does not contain any reference to a user, we assume that $a$ is attacking the previous post in the conversation authored by a different user than the one who wrote $a$. \item \emph{Defence rule}: a post $a$ is defending another post $c$ whenever there exist a post $b$ such that $a$ attacks $b$ and $b$ attacks $c$. \end{itemize} As a corollary of the above rules, our approach will not include self-attacking nodes, this choice is motivated by the fact that social media conversations tend to include sequences of posts authored by the same user which would be an obvious error to model as a series of self-attacks. Self-defencing edges, on the other hand are allowed since it appears reasonable that a user may support his or her own opinions. On the other hand, the same user cannot be both attacked and defended within a single post, since if $a$ attacks $b$ and defends $c$, $b$ and $c$ cannot be authored by the same user. \begin{figure}[ht] \centering \includegraphics[width=0.75\textwidth]{aaf_build.png} \caption{A fragment of an actual Reddit conversation and the graph it generates. Replies are modelled as attacks as long as the reply is not authored by the same user that wrote the original post, and defence edges are drawn between two posts a and c whenever there exist a post b such as a attacks b and b attacks c} \label{} \end{figure} In Figure \ref{} is presented an example of application of the above rules to a fragment of a conversation taken from Reddit\footnote{User names in the figure are purposely changed and some texts have been truncated to fit into the picture.}. The above described rules are used in this work to build automatically graphs from conversation crawled from Social Media. \section{Considered Features and Classification Algorithms} \label{} To identify the users who are likely to attract most positive feedback, we leverage the network structure provided by the graph obtained with our approach to extract information about the interactions between users and use such information to train a classifier. Different combinations of classifiers and graph features have been experimented to assess how effectively highly appreciated users can be identified by relying solely on the structure of reply chains between social media posts. In the first part of the section we are describing the considered graph and node features, while in the second part we are briefly introducing the various classification techniques tested. \subsection{Considered Features} \label{} For each user involved in a conversation, we extracted a set of features representing his o her involvement and the reactions his or her posts generated. We considered a broad set of features, shown in Table \ref{}, to model several different aspects of the obtained graphs. We included among the considered features some basic distributional information, such as the number of posts authored by the considered user within the considered conversation, called \emph{Post Count}. We can normalise the Post Count with respect to the total number of posts occurring in the considered conversation, obtaining the \emph{Conversation Coverage} feature. Moving onto the graph structure, we outline four features describing a user's observed activity: \emph{Received Attacks} defined as the cumulative number of attacks received by the posts of the user, \emph{Performed Attacks} defined as the cumulative number of attacks performed by the user through his or her posts, \emph{Received Defences} defined as the cumulative number of defences received by the posts of the user, and \emph{Performed Defences} defined as the cumulative number of defences performed by the user through his or her posts. The above four features can be averaged over the Post Count, resulting in another four features: \emph{Average Received Attacks}, \emph{Average Performed Attacks}, \emph{Average Received Defences}, and \emph{Average Performed Defences}. Further features can be obtained by combining the above ones to quantify some user behavioural aspects. In particular we focus on three possible combinations we call \emph{Aggressiveness}, \emph{Community Disapproval Ratio}, and \emph{Engagement}. We define the Aggressiveness as the Performed Attacks divided by Performed Defences; such a value is greater than 1 when the attacks performed by a user outnumber his or her defences, otherwise it ranges between 0 and 1. We define the Community Disapproval Ratio as the Received Attacks divided by Received Defences; such value represents the proportion between attacking and supporting arguments to the user's post throughout the conversation and therefore can be considered a measure of how much his or her contributions were disputed, with a high value representing a majority of attacks, and a value close to zero a majority of supporting arguments. Finally, we define the Engagement as the sum of the four edge-based features, assuming that users who both receive and perform a large number of attack and defences can be regarded as highly involved in the conversation. The Engagement feature, being a count over the edges in the graph, can be normalised over the total number of such edges, resulting in the \emph{Normalised Engagement}; this feature represents the fraction of edges in the graph that sprout from the user's posts or enter in it. We can wrap up the information provided by the Engagement and the number of post authored by the user into an \emph{Activity Score} which is defined as the product of Engagement and Conversation Coverage. This feature rewards users who post a lot and participate a lot by means of both attacks and defences, while punishing users who detain few hi-impact posts or too many low-impact ones. Finally, we included features intended to describe how central in the graph topology are the posts authored by a user. We use three centrality measures: between centrality, which is, given a node, the number of shortest paths in the graph between two nodes that pass through it, Eigenvector centrality, which represents the number of connections of a node to other highly central nodes in the graph\footnote{Google's famous PageRank algorithm is a variant of Eigenvector centrality}, and Closeness centrality, which is the sum of the lengths of the shortest paths between the considered nodes and all the other nodes in the graph. Centrality is a key concept of graph theory and can be used to estimate the overall relevance of a node in a network, thus it can embed information relevant to our purpose. To represent this information, we introduce the last three features: \emph{Cumulative Between Centrality}, \emph{Cumulative Eigenvector Centrality}, and \emph{Cumulative Closeness Centrality}. Those features consist in the sum of the centralities in the network of all the posts authored by the considered user. \begin{table} \resizebox{\textwidth}{!}{ \begin{centering} \begin{tabular}{@{}llr@{}} \toprule \textbf{Feature Name} & \textbf{Shorthand} & \textbf{Explanation} \\ \midrule Post Count & PC & Number of posts authored by the user \\ Conversation Coverage & CC & PC/ posts in the conversation \\ \midrule Received Attacks & Att-IN & attack edges entering in the posts authored by the user \\ Performed Attacks & Att-OUT & attack edges originating from the posts authored by the user \\ Received Defences & Def-IN & defence edges entering in the posts authored by the user \\ Performed Defences & Def-OUT & defence edges originating from the posts authored by the user \\ Average Performed Attacks & AvgAtt-OUT & Att-OUT/PC \\ Average Received Attacks & AvgAtt-IN & Att-IN/PC \\ Average Performed Defences & AvgDef-OUT & Def-OUT/PC \\ Average Received Defences & AvgDef-IN & Def-IN/PC \\ Aggressiveness & Agr & Att-OUT/Def-OUT \\ Community Disapproval Ratio & Dis & Att-IN/Def-IN \\ Engagement & En & Att-IN + Att-OUT + Def-IN + Def-OUT \\ Normalised Engagement & NEn & En/ cumulative En of all users in the conversation \\ Activity score & As & En * CC \\ Normalised Activity Score & NAs & As / max As in the conversation \\ \midrule Cumulative Betweenness Centrality & CBC & Cumulative Betweenness centrality of all posts authored by the user \\ Cumulative Eigenvector Centrality & CEC & Cumulative Eigenvector centrality of all posts authored by the user \\ Cumulative Closeness Centrality & CClC & Cumulative Closeness centrality of all posts authored by the user \end{tabular} \end{centering} } \label{} \caption{Considered user features.} \end{table} \subsection{Considered Classifiers} \label{} Several classifiers have been considered to tackle the problem of identifying users whose content attract positive feedback. They are: Naive Bayes classifiers, Conditional Inference Trees, Random Forests, and Support Vector Machines. A Naive Bayes classifier is a probabilistic classifier based on Bayes' theorem that leverages the strong independence assumption between the considered features. Such classifiers, albeit simple, are robust and cope well with hi-dimensional problems. Moreover there is evidence in the literature that Naive Bayes classifiers perform well in classification tasks similar to the one presented in this work \cite{arnt2003learning}. A Conditional Inference Tree is a classifier based upon rule induction, similar to a decision tree, built performing recursively univariate splits of the class value based on the values of a set of features. The splits in decision trees are learnt leveraging information metrics, such as the information entropy, while in a conditional inference tree they are chosen according to permutation-based significance tests, thus the selection bias of picking variables that have multiple splits is mitigated, reducing the risk of overfitting the training data \cite{hothorn2006unbiased}. Random Forests are an ensemble learning technique consisting in building at training time several decision trees, with each of them based on a random subset of the training features, and then aggregating the various predictions to generate the final one \cite{breiman2001random}. This method mitigates the problem of training data overfitting to which decision trees are particularly prone. Finally, Support Vector Machines (SVM) are the most established Kernel method algorithm, commonly used to solve a large number of classification and regression problems. While they can be considered as linear classifiers, their usage of kernel functions allows to map the training data points into a high dimensional space wherein the classification problem is linear, thus allowing non-linear classification. While very powerful, Support Vector Machines can suffer greatly biases induced by uncalibrated class membership probability in the training data. In this study we are considering three variants of SVM, differing in the employed kernel function: the \emph{linear} SVM uses a linear kernel function which is the most common SVM setup for NLP tasks, the \emph{polynomial} SVM uses a polynomial kernel function of degree $d=2$, finally the \emph{radial} SVM uses Radial Basis Function kernel as kernel function, which provides in principle a low-band pass filter, selecting out smooth solutions. \section{Data analysis and Evaluation} \label{} In this section we describe the evaluation work performed to assess the effectiveness of our approach which exploits abstract argumentation and machine learning in detecting users whose contributions are most appreciated by the community. More specifically, we will describe the considered data source, the gathered data sets, the importance of the considered features, and the performance of the considered classifiers on the gathered data sets. \subsection{Data Preparation} \label{} To get meaningful real world data to perform both training and validation, we crawled conversations from Reddit\footnote{}. Reddit is an entertainment, social news networking, and news Web site whose community members can post messages or link to other Web pages, and reply to posts of other users. Each contribution can be voted by the community members with upvotes, expressing positive feedback, and downvotes, expressive negative feedback. At the time of this writing, Reddit is among the 30 most viewed sites of the Web, receives over 1 billion visits each month\footnote{}, and in 2015 it saw 725.85 million submissions from 8.7 million total authors that received 6.89 billion upvotes\footnote{}. One of the most notable characteristics of Reddit is the fact that it provides scoring mechanisms for both posts and users based on the upvotes and downvotes provided by the community over time. In particular we are interested in the score associated to the single posts which is determined by the difference between upvotes and downvotes and can be seen as a straightforward approval rating of the considered content. On the other hand, the user score, called \emph{link karma}, is known for taking into account more than the explicit community feedback, but since its evaluation procedure is not clearly described by Reddit itself, it cannot be considered a reliable rating for our purpose. Therefore we are estimating the general approval of a user's content by considering the cumulative scores of his or her posts (herein we will refer to such a score as \emph{cumulative approval}), without taking into account the link karma. It is important to point out how the score of a single post, being the difference between positive and negative community votes, may be a negative number, therefore a user whose post receive mixed feedback is expected to have a near zero cumulative approval. We crawled the 125 longest conversations available on the social network, gathering posts from over 70,000 users. Graphs were constructed for each conversation according to the methodology described in Section \ref{}, and for each user we evaluated the features described in Section \ref{}. We observed the distribution of users with respect to their cumulative approval and we found that it respected the assumptions made in Section \ref{}, with the vast majority of users concentrated around mildly positive values, and a few scoring extremely low or extremely high. Considering the way the cumulative approval is computed and the shape of the distribution we can safely assume that the 5\% highest-scoring users can be regarded as the ones who produced the most appreciated content, herein we will refer to them as \emph{top users}. The drawback of this choice is that it implies a strong unbalance between the considered classes, therefore to prevent statistical bias in the training of the classifiers we sampled the gathered data to obtain a set of 7,000 users equally split into top users and other users. This balanced set, being a huge simplification of reality, is meant for training and preliminary evaluation purposes and we will refer to it in the rest of the paper as the \emph{Evaluation Set}. To achieve a better representation of a real world scenario we crawled 100 more conversations picked among the longest ones hosted on the platform to create a new data set (herein called the Validation Set) unbiased and with a realistic distribution of users. These conversations, like the ones in the previous set, were processed according to the methodology described in Section \ref{} and \ref{}, then the 5\% top users were flagged according to their cumulative approval. The produced validation set contains over 30,000 users, with 1622 of them flagged as top users. \subsection{Exploratory Data Analysis} \label{} As a first step of our evaluation, we assessed which among the considered features are to be considered the most informative. This analysis has two goals: determining if graph based features are actually informative and finding a minimal dimensionality for this problem, thus reducing the risk of the classifiers overfitting the data. To determine whether or not the usage of graph derived features could benefit the identification of highly appreciated users, we considered three among the classifiers described in Section \ref{}: Conditional Inference Tree, Random Forest, and Support Vector Machine with linear kernel. We chose these three because each of them represents a class of classifiers: CIT is an explicit decision model, RF is and ensemble learning strategy, and SVM is a kernel method. These classifiers were trained and tested by means of cross validation upon the Evaluation Set with three different feature sets: a set including no graph derived features, a set including also attack-derived features, and a set including all the 19 features described in Section \ref{}, thus including defence-derived features. We assume that if a statistically significant difference is observable in all these three classifiers, the usage of graph features is to be considered significant as well. Classifiers trained on the first set achieved the worst performance and classifiers trained on the full 19 ones the best performance; detailed results are shown in Table~\ref{}. As expected, graph based features provided a considerable improvement, and in particular, the defence edges played an important role to further increase the results. \begin{table}[] \begin{tabular}{@{}llll@{}} \toprule Classifier & Conditional Inference Tree & Random Forest & SVM \\ \midrule No graph & 0.5376 & 0.5648 & 0.5062 \\ Attacks & 0.8324 & 0.8224 & 0.8257 \\ Attacks and Defences & 0.8505 & 0.8405 & 0.8338 \\ \bottomrule \end{tabular} \label{} \caption{Classification accuracy with no graph derived features, with attack derived features, and with all considered features.} \end{table} To identify the most informative features, an extensive exploratory data analysis was performed. Four techniques were considered: Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Learning Vector Quantization (LVQ), and Recursive Feature Elimination (RFE). PCA is a well-known technique to estimate a number of linearly independent components that represent most of the observed variability in the data, thus providing a better understanding of the minimum dimensionality that could be sufficient to cover the maximum variance. LDA is another well known technique that takes into account information about objects classes, allowing us investigate which are the directions (called "linear discriminants") that represent best the axis that maximise the separation between the ``top users'' and the rest of the user base. Finally, LVQ and RFE leverage a predictive model, such as a Random Forest, to estimate how relevant the considered features are. More precisely, LVQ estimates features importance using a ROC curve analysis conducted for each feature, and RFE generates several sub-sets of the feature set, then it trains the model on such sub-sets, and evaluates the model's performance to identify the most significant among the generated feature sub-sets. For the purposes of these analysis the first gathered data set (the one including 125 conversations) was considered. Without further venturing into the details of the above described analysis, the insights gathered by means of PCA, LDA, and LVQ were coherent and highlighted the following features as the most informative of the considered set: Average Received Attacks (AvgAtt-IN), Average Received Defences (AvgDef-IN), Engagement (En), and Cumulative Betweenness Centrality (CBC). Notably, the Engagement value appeared to be consistently the strongest predictor throughout the performed feature importance analysis. We will refer to these four features together as the \emph{minimal} feature set. The RFE approach, instead suggested that the best performance on the considered data set could be achieved with a slightly broader feature set, including Received Defences (Def-IN), Conversation Coverage (CC), Average Performed Attacks (AvgAtt-OUT), Average Received Attacks (AvgAtt-IN), Average Received Defences (AvgDef-IN), Community Disapproval Ratio (Dis), and Engagement (En). We will refer to these seven features as the \emph{reduced} feature set, in contrast with the \emph{full} feature set including all the 19 considered features. \subsection{Classification Experiments} \label{} We ran two experiments on the two gathered data sets to find the most suitable combination of feature set and classification model. The first experiment consisted in performing a cross-fold evaluation over the Evaluation Set with all the classifiers presented in Section \ref{} using the full feature set. The results of this evaluation are shown in Table \ref{} where, for each considered classifier, its precision ranges are shown. Each range has a lower and an upper bound that enclose all the observed values with 0.95 probability, and the average observed value. \begin{table}[] \resizebox{\textwidth}{!}{ \centering \begin{tabular}{@{}lllllll@{}} \toprule Accuracy & Naive Bayes & Random Forest & Conditional Inference Tree & SVM Linear & SVM Poly & SVM Radial \\ \midrule lower bound & 64.36\% & 83.90\% & 83.45\% & 82.62\% & 81.58\% & 79.74\% \\ average & 66.43\% & 84.05\% & 85.05\% & 83.38\% & 83.52\% & 82.19\% \\ upper bound & 68.45\% & 84.10\% & 86.55\% & 84.87\% & 84.94\% & 82.85\% \\ \bottomrule \end{tabular} } \label{} \caption{Considered classifiers accuracy range with full feature set measured upon cross-validation.} \end{table} \begin{table} \resizebox{\textwidth}{!}{ \centering \begin{tabular}{@{}llllll@{}} \toprule Metric & Random Forest & Conditional Inference Tree & SVM Linear & SVM Poly & SVM Radial \\ \midrule Precision & 0.8762 & 0.7867 & 0.8943 & 0.8657 & 0.8838 \\ Recall & 0.8156 & 0.9017 & 0.7877 & 0.8160 & 0.7864 \\ F1 & 0.8448 & 0.8403 & 0.8433 & 0.8401 & 0.8323 \\ \bottomrule \end{tabular} } \label{} \caption{Precision, Recall and F1 score of the considered classifiers measure upon cross-validation.} \end{table} It can be easily noticed how the Naive Bayes classifier achieves a significantly lower precision (between 64.36\% and 68.45\%) than the other considered models that are relatively tied with scores between 80\% and 86\% precision. This observation led us to the withdrawal of the Naive Bayes classifier from the considered classifiers pool. Moreover, since what we are really interested in the retrieval of appreciated users, we considered some Information Retrieval metrics, namely Precision, Recall, and their harmonic mean the F1 score, all evaluated on the ``top user'' class. The values of such metrics achieved by the five best-performing classifiers is shown in Table \ref{}. All classifiers were trained with the full feature set, and Random Forests and Conditional Inference Trees appear to achieve the most promising results, performing consistently better than SVM in various configurations. These results can be improved up to 2\% average classification accuracy by adopting the reduced feature set instead of the full one, however, being such a feature set selected with RFE on this very same data (see Section \ref{}) such an improvement is expected and might be the outcome of some data overfitting. The second experiment consisted in training the classifiers on the whole Evaluation Set and then run them over the Validation Set, which presents a more realistic distribution of users than the Evaluation Set. The evaluation was performed over the three chosen feature sets to identify the one that describes best the problem with respect to a possible field usage. The results of this second evaluation are presented in Table~\ref{}, where the observed precision values achieved by each combination of feature set and classifier is shown. However, due to the strong unbalance of classes in the Validation Set, with 94.9\% of users not being top users, classification accuracy alone is little informative. To cope with this characteristic of the domain, the aforementioned Precision, Recall, and F1 score, evaluated on the ``top user'' class, are more informative and their values are shown in Table \ref{}, \ref{}, and \ref{}. \begin{table}{\resizebox{\textwidth}{!}{ \centering \begin{tabular}{@{}llllll@{}} \toprule Feature Set & Random Forest & Conditional Inference Tree & SVM Linear & SVM Poly & SVM Radial \\ \midrule minimal & 90.24\% & \textbf{90.96\%} & 87.68\% & 88.93\% & 88.00\% \\ reduced & 86.84\% & 90.75\% & 88.04\% & 75.04\% & 80.33\% \\ full & 83.95\% & 90.85\% & 79.78\% & 81.01\% & 77.66\% \\ \bottomrule \end{tabular} } } \caption{Considered classifiers average accuracy on Validation Set.} \label{} \end{table} \begin{table}{\resizebox{\textwidth}{!}{ \centering \begin{tabular}{@{}llllll@{}} \toprule Metric & Random Forest & Conditional Inference Tree & SVM Linear & SVM Poly & SVM Radial \\ \midrule Precision & 0.3155 & 0.3332 & 0.2651 & 0.2865 & 0.2696 \\ Recall & 0.8021 & 0.7947 & 0.8163 & 0.8046 & 0.8095 \\ F1 & 0.4529 & 0.4695 & 0.4002 & 0.4225 & 0.4045 \\ \bottomrule \end{tabular} } } \caption{Precision, Recall and F1 score of the considered classifiers with the minimal set of features.} \label{} \end{table} \begin{table}{\resizebox{\textwidth}{!}{ \centering \begin{tabular}{@{}llllll@{}} \toprule Metric & Random Forest & Conditional Inference Tree & SVM Linear & SVM Poly & SVM Radial \\ \midrule Precision & 0.2557 & 0.3290 & 0.2711 & 0.1587 & 0.1871 \\ Recall & 0.8452 & 0.8058 & 0.8144 & 0.9199 & 0.8693 \\ F1 & 0.3927 & 0.4673 & 0.4068 & 0.2706 & 0.3079 \\ \bottomrule \end{tabular} } } \caption{Precision, Recall and F1 score of the considered classifiers with the reduced set of features.} \label{} \end{table} \begin{table}{\resizebox{\textwidth}{!}{ \centering \begin{tabular}{@{}llllll@{}} \toprule Metric & Random Forest & Conditional Inference Tree & SVM Linear & SVM Poly & SVM Radial \\ \midrule Precision & 0.2231 & 0.3318 & 0.1872 & 0.1961 & 0.1701 \\ Recall & 0.8816 & 0.8064 & 0.9026 & 0.8940 & 0.8859 \\ F1 & 0.3561 & 0.4702 & 0.3101 & 0.3216 & 0.2853 \\ \bottomrule \end{tabular} } } \caption{Precision, Recall and F1 score of the considered classifiers with the full set of features.} \label{} \end{table} \section{Discussion} \label{} The insights provided by our evaluation activities described in Section \ref{} appear to substantially uphold the claim introduced in Section~\ref{}, highlighting how the structure of the Social Media conversations embed information about the posts appreciation level, and hence posts' author reputation. First of all, the distribution of the Reddit score, as noted in Section~\ref{} and according to what we expected, our assumption that only a small fraction of the user base is able to consistently generate well received content. With respect to such a distribution our choice of considering ``top users'' the highest scoring 5\% of the user base seems reasonable, given that such fraction is likely to include the whole long tail of highly appreciated users. Before discussing the outcomes of the evaluation it is important to stress how the Reddit score, though being strongly correlated with the number of posts, represents a rather qualitative dimension of a user's activity in the community. As a matter of fact users who post a lot tend to accumulate more Reddit Score than users who seldom post, however, due to the possibility of expressing negative feedback, the correlation between the post count and the RS does not go beyond 0.7, a significant, but not outstanding value. This is particularly striking when comparing Reddit to other social networks such as Facebook wherein the amount of ``likes'' received by a user appears to be close to being a function of the number of posts authored~\cite{bessi2015everyday} with little regard for the actual content posted. Considering, as we did, the average Reddit Score, i.e. the average proportion of upvotes and downvotes received by a user's post, moves even forward the qualitative aspect of this score. The usage of AAF to model conversations, despite the numerous simplifying modelling assumptions introduced in Section~\ref{}, appears to be instrumental to the extents of our task. The results shown in Table~\ref{} suggest that without considering information derived from our graph-based modelling, the precision of determining whether or not a user should be considered a ``top'' one is extremely low and dangerously close to a ``coin flip'' regardless of the employed classifier. Attacks edges alone can provide significant information, but defences are informative as well and their addition appears to benefit the overall accuracy disregarding the employed classifier. Since graph derived features appear to be informative, we tried to understand the performance limits of this technique by experimenting on different data sets with different combinations of feature sets and classifiers. The first experiment described in \ref{} consisted in a cross evaluation over a balanced dataset, which means a very controlled scenario meant to avoid introducing statistical bias into the considered classifiers. Moreover, the considered dataset was built crawling Reddit's most popular discussions, which means extremely large ones. Keeping this in mind, the results gathered are nevertheless very encouraging with a high classification accuracy (peaking between 83.45\% and 86.55\%) and with five classifiers tied around a 0.84 F1 score on the ``top user'' class. These results allow us to assume that in such a scenario our approach could achieve satisfying performance even employing different classifiers. Reality, however, proved to not be so bright when the classifiers trained on the full Evaluation set were benchmarked over the more various Validation Set. The Validation Set most notable features are: inclusion of smaller and less popular discussions, way larger number of considered users, and a realistic proportion between highly appreciated users and mildly or not appreciated ones. In this scenario our approach achieved an higher performance, scoring at best 90.96\% accuracy when considering only the minimal set of features which appears to be less prone to overfitting. On the other hand, the thin numbers of highly appreciated users imply that a naive classifier opting always for the negative class would score an extremely high accuracy, scoring 94.9\%. However, in such an unbalanced scenario, classification accuracy alone cannot be considered truly informative, especially when the goal is to retrieve only the members of one class. Precision, Recall and F1 score, then, provide a better picture of the situation. The last one, in particular, allows us to assess the robustness of our classifiers being the harmonic mean of Precision and Recall; with respect to this metric, a naive classifier returning always the negative class would score zero, because its recall would be zero. Another naive classifier returning the positive class (top user) would score instead 0.09, having a 100\% recall, but a very low precision. The five classifiers that in the previous experiment were tied in terms of F1 score, in this second trial appear to achieve significantly different results. SVMs in particular seem to perform consistently worse than Conditional Inference Trees and Random Forests. All classifiers, except for Conditional Inference Trees, achieve a better performance on the minimal feature set, providing evidence that the larger feature sets give in to overfitting. With respect to the results displayed in Table \ref{}, the ones in Table \ref{}, \ref{}, and \ref{} show a significantly lower Precision, but retain a similar Recall, implying a very thin number of false negative ``top users''. Overall, Conditional Inference Trees appear to be the most robust kind of classifier for this task, consistently achieving best or near-best scores and showing little performance variation over different feature sets. As a final note, the percentage of users flagged as ``top users'' by the five considered classifiers ranged between 12\% and 9\%, implying that, though being somewhat ``optimistic'' the considered classifiers come rather close to the actual split in the data. \section{Conclusions} \label{} In this work we presented a novel approach towards the detection of users who create, from their community's point of view, quality content. With respect to related work which focus on the detection of malicious or unappreciated behaviour, we shifted towards the ``best'' part of the user base that recent literature suggests to be more influential and therefore interesting for trend detection tasks and thus for opinion mining and marketing purpose. With respect to this problem, our main contribution is in the content agnostic nature of the proposed approach, which by solely observing the flow of a discussion, i.e. the sequence of posts and replies, can identify users who are likely to attract the support of their online community. This approach, in fact, allowed us to generate reasonably small sets of potentially highly appreciated users that included up to 90\% of the actual ``top users''. Unlike widespread content-based Sentiment Analysis techniques that leverage the content of the posts to detect positive attitudes, our approach abstracts over several non trivial problems such as irony detection by looking at the whole structure of the discourse. With respect to argument mining, our approach does not investigate the content of messages included in the conversation, but rather their sequence and the order in which users reply to each other, suggesting that this sole information can embed insights about users' attitude. The results gathered so far suggest that tackling the problem at discourse level allows to significantly restrict the search space for this problem, retaining a very high recall. Our future work will be aimed at further exploring the possibilities of applying AAF to social media mining. In particular, we will investigate the usage of the approach described in this paper as a preliminary step in more complex analysis, such as the detection of opinion leaders and other kinds of social influencers. We will also experiment extensions to the approach herein described, with different definitions of defences and attacks. Our discourse-based approach will also be coupled with more traditional message content-based Sentiment Analysis techniques to overcome the limitations of both techniques and to further investigate the relationships among social media users. Finally, we believe that this work suggests a deeper connection between formal Artificial Intelligence techniques and statistical ones. More specifically, we believe that the former ones could be used to achieve a better modelling of the data, thus extracting more abstract, meaningful, and understandable features that can substantially help describing complex phenomena like social media interaction. \bibliographystyle{elsarticle-harv}

\section{Introduction} The observations of gravitational waves emitted by the coalescence of binary black holes by the LIGO-Virgo collaboration \cite{}, together with the recent joint detection of the coalsecence of a binary neutron star system with electromagnetic counterparts \cite{}, have opened a new era in gravitational wave astronomy. In particular, these detections provide the opportunity to challenge general relativity (GR), as well as modified gravities (i.e., gravity theories beyond GR), in the strong field regime of the coalescence of two compact objects, in diversified configurations, and at various redshifts.\\ In GR, post-Newtonian (PN) expansions of Einstein's equations (in the small orbital velocity, weak field limit) are suitable to describe analytically the inspiral phase of binary systems and associated gravitational waveforms. They often rely on the very convenient ``skeletonization" of compact bodies, that is, on their phenomenological reduction to effective point particles, that follow the geodesics of the metric they produce (although necessitating regularization schemes due to the artificial introduction of Dirac distributions in Einstein's equations), see, e.g., \cite{lesHouches} and \cite{ The generalization of skeletonization to modified gravities was introduced by Eardley in the simplest case of (massless) scalar-tensor theories in \cite{Eardley}. There, he proposed to endow each point particle with and effective ``mass function", $m_A(\varphi)$, that depends on the (regularized) value of the scalar field evaluated at its location, which in turn identifies to the scalar environment of the skeletonized body. The resultant simplification of the formal treatment of the two-body problem enabled to address it up to 2.5 PN order \cite{, , }, or, adopting the terminology of \cite{}, 2.5 post-Keplerian (PK) order, to highlight the fact that (strong) self-gravity effects are encompassed in the mass function $m_A(\varphi)$. The explicit computation of $m_A(\varphi)$ for a given body is well-known in the case of neutron stars. It is based on the (numerical) integration of the field equations inside the star, depending on its internal structure (e.g., its equation of state), up to the exterior of the star, where the fields are matched to the near-worldine fields sourced by the effective point particle, see, e.g., \cite{ , , }; see also \cite{}. On the other hand, in that class of theories, its is known that black holes cannot carry scalar ``hair", and hence, are reduced to Schwarzschild's \cite{ In the present paper, we propose to go beyond what was done so far in scalar-tensor theories (ST), and consider the class of Einstein-Maxwell-dilaton theories (EMD). They consist in supplementing ST theories with a (massless) vector gauge field that is non-minimally coupled to the scalar field. Note that this vector does not necessarily correspond to the Maxwell field of electrodynamics, but must rather be considered as a new ``graviphotonic" degree of freedom of gravity. These EMD theories will be at the center of our attention, since, contrarily to ST theories, they predict the existence of black holes that differ from the general relativistic ones through the presence of vector and induced scalar ``hairs" \cite{, , , }. Our aim will therefore be to address the problem of motion of two compact bodies, including these ``hairy" black holes, by generalizing the Eardley-type point particle action to EMD theories.\\ The dynamics of binary black holes in EMD theories has been studied numerically in \cite{}. There, the authors found that for very small values of their scalar cosmological environment $\varphi_0$, the influence of the scalar interaction can be largely neglected in understanding their dynamics, which is therefore hardly distinguishable from its general relativistic counterpart. Now, although these conclusions will coincide with the (analytical) results obtained in the present paper, we shall hint that, when $\varphi_0$ is increased (which, as we shall argue, is physically reasonnable), black holes can undergo steep ``scalarization" transitions, i.e. become strongly coupled to the scalar and vector fields, inducing large deviations to the general relativistic two-body dynamics.\\ The paper is organized as follows: in section \ref{sectionLagrangian} we introduce EMD theories and build an appropriate point particle action, endowing them with a mass function $m_A(\varphi)$ (``\`a la" Eardley) and a constant charge $q_A$. We then compute the resultant two-body Lagrangian at post-Keplerian order (1PK), which describes the conservative part of the dynamics of arbitrary compact bodies in interaction. In section \ref{sectionSkeletonization}, we present a class of EMD black hole solutions with an ``electric" charge and dilatonic ``hair". By means of an appropriate matching condition, we then show how to skeletonize these black holes, that is, how to compute explicitely the corresponding mass function and charge. Finally, in section \ref{section_Sensitivities}, we specify the values of the post-Keplerian couplings that enter the two-body Lagrangian for the black holes mentioned above, and highlight that, depending on their cosmological scalar environment $\varphi_0$, black holes can transition between two extremal regimes: a ``Schwarzschild"-like behaviour, which makes them undistinguishable from their general relativistic counterparts, and a ``scalarized" regime, where their couplings to the scalar and vector fields reach the order of unity, inducing large deviations from general relativity. \section{The two-body problem in Einstein-Maxwell-dilaton theories\label{sectionLagrangian}} \subsection{The Einstein-Maxwell-dilaton skeleton action\label{subsectionSkeleton}} The Einstein-Maxwell-dilaton (EMD) theories describe general relativity supplemented by a massless scalar field and a massless vector gauge field. In vacuum, the EMD action is conveniently written in the Einstein frame as (setting $G\equiv c\equiv 1$) \begin{equation} S_{\rm vac}[g_{\mu\nu},A_\mu,\varphi]=\frac{1}{16\pi}\int d^4x\sqrt{-g}\,\bigg(R-2g^{\mu\nu}\partial_\mu\varphi\partial_\nu\varphi-e^{-2a\varphi}F^{\mu\nu}F_{\mu\nu}\bigg)\ ,\label{EMDvacuumAction} \end{equation} where $R$ is the Ricci scalar, $g=det\,g_{\mu\nu}$, and where $F_{\mu\nu}=\partial_\mu A_\nu-\partial_\nu A_\mu$. \\ The fundamental parameter $a$ induces a coupling between the vector and scalar fields.\footnote{Note that $a=0$ reduces to Einstein-Maxwell theory minimally coupled to a scalar field, while $a\neq 0$ is motivated by low-energy limits of string theory, see, e.g., \cite{}; $a=\sqrt{3}$ is equivalent to the dimensional reduction of the Kaluza-Klein theory \cite{} \cite{} In consequence, vacuum EMD theories encompass the usual $U(1)$ gauge symmetry, $A_\mu\rightarrow A_\mu+\partial_\mu\chi$, $\chi$ being an arbitrary scalar function, while breaking the scalar ``shift symmetry", $\varphi\rightarrow\varphi+cst$. As we shall see in the following, this will be crucial in endowing EMD black holes, that is, vacuum solutions of the theory, with scalar hair.\\ \\ In this paper, we shall restrict ourselves to $a>0$ only, since (\ref{EMDvacuumAction}) is symmetric under the simultaneous redefinitions $a\rightarrow -a$ and $\varphi\rightarrow -\varphi$.\\ The vacuum field equations follow from the variation of (\ref{EMDvacuumAction}), \begin{subequations} \begin{align} &R_{\mu\nu}=2\partial_\mu\varphi\partial_\nu\varphi+2e^{-2a\varphi}\left(F_{\mu\alpha}F_\nu^{\ \alpha} -\frac{1}{4}g_{\mu\nu}F^2\right)\ ,\\ &\nabla_\nu\left(e^{-2a\varphi}F^{\mu\nu}\right)=0\ ,\label{maxwellVacuum}\\ &\Box\,\varphi=-\frac{a}{2}e^{-2a\varphi}F^2\ , \end{align}\label{vacuumEqs} \end{subequations} where $F^2\equiv F_{\mu\nu}F^{\mu\nu}$, where $\nabla_\mu$ denotes the covariant derivative associated to $g_{\mu\nu}$ and where $\Box\equiv\nabla_\mu\nabla^\mu$.\\ In presence of matter, the Einstein-frame action becomes \begin{equation} S=S_{\rm vac}+S_{\rm m}[\Psi,\mathcal A^2(\varphi)g_{\mu\nu},A_\mu]\ ,\label{actionEMD} \end{equation} where $\Psi$ generically denotes matter fields, that are minimally coupled to the gauge vector $A_\mu$ and to the Jordan metric, $\tilde g_{\mu\nu}=\mathcal A^2(\varphi)g_{\mu\nu}$, $\mathcal A(\varphi)$ being a scalar function of $\varphi$ that specifies the theory, together with $a$.\\ \\ However, our aim being to address the problem of motion of two compact bodies in interaction, it will prove convenient to ``skeletonize" them, that is to replace $S_{\rm m}$ in (\ref{actionEMD}) by a phenomenological point particle action, $S_{\rm m}^{\rm pp}$. Now, extending the treatment of Eardley in scalar-tensor theories \cite{Eardley} to incorporate the presence of the vector field, we shall consider the most generic covariant ansatz \begin{align} S_{\rm m}^{\rm pp}[g_{\mu\nu},A_\mu,\varphi, \{x_A^\mu\}]=-\sum_A\int\! m_A(\varphi)\,ds_A+\sum_A q_A\int\! A_\mu\, dx^\mu_A\ ,\label{ansatz} \end{align} where $ds_A=\sqrt{-g_{\mu\nu} dx_A^\mu dx_A^\nu}$, $x_A^\mu[s_A]$ being the worldline of particle $A$, and where $m_A(\varphi)$ is a scalar function to be related to the skeletonized body $A$ later, which depends on the value of the scalar field evaluated at its location $x_A^\mu(s_A)$ (substracting divergent self contributions). In constrast, $q_A$ will be considered as a constant parameter. Indeed, as one easily sees from (\ref{ansatz}), the linear coupling of the worldlines to the vector field $A_\mu$ preserves the $U(1)$ gauge symmetry provided that the following current, $j^\mu$, is conserved: \begin{equation} \partial_\mu j^\mu=0\ ,\qquad \text{where}\qquad j^\mu(y)=\sum_A q_A\,\delta^{(3)}(\vec y-\vec{x}_A(t))\frac{dx_A^\mu}{dt}\ . \end{equation} This in turn implies, as usual, that the charge $q_A$ of each body is conserved (provided that it remains separated from its companion), and cannot depend upon the variation of $\varphi(\vec x_A(t))$.\\ Note that our ansatz (\ref{ansatz}) does not depend on the gradients of the fields and, therefore, cannot take into account any finite size (e.g., tidal) effect, that we will neglect in the present paper; see, e.g., \cite{}. The functions $m_A(\varphi)$ must therefore be understood as depending on the homogeneous, adiabatically varying, scalar field environment of the body $A$, created by the faraway and slowly orbiting companion.\\ Finally, in the weakly self-gravitating limit, the bodies follow the geodesics of the Jordan metric, see (\ref{actionEMD}) and below. Their mass functions are hence reduced, in that limit, to $m_A(\varphi)=\tilde m_A \mathcal A(\varphi)$, where $\tilde m_A=cst$ is the Jordan-frame mass. In contrast, general relavitity is recovered when $m_A(\varphi)=cst$, together with $q_A=0$. \subsection{The post-Keplerian Lagrangian\label{sectionLagrangian1PK}} We now have in hands the necessary material to address the dynamics of compact binary systems in EMD theories: our starting point is the ``skeleton" action \begin{align} S[g_{\mu\nu},A_\mu,\varphi,\{x_A^\mu\}]&=\frac{1}{16\pi}\int d^4x \sqrt{-g}\bigg(\! R-2g^{\mu\nu}\partial_\mu\varphi\partial_\nu\varphi-e^{-2a\varphi}F^2\bigg)-\sum_A\int m_A(\varphi)\,ds_A+\sum_A q_A\int A_\mu\, dx^\mu_A\ , \end{align} with $ds_A=\sqrt{-g_{\mu\nu} dx_A^\mu dx_A^\nu}$, that depends on the fondamental parameter $a$, together with two mass functions $m_A(\varphi)$ and two constant charges $q_A$. The field equations are now sourced by the effective particles: \begin{subequations} \begin{align} &R_{\mu\nu}=2\partial_\mu\varphi\partial_\nu\varphi+2e^{-2a\varphi}\left(F_{\mu\alpha}F_\nu^{\ \alpha} -\frac{1}{4}g_{\mu\nu}F^2\right)+8\pi\sum_A\left(T^A_{\mu\nu}-\frac{1}{2}g_{\mu\nu}T^A\right)\ ,\label{twoBody_fieldEqn_Einstein}\\ &\nabla_\nu\left(e^{-2a\varphi}F^{\mu\nu}\right)=4\pi\sum_A q_A\int\! ds_A\,\frac{\delta^{(4)}\left(y-x_A(s_A)\right)}{\sqrt{-g}}\,\frac{dx_A^\mu}{ds_A}\ ,\\ &\Box\,\varphi=-\frac{a}{2}e^{-2a\varphi}F^2+4\pi\sum_A\int\! ds_A\,\frac{dm_A}{d\varphi}\frac{\delta^{(4)}\left(y-x_A(s_A)\right)}{\sqrt{-g}}\ , \end{align}\label{twoBody_fieldEqn}% \end{subequations} where $\delta^{(4)}\left(x-y\right)$ is the 4-dimensional Dirac distribution, and where $T^A_{\mu\nu}$ is the stress-energy tensor associated to the skeletonized body $A$, \begin{equation} T_A^{\mu\nu}=\int\! ds_A\, m_A(\varphi)\frac{\delta^{(4)}\left(y-x_A(s_A)\right)}{\sqrt{-g}}\frac{dx_A^\mu}{ds_A}\frac{dx_A^\nu}{ds_A}\ . \end{equation} In this paper, we shall focus on the conservative part of the dynamics, neglecting all radiation reaction forces. Moreover, for bound orbits, it is convenient to implement relativistic corrections to the Keplerian dynamics in the weak field, slow orbital velocity limit. \vfill\eject \noindent In appendix \ref{Appendix_Lagr_1PK}, we derive the first post-Keplerian (1PK), $\mathcal O(m/R)\sim \mathcal O(V^2)$ corrections to the two-body Lagrangian (where $R$ is the distance separating the bodies, and $V$ denotes their relative orbital velocity), extending to EMD theories the standard methods presented in, e.g., \cite{} in massless scalar-tensor theories (ST). There, the field equations (\ref{twoBody_fieldEqn}) are solved perturbatively around the Minkowski metric $\eta_{\mu\nu}$ and the constant value $\varphi_0$ of the background scalar field which, being associated to no gauge symmetry, see (\ref{EMDvacuumAction}) and below, cannot be set equal to zero, and is imposed by the cosmological environment of the binary system.\\ \\ At 1PK order, it is also useful to define the following body-dependent quantities \begin{align} &\alpha_A(\varphi)\equiv\frac{d\ln m_A}{d\varphi}\ ,\qquad \beta_A(\varphi)\equiv\frac{d\alpha_A}{d\varphi}\ ,\label{sensitivities}\\ \text{such that}\qquad &\ln m_A(\varphi)=\ln m_A^0+\alpha_A^0(\varphi-\varphi_0)+\frac{1}{2}\beta_A^0 (\varphi-\varphi_0)^2+\cdots\ ,\nonumber \end{align} where and from now on, a zero superscript indicates a quantity evaluated at infinity, $\varphi=\varphi_0$, as in $m_A^0=m_A(\varphi_0)$. Note that the parameter $\alpha_A^0$ measures the effective coupling between the scalar field and the skeletonized body $A$, and will play an essential role in the following.\\ With the definitions given above, the post-Keplerian Lagrangian reads, in harmonic coordinates $\partial_\mu(\sqrt{-g}g^{\mu\nu})=0$, and in the Lorenz gauge $\nabla_\mu A^\mu=0$ (see, again, appendix \ref{Appendix_Lagr_1PK} for details): \begin{align} L_{AB}&=-m_A^0-m_B^0+\frac{1}{2}m_A^0V_A^2+\frac{1}{2}m_B^0V_B^2+\frac{G_{AB}m_A^0 m_B^0}{R}\label{Lag1PK_PPN}\\ &+\frac{1}{8}m_A^0V_A^4+\frac{1}{8}m_B^0V_B^4+\frac{G_{AB}m_A^0 m_B^0}{R}\left[\frac{3}{2}(V_A^2+V_B^2)-\frac{7}{2}(V_A.V_B)-\frac{1}{2}(N.V_A)(N.V_B)+\bar \gamma_{AB}(\vec V_A-\vec V_B)^2\right]\nonumber\\ &-\frac{G_{AB}^2 m_A^0 m_B^0}{2R^2}\left[m_A^0(1+2\bar\beta_B)+m_B^0(1+2\bar\beta_A)\right]+\mathcal O(V^6)\ ,\nonumber \end{align} where $R\equiv\vert \vec x_A-\vec x_B\vert$, $\vec N\equiv(\vec x_A-\vec x_B)/R$, and $\vec V_A\equiv d\vec x_A/dt$. We have also introduced the combinations \begin{subequations} \begin{align} &G_{AB}\equiv1+\alpha_A^0\alpha_B^0-e_Ae_B\ ,\label{gAB}\\ &\bar\gamma_{AB}\equiv\frac{-4\,\alpha_A^0\alpha_B^0+3\,e_A e_B}{2(1+\alpha_A^0\alpha_B^0-e_Ae_B)}\ ,\label{PPNgamma}\\ &\bar\beta_A\equiv\frac{1}{2}\frac{\beta_A^0{\alpha_B^0}^2-2\,e_Ae_B(a\,\alpha_B^0-\alpha_A^0\alpha_B^0)+e_B^2(1+a\,\alpha_A^0-e_A^2)}{(1+\alpha_A^0\alpha_B^0-e_Ae_B)^2}\ ,\label{PPNbeta} \end{align}\label{PPN} \end{subequations} and $A\leftrightarrow B$ counterpart, together with the convenient notations \begin{equation} e_A\equiv\frac{q_A}{m_A^0}e^{a\varphi_0}\quad\text{and}\ \quad e_B\equiv\frac{q_B}{m_B^0}e^{a\varphi_0}\ .\label{ratio} \end{equation} The expression (\ref{Lag1PK_PPN}-\ref{ratio}) of the post-Keplerian Lagrangian in EMD theories, which results from our generic ``skeleton" ansatz (\ref{ansatz}), is the first new technical result of this paper.\\ Remarkably, (\ref{Lag1PK_PPN}) has exactly the same structure than the 1PK Lagrangian in ST theories, see, e.g., \cite{}. Indeed, the effects of the scalar \textit{and} vector fields have been entirely gathered in the four body-dependent combinations (\ref{PPN}), which generalize those introduced by Damour and Esposito-Far\`ese in ST theories \cite{}, that we retreive when the charges vanish, $q_{A/B}=0$. The effective (dimensionless, since we set $G=1$) gravitational coupling $G_{AB}$ reflects the addition of the metric, scalar and (repulsive) vector interactions at the linear level, and is reduced to unity in the general relativity limit, $q_{A/B}=0$ and $\alpha^0_{A/B}=\beta^0_{A/B}=0$. The combinations $\bar\gamma_{AB}$ and $\bar\beta_{A/B}$ depend quadratically on the couplings to the scalar field $\alpha_{A/B}^0$ and to the vector field $e_{A/B}$, and vanish in the GR limit.\\ \\ Finally, we note that our Lagrangian also encompasses Einstein-Maxwell's theory, which is retrieved when the bodies decouple from the scalar field, $m_{A/B}=cst$, i.e. $\alpha^0_{A/B}=\beta^0_{A/B}=0$. \vfill\eject In this section, we proposed a generic ``skeleton" ansatz (\ref{ansatz}) that is well-suited to address the problem of motion of two arbitrary compact bodies in EMD theories (e.g., neutron stars, or black holes). At 1PK order, we showed that the resultant two-body Lagrangian (\ref{Lag1PK_PPN}) only differs from the ST one through the redefinition of $G_{AB}$, $\bar\gamma_{AB}$ and $\bar\beta_{A/B}$ given in (\ref{PPN}), which makes their associated two-body dynamics \textit{a priori} undistinguishable at this order.\\ \\ However, EMD theories deserve special attention since, contrarily to ST theories \cite{}, they predict the existence of ``hairy" black holes (i.e. black holes that differ from the general relativistic ones), that are encompassed in the generic approach developped above. Consequently, our aim in the next sections will be to specify our results for systems involving such ``hairy" black holes, that is, to compute explcitely the values for their post-Keplerian parameters $\alpha_A^0$, $\beta_A^0$ and $e_A$ entering our Lagrangian (\ref{Lag1PK_PPN}). \section{Reducting a hairy black hole to a point particle\label{sectionSkeletonization}} In the following, we will restrict our attention to a specific subclass of ``hairy" black holes with an electric charge, presented in subsection \ref{subsHairyBH}. In order to address their motion in presence of a companion, we shall, in a second step, ``skeletonize" them, that is, compute explicitly the effective ``mass function" $m_A(\varphi)$ and parameter $q_A$ that enter the effective point particle action describing them, \begin{align} S_{\rm m}^{\rm pp}(A)=-\int\! m_A(\varphi)\,ds_A+ q_A\int\! A_\mu\, dx^\mu_A\ , \end{align} see (\ref{ansatz}) and the discussion below.\\ To this extent, we will ``zoom in" to the near-wordline region of particle $A$, which will take on the role of the black hole, and match the effective fields it generates to the real black hole solution. In particular, in keeping with neglecting all finite-size effects (see, again, section \ref{subsectionSkeleton}), the only \textit{gauge-invariant} influence of the faraway companion $B$ will be to impose a value for the adiabatically varying scalar environment $\varphi_\infty$ of our black hole. \subsection{Electrically charged dilatonic black holes\label{subsHairyBH}} Isolated black holes, that is, solutions of the vacuum field equations (\ref{vacuumEqs}), have been thoroughly studied in the literature. In particular, static, spherically symmetric black holes with an electric and (or) magnetic charge were introduced in \cite{} and \cite{}, while their axisymmetric counterparts were found in \cite{} and \cite{} in Kaluza-Klein theory (i.e. when $a=\sqrt{3}$ only).\\ In this paper, we shall restrict our attention to the class of electrically charged, non-spinning black hole solutions that read, in Just coordinates, \begin{subequations} \begin{align} &ds^2=-\left(1-\frac{r_+}{r}\right)\left(1-\frac{r_-}{r}\right)^{\frac{1-a^2}{1+a^2}}dt^2+\left(1-\frac{r_+}{r}\right)^{-1}\left(1-\frac{r_-}{r}\right)^{-\frac{1-a^2}{1+a^2}}dr^2+r^2\left(1-\frac{r_-}{r}\right)^{\frac{2a^2}{1+a^2}}d\Omega^2\ ,\\ &A_t=-\frac{Q\,e^{2a\varphi_\infty}}{r}\ ,\quad A_i=0\ ,\\ &\varphi=\varphi_\infty+\frac{a}{1+a^2}\ln\left(1-\frac{r_-}{r}\right)\ , \end{align}\label{TNsolution}% \end{subequations} where $d\Omega^2=d\theta^2+\sin^2\theta\, d\phi^2$ and where $Q$ is not an independent integration constant but rather satisfies \begin{equation} Q^2=\frac{r_+ r_-}{1+a^2}\,e^{-2a\varphi_\infty}\ .\label{Q_relation} \end{equation} Note that $Q$ is the conserved $U(1)$ charge of the black hole, as one easily sees from a direct integration of (\ref{maxwellVacuum}). Note also that we have ``gauged away" the asymptotic value of $A_\mu$ to zero for convenience.\\ The solution (\ref{TNsolution}-\ref{Q_relation}) hence depends on three integration constants: the radius $r_+$ of the horizon, the location $r_-$ of the curvature singularity, and the asymptotic value $\varphi_\infty$ of the scalar field which, in the presence of the faraway companion, identifies to the local, adiabatically varying value of the scalar field it produces.\\ We note that the electric charge is crucial to induce our black hole with a scalar ``hair". Indeed, when $r_-=0$, (\ref{TNsolution}) is reduced to Schwarzschild's black hole. Another important limit is the scalar-vector decoupling, $a=0$, see (\ref{EMDvacuumAction}). In that case, $\varphi=\varphi_\infty$, $Q^2=r_+r_-$, and (\ref{TNsolution}) is reduced to Reissner-Nordstr\"om's black hole.\\ Finally, as part of the ``skeletonization" to come below, it will prove sufficient to expand the solution (\ref{TNsolution}) at infinity and in isotropic coordinates ($r=\tilde{r}+[r_++r_-]/2+\cdots$) as \begin{subequations} \begin{align} &\tilde{g}_{\mu\nu}=\eta_{\mu\nu}+\delta_{\mu\nu}\left(\frac{r_++\frac{1-a^2}{1+a^2}\,r_-}{\tilde{r}}\right)+\mathcal{O}\left(\frac{1}{\tilde r^2}\right)\ ,\\ &A_t=-\frac{Q\,e^{2a\varphi_\infty}}{\tilde r}+\mathcal{O}\left(\frac{1}{\tilde r^2}\right)\ ,\\ &\varphi=\varphi_\infty-\frac{a}{1+a^2}\frac{r_-}{\tilde r}+\mathcal{O}\left(\frac{1}{\tilde r^2}\right)\ . \end{align}\label{reel_asympt}% \end{subequations} Indeed, we shall see in the next subsection that the asymptotic expansion (\ref{reel_asympt}), which depends on the three integration constants $r_+$, $r_-$ and $\varphi_\infty$, encodes all the necessary information to skeletonize the black hole, i.e. to fix uniquely the function $m_A(\varphi)$ and constant $q_A$ of the effective particle that will take on the role of the black hole. \subsection{The matching conditions\label{subsectionMatching}} In the near-wordline region of the effective particle $A$, and at leading order in the large separation limit, the effective field equations (\ref{twoBody_fieldEqn}) are reduced to \begin{subequations} \begin{align} &R_{\mu\nu}=2\partial_\mu\varphi\partial_\nu\varphi+2e^{-2a\varphi}\left(F_{\mu\alpha}F_\nu^{\ \alpha} -\frac{1}{4}g_{\mu\nu}F^2\right)+8\pi\left(T^A_{\mu\nu}-\frac{1}{2}g_{\mu\nu}T^A\right)\ ,\\ &\nabla_\nu\left(e^{-2a\varphi}F^{\mu\nu}\right)=4\pi q_A\int\! ds_A\,\frac{\delta^{(4)}\left(y-x_A(s_A)\right)}{\sqrt{-g}}\,\frac{dx_A^\mu}{ds_A}\ ,\\ &\Box\,\varphi=-\frac{a}{2}e^{-2a\varphi}F^2+4\pi\int\! ds_A\,\frac{dm_A}{d\varphi}\frac{\delta^{(4)}\left(y-x_A(s_A)\right)}{\sqrt{-g}}\ , \end{align}\label{skelFieldEqn} \end{subequations} where we recall that $ds_A=\sqrt{-g_{\mu\nu}dx_A^\mu dx_A^\nu}$ and that \begin{equation} T_A^{\mu\nu}=\int\! ds_A\, m_A(\varphi)\frac{\delta^{(4)}\left(y-x_A(s_A)\right)}{\sqrt{-g}}\frac{dx_A^\mu}{ds_A}\frac{dx_A^\nu}{ds_A}\ .\label{skelStressEnergyTensor} \end{equation} Let us solve these field equations perturbatively around Minkowski's metric $\eta_{\mu\nu}$, a vector field that can also be ``gauged away" to zero, and a scalar background, $\varphi_\infty$, which is the remaining influence of body $B$, as in (\ref{reel_asympt}). In harmonic coordinates (that identify to isotropic coordinates at this order), $\partial_\mu(\sqrt{-\tilde g}\tilde g^{\mu\nu})=0$, and in the rest-frame of the particle $A$, i.e. setting $\tilde{\vec x}_A=\vec 0$, this yields, at linear order, \begin{subequations} \begin{align} &\tilde{g}_{\mu\nu}=\eta_{\mu\nu}+\delta_{\mu\nu}\left(\frac{2m_A(\varphi_\infty)}{\tilde{r}}\right)+\mathcal O\left(\frac{1}{\tilde{r}^2}\right)\ ,\\ &A_t=-\frac{q_A\,e^{2a\varphi_\infty}}{\tilde{r}}+\mathcal O\left(\frac{1}{\tilde{r}^2}\right)\ ,\\ &\varphi=\varphi_\infty-\frac{1}{\tilde r}\frac{dm_A}{d\varphi}(\varphi_\infty)+\mathcal O\left(\frac{1}{\tilde{r}^2}\right)\ , \end{align}\label{skel_asympt} \end{subequations} which depends on the four effective parameters $q_A$, $\varphi_\infty$, $m_A(\varphi_\infty)$ and $m_A'(\varphi_\infty)$.\\ Now, by comparing (\ref{reel_asympt}) to (\ref{skel_asympt}), one obtains the following matching conditions: \begin{subequations} \begin{align} &m_A(\varphi_\infty)=\frac{1}{2}\left(r_++\frac{1-a^2}{1+a^2}\,r_-\right)\ ,\label{matching1}\\ &q_A=Q\ ,\label{matching2}\\ &\frac{dm_A}{d\varphi}(\varphi_\infty)=\frac{a}{1+a^2}\,r_-\ ,\label{matching3} \end{align}\label{matching}% \end{subequations} where we recall that the electric charge $Q$ satisfies \begin{equation} Q^2=\frac{r_+ r_-}{1+a^2}\,e^{-2a\varphi_\infty}\ ,\label{rappelQ} \end{equation} see (\ref{Q_relation}). Therefore, one ``skeletonizes" the real black hole (\ref{TNsolution}) as an effective point particle, provided that $q_A$, $m_A$ and its derivative evaluated at $\varphi_\infty$ satisfy the three matching conditions (\ref{matching}).\\ \\ Note that as expected, $q_A$ identifies to the electric charge $Q$ of the black hole, see (\ref{matching2}), while $m_A(\varphi_\infty)$ coincides with its standard, say, ADM mass \cite{}, $M_{\rm ADM}=\left(r_++\frac{1-a^2}{1+a^2}\,r_-\right)/2$, see (\ref{matching1}), which will therefore not be conserved when the scalar environment $\varphi_\infty$ of the black hole $A$, as created by its companion $B$, varies along its orbit.\\ Moreover, the system (\ref{matching}) is integrable. Indeed, inverting (\ref{matching}) to substitute $m_A$, $dm_A/d\varphi$ and $q_A$ to $r_+$, $r_-$ and $Q$ in (\ref{rappelQ}), yields the first order differential equation to be satisfied by the function $m_A(\varphi)$: \begin{equation} \left(\frac{dm_A}{d\varphi}\right)\left(m_A(\varphi)-\frac{1-a^2}{2a}\frac{dm_A}{d\varphi}\right)=\frac{a}{2}\,q_A^2\,e^{2a\varphi}\ . \label{eq_non_lineaire} \end{equation} For a given theory $a$, the solution of (\ref{eq_non_lineaire}) depends on the charge $q_A$ of the black hole, together with a unique integration constant, that we shall denote $\mu_A$. The obtention of the matching conditions (\ref{matching}), together with the differential equation (\ref{eq_non_lineaire}) is the second new technical result of this paper, which shows that the dynamics of the black hole (\ref{TNsolution}) is described by two constant parameters only, $q_A$ and $\mu_A$. \section{The sensitivity of a hairy black hole\label{section_Sensitivities}} In this section, we solve the differential equation (\ref{eq_non_lineaire}), that is, we compute explicitely the function $m_A(\varphi)$ that describes the dynamics of our ``hairy" black hole (\ref{TNsolution}) through the point particle action \begin{align} S_{\rm m}^{\rm pp}(A)=-\int\! m_A(\varphi)\,ds_A+ q_A\int\! A_\mu\, dx^\mu_A\ . \end{align} This will, in a second step, allow us to come back to the two-body Lagrangian (\ref{Lag1PK_PPN}) obtained in section \ref{sectionLagrangian1PK}, and compute the values of the fundamental post-Keplerian parameters $\alpha_A^0$, $\beta_A^0$ and $e_A$ that enter it. In particular, we will find that the dynamics of our black hole might depart significantly from that of black holes in general relativity. \subsection{The theory $a=1$} In the particular case $a=1$, (\ref{eq_non_lineaire}) is easily integrated as \begin{equation} m_A(\varphi)=\sqrt{\mu_A^2+ q_A^2 \frac{e^{2\varphi}}{2}}\ ,\label{exemple_typique} \end{equation} where $\mu_A$ is a positive integration constant. We hence have an explicit expression for the function $m_A(\varphi)$, which characterizes our black hole by means of two constant parameters:\\ \\ (i) its electric charge $q_A=Q$ (see (\ref{matching2})), given in (\ref{rappelQ}), which is essential for our black hole to carry scalar ``hair". Indeed, when $q_A=0$, (\ref{exemple_typique}) is reduced to $m_A(\varphi)=\mu_A$ and the black hole decouples both from the vector and the scalar fields; \\ \\ (ii) the constant $\mu_A$, whose physical interpretation can be obtained thus: expressing it in terms of $r_+$, $r_-$, $\varphi_\infty$, using the matching conditions (\ref{matching}), yields \begin{equation} \mu_A^2=\frac{r_+(r_+-r_-)}{4}=\frac{A_H}{16\pi}\ .\label{relationMuAire} \end{equation} Therefore, $\mu_A^2$ is proportional to the area $A_H$ of the horizon of the black hole.\\ \\ In other words, the skeletonization describes a black hole that, when submitted to an adiabatic variation of its scalar field environment $\varphi_\infty$ (created by the slowly orbiting, faraway companion), readjusts its equilibrium configuration (that is, $r_+$ and $r_-$) in order for both its electric charge and horizon area to remain constant (while we recall that its ADM mass is not conserved, see (\ref{matching1}) and comments below).\\ Moreover, given a specific black hole (i.e. specific values for $q_A$ and $\mu_A$), one can come back to the two-body problem, and compute the fundamental parameters that enter the 1PK Lagrangian (\ref{Lag1PK_PPN}), see (\ref{sensitivities}) and (\ref{ratio}), \begin{equation} \alpha_A^0=\frac{d\ln m_A}{d\varphi}(\varphi_0)\ ,\qquad \beta_A^0=\frac{d\alpha_A}{d\varphi}(\varphi_0)\ ,\qquad e_A=\frac{q_A}{m_A^0}e^{a\varphi_0}\ ,\label{rappelSensi} \end{equation} in terms of $q_A$, $\mu_A$, and $\varphi_0$ which is the value of the scalar field far from the binary system, imposed by cosmology. Injecting the explicit mass function (\ref{exemple_typique}) in (\ref{rappelSensi}), we find \begin{align} &\alpha_A(\varphi_0)=\frac{1}{1+e^{2\left(\ln\left|\frac{\mu_A\sqrt{2}}{q_A}\right|-\varphi_0\right)}}\ ,\label{fermiDirac}\\ \nonumber\\ &\text{together with}\qquad \beta_A^0=2\alpha_A^0(1-\alpha_A^0)\qquad \text{and}\qquad (e_A)^2= 2\alpha_A^0\ .\label{betaA_eA} \end{align} \\ The simplicity of (\ref{fermiDirac}) is striking: the coupling $\alpha_A(\varphi_0)$ between the black hole and the scalar field is given by a ``Fermi-Dirac" distribution, as shown in figure \ref{}, which highlights the crucial influence of the scalar cosmological background $\varphi_0$ on the dynamics of a black hole. Indeed, when $\varphi_0$ is increased through $\varphi^{\rm crit}_A=\ln\left|\mu_A\sqrt{2}/q_A\right|$, the black hole $A$ transitions steeply between two extremal regimes:\\ \\ (a) a decoupled regime, $\alpha_A^0\rightarrow 0$, where it is moreover undistinguishable from the general relativistic Schwarzschild black hole, since $\alpha_A^0$ \textit{as well as} $\beta_A^0$ (and higher order derivatives of $\alpha_A$) and $e_A$ vanish, see (\ref{betaA_eA});\\ \\ (b) a regime where it is strongly coupled both to the scalar and to the vector fields, $\alpha_A^0\rightarrow 1$, together with $\beta_A^0\rightarrow 0$ and $(e_A)^2 \rightarrow 2$, which by definition, induces large deviations to the general relativistic two-body Lagrangian through the combinations $G_{AB}$, $\bar\gamma_{AB}$ and $\bar\beta_{A/B}$ that enter it, see (\ref{PPN}).\\ We note that the specific black hole considered (i.e., $q_A$ and $\mu_A$) only influences the location $\varphi^{\rm crit}_A=\ln\left|\mu_A\sqrt{2}/q_A\right|$ of the transition, while the values reached by $\alpha_A^0$, $\beta_A^0$ and $|e_A|$ in the extremal regimes (a) and (b) are \textit{universal}. \begin{figure}[!h] \caption{The ``Fermi-Dirac" coupling $\alpha_A(\varphi_0)$ between the black hole $A$ and the scalar field, as a function of its cosmological environment $\varphi_0$, when $a=1$. The ``scalarization" is centered on $\varphi^{\rm crit}_A=\ln\left|\mu_A\sqrt{2}/q_A\right|$. Three distinct black holes are represented, $|\mu_A/q_A|=\{10,10^{3},10^{5}\}$, $q_A$ and $16\pi\mu_A^2$ being respectively their conserved charges and areas.\label{}} \includegraphics[width=0.5\linewidth]{plot_transition_alpha_aEq1} \end{figure} From now on, this transition phenomenon will be referred to as ``scalarization", by analogy with the terminology introduced in the context of scalar-tensor theories to describe the spontaneous \cite{} and dynamical \cite{} scalarization of neutron stars (altough it must be noted that the phenomenon we highlight here is not a ``phase transition", and that the scalar coupling $\alpha_A^0$ transitions together with the vector coupling, $e_A$). \subsection{Generic EMD theories\label{subsubTransition}} The remarkable features presented above are now easily transposed to generic EMD theories, that is, to an arbitrary value for the fundamental parameter $a$. First, one solves numerically the differential equation (\ref{eq_non_lineaire}), holding fixed the values for $q_A$ and for the integration constant $\mu_A$, in keeping with considering one specific black hole\footnote{Note that although we solve (\ref{eq_non_lineaire}) numerically, this differential equation admits an exact analytical solution in the form of a parametric equation on $m_A(\varphi)$, which is not enlightening to show here. It is however easy to find that the asymptotic behaviour of the solution is $m_A(\varphi)\underset{-\infty}{\sim} \mu_A$ and $m_A(\varphi)\underset{+\infty}{\sim} [q_A^2 e^{2a\varphi}/(1+a^2)]^{1/2}$, which is sufficient for the purpose of this paper.\\}. Then, one deduces the resultant coupling to the scalar field, $\alpha_A^0=d\ln m_A/d\varphi(\varphi_0)$, together with $\beta_A^0$ and $e_A$ through the relations\footnote{The obtention of (\ref{betaA_eA_generic}) is straightforward if one injects, by definition of $\alpha_A^0$ and $\beta_A^0$ (\ref{rappelSensi}), the expansion $m_A(\varphi)=m_A^0[1+\alpha_A^0(\varphi-\varphi_0)+\frac{1}{2}({\alpha_A^0}^2+\beta_A^0)(\varphi-\varphi_0)^2+\cdots]$ into the differential equation (\ref{eq_non_lineaire}), and solves it order by order to get \begin{equation} \alpha_A^0=\frac{a}{1-a^2}\left(1-\sqrt{1-e_A^2(1-a^2)}\right)\quad ,\qquad\beta_A^0=\frac{a^2e_A^2}{1-a^2} \bigg(1-a^2/\sqrt{\strut 1-e_A^2(1-a^2)}\bigg)\ ,\nonumber \end{equation} which is easily inverted to give (\ref{betaA_eA_generic}).\\} \begin{equation} \beta_A^0=\alpha_A^0\, (a-\alpha_A^0)\left[\frac{(1-a^2)\alpha_A^0-2a}{(1-a^2)\alpha_A^0-a}\right]\quad ,\qquad (e_A)^2=\alpha_A^0\left[\frac{2a-(1-a^2)\alpha_A^0}{a^2}\right]\ .\label{betaA_eA_generic} \end{equation} The resulting parameter $\alpha_A^0$ is shown in figure \ref{}, for $a\in \llbracket 0,10 \rrbracket$, setting $|\mu_A/q_A|=10^3$. Again, for all $a$, the black hole transitions between two extremal regimes:\\ \\ (a) a decoupling, ``Schwarzschild-like" regime $\alpha_A^0\rightarrow 0$, together with $\beta_A^0\rightarrow 0$ and $e_A\rightarrow 0$, see (\ref{betaA_eA_generic});\\ \\ (b) a \textit{universally} scalarized regime, $\alpha_A^0\rightarrow a$, with $\beta_A^0\rightarrow 0$ and $(e_A)^2\rightarrow 1+a^2$.\\ Note that the curves $\alpha_A(\varphi_0)$ shown in figure \ref{} are slightly deformed ``Fermi-Dirac" distributions (with an exact identification when $a=1$ only, see discussion above), the steepness of the transition increasing with $a$. Note also that, for a fixed theory $a$, the value of $q_A$ and $\mu_A$ only influences the location $\varphi_A^{\rm crit}$ of the transition (which is shifted to greater $\varphi_0$ values when $|\mu_A/q_A|$ is increased) and the sign of $e_A$.\\ \begin{figure}[!h] \caption{The coupling $\alpha_A(\varphi_0)$ between the black hole and the scalar field, as a function of its cosmological environment $\varphi_0$, for the theories $a\in \llbracket 0,10 \rrbracket$. The specific class of black holes such that $|\mu_A/q_A|=10^3$ is represented. When $a=0$, $\alpha_A=0$ and the black hole is Reissner-Nordst\"om's. For non-zero values of $a$, $\alpha_A(\varphi_0)$ becomes a function of $\varphi_0$ that transitions between $0$ and $a$, the steepness of the ``scalarization" increasing with $a$.\label{}} \includegraphics[width=0.55\linewidth]{plot_alpha_phi_q_mu_generic} \end{figure} \vfill\eject A numerical study of the dynamics of binary black holes in EMD theories has been performed recently in \cite{}. There, the authors found that when $|\varphi_0|=10^{-10}$ and $a=\{1,10^3,3\times 10^3\}$, the influence of the scalar field on binary black holes can be neglected in understanding their dynamics, which in turn becomes hardly distinguishable from its general relativistic counterpart. These features are consistent with the approach developped troughout the present paper: for example, when $|\mu_A/q_A|=10^3$, $|\varphi_0|=10^{-10}$ and $a=3\times 10^3$, we find vanishingly small values for $\alpha_A^0$, $\beta_A^0$ and $e_A$, see, again, figure \ref{ \\ However, our results above hint that this Schwarzschild-like behaviour might be considerably revised when the value of $\varphi_0$ is increased, as shown by the ``scalarization" phenomenon discussed above. Now, gravitational wave detectors such as LIGO-Virgo (or the forthcoming LISA detector) are designed to detect highly redshifted sources, where, indeed, the cosmological environment $\varphi_0$ might have had a non-negligible value (and that differs from today). In such situations, one could expect large deviations to the general relativistic black hole dynamics.\\ The obtention of the explicit mass function $m_A(\varphi)$ and charge $q_A$ that describe EMD black holes, together with the identification of the transition phenomenon regarding their coupling to the scalar and vector fields, or ``scalarization", is the third, and main result of this paper. \subsection{Scalarized binary black holes\label{appendix_BBH_examples}} As discussed above, in the presence of one (or two) black holes, the dynamics of binary systems might depart significantly from its general relativistic counterpart. For the sake of illustration, let us consider a binary system composed of two ``scalarized" black holes , i.e. in the \textit{idealistic} limit when both are described by the regime (b) presented above, \textit{and} with electric charges $q_A$ and $q_B$ of the same sign: \begin{equation} \alpha_{A/B}^0\rightarrow a\ ,\quad \beta_{A/B}^0\rightarrow 0\ ,\quad\text{and}\quad e_{A/B} \rightarrow \sqrt{1+a^2}\ ,\label{rappelCouplings} \end{equation} the value of the parameters ($q_A$, $\mu_A$) and ($q_B$, $\mu_B$) influencing only the location $\varphi_{A/B}^{\rm crit}$ of their transition, see the discussions above, and where an irrelevant sign was chosen in the definition of $e_{A/B}$.\\ Injecting (\ref{rappelCouplings}) into the combinations $G_{AB}$, $\bar\gamma_{AB}$ and $\bar\beta_{A/B}$, given in (\ref{PPN}), that enter the post-Keplerian Lagrangian (\ref{Lag1PK_PPN}) presented in section \ref{sectionLagrangian}, yields $G_{AB}\rightarrow 0$, $G_{AB}\,\bar\gamma_{AB}\rightarrow (3-a^2)/2$, and $G_{AB}^2\,\bar\beta_{A/B}\rightarrow 0$.\\ \\ Therefore, all in all, the final two-body Lagrangian is reduced, at 1PK order, to \begin{align} L_{AB}&\rightarrow-m_A^0\sqrt{1-V_A^2}-m_B^0\sqrt{1-V_B^2}+\left(\frac{3-a^2}{2}\right)\frac{m_A^0 m_B^0}{R}(\vec V_A-\vec V_B)^2+\mathcal O(V^6)\ .\label{PapapetrouMajumdar} \end{align} The simplification (\ref{PapapetrouMajumdar}) is spectacular: among all the interaction terms present in the generic Lagrangian (\ref{Lag1PK_PPN}), only the $\bar\gamma_{AB}$-driven one remains. Moreover, this last term vanishes as well when the bodies are at relative rest ($\vec V_A-\vec V_B=\vec 0$). In other words, EMD binary black holes can transition to a universally ``scalarized" regime where their (attractive) metric, scalar and (repulsive) vector interactions compensate to allow for configurations at rest, at 1PK level at least.\\ This result is consistent with the Majumdar-Papapetrou spacetime\footnote{I am grateful to Thibault Damour for mentioning and sharing his expertise on the Majumdar-Papapetrou binary black hole solutions.}, which is a solution of the Einstein-Maxwell field equations that describes two \textit{extremal} Reissner-Nordstr\"om black holes at rest, see \cite{, , }, and their extensions to EMD theories \cite{}. For a ``scalarizing" black hole, $e_A$ behaves as \begin{equation} (e_A)^2=\left(\frac{q_A}{m_A^0}\right)^2e^{2 a\varphi_0}\rightarrow 1+a^2\ ,\label{eq_extremalization} \end{equation} where $q_A=Q$ is the electric charge of the black hole, and $m_A^0=M_{\rm ADM}$ its ADM mass, see (\ref{matching}) and the discussion below, which generalizes the notion of an extremal black hole to generic EMD theories, $a\neq 0$: indeed, (\ref{eq_extremalization}) is equivalent to $r_-\rightarrow r_+$ (as proven using again the matching conditions (\ref{matching})).\\ \\ In other words, our ``scalarizing" black holes are in fact transitioning towards a ``quasi-extremal" regime (but never reach an exact extremal, ``naked singularity" configuration, since $\varphi_0$ is always finite).\\ Another striking consequence of (\ref{PapapetrouMajumdar}) is that in the context of Kaluza-Klein theory, i.e., $a=\sqrt{3}$, the last interaction term vanishes as well. Therefore, each component of the scalarized black hole binary behaves as a free particle at 1PK order. The generalization of this phenomenon to higher PK order remains to be investigated. \subsection{General relativity and Einstein-Maxwell theory\label{subsection_RN_S}} Let us conclude this section by mentioning that, as it should, the approach developped above is consistent with the well-known general relativity and Einstein-Maxwell theory limits.\\ \\ (i) Firstly, for any EMD theory, i.e. any value for $a$, but when the charge of the black hole vanishes, $q_A=0$, the three matching conditions (\ref{matching}) obtained in section \ref{subsectionMatching} are reduced to \begin{equation} m_A=\frac{r_+}{2}=cst.\quad , \qquad q_A=Q=0\ . \end{equation} Therefore, the mass function $m_A$ identifies to the constant ``Schwarzschild" mass, our skeletonized black hole decouples both from the vector and scalar fields, and describes a Schwarzschild black hole, as expected from (\ref{TNsolution}) and the discussion below. This corresponds to the general relativity limit.\\ \\ (ii) Secondly, for any charge $q_A$, but in the scalar-vector decoupling limit $a=0$, see (\ref{EMDvacuumAction}), the matching conditions (\ref{matching}) become \begin{equation} m_A=\frac{r_++r_-}{2}=cst.\quad ,\qquad q_A=Q\ . \end{equation} This time, our effective particle describes a Reissner-Nordstr\"om black hole, as predicted in Einstein-Maxwell theory, see again discussion below (\ref{TNsolution}). \section{Concluding remarks} Reducing compact bodies to effective point particles has proven to be a very efficient tool to address analytically the many-body problem in general relativity and scalar-tensor theories. In this paper, we generalized this ``skeletonization" to Einstein-Maxwell-dilaton (EMD) theories, and showed that the corresponding most generic skeleton ansatz was the combination of two well-known point particle actions: that introduced by Eardley in scalar-tensor theories, which consists in endowing point particles with a specific ``mass function" $m_A(\varphi)$, and that of a charged particle in electrodynamics, endowed with a constant charge $q_A$.\\ More importantly, we computed, for the first time, the mass functions $m_A(\varphi)$ to be assigned to black holes, and shed light on two major properties regarding their dynamics: firstly, that it is encoded into two parameters (per black hole) only, $q_A$ and $\mu_A$, which are related to their charge and area\footnote{Note that we proved that $\mu_A$ is related to the area of the black hole for the specific theory $a=1$, see (\ref{relationMuAire}). Its generalization to arbitrary EMD theories will be the topic of \cite{CardenasJulie}.}; secondly, that, in keeping their charge and area fixed, black holes can undergo a new type of ``scalarization" phenomenon, leading to a transition between a regime where their dynamics is undistinguishable from that of Schwarzschild's black holes, and a ``quasi-extremal" regime where they strongly and universally couple to the scalar and vector fields (to within the sign of their electric charge). We note that the simplicity of our results, as illustrated by, e.g., figure \ref{}, is inherent to black holes, and contrasts with the rather involved study of neutron stars and their scalarization, see \cite{}, \cite{} or \cite{} (which, for example, depends on their equation of state, and on their coupling to the Jordan metric, $\tilde g_{\mu\nu}=\mathcal A^2(\varphi)g_{\mu\nu}$, see (\ref{actionEMD})). \vfill\eject As already discussed in section \ref{section_Sensitivities}, the ``scalarization" of a given black hole depends drastically on the value $\varphi_0$ of its cosmological environment, which, in turn, is expected to vary in the large range of redshifts that will be reachable with the gravitational wave detectors LIGO-Virgo (and forthcoming LISA); see, for example, \cite{} for the cosmological evolution of $\varphi_0$ in scalar-tensor theories (to which EMD theories can be fairly supposed to reduce to on cosmological scales). The fact that, depending on the redshift from which, say, a binary black hole system emitted, its dynamics can be undistinguishable from the general relativistic one, or, rather, significantly depart from it, supports the importance to investigate sources emitting from the broadest possible range of redshifts in the future.\\ In another paper \cite{}, we extended the effective-one-body (EOB) approach (see, e.g., \cite{} and \cite{}) to massless scalar-tensor theories (ST); that is, we reduced the ST two-body dynamics, which is known at second post-Keplerian order \cite{}, to the geodesic motion of a test particle in an effective static, spherically symmetric metric. In doing so, we implemented the impact of ST theories on the coalescence of compact binary systems as parametrized corrections to the best available general relativistic (5PN) EOB results. Now, in the present paper, we computed the two-body Lagrangian at 1PK level and showed that it only differs from the 1PK ST one in redefining the specific combinations $G_{AB}$, $\bar\gamma_{AB}$, and $\bar\beta_{A/B}$, see (\ref{PPN}) and discussions below. In consequence, the ST-EOB results, presented in \cite{}, are trivially extended to EMD theories (including black holes). For example, the EMD correction to the orbital frequency of a compact binary system at the innermost stable circular orbit (ISCO), presented in \cite{} figure 1 in ST theories, can be obtained replacing simply $G_{AB}$, $\bar\gamma_{AB}$, and $\bar\beta_{A/B}$ by their EMD values.\\ Finally, we note that, although the point particle action (\ref{ansatz}) from which we started, together with the matching conditions (\ref{matching}), have fixed uniquely the mass function $m_A(\varphi)$ and charge $q_A$ to be assigned to a specific EMD black hole, their validity remain to be justified. Now, as discussed in section \ref{subsectionSkeleton}, our point particle action (\ref{ansatz}) does not depend on the gradients of the fields; this presupposes in turn that, for a well-separated binary system, each body is adiabatically readjusting its equilibrium configuration when submitted to the influence of its slowly orbiting, faraway companion. In \cite{CardenasJulie}, the thermodynamics of EMD black holes in equilibrium will provide a convenient framework to interpret and justify their skeletonization. \section*{Acknowledgments} I am very grateful to Nathalie Deruelle for bringing EMD theories to my attention, for enlightening discussions and for her constant support. I also acknowledge stimulating discussions with Marcela C\'ardenas. Finally, I would like to thank Gilles Esposito-Far\`ese and Thibault Damour for very fruitful suggestions.

\section{Introduction}\label{Sec-Introduction} The aim of the present article is to restrict the scalar Laplace-Beltrami operator defined on $\setR^6$, endowed with the metric $\eta~=~\mbox{diag}(+,-,-,-,-,+)$, to (locally) conformally flat four dimensional Lorentzian manifolds $X_f$. These manifolds are obtained as the intersection between the null cone $\mathcal{C}$ and the submanifolds $P_f$ of $\setR^6$, defined through functions $f$ homogeneous of degree one. The set of the manifolds $X_f$ contains in particular the Friedmann-Lema\^{\i}tre-Robertson-Walker (FRLW) spacetimes. The definition of a manifold $X_f$, as the intersection of two hypersurfaces, leads to the definition of two natural vector fields: schematically the normals to the hypersurfaces. From a geometrical viewpoint these two vector fields operate on scalar fields: they are constraints associated to the manifold $X_f$. These two constraints can be used to obtain the restriction of the scalar Laplace-Beltrami operator to the manifold $X_f$. Our strategy to obtain that restriction is to build an orthonormal basis (a sextrad) in the neighborhood of $X_f$ which contains the constraints and to use a geometrical identity to split the Laplace-Beltrami operator, which in this case identifies with the Laplace-de~Rham operator up to a sign, into the Laplace-Beltrami operator on $X_f$ and remaining terms related to the constraints. On a physical side, the restriction from $\setR^6$ to the submanifold $X_f$ maps a scalar field solution of the free ``massless'' Klein-Gordon equation in $\setR^6$ to a solution of the corresponding restricted equation on $X_f$. Now, it appears that sufficient freedom is left in the constraints (which apply on the scalar field) to choose them in such a way that the restricted field on $X_f$ satisfies the Klein-Gordon-like equation (possibly massive and coupled to the curvature): $(\square_{f} - \xi R^f + m^2)\phi^f=0,$ on $X_f$. That is in particular on FLRW spacetimes. Note that, the term $m^2$ (when present) is only a coefficient in a curved space, we nevertheless keep the name of Klein-Gordon (instead of Klein-Gordon-like) for short, in the remaining of the text. At the origin we developed part of the geometric formalism we use here to study the properties of conformally invariant fields on (A)dS spacetimes and how these properties relate to their Minkowskian counterparts if any \cite{}. One of the main advantage of this geometric setup is that it provides us with a (local) continuous mapping of mathematical objects (functions, operators, etc.) between spaces. Indeed, this formalism applies not only to maximally symmetric spaces, but to all locally conformally flat spaces \cite{, }. As a consequence, non-maximally symmetric FLRW spacetimes are concerned, which makes contact with cosmology (FLRW spacetimes, are currently favored, on large scale, by recent obervations \cite{}, with possible deviation from spatial flatness \cite{}). In that context it is of interest to consider, beside conformally invariant massless field, the possibility for non-conformally coupled fields and massive fields. Finally, it should be emphasized that, if the realization of some spacetimes as submanifolds of some ambient spaces is known for some time (see for instance \cite{}, \cite{} for conformal field in $\setR^6$ and more generally references in \cite{}), the formalism we use here differs from these methods in a notable way. Its specificity lies in the realization of the four dimensional space $X_f$ as an intersection between the null cone of $\setR^6$ and the submanifold $P_f$ defined by the function $f$. This submanifold $P_f$ can be considered as an embedding space which is adapted for each spacetime $X_f$. This differs, from this point of view, with so-called ambient space formalism (see for instance \cite{, }) in which the embedding space is fixed. The paper is organized as follows. The restriction of the scalar Laplace-Beltrami operator from $\setR^6$ to $X_f$ is done in Section \ref{SEC-RestrictioToXf}. As a byproduct one obtains a formula relating the Ricci scalar $R^f$ on the spacetime and the function $f$ defining that spacetime in $\setR^6$. The Section \ref{SEC-KGForm} is devoted to the relation with the Klein-Gordon equation on the spacetime $X_f$. We conclude in Section \ref{SEC-Conclusion}. Some topics and additional materials are collected in appendices~: appendix \ref{App-DefAndForm} details the conventions and notations used in this work, a somewhat technical proof used for the restriction of the Laplace-Beltrami operator is given in appendix \ref{App-Proof-deld6-deld4}, an explicit realization to general FLRW spacetimes is given in appendix \ref{App-ConstrainFLRW}, the direct application of our results to the maximally symmetric spaces is tackled in appendix \ref{App-ConstraintMaxSym}. Our conventions for indexes are: $\alpha, \beta, \gamma,\ldots = 0,\ldots, 5$, and $\mu,\nu,\rho,\ldots = 0, 1, 2, 3$. In addition, except otherwise stated, a superscript (or subscript) $f$ indicates a quantity on the $X_f$ space (see Section \ref{SEC-RestrictioToXf} for the definition). The space $\setR^6$ is provided with the natural orthogonal coordinates $\{y^\alpha\}$ and the metric $\eta~=~\mbox{diag}(+,-,-,-,-,+)$ which is left invariant under the group O$(2,4)$ and its connected subgroup SO$_0(2,4)$. Our notations for differential geometry follow those of Fecko \cite{}. Certain definitions and formulas particularly used in this work are gathered in appendix \ref{App-DefAndForm}. \section{The restriction of $\square_6$ to the spacetime $X_f$}\label{SEC-RestrictioToXf} Let us first recall the geometrical framework we use: a four dimensional manifold $X_f$ is obtained as the intersection of the five dimensional null cone of $\setR^6$: $\mathcal{C}:=\{y\in \setR^6 : C(y) = 0\}$, where $C(y):=y^\alpha y_\alpha$, and the hyper-surface $P_f:= \{y\in \setR^6 : f(y) = 1\}$, in which $f$ is homogeneous of degree one. Locally conformally flat spaces, including FLRW, de Sitter and anti-de Sitter spaces, can be obtained by suitable choices of the homogeneous function $f$. Details about this construction may be found in \cite{}. In this section we explicitly restrict the free scalar equation $\square_6 \phi = 0$ to a four dimensional space $X_f$. This is obtained thanks to a geometrical method which refers to Cartan's moving frame (``rep\`ere mobile'') theory. Indeed, instead of the common procedure which consists in using a specific coordinate system including constraints (for instance $\{C(y), y^\mu/f(y), f(y)\}$), we relate constraints to vectors (differential operators) on which we build the moving frame. As a consequence, no coordinate system other than the natural Cartesian system of $\setR^6$: $\{y^\alpha\}$ is needed. Beside the restriction of the scalar operator $\square_6$, we obtain, as a byproduct, a formula which allows us to compute the Ricci scalar of the space $X_f$ from the function $f$ which defines it as a submanifold embedded in $\setR^6$, namely \begin{equation*} \frac{R^f}{6} = (\square_6f)^f - 2 F_f^2, \end{equation*} where $F=\widetilde{df}=\partial^\alpha f\,\partial_\alpha$ is the gradient vector of $f$. \subsection{A general expression for $(\square_6\phi)^f$}\label{closed formula} The basic idea in order to restrict the scalar operator $\square_6 = -\delta d$ to the submanifold $X_f$ is to use the formula (\cite{}, \S 5.8.11) \begin{equation}\label{EQ-BoxSpliting} j^a i_a = p, \end{equation} which applies to any p-form in an orthonormal basis $\{e_a\}$, to split the operator into two parts: \begin{equation*} \square_6 \phi = -\delta d\phi = -\delta j^\mu i_\mu d\phi - \delta(j^4 i_4 + j^5 i_5) d\phi, \end{equation*} in such a way that the first term in the r.h.s. can be restricted to be the Laplace-Beltrami operator of $X_f$ acting on the restricted field $\phi^f$ and the second term can be expressed in terms of constraints related to those defining the geometry ($f$ and $C$) and intrinsic quantities of $X_f$. In order to complete that scheme, we need a basis $\{e_a\}$ in which we perform the splitting in a neighborhood of $X_f$ and to translate the constraints defining $X_f$ in operators acting on the scalar field. Let us begin by considering the constraints on the field. Since they account for the geometry, that is to the surfaces defined by $f$ and $C$ whose intersection is the space $X_f$, a natural choice is to use the two gradient vectors \begin{equation}\label{EQ-defFandD} F := \widetilde{df}=\partial^\alpha f\,\partial_\alpha ,~\mathrm{and}~D:= \frac{1}{2} \widetilde{dC} = y^\alpha \partial_\alpha, \end{equation} as constraint operators on the field. Note that $D = y^\alpha \partial_\alpha$ is the generator of dilations in $\setR^6$. The commutators \begin{equation*} [\square_6, D] = 2 \square_6,~~~[F, D] = F \end{equation*} show us that, since $\square_6 \phi = 0$, only $F\phi$ has to be given. The operator $D$ can be left unspecified, which means that the function $\phi$ can be choosen homogeneous of some unspecified weight $r$, namely $D\phi = r \phi$. We now build the orthonormal basis. Since it must be defined in a neighborhood of $X_f$ let us define the space \begin{equation}\label{EQ-Def-Mpq} M_{\lambda,q}=\{y:\; f(y)=\lambda,\, C(y)=q,\; \lambda,q \in \setR, \lambda > 0\}. \end{equation} We then have $X_f = M_{1,0}$ and since $X_f$ is a pseudo-Riemannian manifold so is $M_{\lambda,q}$ at least in a neighborhood of $X_f$. Let us choose for each $M_{\lambda,q}$ an orthogonal bases field of T$M_{\lambda,q}$: $\{\partial_\mu\}$. A vector $\partial_\mu$ is tangent to the surfaces defined by $f(y)=\lambda$ and to the surfaces defined by $C(y)=q$ it is thus orthogonal to $F$ and $D$. We can then build an orthonormal basis $\{e_\alpha\}$ from the six vectors $\{\partial_\mu, F, D\}$. We consider two separate cases depending on the nature of $F$. Let us first assume $F^2\neq0$, and set \begin{equation}\label{EQ-eFe_muFneq0} e_\mu := \frac{\partial_\mu}{\vert\partial_\mu\vert},~~e_{\sss 5} := \frac{F}{\vert F \vert}. \end{equation} Then $e_4$ is obtained as a normed linear combination of $D$ and $F$, we find \begin{equation}\label{EQ-e4-F2neq0} e_4 =\frac{1}{\sqrt{\Delta}}\left( f e_5-\vert F\vert D\right) , \end{equation} with $\Delta := f^2-F^2 D^2 $. Note that, these expressions apply whatever the sign of $F^2$. When $F^2=0$ we redefine $e_5$ as \begin{equation}\label{EQ-e5 F=0} e_{\sss 5} := \frac{F+D}{\vert F + D\vert}. \end{equation} Then, we can complete the two orthonormal vectors $\{e_\mu, e_5\}$ and obtain the orthonormal basis $\{e_\alpha\}$ with \begin{equation}\label{EQ-e4-F2nq0} e_4 =\frac{1}{\sqrt{\Delta_0}}\left( (1 + \frac{D^2}{f}) F - D\right), \end{equation} where $\Delta_0 := \vert F + D\vert = 2f + D^2$. Having determined an orthonormal basis we can now split $\square_6 \phi$ in that basis using (\ref{EQ-BoxSpliting}) making the constraints apparent. As mentioned before, the term $-\delta j^\mu i_\mu d\phi$ reduces to $\square_f$ on $X_f$, the proof of this statement is given in Appendix \ref{App-Proof-deld6-deld4}. The calculation of the second term of the r.h.s. of (\ref{EQ-BoxSpliting})~: $\delta(j^4 i_4 + j^5 i_5) d\phi$, follows the same lines in the three cases $F^2>0$, $F^2<0$ or $F^2=0$ and leads to the same final formula. Here, we perform the calculation for $F^2>0$. We first express the operators $i$ and $j$ as \begin{align*} i_5&=\frac{1}{|F|}i_{\sss F},\\ i_4&=\frac{f}{\sqrt{\Delta}{|F|}}i_{\sss F} -\frac{|F|}{\sqrt{\Delta}}i_{\sss D},\\ j^5&=\frac{1}{|F|}j_{\sss F},\\ j^4&=-j_4= -\frac{f}{{\sqrt{\Delta}}|F|}j_{\sss F} +\frac{|F|}{\sqrt{\Delta}}j_{\sss D}. \end{align*} Then, one obtains \begin{equation*} \begin{split} &\delta (j^5i_5+j^4i_4)d\phi=\\ &\delta\left(-\frac{D^2}{\Delta}j_{\sss F} F+ \frac{f}{\Delta}j_{\sss F} D+\frac{f}{\Delta}j_{\sss D}F-\frac{F^2}{\Delta}j_{\sss D} D\right)\phi. \end{split} \end{equation*} This expression can be explicitly calculated in the natural Cartesian coordinate system of $\setR^6$ in which the codifferential operator applied on a one-form $\omega$ is minus the divergence of the co-vector $\delta \omega = -\partial_\alpha \omega^\alpha$. A straightforward calculation of each term leads to \begin{equation}\label{EQ-delta-ji-45} \begin{split} &\delta (j^5i_5+j^4i_4)d\phi=\\ &\frac{1}{\Delta}\left\{\left(-2(r+1)f+D^2\square_6f+\frac{D^4}{\Delta}F(F^2)\right)F(\phi)\right.\\ &\left. - \left(f\square_6 f - (r+3)F^2 + \frac{D^2}{\Delta} f F(F^2)\right) r \phi\right.\\ &\left. + D^2 F(F(\phi))\right \}. \end{split} \end{equation} Restricting to $X_f$ gives us \begin{equation*} \begin{split} \delta (j^5i_5+j^4i_4)d\phi=\\ - 2(r+1)\left(F\phi\right)^f - r \left((\square_6f)^f - (r+3)F_f^2\right)\phi^f. \end{split} \end{equation*} Finally, the formula for $(\square_6 \phi)^f$ becomes \begin{align}\label{EQ-Mainequation} \left(\square_{6}\phi\right)^f = \square_{f}\phi^f \nonumber\\ + 2(r+1)\left(F\phi\right)^f + r \left((\square_6f)^f - (r+3)F_f^2\right)\phi^f. \end{align} This is the main formula, announced before, linking explicitly the field equations of $\setR^6$ and $X_f$. As a byproduct it leads to an expression relating $f$ and the Ricci scalar of $X_f$, which is derived in the next section. \subsection{A relation between $R^f$ and $f$}\label{SEC-curvature} The expression (\ref{EQ-Mainequation}) applies in particular to the massless conformally coupled scalar field for which $r=-1$. With this particular value it becomes \begin{equation*} \left(\square_{6}\phi\right)^f = \square_{f}\phi^f - \left((\square_6f)^f - 2 F_f^2\right)\phi^f. \end{equation*} Now, the Klein-Gordon operator acting on the conformal scalar field, on a (locally conformally flat) manifold $X_f$, is well known and reads \begin{equation*} \left(\square_{f} - \frac{R^f}{6}\right)\phi^f, \end{equation*} where $R^f$ is the Ricci scalar of $X_f$. Apart from its homogeneity, $f$ is otherwise an arbitrary function, thus one must have \begin{equation}\label{EQ-R from f} \frac{R^f}{6} = (\square_6f)^f - 2 F_f^2. \end{equation} Since $F:=\widetilde{df}$, this relation allows us to compute the Ricci scalar $R^f$ from the function $f$ only. Finally, we remark that for the conformal scalar field, the constraint $F$ disappears of the Eq. (\ref{EQ-Mainequation}) due to the degree of homogeneity, $-1$, of the field. This is due to the fact that the conformal scalar field evolves, indeed, on the whole null cone $\mathcal{C}$. \section{The Klein-Gordon form of $(\square_6\phi)^f$}\label{SEC-KGForm} In the general expression (\ref{EQ-Mainequation}), neither the choice of the manifold defined by $f$, nor the constraints on the field, the action of $F$ and the degree of homogeneity (action of $D$), have been specified. In this section we examine, in the non-conformal case ($r\ne -1$), how constraints can be set in order to obtain on $X_f$ the Klein-Gordon operator acting on the scalar field $\phi^f$: \begin{equation*} (\square_{6}\phi)^f = (\square_{f} - \xi R^f + m^2)\phi^f. \end{equation*} The result is that, in order to solve the Klein-Gordon equation on a spacetime $X_f$ the homogeneous field $\phi$ solution of $\square_6 \phi = 0$ has to satisfy the constraint \begin{equation}\label{EQ-CnstraintFLRW} F\phi = G(y)\phi, \end{equation} where $G(y)$ is a function, homogeneous of degree $-1$, determined by $f$. For a FLRW spacetime, the vector $F$ and the function $G$ are explicitly determined in function of the scale factor $a$. An explicit realization in a convenient global coordinate system of this constraint is given in Appendix \ref{App-ConstrainFLRW}. In the following, we first consider the general FLRW spacetimes, then we re-examine the (A)dS spacetimes for which a specific and complete treatment is possible. \subsection{Robertson-Walker spaces}\label{SEC-RW} When $f$ defines a space which is not maximally symmetric, the terms $(\square_6 f)^f$ and $F^2_f$, which appear in the last term of (\ref{EQ-Mainequation}), are not constant. This reflects from (\ref{EQ-R from f}) the non-constancy of the scalar curvature $R^f$. As a consequence, the massive Klein-Gordon operator: $(\square_{6}\phi)^f = (\square_{f} - \xi R^f + m^2)\phi^f $, could be recovered only for a proper choice of the constraint $F\phi$. One can verify that unfortunately the simplest choice: $F\phi=0$, doesn't work, for instance, for a FLRW spacetime with scale factor $a(t) = t^\alpha$, which is an approximation of common use in cosmology (see Appendix \ref{App-ConstrainFLRW}). Indeed, the choice $F\phi = 0$ can only be made for the (A)dS spaces as we will see in Section \ref{SEC-AdS}. The choice of the constraint $F\phi$ is indicated by the Eq. (\ref{EQ-Mainequation}) which can be recast under the form of the Klein-Gordon equation $ (\square_{f} - \xi R^f + m^2)\phi^f =0 $ if \begin{equation}\label{EQ-Condition-KG-FLRW} \begin{split} 2(r+1)\left(F\phi\right)^f + r \left((\square_6f)^f - (r+3)F_f^2\right)\phi^f =\\ (-\xi R^f + m^2) \phi^f. \end{split} \end{equation} Then, $F\phi$ must be set in order to have $m^2$ constant. The simplest choice is \begin{equation}\label{EQ-Constaint-Fphi=Gphi} F\phi = G(y)\phi, \end{equation} where $G$ is a function homogeneous of degree $-1$ satisfying \begin{equation*} \begin{split} G^f = \frac{1}{2(r+1) }\left\{ m^2- \left( (r + 6 \xi) (\square_6f)^f \right. \right.\\ + \left.\left. \left(r(r+3) + 12 \xi\right) F^2_f \right)\right\}, \end{split} \end{equation*} where we have introduced the coupling constant $\xi$ to the curvature $R^f$ and used Eq. (\ref{EQ-R from f}). Apart from the case of maximally symmetric spaces considered below, no condition is present to determine the function $G$. We then resort to a simplicity criterion to choose that function: using the Eq. (\ref{EQ-delta-ji-45}) restricted to the cone ($D^2 = 0$), one can set \begin{equation}\label{EQ-G-FLRW} G = \alpha \square_6 f + \beta \frac{F^2}{f} + G_0 \frac{1}{f}, \end{equation} with \begin{equation*} \alpha := \frac{(r+6\xi)}{2(r+1)},~ \beta := \frac{r((r+3) + 12\xi)}{2(r+1)},~G_0:=\frac{m^2}{2(r+1)}. \end{equation*} The constraint (\ref{EQ-Constaint-Fphi=Gphi}) is then completely specified as soon as the function $f$ is known. In addition, as shown in Appendix \ref{App-ConstrainFLRW}, given the metric of some FLRW spacetime in some standard form (see \cite{}), one can relate the scale factor for that four dimensional space to the function $f$ in $\setR^6$ which realizes it as a space $X_f$. Thus, the constraint (\ref{EQ-Constaint-Fphi=Gphi}) can be recast in terms of the scale factor of that FLRW space extended to a function of $\setR^6$. All possible extensions of the scale factor obtained in that way restrict of course to the same FLRW metric. An explicit expression of (\ref{EQ-Constaint-Fphi=Gphi}) is obtained in Appendix \ref{App-ConstrainFLRW} by that procedure using the Minkowskian form of the FLRW metric $ds^2 = a^2ds^2_{{\sss M}ink}$ given in \cite{}. For the maximally symmetric spaces, (A)dS and Minkowski spaces, the situation is somehow different. Indeed, the method we use in this section, which grounds on the form (\ref{EQ-Constaint-Fphi=Gphi}) of the constraint, is not the most straightforward (nor probably the most useful for the well-known Minkowski space). It can nevertheless be adapted to those spaces. Such a modification is done for completeness in Appendix \ref{App-ConstraintMaxSym}. We now turn to the more direct treatment of (A)dS spaces. \subsection{(A)dS spaces}\label{SEC-AdS} We retrieve here a result obtained in \cite{} using a coordinates based derivation for (\ref{EQ-Mainequation}) specific to (A)dS spaces. Let us recall here, for completeness, these results in the present context. The de Sitter space can be obtained by setting \begin{equation}\label{Eq-fdSitter} f(y) := \frac{1}{2}(1+H^2)y^5 + \frac{1}{2}(1-H^2)y^4, \end{equation} $H$ being related to the Ricci scalar through $R= -12 H^2$. The Anti-de~Sitter space is obtained by changing the parameter $H^2$ to $-H^2$ in the above equation. Note that the Minkowski space is obtained by setting $H=0$. By contrast with the case of the FLRW spacetimes with non-constant scalar curvature, the choice $(F\phi)^f = 0$ can be made in Eq. (\ref{EQ-Mainequation}). Taking the above expression of $f$ into account Eq. (\ref{EQ-Mainequation}) becomes \begin{equation}\label{EQ-AdSMink} \left(\square_{6}\phi\right)^f = \square_{f}\phi^f - r(r+3)F^2 \phi^f, \end{equation} with $F^2 = F^2_f= H^2, - H^2$ for dS and AdS spaces respectively (note that $F^2 = 0$ corresponds to the Minkowski space). The above equation can be seen as the Klein-Gordon equation for all scalar representations of the (A)dS group for a suitable choice of the degree of homogeneity $r$. Precisely, since on (A)dS space, the Laplace-Beltrami operator $\square_f$, and the first order Casimir operator $\mathcal{Q}_1$, are related through $\square_f = - F^2 \mathcal{Q}_1$, all scalar UIRs can be retrieved. To this end, starting from Eq. (\ref{EQ-AdSMink}), let us set the parametrization \begin{equation*} - r(r+3)F^2\phi^f = (m^2 + 12 \xi F^2)\phi^f, \end{equation*} $m$ and $\xi$ being parameters which, in the massive representations, are identified with mass and coupling to the curvature. This equation is solved for $r$ and gives \begin{equation* r = -\frac{3}{2} + \sqrt{\frac{9}{4} - \left(12 \xi +\frac{m^2}{F^2}\right)}. \end{equation*} where the positive root is selected, in the AdS case because it gives the right degree of homogeneity to the conformally coupled scalar field, and in the dS case because of the symmetry $r\rightarrow -(r+3)$. The value of $12 \xi + m^2/F^2$ can be used in order to classify the involved representations of the (A)dS group (see for instance \cite{}). {We finally note that for Minkowski space, for which $F^2 =0$, the Eq. (\ref{EQ-Mainequation}) reduces, for $F\phi = 0$, to that of the conformal scalar. So that no mass term can be obtained by setting the constraint $F$ to zero.} \section{Conclusions}\label{SEC-Conclusion} In the present paper we have explicitly restricted the scalar operator $\square_6$ defined on $\setR^6$ to locally conformally flat four dimensional submanifolds of $\setR^6$ obtained as the intersection of hypersurfaces defined through a function $f$, homogeneous of degree one, and the null cone $\mathcal{C}$ of $\setR^6$. One consequence is that the determination of the solution of the Klein-Gordon equation: ${\displaystyle (\square_{f} - \xi R^f + m^2)\phi^f = 0}$, on a FLRW spacetime is reduced to the determination of the solution of the equations \begin{equation*} \left\{ \begin{aligned} & \square_6 \phi = 0\\ & F\phi = G(y) \phi, \end{aligned} \right. \end{equation*} in $\setR^6$, where $\phi$ is homogeneous of an unspecified degree $r$, $F=\sharp df$ is the operator gradient of $f$ (that is $F^\alpha = \partial^\alpha f$ in the basis associated with the Cartesian coordinates $\{y^\alpha\}$ of $\setR^6$) and $G$ a function, both explicitly determined by the scale factor of the FLRW spacetime (see Sec. \ref{SEC-RW}).

\section{Introduction}\label{intro} The observation of exotic stable multi-charge objects would represent striking evidence for physics beyond the Standard Model. Cosmological arguments put severe constraints on possible properties of such objects. Such particles (see e.g. Ref. \cite{DMRev} for review and reference) should be stable, provide the measured dark matter density and be decoupled from plasma and radiation at least before the beginning of matter dominated stage. The easiest way to satisfy these conditions is to involve neutral elementary weakly interacting massive particles (WIMPs). SUSY Models provide a list of possible WIMP candidates: neutralino, axino, gravitino etc., However it may not be the only particle physics solution for the dark matter problem. One of such alternative solutions is based on the existence of heavy stable charged particles bound in neutral "dark atoms". Dark atoms offer an interesting possibility to solve the puzzles of dark matter searches. It turns out that even stable electrically charged particles can exist hidden in such atoms, bound by ordinary Coulomb interactions (see \cite{DMRev,mpla,DDMRev} and references therein). Stable particles with charge -1 are excluded due to overproduction of anomalous isotopes. However, there doesn't appear such an evident contradiction for negatively doubly charged particles. There exist several types of particle models where heavy stable -2 charged species, $O^{--}$, are predicted: \begin{itemize} \item[(a)] AC-leptons, predicted as an extension of the Standard Model, based on the approach of almost-commutative geometry \cite{}. \item[(b)] Technileptons and anti-technibaryons in the framework of Walking Technicolor (WTC) \cite{KK,}. \item[(c)] stable "heavy quark clusters" $\bar U \bar U \bar U$ formed by anti-$U$ quark of 4th generation \cite{} \item[(d)] and, finally, stable charged clusters $\bar u_5 \bar u_5 \bar u_5$ of (anti)quarks $\bar u_5$ of 5th family can follow from the approach, unifying spins and charges\cite{Norma}. \end{itemize} All these models also predict corresponding +2 charge particles. If these positively charged particles remain free in the early Universe, they can recombine with ordinary electrons in anomalous helium, which is strongly constrained in terrestrial matter. Therefore a cosmological scenario should provide a mechanism which suppresses anomalous helium. There are two possible mechanisms than can provide a suppression: \begin{itemize} \item[(i)] The abundance of anomalous helium in the Galaxy may be significant, but in terrestrial matter a recombination mechanism could suppress this abundance below experimental upper limits \cite{}. The existence of a new U(1) gauge symmetry, causing new Coulomb-like long range interactions between charged dark matter particles, is crucial for this mechanism. This leads inevitably to the existence of dark radiation in the form of hidden photons. \item[(ii)] Free positively charged particles are already suppressed in the early Universe and the abundance of anomalous helium in the Galaxy is negligible \cite{mpla,I}. \end{itemize} These two possibilities correspond to two different cosmological scenarios of dark atoms. The first one is realized in the scenario with AC leptons, forming neutral AC atoms \cite{FKS}. The second assumes a charge asymmetry of the $O^{--}$ which forms the atom-like states with primordial helium \cite{mpla,I}. If new stable species belong to non-trivial representations of the SU(2) electroweak group, sphaleron transitions at high temperatures can provide the relation between baryon asymmetry and excess of -2 charge stable species, as it was demonstrated in the case of WTC \cite{KK,KK2,unesco,iwara}. After it is formed in the Standard Big Bang Nucleosynthesis (BBN), $^4He$ screens the $O^{--}$ charged particles in composite $(^4He^{++}O^{--})$ {\it $OHe$} ``atoms'' \cite{I}. In all the models of $OHe$, $O^{--}$ behaves either as a lepton or as a specific ``heavy quark cluster" with strongly suppressed hadronic interactions. The cosmological scenario of the $OHe$ Universe involves only one parameter of new physics $-$ the mass of O$^{--}$. Such a scenario is insensitive to the properties of $O^{--}$ (except for its mass), since the main features of the $OHe$ dark atoms are determined by their nuclear interacting helium shell. In terrestrial matter such dark matter species are slowed down and cannot cause significant nuclear recoil in the underground detectors, making them elusive in direct WIMP search experiments (where detection is based on nuclear recoil) such as CDMS, XENON100 and LUX. The positive results of DAMA experiments (see \cite{DAMAtalk} for review and references) can find in this scenario a nontrivial explanation due to a low energy radiative capture of $OHe$ by intermediate mass nuclei~\cite{mpla,DMRev,DDMRev}. This explains the negative results of the XENON100 and LUX experiments. The rate of this capture is proportional to the temperature: this leads to a suppression of this effect in cryogenic detectors, such as CDMS. OHe collisions in the central part of the Galaxy lead to OHe excitations, and de-excitations with pair production in E0 transitions can explain the excess of the positron-annihilation line, observed by INTEGRAL in the galactic bulge \cite{DMRev,DDMRev,KK2,CKWahe}. One should note that the nuclear physics of OHe is in the course of development and its basic element for a successful and self-consistent OHe dark matter scenario is related to the existence of a dipole Coulomb barrier, arising in the process of OHe-nucleus interaction and providing the dominance of elastic collisions of OHe with nuclei. This problem is the main open question of composite dark matter, which implies correct quantum mechanical solution \cite{CKW}. The lack of such a barrier and essential contribution of inelastic OHe-nucleus processes seem to lead to inevitable overproduction of anomalous isotopes \cite{CKW2}. Production of pairs of elementary stable doubly charged heavy leptons is characterized by a number of distinct experimental signatures that would provide effective search for them at the experiments at the LHC and test the atom-like structure of the cosmological dark matter. Moreover, astrophysical consequences of composite dark matter model can reproduce experimentally detected excess in positron annihilation line and high energy positron fraction in cosmic rays only if the mass of double charged $X$ particles is in the 1 TeV range (Section 2). We discuss confrontation of these predictions and experimental data in Section 3. The current status and expected progress in experimental searches for stable double charged particles as constituents of composite dark matter are summarized in the concluding Section 4. \section{Indirect effects of composite dark matter} \label{astro} The existence of such form of matter as O-helium should lead to a number of astrophysical signatures, which can constrain or prove this hypothesis. One of the signatures of O-helium can be a presence of an anomalous low Z/A component in the cosmic ray flux. O-helium atoms that are present in the Galaxy in the form of the dark matter can be destroyed in astrophysical processes and free $X^{−−}$ can be accelerated as ordinary charged particles. O-helium atoms can be ionized due to nuclear interaction with cosmic rays or in the front of a shock wave in the Supernova explosions, where they were effectively accumulated during star evolution \cite{I}. If the mechanisms of $X^{−−}$ acceleration are effective, the low Z/A component with charge −2 should be present in cosmic rays at the level of $F_X/F_p \sim 10^{-9}m^{-1}_o $ \cite{KK2}, and might be observed by PAMELA and AMS02 cosmic ray experiments. Here $m_o$ is the mass of O-helium in TeV, $F_X$ and $F_p$ are the fluxes of $X^{−−}$ and protons, respectively. \subsection{Excess of positron annihilation line in the galactic bulge} \label{bulge} Another signature of O-helium in the Galaxy is the excess of the positron annihilation line in cosmic gamma radiation due to de-excitation of the O-helium after its interaction in the interstellar space. If 2S level of O-helium is excited, its direct one-photon transition to the 1S ground state is forbidden and the de-excitation mainly goes through direct pair production. In principle this mechanism of positron production can explain the excess in positron annihilation line from the galactic bulge, measured by the INTEGRAL experiment. Due to the large uncertainty of DM distribution in the galactic bulge this interpretation of the INTEGRAL data is possible in a wide range of masses of O-helium with the minimal required central density of O-helium dark matter at $m_o = 1.25 \,{\rm TeV}$ \cite{integral1,integral2 For the smaller values of $m_o$ on needs larger central density to provide effective excitation of O-helium in collisions Current analysis favors lowest values of central dark matter density, making possible O-helium explanation for this excess only for a narrow window around this minimal value (see Fig. \ref{integral}) \begin{figure}[htbp] \begin{center} \includegraphics[scale=1]{integral.png} \caption{Dark matter is subdominant in the central part of Galaxy. It strongly suppresses it dynamical effect and causes large uncertainty in dark matter density and velocity distribution. At this uncertainty one can explain the positron line excess, observed by INTERGRAL, for a wide range of $m_o$ given by the curve with minimum at $m_o = 1.25 \,{\rm TeV}$. However, recent analysis of possible dark matter distribution in the galactic bulge favor minimal value of its central density.} \label{integral} \end{center} \end{figure} \subsection{Composite dark matter solution for high energy positron excess} \label{HEpositrons} In a two-component dark atom model, based on Walking Technicolor, a sparse WIMP-like component of atom-like state, made of positive and neg- ative doubly charged techniparticles, is present together with the dominant OHe dark atom and the decays of doubly positive charged techniparticles to pairs of same-sign leptons can explain the excess of high-energy cosmic-ray positrons, found in PAMELA and AMS02 experiments [17]. This explana- tion is possible for the mass of decaying +2 charged particle below 1 TeV and depends on the branching ratios of leptonic channels (See Fig. \ref{ams}). \begin{figure}[htbp] \begin{center} \includegraphics[scale=1]{ams.png} \caption{Best fit high energy positron fluxes from decaying composite dark matter in confrontation with the results of AMS02 experiment.} \label{ams} \end{center} \end{figure} Since even pure lepton decay channels are inevitably accompanied by gamma radiation the important constraint on this model follows from the measurement of cosmic gamma ray background in FERMI/LAT experiment. \begin{figure}[htbp] \begin{center} \includegraphics[scale=1.2]{fermi.png} \caption{Gamma ray flux accompanying the best fit high energy positron fluxes from decaying composite dark matter reproducing the results of AMS02 experiment, in confrontation with FERMI/LAT measurement of gamma ray background.} \label{fermi} \end{center} \end{figure} The multi-parameter analysis of decaying dark atom constituent model determines the maximal model independent value of the mass of decaying +2 charge particle, at which this explanation is possible $$m_o<1 TeV.$$ One should take into account that according to even in this range hypothesis on decaying composite dark matter, distributed in the galactic halo, can lead to gamma ray flux exceeding the measured by FERMI/LAT. \subsection{Sensitivity of indirect effect of composite dark matter to the mass of their double charged constituents} \label{mass} We see that indirect effects of composite dark matter strongly depend on the mass of its double charged constituents. To explain the excess of positron annihilation line in the galactic bulge mass of double charged constituent of O-helium should be in a narrow window around $$m_o = 1.25 \,{\rm TeV}.$$ To explain the excess of high energy cosmic ray positrons by decays of constituents of composite dark matter with charge +2 and to avoid overproduction of gamma background, accompanying such decays, the mass of such constituent should be in the range $$m_o < 1 \,{\rm TeV}.$$ These predictions should be confronted with the experimental data on the accelerator search for stable double charged particles. \section{Searches for stable multi-charged particles} \label{experiment} A~new charged massive particle with electric charge $\neq 1e$ would represent a~dramatic deviation from the~predictions of the~Standard Model, and such a~spectacular discovery would lead to fundamental insights and critical theoretical developments. Searches for such kind of particles were carried out in many cosmic ray and collider experiments (see for instance review in~\cite{fair}). Experimental search for double charged particles is of a~special interest because of important cosmological sequences discussed in previous sections. In a~tree approximation, such particles cannot decay to pair of quarks due to electric charge conservation and only decays to the~same sign leptons are possible. The~latter implies lepton number nonconservation, being a~profound signature of new physics. In general, it makes such states sufficiently long-living in a~cosmological scale. Obviously, such experiments are limited to look only for the~``long-lived'' particles, which do not decay within a~detector, as opposed to truly stable particles, which do not decay at all. Since the~kinematics and cross sections for double charged particle production processes cannot be reliably predicted, search results at collider experiments are usually quoted as cross section limits for a~range of charges and masses for well-defined kinematic models. In these experiments, the~mass limit was set at the~level of up to $100$~\,{\rm GeV}(see e.g. for review~\cite{fair}). The~CDF experiment collaboration performed a~search for long-lived double charged Higgs bosons ($H^{++}, H^{--}$) with $292$~pb$^{-1}$ of data collected in $p\bar{p}$ collisions at $\sqrt{s}=1.96$~\,{\rm TeV}{}~\cite{}. The~dominant production mode considered was $p\bar{p}\rightarrow \gamma^{\star}/Z+X\rightarrow H^{++}H^{--}+X$. Background sources include jets fragmenting to high-$\ensuremath{p_{\text{T}}}$ tracks, $Z\rightarrow ee$, $Z\rightarrow \mu \mu$, and $Z\rightarrow \tau \tau$, where at least one $\tau$ decays hadronically. Number of events expected from these backgrounds in the~signal region was estimated to be $<10^{-5}$ at $68\%$ confidence level (CL). Not a~single event with a~$H^{++}$ or $H^{--}$ was found in experimental data. This allows to set cross-section limit shown in Fig.~\ref{CDF_limits} as a~horizontal solid line. Next-to-leading order theoretical calculations of the cross-section for pair production of $H^{\pm\pm}$ bosons for left-handed and right-handed couplings are also shown in this figure. Comparison of expected and observed cross-section limits gives the~following mass constrains: $133$ and $109$~\,{\rm GeV}{} on the~masses of long-lived $H^{\pm\pm}_L$ and $H^{\pm\pm}_R$, respectively, at $95\%$ CL as shown in Fig.~\ref{CDF_limits}. In case of degenerate $H^{\pm\pm}_L$ and $H^{\pm\pm}_R$ bosons, the~mass limit was set to $146$~\,{\rm GeV}{}. \begin{figure}[htbp] \begin{center} \includegraphics[scale=1.2]{CDF_limits.pdf} \caption{The~comparison of the~experimental cross section upper limit with the~theoretical next-to-leading order cross section for pair production of $H^{\pm\pm}$ bosons. The~theoretical cross sections are computed separately for bosons with left-handed ($H^{\pm\pm}_L$) and right-handed ($H^{\pm\pm}_R$) couplings, and summed for the~case that their masses are degenerate,~\cite{}.} \label{CDF_limits} \end{center} \end{figure} \subsection{Searches at Large Hadron Collider} Significant increase of beam energy at the~Large Hadron Collider (LHC) opens a~new era in the~high energy physics and allows accessing uncharted territories of particle masses. In this section the~results of searches for the~multi-charged particles, performed by the~ATLAS and the~CMS collaborations at LHC, will be described. Calculations of the cross sections assume that these particles are generated as new massive spin-$1/2$ ones, neutral under SU(3)$_C$ and SU(2)$_L$. \subsubsection{ATLAS experiment at LHC} \label{ATLAS_search} In Run~1 (2010--2012), the~ATLAS~\cite{} collaboration at LHC performed two searches for long-lived multi-charged particles, including the~double charged particles: one search with $4.4$~fb$^{-1}$ of data collected in $pp$ collisions at $\sqrt{s}=7$~\,{\rm TeV}{}~\cite{}, and another one with $20.3$~fb$^{-1}$ collected at $\sqrt{s}=8$~\,{\rm TeV}{}~\cite{}. Both these searches feature particles with large transverse momentum values, traversing the~entire ATLAS detector. An~energy loss of a~double charged particle is by a~factor of $q^2=4$ higher than that of single charged particle. Such particles will leave a~very characteristic signature of high ionization in the~detector. More specifically, the~searches look for particles with anomalously high ionization on their tracks in three independent detector subsystems: silicon pixel detector (Pixel) and transition radiation tracker (TRT) in the~ATLAS inner detector, and monitoring drift tubes (MDT) in the~muon system. The~estimate of the~expected background originating from the~SM processes and providing input into the~signal region D was calculated to be $0.013\pm0.002 \textrm{(stat.)}\pm0.003\textrm{(syst.)}$ events. In order to define cross section, a~reconstruction efficiency of signal particles has to be known. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.45]{EfficienciesTrends_PAPER.pdf} \caption{The~signal efficiencies for different masses and charges of the~multi-charged particles for the~DY production model. Double charged particles are denoted as ``$z=2$'' (red points and line). The~picture is taken from~\cite{}.} \label{ATLAS_EfficiencyTrend} \end{center} \end{figure} This value is defined as a~fraction of simulated events with at least one multi-charged particle satisfying all of the~aforementioned criteria over the~number of all generated events. In other words, it is a~search sensitivity to find a~hypothetical particle with the~ATLAS experiment. These values are shown in Fig.~\ref{ATLAS_EfficiencyTrend} for each considered signal sample containing particles with charges from $2$ to $6$. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.45]{AllMasses_Xsections_AllChargesOnOneHisto_NoUncertaintyBands.pdf} \caption{Observed $95\%$ CL cross-section upper limits and theoretical cross-sections as functions of the multi-charged particles mass. Again, the~double charged particles are denoted as ``$z=2$'' (red points and lines). The~picture is taken from~\cite{}.} \label{ATLAS_Limits} \end{center} \end{figure} No events with double charged particles were found in neither 2011 or 2012 experimental data sets, setting the~lower mass limits to $430$ and $660$~\,{\rm GeV}{}, respectively, at $95\%$ CL. The~comparison of observed cross-section upper limits and theoretically predicted cross-sections is shown in Fig.~\ref{ATLAS_Limits}. \subsubsection {CMS experiment at LHC} \label{CMS_search} In parallel to the~ATLAS experiment, the~CMS~\cite{} collaboration at LHC performed a~search for double charged particles, using $5.0$~fb$^{-1}$ of data collected in $pp$ collisions at $\sqrt{s}=7$~\,{\rm TeV}{} and $18.8$~fb$^{-1}$ collected at $\sqrt{s}=8$~\,{\rm TeV}{}~\cite{}. The~search features particles with high ionization along their tracks in the~inner silicon pixel and strip tracker. Tracks with specific ionization $I_h>3$~\,{\rm MeV}{}/cm were selected. The muon system was used to measure the~time-of-flight values. Tracks with $1/\beta>1.2$ were considered. For the~part of the~search based on the~$\sqrt{s}=7$~\,{\rm TeV}{} data, the~number of events in the~signal region, expected from SM processes, was estimated to be $0.15\pm0.04$, whereas for the~$\sqrt{s}=8$~\,{\rm TeV}{} part it was $0.52\pm0.11$~events. The~uncertainties include both statistical and systematical contributions. $0$ and $1$~events were observed in the~signal regions for the~$7$ and $8$~\,{\rm TeV}{} analyses, respectively, which is consistent with the~predicted event rate. Comparison between observed upper cross section limits and theoretically predicted cross section values for the~$8$~\,{\rm TeV}{} is shown in Fig.~\ref{CMS_limits_8TeV}. \begin{figure}[htbp] \begin{center} \includegraphics[scale=0.6]{CMS_limits_8TeV.png} \caption{Observed $95\%$ CL cross-section upper limits and theoretical cross-sections as functions of the multi-charged particles mass in CMS search at the~$\sqrt{s}=8$~\,{\rm TeV}{}. The~double charged particles are denoted as ``$|Q|=2e$''. The~picture is taken from~\cite{}.} \label{CMS_limits_8TeV} \end{center} \end{figure} For the~$8$~\,{\rm TeV}{} search, the~mass limit of $665$~\,{\rm GeV}{} was obtained. This result (within uncertainties) is very close to the~ATLAS limit of $660$~\,{\rm GeV}{} for the~$8$~\,{\rm TeV}{} data set. Also, CMS treated the~results obtained at $7$ and $8$~\,{\rm TeV}{} as combined. This allowed to push the~lower mass limit to $685$~\,{\rm GeV}{} at $95\%$ CL. A~combination of the~results of two experiments for $8$~\,{\rm TeV}{} would mean an~increase of statistics by a~factor of $2$. Having said that, one can conclude that the~mass limit based on the~results of both experiment for double charged particles can be set at the~level of about $730$~\,{\rm GeV}{}. \subsubsection {What one expects from LHC Run~2} \label{LHC Run 2} Considering a~recent CMS Physics Analysis Summary~\cite{} and an~ATLAS paper in preparation, both on the~searches for double charged particles in data delivered by LHC to these experiments in 2015--2016, we may conclude that each of these two experiments will be able to set a~lower mass limit on the~double charged particles at $m=1000\pm50$~\,{\rm GeV}{}. Going further and considering the data taking periods of ATLAS and CMS until the end of Run 2 (end of 2018), we can estimate a~low limit on the~double charged particles mass corresponding to the~Run 2 data set. Several assumptions are made in these speculations: \begin{itemize} \item by the~end of 2018, ATLAS and CMS will each record about $120$~fb$^{-1}$ of $\sqrt{s}=13$~\,{\rm TeV}{} data; \item signal efficiency will remain at a~present level in both experiments, without being compromised by high detector occupancy or any other effects; \item no double charged particle candidates will be in observed in the~first place. \end{itemize} Considering all of the~above, the~ATLAS and CMS collaborations may each be expected to set the~lower mass limits at the~level of $1.2$~\,{\rm TeV}{} based on their analyses of the~entire $13$~\,{\rm TeV}{} data set. If these two experiments combined their independently gathered statistics together for this kind of search, the~limits would go as high as up to $1.3$~\,{\rm TeV}{}. \section{Conclusions} The existence of heavy stable neutral particles is one of the popular solutions for the dark matter problem. However, DM considered to be electrically neutral, potentially can be formed by stable heavy charged particles bound in neutral atom-like states by Coulomb attraction. Analysis of the cosmological data and atomic composition of the Universe gives the constrains on the particle charge showing that only -2 charged constituents, being trapped by primordial helium in neutral O-helium states, can avoid the problem of overproduction of the anomalous isotopes of chemical elements, which are severely constrained by observations. Cosmological model of O-helium dark matter can even explain puzzles of direct dark matter searches. Stable charge -2 states ($X^{--}$) can be elementary like AC-leptons or technileptons, or look like technibaryons. The latter, composed of techniquarks, reveal their structure at much higher energy scale and should be produced at colliders and accelerators as elementary species. They can also be composite like ``heavy quark clusters'' $\bar U \bar U \bar U$ formed by anti-U quark in one of the models of fourth generation \cite{I} or $\bar u_5 \bar u_5 \bar u_5$ of (anti)quarks $\bar u_5$ of stable 5th family in the approach \cite{Norma}. In the context of composite dark matter scenario accelerator search for stable doubly charged leptons acquires the meaning of direct critical test for existence of charged constituents of cosmological dark matter. The signature for AC leptons and techniparticles is unique and distinctive what allows to separate them from other hypothetical exotic particles. Composite dark matter models can explain observed excess of high energy positrons and gamma radiation in positron annihilation line at the masses of $X^{--}$ in the range of $1$~\,{\rm TeV}{}, it makes search for double charged particles in this range direct experimental test for these predictions of composite dark matter models. Test for composite $X^{--}$ can be only indirect: through the search for heavy hadrons, composed of single $U$ or $\bar U$ and light quarks (similar to R-hadrons). The~ATLAS and CMS collaborations at the~Large Hadron Collider are searching for the~double charged particles since 2011. The~most stringent results achieved so far exclude the~existence of such particles up to their mass of $680$~\,{\rm GeV}{}. This value was obtained by both ATLAS and CMS collaborations independently. It is expected that if these two collaborations combine their independently gathered statistics of LHC Run 2 (2015--2018), the~lower mass limit of double charged particles could reach the~level of about $1.3$~\,{\rm TeV}{}. \section*{Acknowledgements} We thank K.Belotsky, J.-R. Cudell, C. Kouvaris, M.Laletin for collaboration in development of the presented approach. The work was supported by Russian Science Foundation and fulfilled in the framework of MEPhI Academic Excellence Project (contract 02.a03.21.0005, 27.08.2013)..

\section{Introduction} Modern cosmology relies heavily in the observations and analysis of the cosmic microwave background (CMB) data. If the standard six parameter $\Lambda$CDM model provides a satisfying description of the outcome of the main cosmological probes \citep{planckcmb}, the determination of its parameters appears to be tension when separately considering the CMB angular power spectrum on the one hand, and the abundance of galaxy cluster in the local universe ($z < 1$) on the other \citep[see e.g.,][for a recent CMB-galaxy cluster joint analysis]{sal17}. One current challenge of modern cosmology is thus related to the understanding of the origin of this apparent tension. This could be produced by our lack of knowledge regarding the galaxy cluster mass-observable relations, or by new physics beyond the standard $\Lambda$CDM model affecting the structure growth between the CMB last scattering surface and the local universe \citep{sal17,planckSZC}. In this context, several probes can be used to trace the large scale distribution of matter in the universe. During their propagation along the line of sight, the CMB photons are affected by several physical processes such as the inverse Compton scattering expressed by the thermal Sunyaev-Zel'dovich effect \citep[hereafter tSZ,][]{sun69,sun72}, and the CMB gravitational lensing \citep{bla87} accounting for the deflections induced by the gravitational potential integrated along the line of sight, $\phi$. The hot electrons causing the tSZ effect are also emitting in the X-ray domain through Bremsstrahlung radiation, in such a way that the tSZ effect, X-ray emission, and gravitational lensing have been powerful sources of cosmological and astrophysical constraints \citep[see, e.g.,][]{planckSZC,planckphi}. However, the use of galaxy clusters as cosmological probes often requires the determination of their total mass, a complicated step usually relying on the hydrostatic equilibrium hypothesis. While such mass estimates are known to be biased, $M_{\rm hydro}/M_{\rm true} = 1-b_H$, hydrodynamical simulations favor low values for this bias, $b_H < 0.2$ \citep{lau13,bif16}. At the same time, on the observational side a significant amount of analyses combining X-ray, tSZ and weak lensing observations have been conducted in the last five years, either based on an object by object approach, or resorting to stacking algorithms \citep[see, e.g.,][]{hur17c,med17,ser17,jim17,par17,oka16,bat16,app16,smi16,hoe15,sim15,isr15,van14,don14,gru14,mah13}. These analyses obtain an hydrostatic mass bias in the range $b_H = 0.20 \pm 0.08$, which seem to be compatible with the measurement obtained after combining CMB weak lensing and tSZ measurements towards SDSS DR8 redMaPPer \citep{ryk14} galaxy clusters, $b_H = 0.26 \pm 0.07$ \citep{hur17c} . The statistical correlations of these tracers on the sky have also driven a lot of attention. For instance, the correlation between tSZ-X and tSZ-$\phi$ cross-analyses have been used to set cosmological constraints \citep[see, e.g.,][]{hil14,hur15a,hur15b}. These measurements present different sensitivities to cosmological parameters and mass-observable relations. Consequently, a coherent statistical analysis of $\phi$, tSZ, and X-ray emission is a powerful tool to identify the origin of the tension between the CMB and the low redshift universe. In the present paper, we present the first measurement of the CMB weak lensing and X-ray emission cross-correlation power spectrum. We model this cross-correlation using a halo-model formalism to derive cosmological constraints. Finally, we combine this result with previous studies to derive constraints on the hydrostatic mass bias. \section{Modelling the tSZ, X-ray, and $\phi$ cross-correlations} \label{crossth} \begin{figure*}[!h] \begin{center} \includegraphics[width=0.9\linewidth]{./figs/powspec_xszphi_dcl.eps} \caption{Power density for the one-halo term, $\frac{{\rm d}^2\ln{\rm (C_\ell)}}{{\rm d}{\rm \ln}(M_{500}){\rm d}\ln(z)}$, as a function of halo mass, $M_{500}$, and redshift, $z$ at $\ell = 200$. On the diagonal we display power density for the auto-correlation power spectra. From left to right: contribution to X-ray, CMB weak lensing, and tSZ angular power spectrum at $\ell=200$. Off-diagonal panels represent power density of cross-power spectra, from left to right and top to bottom referring to: X-ray--lensing, X-ray--tSZ, and lensing--tSZ cross power spectra. The color scale is the same for all panels and represents the total amount of correlation between probes.} \label{dcl} \end{center} \end{figure*} We refer to \citet{hur15a} for a detailed modeling of tSZ and $\phi$ cross-power spectra, and to \citet{hur14} for the modeling of the X-ray emission. We relate cosmological parameters to the dark matter halo number per unit of mass and redshift, $\frac{{\rm d^2N}}{{\rm d}M {\rm d}V}$, using the mass function from \citet{tin08}. \subsection{Poissonian term} \label{1-halo} Using the Limber approximation, we can write the one-halo term as \begin{equation} C_{\ell}^{\rm A\times B-{P}} = 4 \pi \int {\rm d}z \frac{{\rm d}V}{{\rm d}z {\rm d}\Omega}\int{\rm d}M \frac{{\rm d^2N}}{{\rm d}M {\rm d}V} W^{\rm P}_{\rm A} W^{\rm P}_{\rm B}. \end{equation} The tSZ contribution can be written as \begin{align} W^{\rm P}_{\rm tSZ} = Y_{500} y_{\ell}, \end{align} where $Y_{500}$ represents the tSZ flux of the clusters within a radius where the matter density equals 500 times the critical density at the clusters' redshift, related to the mass contained in the same volume ($M_{500}$) via the scaling law presented in \citet{planckSZC}. In this same expression, $y_\ell$ represents the Fourier transform on the sphere of the cluster pressure profile per unit of tSZ flux from \citet{planckSZC} . We consider a GNFW Universal pressure profile \citep{arn10,planckppp}. We model the lensing contribution as \begin{align} W^{\rm P}_{\phi}~=~-2 \psi_\ell \frac{(\chi' - \chi) \chi}{\chi'}, \end{align} with $\chi$ the comoving distance, $\chi'$ the comoving distance to the surface of the last scattering of the CMB, and $\psi_\ell$ the 3D lensing potential Fourier transform on the sky. We can express the potential $\psi$ as a function of the density contrast, \begin{align} \Delta \psi = \frac{3}{2} \Omega_m H_0^2 \frac{\delta_{\rm 3D}}{a}, \end{align} with $a$ the universe scale factor and $\delta$ the density contrast. From this, the lensing contribution reads \begin{align} W^{\rm P}_{\phi} = \frac{3 \Omega_m H_0^2 (1+z)}{c^2 \ell (\ell+1)} \frac{(\chi' - \chi) \chi}{\chi'} \delta_\ell, \end{align} where $\delta_{\ell}$ is the Fourier transform of the density contrast profile, $\delta_{\rm 3D}(u)$, computed as, \begin{equation} \delta_{\ell} = \frac{4 \pi r_{\rm 500}}{l^2_{\rm 500}} \int_0^{\infty} {\rm d}u \, u^2 \delta_{\rm 3D}(u) \frac{{\rm sin}(\ell u / \ell_{\rm 500})}{\ell u / \ell_{\rm 500}}, \end{equation} where $u = r/r_{\rm 500}$ is the normalized radius of the profile, $\ell_{\rm 500} = D_{\rm A}/r_{500}$, $D_{\rm A}$ is the angular diameter distance, and $r_{500}$ is the radius within which the matter density is 500 times the critical density of the universe.\\ Finally, the X-ray contribution can be written as \begin{align} W^{\rm P}_{\rm X} = S_{500} x_\ell, \end{align} with ${S}_{500} = \overline{C} L_{500}$, the X-ray count-rate in the [0.5-2.0] keV energy band of the host halo, $L_{500}$ the unabsorbed X-ray luminosity in the [0.1-2.4] keV energy range, $\overline{C}$ the average luminosity to count-rate conversion factor described in \citet{hur14}, and $x_\ell$ the Fourier transform of the X-ray number count profile. To model the $L_{500}-M_{500}$ relation, we used the relation derived by \citet{arn10} from the \textsc{REXCESS} sample \citep{boh07}. We considered a polytropic equation of state with a polytropic index of 1.2 to compute the density and the temperature profiles from the pressure profile. \\ \subsection{Large-scale correlation terms} \label{2halos} We express the large scale correlations, the two-halo term, contribution as \begin{align} C_{\ell}^{\rm A\times B-C} = 4 \pi \int {\rm d}z \frac{{\rm d}V}{{\rm d}z{\rm d}\Omega} W^{\rm C}_{\rm A} W^{\rm C}_{\rm B} P_k, \end{align} with $P_k$, the matter power-spectrum computed using {\tt CLASS} \citep{les11}.\\ For the tSZ effect, the CMB weak lensing, and the X-ray count rate we can express the window functions as, \begin{align} W^{\rm C}_{\rm tSZ} &= \int{\rm d}M \frac{{\rm d^2N}}{{\rm d}M {\rm d}V} Y_{500} y_{\ell} b_{\rm lin}, \nonumber \\ W^{\rm C}_{\phi} &= \frac{3 \Omega_m H_0^2}{c^2 \ell (\ell+1)} (1+z) \frac{(\chi' - \chi)}{\chi' \chi}. \nonumber \\ W^{\rm C}_{\rm X} &= \int{\rm d}M \frac{{\rm d^2N}}{{\rm d}M {\rm d}V} S_{500} x_{\ell} b_{\rm lin}. \end{align} where $b_{\rm lin}$ is the linear bias relating the halo distribution to the overdensity distribution. We considered the bias from \citet{mo96}, which is realistic on galaxy cluster scales. \\ We present in Fig.~\ref{dcl} the power density distribution in the $M_{500}$-$z$ plane for the (one-halo) X-ray, weak lensing, and tSZ auto- and cross-correlation power spectra at $\ell = 200$. We observe that weak-lensing favors higher redshift objects compared to the tSZ effect and X-ray emission. The X-ray and tSZ are highly correlated (at $\simeq 76\,$\%), while the tSZ effect and the weak lensing one-halo terms are moderately correlated ($\simeq 46$\,\%), and the X-ray emission and the weak lensing one halo-terms show a lower correlation at the level of $\simeq $\,20\%. We refer the reader to \citet{hur15a} and \citet{hur15b} for a detailed description of the mass and redshift dependence of the CMB weak lensing and X-ray window functions. In the light of Fig.~\ref{dcl}, it turns clear that detecting the X-ray--lensing cross-correlation is thus particularly challenging, considering the high-noise level of {\it Planck} CMB weak lensing maps, and the AGN-dominated X-ray sky. \section{First measurement of the cross-correlation between X-rays and weak-lensing} \label{secxphi} We use the ROSAT all-sky survey (RASS) public data\footnote{\url{ }}, which covers 99.8\% of the sky, including 97\% that has an exposure time longer than 100s \citep{vog99}. A description of the reprojection of the RASS data on the Healpix full-sky map can be found in \citet{hur15b}. We also use the {\it Planck} full-sky CMB weak lensing map \citep{planckphi}. We mask all regions with less than 100s exposure time in RASS full-sky map, while using the mask associated to the {\it Planck} all-sky weak lensing map. This combination of masks results in an effective sky fraction of $f_{\rm sky} \simeq 75$\% in the cross-correlation. We have also verified that using more aggressive galactic masks (down to a sky coverage of 50\%) does not modify the results significantly. \begin{figure}[!h] \begin{center} \includegraphics[width=0.9\linewidth]{./figs/powspec_xphi_data.eps} \caption{X-ray-$\phi$ cross-correlation angular power spectrum, measured between {\it Planck} CMB lensing full-sky map and a RASS data reprojected on the full-sky (black sample), the same cross-spectra when masking NVSS sources is shown as grey samples. The red solid line shows the theoretical prediction assuming ($\sigma_8 = 0.8$, $\Omega_{\rm m} = 0.3$, $b_H = 0.2$), the green dashed line shows the two-halo term contribution, and the blue dashed line the one halo term contribution. The solid cyan line displays the contribution from AGN-$\phi$ cross-correlation.} \label{xphi} \end{center} \end{figure} We compute the uncertainty in the weakly non-Gaussian limit as \begin{align} V(C_\ell^{x\phi}) = \frac{\left(C_\ell^{x\phi}\right)^2 + C_\ell^{xx}C_\ell^{\phi \phi}}{(2\ell +1)f_{\rm sky}}, \end{align} where $C_\ell^{x\phi}$, $C_\ell^{xx}$, are $C_\ell^{\phi \phi}$ are the X-ray-lensing, X-ray, and weak lensing power spectra. We obtained a significance of 9.1$\sigma$ for the X-ray-weak lensing cross-power spectrum in the range $\ell \in [0,400]$. We verified that AGNs do not produce a significant contribution to the total signal by masking BOSS AGNs \citep{alb16} and performing our analysis on the BOSS footprint ($f_{\rm sky} = 25$\%). We also masked NVSS bright sources, $S > 0.03$ Jy \citep{con98} and restrict our analysis to the NVSS footprint ($\delta > -40^{\rm o}$). We do not observe any significant modification of our results when masking the BOSS AGNs or NVSS sources. We also modeled the AGN contribution following the AGN mass and luminosity functions from \citet{hut14}. This contribution is about two orders of magnitude smaller than the X-ray-$\phi$ cross-correlation, as shown on Fig.~\ref{xphi}. Consequently, AGN impact on our results is small compared to the uncertainty level. We thus neglect the AGN contribution to the total X-ray-lensing cross-correlation. In Fig.~\ref{xphi}, we present the derived angular cross-power spectrum, compared to our modeling for ($\sigma_8 = 0.8$, $\Omega_{\rm m} = 0.3$, $b_H = 0.2$). We observe that the two halo term dominates for $\ell < 100$ and that the one halo term dominates at higher multipoles. This illustrates how {\it Planck} weak lensing map contains a significant signal produced by compact objects like galaxy clusters. Assuming a Gaussian prior of $b_H = 0.20 \pm 0.05$, consistent with hydrodynamical simulations, we derive $\Sigma_8 = \sigma_8 \left(\Omega_{\rm m}/0.30\right)^{0.27} = 0.80 \pm 0.03$. \section{Combined analysis of thermal Sunyaev-Zel'dovich effect, X-ray, and weak-lensing signals} \label{secdis} We next combine our results with tSZ--$\phi$, tSZ--tSZ, tSZ--X-ray, $\phi$--$\phi$, and CMB--CMB power spectra results. We assume that the CMB angular power spectrum is uncorrelated with all other probes\footnote{This is a reasonable approximation provided the low level of correlation induced by the integrated Sachs-Wolfe effect, which is restricted to very low multipoles ($\ell < 50$). We are also assuming that residual AGN and galaxy cluster tSZ residual contamination on the CMB should be of relevance at much higher multipoles than those considered here ($\ell > 1000$)}, so we compute the covariance between all large-scale structure tracers in the weakly non-Gaussian limit as, \begin{align} {\rm COV}(C_\ell^{\rm AB},C_\ell^{\rm CD}) = \frac{C_\ell^{\rm AC}C_\ell^{\rm BD} + C_\ell^{\rm AD}C_\ell^{\rm BC}}{(2\ell +1)f_{\rm sky}}, \end{align} where A, B, C and D stand for the tSZ effect, the CMB weak lensing signal, and the X-ray emission. We use the tSZ analysis from \citet{hur17a}, the tSZ--$\phi$ measurement corrected for cosmic infra-red background contamination from \citet{hur15a}, the tSZ--X-ray results from \citet{hur15b}, the {\it Planck} weak lensing results from \citet{planckphi}, and the {\it Planck} CMB results from \citet{planckcmb}. \begin{figure}[!h] \begin{center} \includegraphics[width=0.9\linewidth]{./figs/powspec_xphi_like.eps} \caption{Likelihood function of $\Sigma_8 = \sigma_8 \left(\Omega_{\rm m}/0.30\right)^{0.27}$ from different analyses: tSZ angular power spectrum, bispectrum, and number count \citep[red,][]{hur17a}, CMB weak lensing \citep[dark blue,][]{planckphi}, CMB angular power spectrum \citep[green,][]{planckcmb}, tSZ--weak lensing cross correlation \citep[cyan,][]{hur15a}, tSZ--X-ray cross-correlation \citep[orange,][]{hur15b}, and X-ray--$\phi$ cross-correlation (black, this work)} \label{like} \end{center} \end{figure} We performed a join fit of these results to derived cosmological constraints on $\sigma_8$ and $\Omega_{\rm m}$ and constraints on the hydrostatic mass bias, $b_H$. Both weak-lensing and CMB results are not sensitive to the hydrostatic mass bias. Consequently, weak lensing and CMB constraints set the cosmological parameters: $\sigma_8$ and $\Omega_{\rm m}$, whereas the large scale structure tracers, namely the tSZ effect and the X-ray emission, set the hydrostatic mass-bias value. From this combined analysis, we derived $b_H = 0.30 \pm 0.05$. We present the resultant likelihoods for tSZ, tSZ--X, tSZ--$\phi$, X--$\phi$, $\phi$, and CMB analyses for $b_H = 0.30 \pm 0.05$ in Fig.~\ref{like}. The X-ray auto-correlation power spectrum is not shown in this figure as it is particularly difficult to derive robust constraints from it, considering that the X-ray sky is dominated by AGN contribution. We observe that some tension remains, especially with tSZ derived constraints, but all large scale structure analyses presented here are consistent within 2\,$\sigma$ with the {\it Planck} CMB results. \section{Conclusion and discussion} \label{seccon} We have produced the first detection of the X-ray--$\phi$ cross-correlation angular power spectrum, with a significance of 9.1\,$\sigma$. We have established cosmological constraints on $\sigma_8$ and $\Omega_m$ from this cross-correlation, that we find consistent with previous large-scale structure \citep{hur15a,hur15b,hur17a,hur17c} and CMB analyses \citep{planckcmb,planckphi}. Similarly to the tSZ--X cross-correlation, the X--$\phi$ correlation favors values of the hydrostatic mass bias lower than those suggested in tSZ-CMB combined analyses \citep{sal17}. It also favors a higher value for $b_H$ than most of the weak-lensing based analyses of the last four years \citep[see e.g.,][]{med17,ser17,jim17,par17,oka16,bat16,app16,smi16,hoe15,sim15,isr15,don14,gru14,mah13}. This analyses prefer the range $b_H = 0.20 \pm 0.08$. With the constraint inferred here ($b_H = 0.30 \pm 0.05$), large scale structure cosmological constraints from the local universe on $\sigma_8$ and $\Omega_m$ now surround the CMB-based cosmological constraints. This result favors high value of $b$ compared to hydrodynamical simulations that prefer $b_H < 0.2$. Additionally, under the assumption that these measurements are systematic-free, the significant difference between tSZ and X-ray based results may indicate that these two probes select significantly different populations of galaxy clusters in terms of hydrostatic mass bias. \section*{Acknowledgment} \thanks{We acknowledge the use of HEALPix \citep{gor05}} \bibliographystyle{aa}

\section{Introduction}\label{}} \IEEEPARstart{A}{pproximate} nearest neighbor (ANN) search in a static database has achieved great success in supporting many tasks, such as information retrieval, classification and object detection. However, due to the massive amount of data generation at an unprecedented rate daily in the era of big data, databases are dynamically growing with data distribution evolving over time, and existing ANN search methods would achieve unsatisfactory performance without new data incorporated in their models. In addition, it is impractical for these methods to retrain the model from scratch for the continuously changing database due to the large scale computational time and memory. Therefore, it is increasingly important to handle ANN search in a dynamic database environment. ANN search in a dynamic database has a widespread applications in the real world. For example, a large number of news articles are generated and updated on hourly/daily basis, so a news searching system \cite{} requires to support news topic tracking and retrieval in a frequently changing news database. For object detection in video surveillance \cite{}, video data is continuously recorded, so that the distances between/among similar or dissimilar objects are continuously changing. For image retrieval in dynamic databases \cite{}, relevant images are retrieved from a constantly changing image collection, and the retrieved images could therefore be different over time given the same image query. In such an environment, real-time query needs to be answered based on all the data collected to the database so far. In recent years, there has been an increasing concern over the computational cost and memory requirement dealing with continuously growing large scale databases, and therefore there are many online learning algorithm works \cite{} proposed to update the model each time streaming data coming in. Therefore, we consider the following problem. Given a dynamic database environment, develop an online learning model accommodating the new streaming data with low computational cost for ANN search. Recently, several studies on online hashing \cite{} show that hashing based ANN approaches can be adapted to the dynamic database environment by updating hash functions accommodating new streaming data and then updating the hash codes of the exiting stored data via the new hash functions. Searching is performed in the Hamming space which is efficient and has low computational cost. However, an important problem that these works have not addressed is the computation of the hash code maintenance. To handle the streaming fashion of the data, the hash functions are required to be frequently updated, which will result in constant hash code recomputation of all the existing data in the reference database. This will inevitably incur an increasing amount of update time as the data volume increases. In addition, these online hashing approaches require the system to keep the old data so that the new hash code of the old data can be updated each time, leading to inefficiency in memory and computational load. Therefore, computational complexity and storage cost are still our major concerns in developing an online indexing model. \begin{figure*} \centering \includegraphics[scale=0.43]{hashing_vs_pq.png} \caption{\label{hashing_vs_pq} Hashing vs PQ in online update. The hash codes of the data points in the reference database will get updated if the hash functions get updated by the new data. \emph{The index of the codewords in the PQ codebook, on the other hand, will remain the same even though the codebook gets updated by the new data.} Thus online PQ is able to save severely much time by avoiding codewords maintenance of the reference database. (Best viewed in colors)} \end{figure*} Product quantization (PQ) \cite{} is an effective and successful alternative solution for ANN search. PQ partitions the original space into a Cartesian product of low dimensional subspaces and quantizes each subspace into a number of sub-codewords. In this way, PQ is able to produce a large number of codewords with low storage cost and perform ANN search with inexpensive computation. Moreover, it preserves the quantization error and can achieve satisfactory recall performance. Most importantly, unlike hashing-based methods representing each data instance by a hash code, which depends on a set of hash functions, quantization-based methods represent each data instance by an index, which associates with a codeword that is in the same vector space with the data instance. However, PQ is a batch mode method which is not designed for the problem of accommodating streaming data in the model. Therefore, to address the problem of handling streaming data for ANN search and tackle the challenge of hash code recomputation, we develop an online PQ approach, which updates the codewords by streaming data without the need to update the indices of the existing data in the reference database, to further alleviate the issue of large scale update computational cost. Figure \ref{hashing_vs_pq} compares hashing method and PQ in the code representation and maintenance, which illustrates the advantage of PQ over hashing in computational cost and memory efficiency. Once the index models get updated by the streaming data, the updated hash functions in hashing methods will produce new hash codes for each data point in the reference database, which will incur expensive cost for large scale databases. The updated product quantizer in PQ, on the other hand, updates the codewords in the codebook, but it does not change the index of the updated codewords of each data point in the reference database. To further reduce the update computational cost, we illustrate the idea of partial codebook update \cite{} and present two budget constraints for the model to update the codebook partially instead of all. Furthermore, we derive a loss bound which guarantees the performance of online PQ. Unlike traditional analysis, our model is a non-convex problem with matrices as the variables, so its theoretical analysis is not trivial to be handled. To emphasize the real-time data for querying, we also propose an online PQ model over a sliding window, which support both data insertion and deletion. \section{Related Works} \begin{table*}[] \centering \caption{Comparison between the existing methods and ours} \label{attribute-table} \scalebox{0.8}{ \begin{tabular}{|c|c|c|c|c|c|} \hline Method & \begin{tabular}[c]{@{}c@{}}Applicable to\\ streaming data\end{tabular} & \begin{tabular}[c]{@{}c@{}}Preserves the\\ quantization error\end{tabular} & \begin{tabular}[c]{@{}c@{}}Does not require \\ codewords maintenance\end{tabular} & \begin{tabular}[c]{@{}c@{}}Does not \\ require labels\end{tabular} & \begin{tabular}[c]{@{}c@{}}Does not require \\ to keep old data\end{tabular} \\ \hline \begin{tabular}[c]{@{}c@{}}Supervised data-dependent hashing\\ (\cite{})\end{tabular} & \xmark & \xmark & \xmark & \xmark & \xmark \\ \hline \begin{tabular}[c]{@{}c@{}}Unsupervised data-dependent hashing\\ (\cite{})\end{tabular} & \xmark & \xmark & \xmark & \cmark & \xmark \\ \hline \begin{tabular}[c]{@{}c@{}} Data-independent hashing\\ (\cite{})\end{tabular} & \cmark & \xmark & \xmark & \cmark & \cmark \\ \hline \begin{tabular}[c]{@{}c@{}} Online supervised hashing\\ (\cite{})\end{tabular} & \cmark & \xmark & \xmark & \xmark & \xmark \\ \hline \begin{tabular}[c]{@{}c@{}} Online unsupervised hashing\\ (\cite{})\end{tabular} & \cmark & \xmark & \xmark & \cmark & \xmark \\ \hline \begin{tabular}[c]{@{}c@{}} Quantization that requires codewords maintenance: \\AQ, CQ, SQ, TQ, KMH, ABQ, SSH (\cite{, })\end{tabular} & \xmark & \cmark & \xmark & \cmark & \xmark \\ \hline \begin{tabular}[c]{@{}c@{}} Quantization that does not require codewords maintenance: \\PQ, OPQ (\cite{})\end{tabular} & \xmark & \cmark & \cmark & \cmark & \xmark \\ \hline Proposed model & \cmark & \cmark & \cmark & \cmark & \cmark \\ \hline \end{tabular} } \end{table*} Hashing methods generate a set of hash functions to map a data instance to a hash code in order to facilitate fast nearest neighbor search. Existing hashing methods are grouped in data-independent hashing and data-dependent hashing. One of the most representative work for data-independent hashing is Locality Sensitive Hashing (LSH) \cite{}, where its hashing functions are randomly generated. LSH has the theoretical performance guarantee that similar data instances will be mapped to similar hash codes with a certain probability. Since data-independent hashing methods are independent from the input data, they can be easily adopted in an online fashion. Data-dependent hashing, on the other hand, learns the hash functions from the given data, which can achieve better performance than data-independent hashing methods. Its representative works are Spectral Hashing (SH) \cite{}, which uses spectral method to encode similarity graph of the input into hash functions, IsoH \cite{} which finds a rotation matrix for equal variance in the projected dimensions and ITQ \cite{} which learns an orthogonal rotation matrix for minimizing the quantization error of data items to their hash codes. To handle nearest neighbor search in a dynamic database, online hashing methods \cite{} have attracted a great attention in recent years. They allow their models to accommodate to the new data coming sequentially, without retraining all stored data points. Specifically, Online Hashing \cite{}, AdaptHash \cite{} and Online Supervised Hashing \cite{cakir2016online} are online supervised hashing methods, requiring label information, which might not be commonly available in many real-world applications. Stream Spectral Binary Coding (SSBC) \cite{} and Online Sketching Hashing (OSH) \cite{} are the only two existing online unsupervised hashing methods which do not require labels, where both of them are matrix sketch-based methods to learn to represent the data seen so far by a small sketch. However, all the online hashing methods suffer from the existing data storage and the high computational cost of hash code maintenance on the existing data. Each time new data comes, they update their hash functions accommodating to the new data and then update the hash codes of all stored data according to the new hash functions, which could be very time-consuming for a large scale database. Multi-codebook quantization (MCQ) methods \cite{, } are derived by minimizing the quantization error between the original input data and their corresponding codewords. Each codeword is represented by a set of sub-codewords selected from multiple codebooks. Please refer to \cite{matsui2018survey} for a comprehensive literature survey. To the best of our knowledge, no MCQ methods have been explored to an online fasion. Some of the popular works such as Composite Quantization (CQ) \cite{}, Sparse Composite Quantization (SQ) \cite{}, Additive Quantization (AQ) \cite{} and Tree Quantization (TQ) \cite{} require the codeword maintenance of the old data to update the codewords due to the constraints or structure of their models. Product Quantization (PQ) \cite{}, as one of the most classical MCQ method for fast nearest neighbor search, decomposes the input space into a Cartesian product of subspaces. The codeword of a data instance is represented by the concatenation of the sub-codeword of the data in all subspaces. More specifically, each sub-codeword is represented by an index. Thus the codeword of a data instance is the concatenation of the sub-codeword indices of all subspaces. Assuming that the codebook will not change much with an update by a streaming data, an online PQ model can be proposed. Under this assumption, there is no need of codewords maintenance of the existing data, as the indices of the sub-codewords of the data remain the same even though the actual sub-codewords are updated. It is different from online hashing methods, which requires the hash code update for the existing data each time the hash functions get updated. Extension works of PQ such as Optimized Product Quantization (OPQ) \cite{}, K-means Hashing (KMH) \cite{}, Adaptive Binary Quantization (ABQ) \cite{} and Structure Sensitive Hashing (SSH) \cite{} can also be developed to an online fashion. Specifically, KMH, ABQ and SSH approximate the distance between codewords by their Hamming distance, so they will require codewords maintenance in the online setting. In this paper, we focus on PQ method, so a novel online paradigm for PQ is proposed. The difference between existing methods and ours are summarized in Table \ref{attribute-table}. \section{Online Product Quantization with Budget Constraints} \begin{figure} \centering \includegraphics[scale=0.5]{online_update.png} \caption{\label{online_update} A general procedure for Online Product Quantization update. At each iteration, the new codebook gets updated by the streaming data. Searching can be performed using the latest codebook.} \end{figure} The traditional PQ method assumes that the database is static and hence it is not suitable for non-stationary data setting especially when the data is time-varying. Therefore, it is crucial to develop an online version of PQ to deal with dynamic database. A general procedure of our online product quantization framework is illustrated in Figure \ref{online_update}. The codebook at each iteration gets updated by the streaming data without retraining all the collected data. ANN search can be conducted against the latest codebook in terms of user queries. Unlike online hashing methods which update hashing functions and hash codes of the existing data, online PQ updates codebooks only and the codeword index of the existing data remains the same. \subsection{Preliminaries} We first define the vector quantization approach \cite{gray1998quantization} and the concept of quantization error, and then introduce Product quantization. Table \ref{notations} summarizes the notations frequently used in this paper. \begin{table}[] \centering \caption{Summary of Notations} \label{notations} \scalebox{0.87}{ \begin{tabular}{|l|l|} \hline \textbf{Notation} & \textbf{Definition} \\ \hline $\mathcal{X}$ & a set of data points \\ \hline $\mathcal{Z}$ & codebook \\ \hline $||\cdot||$ & $l_2$-norm \\ \hline $||\cdot||_{F}$ & Frobenius norm \\ \hline $t$ & iteration number. $t = 1, 2, ..., T$ \\ \hline $M$ & the number of subspaces (subquantizers) \\ \hline $K$ & the number of sub-codewords in each subspace \\ \hline $x^t \in \mathbb{R}^{D}$ & the streaming data in iteration $t$ \\ \hline $x_m^t \in \mathbb{R}^{\frac{D}{M}}$ & the subvector of the data in the $m$th subspace in iteration $t$ \\ \hline $z_{m,k}^t \in \mathbb{R}^{\frac{D}{M}}$ & the $k$th sub-codeword of the $m$th subspace in iteration $t$ \\ \hline $Z_{m}^{t} \in \mathbb{R}^{\frac{D}{M} \times K}$ & sub-codewords concatenation $[z_{m,1}^{t}, ..., z_{m,K}^{t}]$ \\ \hline $\mathcal{C}_{m,k}^{t}$ & the set of vector index assigned to the sub-codeword $z_{m,k}^t$ \\ \hline $n_{m,k}$ & the number of vectors assigned to the sub-codeword $z_{m,k}^t$ \\ \hline $\Delta z_{m,k}^t \in \mathbb{R}^{\frac{D}{M}}$ & \begin{tabular}[c]{@{}l@{}}the difference of the $z_{m,k}^t$ in two consecutive iterations\\ i.e., $z_{m,k}^{t+1} = z_{m,k}^{t} + \Delta z_{m,k}^t$\end{tabular} \\ \hline $\Delta Z_m^t \in \mathbb{R}^{\frac{D}{M} \times K}$ & differences concatenation $[\Delta z_{m,1}^t,...,\Delta z_{m,K}^t]$ \\ \hline $B$ & size of the mini-batch \\ \hline $\alpha$ & the number of subspaces to be updated \\ \hline $\lambda$ & the percentage of sub-codewords to be updated \\ \hline $L$ & size of the sliding window \\ \hline \end{tabular}} \end{table} \begin{definition}[Vector quantization \cite{gray1998quantization}] \label{def1} \emph{Vector quantization} approach quantizes a vector $x \in {\rm I\!R}^D$ to its codeword $z_k$ in a codebook $\mathcal{Z} = \{z_k\}$ where $k \in \{1,...,K\}$. \end{definition} \begin{definition}[Quantization error] \label{def2} Given a finite set of data points $\mathcal{X}$, vector quantization aims to minimize the \emph{quantization error} which is defined in the following: \begin{align*} \min_{\substack{\mathcal{C}_{1},...,\mathcal{C}_{K}\\ z_{1},...,z_{K}}} \sum_{i=1}^{|\mathcal{X}|}\|x^i - z_{k} \|^2 \end{align*} where $\mathcal{C}_{k}$ is the set of data indices assigned to the codeword $z_k$. \\ \textbf{\emph{Remark.}} In batch methods, ${C_k}$ can be automatically created after ${z_k}$ are decided. In online methods, however, ${C_k}$ is updated over time and it is not recomputed with respect to ${z_k}$ at each iteration. \end{definition} According to the first Lloyd's condition, $x^i$ should be mapped to its nearest codeword $z_k$ in the codebook. A codeword $z_k$ can be computed as the centroid of the vectors with index in $\mathcal{C}_k$. All of the codewords form the codebook $\mathcal{Z}$ with size $K$. Unlike vector quantization which uses one quantizer to map a vector, product quantization (PQ) uses $M$ subquantizers. It represents any $x \in {\rm I\!R}^D$ as a concatenation of $M$ sub-vectors $[x_1,...,x_m,...,x_M]$ where $x_m \in {\rm I\!R}^{D/M}$, assuming that $D$ is a multiple of $M$ for simplicity. The PQ codebook is then composed of $M$ sub-codebooks and each of the sub-codebook contains $K$ sub-codewords quantized from a distinct subquantizer. Any codeword belongs to the Cartesian product of the sub-codewords in each sub-codebook. The codeword of x is constructed by the concatenation of M sub-codewords $z = [z_{1,k_1},...,z_{m,k_m},...,z_{M,k_M}]$, where $z_{m,k_m}$ is the sub-codeword of $x_m$. \subsection{Online Product Quantization} Inspired by product quantization and online learning, the objective function of the online product quantization at each iteration $t$ is shown in the following: \begin{equation}\label{3} \begin{split} \min_{\substack{\mathcal{C}_{1,1}^t,...,\mathcal{C}_{m,k}^t,...,\mathcal{C}_{M,K}^t\\ z_{1,1}^t,...,z_{m,k}^t,...,z_{M,K}^t}} \sum_{m=1}^M\|x_m^{t} - z_{m,k}^t \|^2 \\ \end{split} \end{equation} where $x_m^t$ is the streaming data in the $m$th subspace in the $t$th iteration and its nearest sub-codeword is $z_{m,k}^t$. We expect to minimize the quantization error of the data at the current iteration $t$. Inspired by sequential vector quantization algorithm \cite{} to update the codebook, the solution of online PQ is shown in Algorithm \ref{alg:1}. \begin{algorithm} \caption{\label{alg:1} Online PQ} \begin{algorithmic}[1] \STATE initialize PQ with the $M*K$ sub-codewords $z_{1,1}^0,...,z_{m,k}^0,...,z_{M,K}^0$ using a initial set of data \\ \STATE initialize $C_{1,1}^0,...,C_{m,k}^0,...,C_{M,K}^0$ to be the cluster sets that contain the index of the initial data that belong to the cluster \STATE create counters $n_{1,1},...,n_{m,k},...,n_{M,K}$ for each cluster and initialize each $n_{m,k}$ to be the number of initial data points assigned to the corresponding $C_{m,k}^0$ \\ \FOR{$t=1, 2, 3, ...$} \STATE get a new data $x^t$\\ \STATE partition $x^t$ into $M$ subspaces $[x_1^t,...,x_M^t]$\\ \STATE in each subspace $m \in \{1,...,M\}$, determine and assign the nearest sub-codeword $z_{m,k}^t$ for each sub-vector $x_m^t$\\ \label{} \STATE update the cluster set $C_{m,k}^t \leftarrow C_{m,k}^{t-1} \cup \{ind\}$ $\forall m \in \{1,...,M\}$ where $ind$ is the index number of $x^{t}$ \\ \STATE update the number of points for each sub-codeword: $n_{m,k} \leftarrow n_{m,k} + 1$ $\forall m \in \{1,...,M\}$ \\ \label{} \STATE update the sub-codeword: $z_{m,k}^{t+1} \leftarrow z_{m,k}^t + \frac{1}{n_{m,k}}(x_m^t - z_{m,k}^t)$ $\forall m \in \{1,...,M\}$ \\ \label{} \ENDFOR \end{algorithmic} \end{algorithm} \subsection{Mini-batch Extension} In addition to processing one streaming data at a time, our framework can also handle a mini-batch of data at a time. In the case of processing mini-batch of data, we assume that each time we get a new batch of data points $X^t \in {\rm I\!R}^{B \times D}$ where $B$ is the size of the mini-batch. Its objective function is stated as the following: \begin{equation}\label{3-1} \begin{split} \min_{\substack{\mathcal{C}_{1,1}^t,...,\mathcal{C}_{m,k}^t,...,\mathcal{C}_{M,K}^t\\ z_{1,1}^t,...,z_{m,k}^t,...,z_{M,K}^t}} \sum_{m=1}^M\sum_{i=1}^{B}\|x_m^{t,i} - z_{m,k}^t \|^2 \\ \end{split} \end{equation} where $x_m^{t,i}$ is the $i$th streaming data of the current mini-batch in the $m$th subspace at the $t$th iteration and its nearest sub-codeword is $z_{m,k}^t$. Follow the aforementioned algorithm, we determine the sub-codeword for each sub-vector in each subspace for all data in the mini-batch, and update the counters accordingly for the determined sub-codewords. Finally, each determined sub-codewords can be updated as $z_{m,k}^{t+1} \leftarrow z_{m,k}^t + \frac{1}{n_{m,k}}\sum_{\{i \in \mathcal{C}_{m,k}^t\}}(x_m^{t,i} - z_{m,k}^t)$ $\forall m \in \{1,...,M\}$ where $x_m^{t,i}$ is a streaming data point in the $m$th subspace in the $t$th iteration with $z_{m,k}^t$ as its nearest sub-codeword. \subsection{Partial Codebook Update} \begin{figure} \centering \includegraphics[scale=0.31]{budget_constraint.png} \caption{\label{budget_constraint} A schematic figure of online product quantization with budget constraints. There are two subspaces where each subspace has two sub-codewords. After the codebook adapting to the new data, two of the four sub-codewords get hugely changed (highlighted in a red dashed rectangle) and the rest two sub-codewords barely changed.} \end{figure} As we mentioned in the introduction, one of the issues is that online indexing model might incur high computational cost in update. Each new incoming data point might contribute in different significance of changes in different subspaces of nearest sub-codeword update. An obvious example of this is that, given a new streaming data, one of its sub-vector is far from its nearest sub-codeword and another of its sub-vector is close to its nearest sub-codeword, then the first one contributes more in PQ index update than the second one. Moreover, mini-batch streaming data sometimes would result in a number of sub-codewords to be updated across different subspaces, and different sub-codewords (within or outside the same subspace) would have different significance of changes in update. It is worthless to update the sub-codeword when the update change is minimal. To better illustrate the idea, we show the update of a mini-batch of streaming data in Figure \ref{budget_constraint}. After assigning the nearest codeword to the new data, it shows that there is one sub-codeword in each subspace that is hugely different from its previous sub-codeword. The other two sub-codewords barely changed. Therefore, the update cost can be further reduced by ignoring the update of these sub-codewords that have less significant changes. Thus we can tackle the issue of possible high computational cost of update as we mentioned in the introduction by employing partial codebook update strategy, which can be achieved by adding one of the two budget constraints: the number of subspaces and the number of sub-codewords to be updated. \subsubsection{Constraint On Subspace Update} Each streaming data point is assigned to a sub-codeword in each subspace, so at least one sub-codeword in each subspace needs to be updated at each iteration. It is possible that the features in some subspaces of the new data have a vital contribution in their corresponding sub-codeword update and the features in some other subspaces have trivial contribution. Therefore, we target at updating the sub-codewords in the subspaces with significant update changes only. The subspace update constraint we add to our framework is used to update a subset of subspaces $\phi$. \begin{equation*} \phi \subseteq \{1,...,M\}, |\phi| \leq \alpha \end{equation*} where $\alpha$ represents the number of subspaces to be updated and $1 \leq \alpha \leq M$. We apply heuristics to get the optimal solution by selecting the top $\alpha$ subspaces with the most significant update. The significance of the subspace update can be computed by the sum of the quantization errors of streaming data. Thus Steps~\ref{} and ~\ref{} in Algorithm \ref{alg:1} are applied for determined sub-codewords in the selected top $\alpha$ subspaces. \subsubsection{Constraint On Sub-codeword Update} Specifically in the case of mini-batch streaming data update, it is likely that each mini-batch consists of different classes of data, which results in a number of sub-codewords in each subspace to be updated. Similar to subspace update constraint, we propose a sub-codeword update constraint to select a subset of sub-codewords $\psi$ to update. \begin{equation*} \psi \subseteq \{(1,1), ...,(m,k),...,(M,K)\}, |\psi| \leq \lambda MK \end{equation*} \noindent where $\lambda$ represents the percentage of sub-codewords to be updated and $0 \leq \lambda \leq 1$. Each $(m,k)$ tuple in $\psi$ represents the index of the sub-codeword $z_{m,k}$. Similar to the solution for handling the subspace update constraint, we select the top $\lambda MK$ sub-codewords with the highest quantization error and apply the update steps ~\ref{} and ~\ref{} to the selected sub-codwords. \subsection{Complexity Analysis} For our online PQ model with $M$ subspaces and $K$ sub-codewords in each subspace, the codebook update complexity for each iteration by a mini-batch of streaming data with size $B$ ($B \ge 1$) and $D$ dimensions is $\mathcal{O}(BKD + BM + BD)$, where these three elements represent the complexity of streaming data encoding, codewords counter update and codewords update respectively. This complexity can be reduced in two ways. First, the complexity of streaming data encoding, requiring the distance computation between data and each of the codewords, can be reduced by applying random projection methods such as LSH. In this case, its complexity can be reduced from $\mathcal{O}(BKD)$ to $\mathcal{O}(BKb)$, where $b$ is the number of bits learned by the random projection method. Thus both $\mathcal{O}(BKb)$ and $\mathcal{O}(BD)$ are dominant in the online model, especially for high-dimensional dataset (large $D$) with short hash codes (small $b$). Second, by applying the budget constraints in our model, the complexity of the codewords update $\mathcal{O}(BD)$ will get proportionally decreased to the constraint parameters. Note the overall update complexity does not depend on the volume of the database at the current iteration. \section{Loss Bound} In this section we study the relative loss bounds for our online product quantization, assuming our framework processes streaming data one at a time. Traditional analysis for online models are convex problems with vectors as variables. Our model, on the other hand, is non-convex and has matrices as variables, which makes the analysis non-trivial to be handled. Moreover, each of the continuously learned codewords may not be consistently matching with each codeword in the best fixed batch model. For example, a new incoming data may be assigned to the codeword with index 1 in our model but to the codeword with index 3 in the best batch model. This will make the loss bound even more difficult to study. Without using the properties of the convex function, we derive the loss bound of our model. Here we define ``loss" and ``codeword" analogous to ``prediction loss" and ``prediction", and follow the analysis in \cite{}. Since all $M$ subquantizers are independent to each other, we focus on the loss bound in a subquantizer $m$. The instantaneous quantization error function for our algorithm using the $m$th subquantizer during iteration $t$ is defined as: \begin{equation}\label{loss1} \begin{split} & \ell^{t}(Z_{m}^{t}) = ||x_{m}^{t} - z_{m,k}^t||^2 \end{split} \end{equation} where $Z_{m}^{t} \in \mathbb{R}^{\frac{D}{M} \times K}$ is the concatenation of all sub-codewords in the $m$th subspace $[z_{m,1}^{t}, ..., z_{m,K}^{t}]$ and $z_{m,k}^t$ is the closest sub-codeword to $x_{m}^{t}$. For convenience, we use $\ell_{Z_{m}}^{t}$ to denote $\ell^{t}(Z_{m}^{t})$ given $Z_{m}^t$ as the sub-codewords. To study the relative loss bound of our online model, by following \cite{}, we introduce an arbitrary matrix $U$, and we will prove that if there exists a matrix $U$ as the concatenated sub-codewords learned from the best batch model in hindsight (with the minimum quantization error for any $t$), our online model can converge to this batch model. Here we assume that $U \in {\rm I\!R}^{\frac{D}{M} \times K}$ concatenates $K$ sub-codewords $[u_{1}, ..., u_{K}]$. We arrange the order of these sub-codewords in the way that the column vectors of $Z_{m}^t$ and $U$ are paired by minimum distance, \emph{i.e.} $z_{m,k}^t$ matches $u_{k}$ by sub-codeword index to achieve a minimum $\sum_{k=1}^{K} {||z_{m,k}^t - u_{k}||^2}$, so that we will have streaming data assigned to the same index of the sub-codewords in $Z_{m}^t$ and $U$ as likely as possible. We use $\ell_{U}^{t}$ to denote the quantization error given $U$ as the sub-codewords: \begin{equation}\label{loss2} \begin{split} & \ell_{U}^{t} = \ell^t(U) = ||x_{m}^{t} - u_{k*}||^2 \end{split} \end{equation} where $u_{k*}$ is the closest sub-codeword to $x_{m}^{t}$. Following Lemma 1 in \cite{}, we derive the following lemma: \begin{lemma}\label{lemma1} Let $x_{m}^{1}, ..., x_{m}^{T}$ be a sequence of examples for the $m$th subspace where $x_{m}^{t} \in {\rm I\!R}^{\frac{D}{M}}$. Assume $||x_{m}^{t} - u_{k}|| - ||x_{m}^{t} - u_{k*}|| \leq \beta$ where $\beta$ is a constant. Let $\mathcal{T}_{n}$ be the set of iteration numbers where $z_{m,k}^{t}$ does not match $u_{k*}$, \emph{i.e.} $||u_{k} - u_{k*}|| \ne 0$. Using the notation provided in Eq.\ref{loss1} and Eq.\ref{loss2}, then \begin{equation*} \begin{split} & \sum_{t=1}^{T}\frac{1}{n_{m,k}^t} ((1 - \frac{1}{n_{m,k}^t})\ell_{Z_{m}}^{t} - \ell_{U}^{t} - ||x_{m}^{t}||^2 - ||u_{k}||^2) \\ & \indent \indent \indent \indent \indent \indent - \beta \sum_{t \in \mathcal{T}_{n}}\frac{1}{n_{m,k}^t} \leq ||Z_{m}^1 - U||^2 \end{split} \end{equation*} \end{lemma} where $Z_{m}^1$ is initialized to be nonzero matrix. \begin{proof} Following the proof of Lemma 1 in \cite{}, define $\Delta_{t}$ to be $||Z_m^t - U||_F^2 - ||Z_{m}^{t+1} - U||_F^2$. We obtain that, \begin{equation*} \sum_{t=1}^{T} \Delta_t \leq ||Z_m^1 - U||_F^2 \end{equation*} We now try to bound $\Delta_t$. If a sub-codeword of a streaming data point is not selected to be updated due to the budget constraint during iteration $t$, \emph{i.e.} $Z_m^{t+1}$ is the same as $Z_m^t$, then this $\Delta_t = 0$. Therefore, we focus on the iterations for which $||\Delta {Z_m^t}||_F \textgreater 0$. As $z_{m,k}^{t+1} = z_{m,k}^t + \Delta z_{m,k}^t$ where $\Delta z_{m,k}^t = \frac{x_m^t - z_{m,k}^t}{n_{m,k}^t}$ when $z_{m,k}^t$ is the codeword for $x_m^t$, we define $\Phi_{k} (\Delta z_{m,k}^t)$ as the difference between $Z_m^{t+1}$ and $Z_m^{t}$, where $\Delta z_{m,k}^t$ is the difference vector between the $k$th column vector of $Z_m^{t+1}$ and $Z_m^{t}$. We can therefore write $\Delta_t$ as, \begin{equation*} \begin{split} \Delta_t & = ||Z_m^t - U||_F^2 - ||Z_{m}^{t+1} - U||_F^2 \\ & = ||Z_m^t - U||_F^2 - ||Z_{m}^{t} - U + \Phi_{k} (\Delta z_{m,k}^t)||_F^2 \\ & = -2trace((Z_{m}^{t} - U)^{T}\Phi_{k} (\Delta z_{m,k}^t)) \\ & \indent \indent \indent \indent \indent \indent - trace(\Phi_{k} (\Delta z_{m,k}^t)^T \Phi_{k} (\Delta z_{m,k}^t)) \\ & = -2(z_{m,k}^{t} - u_k) \cdot \Delta z_{m,k}^t - ||\Delta z_{m,k}^t||^2 \\ & = \frac{1}{n_{m,k}^t}(-2 z_{m,k}^t \cdot x_m^t + 2 x_m^t \cdot u_k + 2||z_{m,k}^t||^2 \\ & \indent \indent \indent \indent \indent \indent \indent - 2 z_{m,k}^t \cdot u_k - \frac{||x_m^t - z_{m,k}^t||^2}{{n_{m.k}^t}}) \end{split} \end{equation*} From Eq.\ref{loss1} and Eq.\ref{loss2}, we obtain that $\ell_{Z_{m}}^{t} - ||x_m^t||^2 = ||z_{m,k}^t||^2 - 2 x_m^t \cdot z_{m,k}^t$ and $-\ell_{U}^{t} \leq 2x_m^t\cdot u_{k*}$. In addition, $-||u_k||^2 \leq ||z_{m,k}^t||^2 - 2z_{m,k}^t \cdot u_k$, If $z_{m,k}^t$ matches $u_{k*}$, then $u_{k*}$ is the same as $u_{k}$, then \begin{equation*} \Delta_t \geq \frac{1}{n_{m,k}^t}(\ell_{Z_{m}}^{t} - \frac{\ell_{Z_{m}}^{t}}{n_{m,k}^t} - \ell_{U}^{t} - ||x_m^t||^2 - ||u_k||^2) \end{equation*} If $z_{m,k}^t$ does not match $u_{k*}$, then based on the assumption $||x_{m}^{t} - u_{k}|| - ||x_{m}^{t} - u_{k*}|| \leq \beta$, \begin{equation*} \Delta_t \geq \frac{1}{n_{m,k}^t}(\ell_{Z_{m}}^{t} - \frac{\ell_{Z_{m}}^{t}}{n_{m,k}^t} - \ell_{U}^{t} - \beta - ||x_m^t||^2 - ||u_k||^2) \end{equation*} Overall, we obtain our conclusion. \end{proof} Following the Theorem 2 in \cite{} and our Lemma 1, we derive our theorem. \begin{theorem} \label{the2} Let $x_{m}^{1}, ..., x_{m}^{T}$ be a sequence of examples for the $m$th subspace where $x_{m}^{t} \in {\rm I\!R}^{\frac{D}{M}}$ and $||x_{m}^{t}||^2 \leq R^2$. Assume that there exists a matrix $U$ such that $\ell_{U}^{t}$ is minimized for all $t$, and $\max_{1\le k \le K} ||u_k||^2 \le F^2$. Then, the cumulative quantization error of our algorithm is bounded by \begin{equation}\label{bound} \begin{split} & \sum_{t=1}^{T} \ell_{Z_{m}}^{t} \textless 4(||Z_m^1 - U||_F^2 + \beta \sum_{t \in \mathcal{T}_{n}}\frac{1}{n_{m,k}^t}) \\ & \indent \indent \indent + 4T(R^2 + F^2) + 8 + 4\sum_{t=1}^{T} \ell_{U}^{t} \end{split} \end{equation} \end{theorem} \begin{proof} Since $1 \leq n_t^{m,k} \leq t$, then $\frac{1}{n^t_{m,k}} \leq 1$. Using the facts that $||x_{m}^{t}||^2 \leq R^2$ and $\max_{1\le k \le K} ||u_k||^2 \le F^2$, Lemma 1 implies that, \begin{equation*} \begin{split} \sum_{t=1}^{T}\frac{1}{n_{m,k}^t} (1 - \frac{1}{n_{m,k}^t})\ell_{Z_{m}}^{t} - \sum_{t=1}^{T}\ell_{U}^{t} \leq ||Z_{m}^1 - U||_F^2 \\ + \beta \sum_{t \in \mathcal{T}_{n}}\frac{1}{n_{m,k}^t} + T(R^2 + F^2) + 2 \end{split} \end{equation*} Since $\frac{1}{n_{m,k}^t} (1 - \frac{1}{n_{m,k}^t}) \leq \frac{1}{4}$, we get our relative loss bound. \end{proof} \begin{remark} If there exists a concatenated sub-codewords $U$ that is produced by the best fixed batch algorithm in hindsight for any $t$, then $\ell_{U}^{t}$, representing the quantization error of the batch method, can be minimal. The first term in the inequality (\ref{bound}) is the difference between the initialized codewords $Z_m^1$ of our online model and the best batch algorithm solution $U$. The second term represents the summation of the reciprocal of the counter $n_{m,k}$ that belongs to the updated cluster $C_{m,k}^t$ in the iterations when $z_{m,k}^t$ does not match $u_{k*}$, \emph{i.e.}, the streaming data belongs to two different indices of the clusters from our online model and the best batch model. A tighter bound can be achieved if the initialized sub-codewords is close to the optimal sub-codewords and $z_{m,k}^t$ matches $u_k$ for each iteration. Since all the terms except for the third one in the inequality (\ref{bound}) are constants, and $(R^2 + F^2)$ is also constant, the cumulative loss bound scales linearly with the number of iterations $T$. Thus the performance of online PQ model is guaranteed for unseen data. \end{remark} \begin{remark} The proved online loss bound above can be used to obtain the generalization error bound \cite{}, and the generalization error bound will show that the codewords of our online method are similar to the ones of the batch method. Therefore, the quantization error of our online model can converge to the error of the batch algorithm. Its corresponding experimental result is shown in Section~\ref{convergence-section}. \end{remark} \section{Online Product Quantization over a Sliding Window} From the cumulative loss bound proved in Theorem~\ref{the2}, we can see that it is important to reduce the number of iterations when $z_{m,k}^t$ does match $u_k$ in order to achieve a tighter bound. One reason for the mismatched case is the tradeoff between update efficiency and quantization error. The old data and the new data that belong to the same codeword in the online model may be generated from different data distributions. Thus it gives us the insight to consider the sliding window approach of handling the streaming data, where the updated codewords will be emphasized on the real-time data and the expired data will be removed. Using this approach allows us to reduce the effect of the codewords mismatching iterations. In the context of data streams, data distribution evolves over time. Some applications may aim at capturing the real-time behaviour of the data streams and thus emphasize the most recent data points in the stream. As the data stream evolves, some data will be expired based on their arrival times. Leveraging a sliding window in the model allows us to reflect the real-time and evolving behaviour of the data streams in order to facilitate the most recent data maintenance. Data insertion is performed to the online indexing model using the recent data in the sliding window, and data deletion removes the contributions of the codebook made by the expired data. In this section, we present the online PQ model over a time-based sliding window with both data insertion and deletion supported. \begin{figure} \centering \includegraphics[scale=0.45]{streaming_comparison.png} \caption{\label{streaming_comparison} Approaches of Handling Data Streams. Streaming: data streams one at a time. Mini-batch: a mini-batch of data with size 3 is processed by the model at each iteration t. Sliding window: a moving window with size 3 applied on continuously changing data.} \end{figure} \subsection{Online Product Quantization with Data Insertion and Deletion} Assume we are given a sliding window of size $L$. For window at the $t$th iteration, it consists of a stream of data $x^{t,1}, x^{t,2}, ...,x^{t,L}$, where $x^{t,L}$ is the newly inserted data point to the window at the current iteration, and $x^{t-1,1}$ is just expired and removed from the window. Therefore, the objective function of the Online PQ over a sliding window at the $t$th iteration is stated as the following: \begin{equation}\label{5} \begin{split} \min_{\substack{\mathcal{C}_{1,1}^t,...,\mathcal{C}_{m,k}^t,...,\mathcal{C}_{M,K}^t\\ z_{1,1}^t,...,z_{m,k}^t,...,z_{M,K}^t}} \sum_{m=1}^M\sum_{i=1}^{L}\|x_m^{t,i} - z_{m,k}^t \|^2 \\ \end{split} \end{equation} where $x_m^{t,i}$ is the $i$th streaming data of the window in the $m$th subspace at the $t$th iteration and its nearest sub-codeword is $z_{m,k}^t$. To emphasize the real-time data in the stream, we want our model only affected by the data in the sliding window at the current iteration. Therefore, each time a new data streaming into the system, it moves to the sliding window. We first update the codebook by adding the contributions made by the new data. Correspondingly, the oldest data in the sliding window will be removed. We tackle the issue of data expiry by deleting the contribution to the codebook made by the data point that is just removed from the window. The solution of online PQ over a sliding window to handle insertion and deletion to the codebook is shown in Algorithm \ref{alg:2}: \begin{algorithm} \caption{\label{alg:2} Online PQ over a Sliding Window} \begin{algorithmic}[1] \STATE initialize PQ with the $M*K$ sub-codewords $z_{1,1}^0,...,z_{m,k}^0,...,z_{M,K}^0$ \\ \STATE initialize $C_{1,1}^0,...,C_{m,k}^0,...,C_{M,K}^0$ to be the cluster sets that contain the index of the initial data that belong to the cluster \STATE create counters $n_{1,1},...,n_{m,k},...,n_{M,K}$ for cluster and initialize each $n_{m,k}$ to be the number of initial data points assigned to the corresponding $C_{m,k}^0$ \\ \STATE initialize the size of the sliding window $L$ \FOR{$t=1, 2, 3, ...$} \STATE remove $x^{t-1,1}$ from the sliding window \STATE $x^{t-1,i} (2 \le i \le L)$ from the $(t-1)$th sliding window is now represented as $x^{t,i-1}$ at the current iteration \STATE insert the new data $x^{t,L}$ to the sliding window at position $L$\\ \STATE partition the newly inserted data $x^{t,L}$ and the deleted one $x^{t-1,1}$ into $M$ subspaces $[x_1^{t,L},...,x_M^{t,L}]$ and $[x_1^{t-1,1},...,x_M^{t-1,1}]$, respectively\\ \textbf{Insertion:} \FOR{$m=1,...,M$} \STATE determine and assign the nearest sub-codeword $z_{m,k}^t$ for each sub-vector $x_m^{t,L}$\\ \STATE update the cluster set $C_{m,k}^t \leftarrow C_{m,k}^{t-1} \cup \{ind\}$ where $ind$ is the the index number of $x^{t,L}$ \STATE update the counter for $C_{m,k}^t$: $n_{m,k} \leftarrow n_{m,k} + 1$ \\ \STATE update the sub-codeword: $z_{m,k}^{t+1} \leftarrow z_{m,k}^t + \frac{1}{n_{m,k}}(x_m^{t,L} - z_{m,k}^t)$ \\ \ENDFOR \\ \textbf{Deletion:} \FOR{$m=1,...,M$} \STATE determine and assign the nearest sub-codeword $z_{m,k}^t$ for each sub-vector $x_m^{t-1,1}$\\ \STATE update the cluster set $C_{m,k}^t \leftarrow C_{m,k}^{t-1} \setminus \{ind\}$ where $ind$ is the the index number of $x^{t-1,1}$ \STATE update the counter for $C_{m,k}^t$: $n_{m,k} \leftarrow n_{m,k} - 1$ \\ \STATE update the sub-codeword: $z_{m,k}^{t+1} \leftarrow z_{m,k}^t - \frac{1}{n_{m,k}}(x_m^{t-1,1} - z_{m,k}^t)$ \\ \ENDFOR \ENDFOR \end{algorithmic} \end{algorithm} \subsection{Connections among Online PQ Algorithms} Figure \ref{streaming_comparison} compares three different approaches of handling data streams for our online PQ model. Streaming Online PQ processes new data one at a time. Mini-batch Online PQ processes a mini-batch of data at a time. If the size of the mini-batch is set to 1, Streaming Online PQ is the same as Mini-batch Online PQ. Online PQ over a Sliding Window involves the deletion of the expired data that is removed from the moving window at the current iteration. If the size of the sliding window is set to be infinite, no data deletion will be performed and it is the same as Streaming or Mini-batch Online PQ depending on the size of the new data to be updated at each iteration. \section{Experiments} We conduct a series of experiments on several real-world datasets to evaluate the efficiency and effectiveness of our model. In this section, we first introduce the datasets used in the experiments. We then show the convergence of our online PQ model to the batch PQ method in terms of the quantization error, and then compare the online version and the mini-batch version of our online PQ model. After that, we analyze the impact of the parameters $\alpha$ and $\lambda$ in update constraints. Finally, we compare our proposed model with existing related hashing methods for different applications. \subsection{Datasets and evaluation criterion} There are one text dataset, four image datasets and two video datasets employed to evaluate the proposed method. \textbf{20 Newsgroups Data} (News20) \cite{} consists of chronologically ordered 18,845 newsgroup messages. \textbf{Caltech-101} \cite{} consists of 9144 images and each image belongs to one of the 101 categories. \textbf{Half dome} \cite{} includes 107,732 image patches obtained from Photo Tourism reconstructions from Half Dome (Yosemite). \textbf{Sun397} \cite{} contains around 108K images in 397 scenes. \textbf{ImageNet} \cite{} has over 1.2 million images with a total of 1000 classes. \textbf{YoutubeFaces}\footnote{} contains 3,425 videos of 1,595 different people, with a total of 621,126 frames. \textbf{UQ\_VIDEO}\footnote{} consists of 169,952 videos with 3,305,525 frames in total. We use 300-D doc2vec features to represent each news article in News20 and 512-D GIST features to represent each image in the four image datasets. We use two different features, 480-D Center-Symmetric LBP (CSLBP) and 560-D Four-Patch LBP (FPLBP) to represent each frame in YoutubeFaces. 162-D HSV feature is used in UQ\_VIDEO dataset. Table \ref{} shows detailed statistical information about datasets used in evaluation. We measure the performance of our proposed model by the model update time and the search quality measurement recall@R adopted in \cite{}. We use recall@20 which indicates that fraction of the query for which the nearest neighbor is in the top 20 retrieved images by the model. \begin{table}[] \centering \caption{Detailed datasets information} \label{} \begin{tabular}{|c|c|c|l|c|} \hline \textbf{Dataset} & \textbf{Class no.} & \multicolumn{2}{c|}{\textbf{Size}} & \textbf{Feature} \\ \hline News20 & 20 & \multicolumn{2}{c|}{18,845} & Doc2vec (300) \\ \hline Caltech-101 & 101 & \multicolumn{2}{c|}{9,144} & GIST (512) \\ \hline Half dome & 28,086 & \multicolumn{2}{c|}{107,732} & GIST (512) \\ \hline Sun397 & 397 & \multicolumn{2}{c|}{108,753} & GIST (512) \\ \hline ImageNet & 1000 & \multicolumn{2}{c|}{1,281,167} & GIST (512) \\ \hline \textbf{Video Dataset} & \textbf{Video} & \multicolumn{2}{c|}{\textbf{Frame}} & \textbf{Feature} \\ \hline YoutubeFaces (CSLBP) & 3,425 & \multicolumn{2}{c|}{621,126} & CSLBP (480) \\ \hline YoutubeFaces (FPLBP) & 3,425 & \multicolumn{2}{c|}{621,126} & FPLBP (560) \\ \hline UQ\_VIDEO & 169,952 & \multicolumn{2}{c|}{3,304,554} & HSV (162) \\ \hline \end{tabular} \end{table} \subsection{Convergence}\label{convergence-section} \begin{figure}[!htbp] \centering \includegraphics[scale=0.5]{convergence.png} \caption{\label{convergence} Convergence of online PQ using ImageNet dataset. Effective iterations are shown on the x-axis.} \end{figure} The data instances in the entire dataset are input sequentially to our online PQ model. We run our algorithm for 50 effective iterations\footnote{Running one pass using the entire dataset is an effective iteration}. To show the convergence of our online model, we compare its training loss at each iteration with the one of the batch PQ method. The training loss is computed as the averaged quantization error for all data points in one pass. Figure \ref{convergence} shows that the training loss of our online model converges to the one of the batch model, implying that codewords learned from the online PQ model are similar to the ones learned from the batch PQ approach. Therefore, the performance of the online PQ model converges to the batch PQ performance. \begin{figure}[!htbp] \centering \includegraphics[scale=0.41]{online_vs_mini_batch.png} \caption{\label{online_vs_minibatch} The left figure shows the update time for each iteration of update. The time of the online version for each iteration sums up the update time of the streaming data corresponding to the ones in the mini-batch. The right figure shows the recall@1, 20 and 100 for each iteration. } \end{figure} \subsection{Online vs mini-batch} In real-world applications, streaming data might be processed one at a time or in small batches. For example, real-time topic detection in streaming media can be applied on texts, images or videos. If the streaming data is Twitter post, it might be processed one at a time. If the media is video, then the streaming data can be processed in mini-batches of video frames. Our model can process streaming data either one at a time or in mini-batches. We compare these two versions of our model on Caltech-101 dataset. For our model, we use $M = 8$ and $K = 256$. Since the memory cost of storing each codeword is $M\ceil{\log_{2} K}$ bits \cite{}, where $\ceil{.}$ is the ceiling function that maps the value to its nearest integer up, then the number of bits used in our model is 64. We split the data into twelve groups, where one of the groups is used for learning the codebook, one of the groups is used as the query and each one of the rest of the ten groups is used to update the original codebook, so that we have the performance for ten iterations. Figure \ref{online_vs_minibatch} shows the comparison of the online version and the mini-batch (MB) version in update time and recall@R measurements. It indicates that the mini-batch version takes much less update time than the online version but they have similar search quality. Therefore, we adopt mini-batch version of our model in the rest of the experiments. \begin{figure}[!htbp] \centering \includegraphics[scale=0.565]{tradeoff.png} \caption{\label{tradeoff} Trade-offs between update time cost and the search accuracy. The first column shows the impact of the subspace update constraint. The second column shows the impact of the sub-codeword update constraint. The red line is the reference line for the scatter plot.} \end{figure} \subsection{Update time complexity vs search accuracy} There are two budget constraints we proposed for the codebook update to further reduce the time cost: number of subspaces and number of sub-codewords to be updated. In this experiment, we evaluate the impact of these two constraints and the trade-offs between update time cost and the search quality using a synthetics dataset. We randomly sampled 12000 data points with 128-D features from multivariate normal distribution. We use recall@50 as the performance measurement which indicates that fraction of the query for which the nearest neighbor is in the top 50 retrieved data points by the model. We set $M = 16$ and $K = 256$, and vary the number of updated subspaces from 1 to 16 and the portion of updated sub-codewords from 0.1 to 1 respectively. We split the dataset evenly into twelve groups and set one of the groups as the learning set to learn the codebook and another one as the query set, and use each of the rest of ten groups to update the learned codebook and record the update time cost and the search accuracy for 10 times while varying the update constraints. From Figure \ref{tradeoff}, we observe that the search quality and update time cost strongly depend on these two update constraints. As we increase the number of subspaces or the portion of sub-codewords to be updated, the update time cost is increasing, along with the search accuracy. Therefore, higher update time cost is required for better search accuracy. \subsection{Baseline methods} To verify that our online PQ model is time-efficient in update and effective in nearest neighbor search, we make comparison with several related online indexing and batch learning methods. We evaluate the performance in both search accuracy and update time cost. We select SSBC \cite{} and OSH \cite{} as two of the baseline methods, as they are the only unsupervised online indexing methods to the best of our knowledge. Specifically, SSBC is only applied on two of our smallest datasets, News20 and Caltech-101, as it takes a significant amount of time to train. In addition, two supervised online indexing methods OH \cite{} and AdaptHash \cite{} are selected, with the top 5 percentile nearest neighbors in Euclidean space as the ground truth following the setting as in \cite{}. Further, five batch learning indexing methods are selected. They are all unsupervised data-dependent methods: PQ \cite{}, spectral hashing (SH) \cite{}, IsoH \cite{}, ITQ \cite{} and KMH \cite{}. Each of these methods is compared with online PQ in two ways. The first way uses all the data points seen so far to retrain the model (batch) at each iteration. The second way uses the model trained from the initial iteration to assign the codeword to the streaming data in all the rest of the iterations (no update). We do not apply the batch learning methods on our large-scale datasets, UQ\_VIDEO and ImageNet, as it takes too much retraining time at each iteration. \subsection{Object tracking and retrieval in a dynamic database} In many real-world applications, data is continuously generated everyday and the database needs to get updated dynamically by the newly available data. For example, news articles can be posted any time and it is important to enhance user experience in news topic tracking and related news retrieval. New images with new animal species may be inserted to the large scale image database. Index update needs to be supported to allow users to retrieve images with expected animal in a dynamically changing database. A live video or a surveillance video may generate several frame per second, which makes the real-time object tracking or face recognition a crucial task to solve. In this experiment, we evaluate our model on how it handles dynamic updates in both time efficiency and search accuracy in three different types of data: text, image and video. \subsubsection{Setting}\label{setting} \begin{table*}[!htbp] \centering \caption{\small Number of iterations and average mini-batch size for each dataset} \label{} \begin{tabular}{|c|c|c|c|c|c|c|c|} \hline \textbf{Dataset} & \textbf{News20} & \textbf{Caltech-101} & \textbf{Half dome} & \textbf{Sun397} & \textbf{ImageNet} & \textbf{YoutubeFaces} & \textbf{UQ\_VIDEO} \\ \hline \textbf{Iteration No.} & 20 & 12 & 12 & 21 & 101 & 67 & 25 \\ \hline \textbf{Avg mini-batch size} & 942.25 & 762 & 8,977.7 & 5,178.7 & 12,684.8 & 8,628.4 & 132,182 \\ \hline \end{tabular} \end{table*} \begin{figure} \centering \begin{subfigure}[b]{0.5\textwidth} \centering \includegraphics[width=1\textwidth]{online_legend.png} \end{subfigure} \begin{subfigure}[b]{0.48\textwidth} \centering \centerline{\includegraphics[width=9cm]{news20_diff_bits.png}} \end{subfigure} \caption[] {\small Recall@20 performance (1st row) and Update time cost (2nd row) comparison against online hashing methods at each iteration for News20 using different number of bits. 1st column: 32 bits. 2nd column: 48 bits. 3rd column: 64 bits. Time cost is in log scale.} \label{news20_diff_bits} \end{figure} For each dataset, we split data into several mini-batches, and stream each mini-batch into the models in turn. Since the news articles in News20 dataset are ordered by date, we stream the data to the models in its chronological order. Image datasets consist of different classes. To simulate data evolution over time, we stream images by classes and each pair of two consecutive mini-batches have half of the images from the same class. In the video datasets, videos are ordered by their labels, such as the videos belonging to the same person are grouped together and then sets of videos stream to the models in turn. For text and image datasets, we have dynamic query set in each iteration. After the initial mini-batch of data used to initialize the model, each time a mini-batch of streaming data comes, we use each of them as the query to search for the nearest neighbors from the existing database, and then update the current model by accommodating these streaming data. For YoutubeFace video dataset, we have a randomly sampled fixed set of queries consisting 226 videos with 43,020 frames in total. UQ\_VIDEO dataset provides a fixed set of 24 videos with 902 frames in total. Table \ref{} shows detailed information about the data streams. In this experiment, we compare our method with subspace update constraint (Online PQ - SS) and sub-codeword update constraint (Online PQ - SC) to several batch mode methods and online methods for the task of continuous update in a dynamic setting. In our model, we set $M = 8$ and $K = 256$. then the number of bits used for vector encoding is 64. We constraint the number of the updated subspaces $\alpha$ to be 4 and the portion of the updated sub-codewords $\lambda$ to be 0.5 respectively. The first batch is used for codebook initialization and the rest of the batches are used to update the existing codebook one at a time. All the key parameters in the baseline methods are set to the ones recommended in the corresponding papers. All the methods compared are implemented in Matlab provided by the authors and all experiments are conducted on a workstation with a 3.10GHZ Intel CPU and 120GB main memory running on a Linux platform. We show that how our method performs using different number of bits (32, 48 and 64 bits) compared with online baseline methods for News20 dataset. For all other experiments, we use 64 bits for vector encoding in all of the comparison models for fair comparisons. \begin{figure*} \centering \begin{subfigure}[b]{1\textwidth} \centering \includegraphics[width=0.8\textwidth]{online_legend.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_online_News20_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_online_Caltech-101_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_online_Half_Dome_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_online_Sun397_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_online_News20_time.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_online_Caltech-101_time.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_online_Half_Dome_time.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_online_Sun397_time.png} \end{subfigure} \caption[ Results for news and image retrieval in a dynamic database comparison against online hashing methods. Recall@20 performance (1st row) and Update time cost (2nd row). 1st column: News20. 2nd column: Caltech-101. 3rd column: Sun397. 4th column: Half dome. Time cost is in log scale.] {\small Results for news and image retrieval in a dynamic database comparison against online hashing methods. Recall@20 performance (1st row) and Update time cost (2nd row). 1st column: News20. 2nd column: Caltech-101. 3rd column: Half dome. 4th column: Sun397. Time cost is in log scale.} \label{knn_img_online} \end{figure*} \begin{figure*} \centering \begin{subfigure}[b]{1\textwidth} \centering \includegraphics[width=0.6\textwidth]{online_vid_legend.png} \end{subfigure} \vskip\baselineskip \begin{subfigure}[b]{0.49\textwidth} \centering \includegraphics[width=9cm]{streaming_online_vid_plots.png} \end{subfigure} \begin{subfigure}[b]{.48\textwidth} \includegraphics[width=8.5cm]{streaming_online_ImageNet_perf.png} \vspace{2ex} \includegraphics[width=8.3cm]{streaming_online_ImageNet_time.png} \end{subfigure} \caption[ ] {\small Results for YoutubeFaces on CSLBP and FPLBP features, UQ\_VIDEO and ImageNet in a dynamic database comparison against online hashing methods. Recall@20 (1st row) and Update time cost (2nd row). 1st column: YoutubeFaces CSLBP feature. 2nd column: YoutubeFaces FPLBP feature. 3rd column: UQ\_VIDEO. 4th column: ImageNet. Time cost is in log scale.} \label{knn_video_online} \end{figure*} \subsubsection{Online methods comparison} Figure \ref{news20_diff_bits} demonstrate the performance of indexing update using different number of bits of the model on News20 dataset compared with four online hashing methods. It clearly shows that our proposed models consistently outperforms other online hashing methods, with the lowest update time cost in different number of bits. In particular, when the number of bits increases, the difference between online PQ and other online methods gets increasing. SSBC achieves comparable search accuracy with online PQ in some of the iterations using 32 bits, but its update time cost is significantly higher than other methods. The curve trends of all methods over different number of bits used are consistent. As shown in Figure \ref{knn_img_online} and Figure \ref{knn_video_online}, it is evident that our proposed method with two different budget constraints can achieve superior performance in both efficiency and effectiveness compared to other online methods. Specifically, online PQ significantly outperforms the second best online method, OSH, in search accuracy for all datasets and is much faster in model update. These two figures are quite revealing in several ways. \begin{enumerate}[label={\alph*)},wide, labelwidth=!, labelindent=0pt] \item Interestingly, there is a sharp recall drop at iteration 13 in the News20 dataset. Since the distribution between our dynamic query sets changes frequently, this drastic change is caused by the significant query distribution change. All of the methods can not respond quickly to this effect, which reflects the importance and the necessity of developing an online approach that accommodates streaming data with changing distributions. \item Similarly for the Half dome dataset, the sudden recall improvement for all methods at iteration 11 implies the similar data distribution between the query set and all the existing stored data sets at iteration 11. \item The update cost for each method is the total update time for the mini-batch of new data at each iteration and it is relatively stable through most of the iterations because they have similar sizes of the mini-batch data to update. The last iteration only updates around half of the mini-batch size to the previous iterations, so its update cost drops with respect to the number of the data instances to be updated. \item As OH and AdaptHash are supervised online hashing methods, and OSH performs the best over all baseline methods, we compare our model with two different budget constraints with OSH for UQ\_VIDEO dataset. It is obvious that our method achieves better search accuracy with lower update time cost. Moreover, although the performance difference between the two budget constraints of our model is minimum, updating sub-codewords in half of the subspaces performs slightly better than updating half of the sub-codewords of all in both search accuracy and update time. \end{enumerate} \begin{figure*} \centering \begin{subfigure}[b]{1\textwidth} \centering \includegraphics[width=1\textwidth]{batch_legend.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_batch_News20_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_batch_Caltech-101_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_batch_Half_Dome_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_batch_Sun397_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_batch_News20_time.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_batch_Caltech-101_time.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_batch_Half_Dome_time.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{streaming_batch_Sun397_time.png} \end{subfigure} \caption[ ] {\small Results for news and image retrieval in a dynamic database comparison against batch methods. Recall@20 performance (1st row) and Update time cost (2nd row). 1st column: News20. 2nd column: Caltech-101. 3rd column: Sun397. 4th column: Half dome. Part of the recall plots of online and baseline PQ methods for Sun397 and Half dome are enlarged. Time cost is in log scale.} \label{knn_img_batch} \end{figure*} \subsubsection{Batch methods comparison} To further evaluate the performance of nearest neighbor search of our online model on how well it approaches to the search accuracy of batch mode methods and to the model update time of ``no update'' methods, we compare our model with each of the batch mode methods in two ways: retrain the model at each iteration (batch) and using the model trained on the initial iteration once for all (no update). The comparison results displayed in Figure \ref{knn_img_batch} implies several interesting observations. First, as the update time cost graphs are plot in log scale, the update time of online PQ is only slightly more than the one of the ``no update'' methods, but significantly lower than the one of the ``batch'' methods. Second, online PQ and PQ methods significantly outperform other batch hashing methods, and online PQ performs slightly worse than ``batch'' PQ and better than ``no update'' PQ. Therefore, we can conclude that our online model achieves good tradeoff between accuracy and update efficiency. Though KMH performs slightly better than ITQ in its original paper, it performs worse in our experiment setting. This is because that we are in the online setting where the data distribution of the query set may be a lot different from the one of the existing database. More results on our online model compared with batch methods for video datasets are in the Supplementary Material. \subsection{Continuous querying on dynamic real-time data} In many emerging application environments, they commonly require to emphasize the most recent data and ignore the expired data in retrieval search. Examples of such applications include network monitoring and online portfolio selection. Furthermore, applications such as hot topic news article retrieval system, object tracking given a recent period of time from social network albums or live videos require the real-time behaviour of the data. Therefore, to reflect this requirement in an online indexing system, we employ sliding window technique. In this experiment, we investigate the comparison between with and without employing the sliding window technique, and presents the comparison results on different hash methods. \begin{figure}[ht] \centering \includegraphics[scale=0.45]{slidingwindow_sun397_nodeletion.png} \caption{\label{slidingwindow_sun397_nodeletion} \small Online PQ over a sliding window approach between deletion and no deletion of the expired data to the model for Sun397. Recall@20 (1st row) and Update time cost (2nd row).} \end{figure} \begin{figure*} \centering \begin{subfigure}[b]{1\textwidth} \centering \includegraphics[width=0.6\textwidth]{slidingwindow_online_legend.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{slidingwindow_online_News20_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{slidingwindow_online_Caltech-101_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{slidingwindow_online_Half_Dome_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{slidingwindow_online_Sun397_perf.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{slidingwindow_online_News20_time.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{slidingwindow_online_Caltech-101_time.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{slidingwindow_online_Half_Dome_time.png} \end{subfigure} \begin{subfigure}[b]{0.23\textwidth} \centering \includegraphics[width=4.2cm,height=3.6cm]{slidingwindow_online_Sun397_time.png} \end{subfigure} \caption[] {\small Results for news and image retrieval over a sliding window comparison against online hashing methods. Recall@20 performance (1st row) and Update time cost (2nd row). 1st column: News20. 2nd column: Caltech-101. 3rd column: Sun397. 4th column: Half dome. Time cost is in log scale.} \label{slidingwindow_img_online} \end{figure*} \begin{figure} \centering \begin{subfigure}[b]{0.5\textwidth} \centering \includegraphics[width=0.8\textwidth]{slidingwindow_online_vid_legend.png} \end{subfigure} \vskip\baselineskip \begin{subfigure}[b]{0.48\textwidth} \centering \centerline{\includegraphics[width=8.9cm]{slidingwindow_online_vid_plots.png}} \end{subfigure} \caption[] {\small Results for YoutubeFaces dataset on CSLBP and FPLBP features and UQ\_VIDEO over a sliding window comparison against online hashing methods. Recall@20 (1st row) and Update time cost (2nd row). 1st column: YoutubeFaces CSLBP feature. 2nd column: YoutubeFaces FPLBP feature. 3rd column: UQ\_VIDEO. Time cost is in log scale.} \label{slidingwindow_video_online} \end{figure} \subsubsection{Setting} We follow the same way as in the setting in Section \ref{setting} to order the data in text and video datasets and the image data is ordered by classes without the overlapping of classes in each pair of two consecutive mini-batches this time. We set the number of iterations to be 11 for text and image datasets, and 65 and 34 for YoutubeFaces and UQ\_VIDEO respectively. The sliding window size is set to be 2000 for News20, 1000 for Caltech-101, 10000 for Sun397, Half dome and YoutubeFaces, and 100000 for UQ\_VIDEO. Except for News20 dataset which contains news articles in chronological order, for the rest of the datasets, the sliding window contains images/videos belonging to a certain amount of classes/people at each iteration, so that the contribution of the classes/people for the expired data are removed from our proposed model. We use the dynamic query set in this setting for all the datasets, so we use each new coming mini-batch of data as the query set first to retrieve similar data from the sliding window of the previous iteration and then use this mini-batch of data to update the model. In our model, we remove the contribution of the data once it is removed from the sliding window. We set $M$ = 8 and $K$ = 256, and update all the codewords over all subspaces. Batch mode baseline methods retrain the model using the data in the sliding window at each iteration. \subsubsection{Online methods comparison} Our proposed method over a sliding window approach adds the contribution of the new incoming data to the index model and removes the contribution of the expired data to the index model at each iteration. To evaluate our approach on data deletion using the sliding window technique, we highlight the difference between the models with and without expired data deletion in nearest neighbor search task in Figure \ref{slidingwindow_sun397_nodeletion}. The update time of online PQ with expired data deletion is reasonably slightly higher than that of online PQ without expired data deletion and the search accuracy of online PQ with expired data deletion is slightly better as it emphasizes on the ``real-time'' data. From Figure \ref{slidingwindow_img_online} and Figure \ref{slidingwindow_video_online}, we can see that online PQ over a sliding window approach achieves the best performance over other online methods in terms of update time efficiency and search effectiveness. Specifically, the search accuracies for YoutubeFaces dataset of CSLBP feature and UQ\_VIDEO are significantly higher than OSH, with low update time cost. \subsubsection{Batch methods comparison} To investigate how well our proposed method over a sliding window technique with expired data deletion approaches the performance of the batch mode methods where the batch models will be retrained on the data from the sliding window at each iteration, we compare our model with batch mode methods. In addition, we compare with ``no update'' model to show the update time complexity of our method. Since the results are similar to the batch methods comparison for online PQ model in Section 6.6.3, we put them in the Supplementary Material. \section{Conclusion and future work} In this paper, we have presented our online PQ method to accommodate streaming data. In addition, we employ two budget constraints to facilitate partial codebook update to further alleviate the update time cost. A relative loss bound has been derived to guarantee the performance of our model. In addition, we propose an online PQ over sliding window approach, to emphasize on the real-time data. Experimental results show that our method is significantly faster in accommodating the streaming data, outperforms the competing online and batch hashing methods in terms of search accuracy and update time cost, and attains comparable search quality with batch mode PQ. In our future work, we will extend the online update for other MCQ methods, leveraging the advantage of them in a dynamic database environment to enhance the search performance. Each of them has challenges to be effectively extended to handle streaming data. For example, CQ \cite{} and SQ \cite{} require the old data for the codewords update at each iteration due to the constant inter-dictionary-element-product in the model constraint. AQ \cite{} requires a high computational encoding procedure, which will dominate the update process in an online fashion. TQ \cite{} needs to consider the tree graph update together with the codebook and the indices of the stored data. Extensions to these methods can be developed to address the challenges for online update. In addition, online PQ model can be extended to handle other learning problems such as multi-output learning \cite{liu2018metric,shen2017multilabel}. Moreover, the theoretical bound for the online model will be further investigated. \ifCLASSOPTIONcompsoc \section*{Acknowledgments} \else \section*{Acknowledgment} \fi Ivor W. Tsang is supported by the ARC Future Fellowship FT130100746, ARC LP150100671 and DP180100106. Ying Zhang is supported by ARC FT170100128 and DP180103096. \ifCLASSOPTIONcaptionsoff \newpage \fi \bibliographystyle{IEEEtran}

\section{Introduction} Machine-to-machine (M2M) communications have been considered as one of the promising technologies to realize the Internet of Things (IoT) in the future 5th generation network. M2M communications can be applied to many IoT applications, which mainly involve new business models and opportunities, such as smart grids, environmental monitoring and intelligent transport systems \cite{Wu2011M2M}. Different from conventional human type communications, M2M communications have many unique features \cite{Zheng2012Radio}. The unique features include massive transmissions from a large number of machine type communication devices (MTCDs), small bursty natured traffic (periodically generated), extra low power consumption of MTCDs, high requirements of energy efficiency and security. A key challenge for M2M communications is access control, which manages the engagement of massive MTCDs to the core network. To tackle this challenge, various solutions have been proposed, e.g., by using wired access (cable, DSL) \cite{6525600}, wireless short distance techniques (WLAN, Bluetooth), and wide area cellular network infrastructure (Long Term Evolution-Advanced (LTE-A), WiMAX)\cite{6845044}. Among all these solutions, an effective approach is to deploy machine type communication gateways (MTCGs) to act as relays of MTCDs \cite{Zheng2012Radio}. With the help of MTCGs, all MTCDs can be successfully connected to the base station (BS) at the additional cost of energy consumption \cite{6786066,7010526,6824817,7420617}. To enable multiple MTCDs to transmit data to the same MTCG, time division multiple access (TDMA) was adopted in \cite{Zhang_2016}. However, since there are a vast number of MTCDs to be served, TDMA leads to large transmission delay and synchronization overhead. By splitting users in the power domain, non-orthogonal multiple access (NOMA) can simultaneously serve multiple users at the same frequency or time resource \cite{saito2013non}. Consequently, NOMA based access scheme yields a significant gain in spectral efficiency over the conventional orthogonal TDMA \cite{saito2013non,Yang2017MultiCell,ding2017application,Chen2014ULNOMA,Imari2014ULNOMA,Zhang2016ULNOMA}. This favorable characteristic makes NOMA an attractive solution for supporting massive MTCDs in M2M networks. Considering NOMA, \cite{Yang2017Energy} investigated an M2M enabled cellular network, where multiple MTCDs simultaneously transmit data to the same MTCG and multiple MTCGs simultaneously transmit the gathered data to the BS. Besides, another key challenge is the energy consumption of MTCDs \cite{7378918,6644241,7263374}. According to \cite{Zheng2012Radio}, the total system throughput of an M2M network is mainly limited by the energy budget of the MTCDs. To improve the system performance, energy harvesting (EH) has been applied to wireless communication networks \cite{4595260,Bi2015Wireless,Huang2015Energy,7001217}. In particular, direct and non-direct energy transfer based schemes for EH were investigated in [23], while in [24], the optimization of green-energy-powered cognitive radio networks was surveyed. Recently, the downlink resource allocation for EH in small cells was studied in \cite{7317521,7562484,7948714}. By using EH, MTCDs are able to harvest wireless energy from radio frequency (RF) signals \cite{6477832,7944540,7555321,Yang2017EnergyM2MEH}, and the system energy can be significantly improved. Consequently, implementing EH is promising in M2M communications especially with MTCDs configured with low power consumption. In previous wireless powered communication networks using relays \cite{6552840,7342973,7470278}, it was assumed that an energy constrained relay node harvests energy from RF signals and the relay uses that harvested energy to forward source information to destination. Due to the extra low power budget of MTCDs in M2M communications, it is reasonable to let the MTCD transmit data to an MTCG, and then the MTCG relays the information while the MTCD simultaneously harvests energy from the MTCG, which is different from existing works, e.g., \cite{6552840,7342973,7470278}. Enabling the source node to harvest energy from the relay node, a power-allocation scheme for a decode-and-forward relaying-enhanced wireless system was proposed in \cite{Huang2016Optical} with one source node, one relay node and one destination node. The above-mentioned energy consumption models, considered in \cite{Yang2017EnergyM2M,6824817,7420617,Zhang_2016,Yang2017Energy,7378918,6644241,6477832,7944540,7555321,Yang2017EnergyM2MEH}, are only concerned with the RF transmission power and ignore the circuit power consumption of MTCDs and MTCGs. However, as stated in \cite{Kim2010Leveraging}, the circuit power consumption is non-negligible compared to RF transmission power. Without considering the circuit power consumption, energy saving can always benefit from longer transmission time \cite{angelakis2014minimum,Chin2015Power}. While considering the circuit power consumption which definitely exists in practice, the results vary significantly in that it can not be always energy efficient for long transmission time due to the fact that the total energy consumption becomes infinity as transmission time goes infinity. Hence, it is of importance to investigate the optimal transmission time when taking into account the circuit power consumption in applications. Given access control and energy consumption challenges in M2M communications, both TDMA and NOMA based M2M networks with EH are proposed in this work. The MTCDs first transmit data to the corresponding MTCGs, and then MTCGs transmit wireless information to the BS and wireless energy to the MTCDs. To prolong the lifetime of the considered network, the harvested energy for each MTCD is set to be no less than the consumed energy in information transmission (IT) stage. The main contributions of this paper are summarized as follows: \begin{itemize} \item We formulate the total energy minimization problem for the M2M enabled cellular network with non-linear energy harvesting (EH) model via joint power control and time allocation. In the non-linear EH model, we consider the receiver sensitivity, on which energy conversion starts beyond a threshold. Besides, we explicitly take into account the circuit energy consumption of both MTCDs and MTCGs. All theses factors are critical in practical applications which inevitably affect the system performance. Specifically, the non-linear EH model leads to a non-smooth objective function and non-smooth constraints, and the circuit energy consumption affects the optimal transmission time of the system. \item For the NOMA strategy, we observe that: 1) it is optimal for each MTCD to transmit with minimal throughput; 2) it is further revealed that the energy consumption of each MTCD is a convex function with respect to the allocated transmission time. Given these observations, it indicates that a globally optimal transmission time always exists that the optimal transmission time equals the maximally allowed transmission time if it does not exceed a quantified threshold derived in closed form. \item To solve the original total energy minimization problem for the NOMA strategy, we devise a low-complexity iterative power control and time allocation algorithm. Specifically, to deal with the non-smooth EH function, we introduce new sets during which MTCDs can effectively harvest energy. Given new sets, the EH function of MTCDs can be presented as a continuous one. Moreover, to deal with nonconvex objective function, nonconvex minimal throughput constraints, and nonconvex energy causality constraints, we transform these nonconvex ones into convex ones by manipulations with the optimal conditions. The convergence of the iterative algorithm is strictly proved. \item For the TDMA strategy, we verify that the two observations for NOMA are also valid. Although the original total energy minimization problem for the TDMA strategy is nonconvex, the problem can be transformed into an equivalent tractable one, which can be iteratively solved to its suboptimality. For the total energy minimization, numerical results identify that NOMA is superior over TDMA at small circuit power regime of MTCDs, while TDMA outperforms NOMA at large circuit power regime of MTCDs. \end{itemize} This paper is organized as follows. In Section $\text{\uppercase\expandafter{\romannumeral2}}$, we introduce the system and power consumption model. Section $\text{\uppercase\expandafter{\romannumeral3}}$ and Section $\text{\uppercase\expandafter{\romannumeral4}}$ provide the energy efficient resource allocation for NOMA and TDMA, respectively. Numerical results are displayed in Section $\text{\uppercase\expandafter{\romannumeral5}}$ and conclusions are drawn in Section $\text{\uppercase\expandafter{\romannumeral6}}$. \section{System and Power Consumption Model} \subsection{System Model} Consider an uplink M2M enabled cellular network with $N$ MTCGs and $M$ MTCDs, as shown in Fig.~\ref{sys1}. Denote the sets of MTCGs and MTCDs by $\mathcal N=\{1, \cdots, N\}$ and $\mathcal M=\{1, \cdots, M\}$, respectively. Each MTCG serves as a relay for some MTCDs. Assume that the decode-and-forward protocol \cite{Karmakar2011DF} is adopted at each MTCG. Denote ${\cal{J}}_i=\{J_{i-1}+1, \cdots, J_i\}$ as the specific set of MTCDs served by MTCG $i\in \mathcal N$, where $J_0=0$, $J_N=M$, $J_i=\sum_{l=1}^i|{\cal{J}}_l|$, and $|\cdot|$ is the cardinality of a set.\footnote{In this paper, we assume that MTCDs are already associated to MTCGs by using the cluster formation methods for M2M communications, e.g., in \cite{6503998,6649207,7558142,7913656}. Joint optimization of cluster formation and resource allocation in M2M communications with NOMA/TDMA and EH can certainly further improve the performance, but we leave it in future work in order to focus on the power control and time allocation in the current submission.} To reduce the receiver complexity at the MTCG, the maximal number of MTCDs associated to one MTCG is set as four. Obviously, we have $\bigcup_{i\in\mathcal N} \mathcal J_i=\mathcal M$. \begin{figure} \centering \includegraphics[width=3in]{fig1.eps} \caption{The considered uplink M2M enabled cellular network.}\label{sys1} \end{figure} \subsection{NOMA Strategy} \begin{figure} \centering \includegraphics[width=3.35in]{fig2.eps} \caption{Time sharing scheme for NOMA strategy during one uplink transmission period.}\label{sys2} \end{figure} In time constraint $T$, each MTCD has some payloads to transmit to the BS. By using superposition coding at the transmitter and successive interference cancellation (SIC) at the receiver, multiple MTCDs (or MTCGs) can simultaneously transmit signals to the corresponding receiver using NOMA. To reduce the receiver complexity and error propagation due to SIC, it is reasonable for the same resource to be multiplexed by a small number (usually two to four) of devices \cite{6464495}. Considering the receiver complexity at the BS, the sets of MTCGs are further classified into multiple small clusters. For MTCGs, the set $\mathcal N$ is classified into $K$ clusters. Let $\mathcal K=\{1, 2, \cdots, K\}$ be the set of clusters. Denote ${\cal{I}}_k=\{I_{k-1}+1, \cdots, I_k\}$ as the specific set of MTCGs in cluster $k\in\mathcal K$, where $I_0=0$, $I_K=N$, $I_k=\sum_{l=1}^k|{\cal{I}}_l|$, and $|\mathcal I_k|\leq 4$.\footnote{A scheme for cluster formation for uplink NOMA is in \cite{Mohammad17uplinkNOMA}. According to \cite{Mohammad17uplinkNOMA} and \cite{7273963}, one common scheme with two devices in each cluster is the strong-weak scheme, i.e., the device with the strongest channel condition is paired with the device with the weakest, and the device with the second strongest is paired with one with the second weakest, and so on.} Note that our major motivation of using NOMA is to enhance the ability of serving more terminals simultaneously \cite{saito2013non}. However, the number of terminals occupying the same resource cannot be arbitrarily large in order to make NOMA effective in practice. Therefore, if there would be an even higher number of IoT terminals, we believe that a number of ways including NOMA should be further incorporated to better address the access problem for higher number of IoT terminals \cite{Shirvanimoghaddam2017Fund}. As depicted in Fig.~\ref{sys2}, time $T$ consists of $N+K$ uplink transmission phases for MTCDs and MTCGs. NOMA is adopted for MTCDs to transmit data to MTCGs in the first $N$ phases. Both NOMA and EH are in operation in the last K phases where MTCGs transmit data to the BS and simultaneously MTCDs harvest energy from all MTCGs. In the $i$-th ($i\leq N$) phase with allocated time $t_{i}$, all MTCDs in $\mathcal J_i$ simultaneously transmit data to MTCG $i$ according to the NOMA principle and MTCG $i$ detects the signal. In the ($N+k$)-th ($k\leq K$) phase with allocated time $t_{N+k}$, all MTCGs in $\mathcal I_k$ simultaneously transmit the decoded data from the served MTCDs to the BS by using the NOMA strategy. As a result, we have the following transmission time constraint \begin{equation}\label{sys1eq1} \sum_{i =1} ^{N+K}t_i \leq T. \end{equation} In the $i$-th ($i\leq N$) phase, all MTCDs in $\mathcal J_i$ simultaneously transmit data to MTCG $i$ following the NOMA principle. The received signal of MTCG $i$ is \begin{equation}\label{sys1eq2} y_i=\sum_{j=J_{i-1}+1}^{ J_i} h_{ij} \sqrt {p_j}s_j + n_i, \end{equation} where $h_{ij}$ is the channel between MTCD $j$ and MTCG $i$, $p_j$ denotes the transmission power of MTCD $j$, $s_j$ is the transmitted message of MTCD $j$, and $n_i$ represents the additive zero-mean Gaussian noise with variance $\sigma^2$. Without loss of generality, the channels are sorted as $|h_{i (J_{i-1}+1)}|^2 \geq \cdots \geq |h_{i J_i}|^2$. By applying SIC to decode the signals \cite{Chen2014ULNOMA,Imari2014ULNOMA,Zhang2016ULNOMA}, the achievable throughput of MTCD $j\in\mathcal J_i$ is \begin{equation}\label{sys1eq3} r_{ij}=B t_{i}\log_2\left( 1+\frac{|h_{ij}|^2 p_j} { \sum_{l=j+1}^{J_i} |h_{il}|^2 p_l +\sigma^2} \right), \end{equation} where $B$ is the available bandwidth for transmission. Note that we consider the case where MTCDs associated to different MTCGs are allocated with orthogonal time resource. Therefore, the interference from other MTCDs associated to different MTCDs is ignored. In the ($N+k$)-th ($k\leq K$) phase with allocated time $t_{N+k}$, after having successfully decoded the messages in the last $N$ phases, MTCGs in $\mathcal I_k$ simultaneously transmit the gathered data to the BS based on the NOMA principle. Denote the channel between MTCG $i$ and the BS by $h_i$. Without loss of generality, the channels are sorted as $|h_{I_{k-1}+1}|^2 \geq \cdots \geq |h_{I_k}|^2$, $\forall k \in \mathcal K$. Hence, the achievable throughput of MTCG $i\in\mathcal I_k$ can be expressed as \cite{Chen2014ULNOMA,Imari2014ULNOMA,Zhang2016ULNOMA} \begin{equation}\label{sys1eq4} r_i=B t_{N+i}\log_2\left( 1+\frac{|h_{i}|^2 q_i} { \sum_{n=i+1}^{I_k} |h_{n}|^2 q_n+\sigma^2} \right), \end{equation} where $q_i$ is the transmission power of MTCG $i$. According to \cite{Zheng2012Radio}, MTCDs are always equipped with finite batteries, which limit the lifetime of the M2M enabled cellular network. To further prolong the lifetime, EH technology is adopted for MTCDs to harvest energy remotely from RF signals radiated by MTCGs \cite{Bi2015Wireless}. Specifically, each MTCD harvests energy when MTCGs transmit data to the BS. Since the noise power is much smaller than the received power of MTCGs in practice \cite{Ju2014Throughput,Ju2014Optimal,zhou2014wireless}, the energy harvested from the channel noise is negligible. Assume that uplink channel and downlink channel follow the channel reciprocity \cite{guillaud2005practical}. The total energy harvested by MTCD $j$ served by MTCG $i$ can be evaluated as \begin{equation}\label{sys1eq5} E_{ij}^{\text H}=\sum_{k=1}^K t_{N+k}u\left(\sum_{n=I_{k-1}+1}^{ I_k}|h_{nj}|^2 q_n \right), \quad \forall i\in\mathcal N, j \in \mathcal J_i, \end{equation} where $\sum_{n=I_{k-1}+1}^{ I_k}|h_{nj}|^2 q_n$ is the received RF power of MTCD $j$ during time $t_{N+k}$, and function $u(\cdot)$ captures the EH model which maps input RF power into harvested power. Two commonly used EH models are shown in Fig.~\ref{syseh}, i.e., linear and non-linear EH models. According to \cite{7264986} and \cite{7511403}, linear EH model may lead to resource allocation mismatch. In order to capture the effects of practical EH circuits on the end-to-end power conversion, we adopt the more practical non-linear EH model proposed in \cite{7264986}: \begin{equation}\label{sys1eq5_2} u(x)=\left\{ \begin{array}{ll} \frac{M(1+\text e^{ab})}{\text{e} ^{ab}+\text{e} ^{-a(x-2b)}} -\frac{M}{\text{e}^{ab}}, & \text{if $x\geq P_0$} \\ 0, & \text{elsewise} \end{array} \right. , \end{equation} where $a$, $b$, $M$ and $P_0$ are positive parameters which capture the joint effects of different non-linear phenomena caused by hardware constraints. Note that $P_0$ is the receiver sensitivity threshold of each MTCD, in which energy conversion starts. Hence, it is possible that some MTCDs cannot effectively harvest energy in some slots, since the received power is below the receiver sensitivity threshold $P_0$. \begin{figure} \centering \includegraphics[width=3in]{fig3.eps} \caption{Comparison between the linear and non-linear EH model.}\label{syseh} \end{figure} The total energy consumption of the M2M enabled communication network consists of two parts: the energy consumed by MTCDs and MTCGs. For each part, the energy consumption of a transmitter consists of both RF transmission power and circuit power due to hardware processing \cite{Wu2015Resource}. According to \cite{Wu2016Energy}, the energy consumption when MTCDs or MTCGs are in idle model, i.e., do not transmit RF signals, is negligible. During the $i$-th ($i\leq N$) phase, MTCD $j \in \mathcal J_i$ served by MTCG $i$ just transmits data to MTCG $i$ with allocated time $t_i$ and transmission power $p_{j}$. Thus, the energy $E_{ij}$ consumed by MTCD $j \in \mathcal J_i$ can be modeled as \begin{equation}\label{sys1eq6} E_{ij}= t_i \left(\frac{p_j}{\eta} + P^{\text C}\right), \quad \forall i\in\mathcal N, j \in \mathcal J_i, \end{equation} where $\eta\in(0,1]$ and $P^{\text C}$ denote the power amplifier (PA) efficiency and the circuit power consumption of each MTCD, respectively. According to the energy causality constraint in EH networks, $E_{ij}$ has to satisfy $E_{ij} \leq E_{ij}^{\text H}$. Summing the energy consumed by all MTCDs in $\mathcal J_i$, we can obtain the energy $E_i$ consumed during the $i$-th phase as \begin{equation}\label{sys1eq6_2} E_i=\sum_{j=J_{i-1}+1}^{J_i}E_{ij}, \quad \forall i\in\mathcal N. \end{equation} During the ($N+k$)-th phase, the system energy consumption, denoted by $E_{N+k}$, is modeled as \begin{eqnarray}\label{sys1eq7} E_{N+k}&&\!\!\!\!\!\!\!\!\!\!=\!\sum_{i =I_{k-1}+1}^{ I_k} t_{N+k} \left(\frac{q_i}{\xi} + Q^{\text C}\right) \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \quad- \sum_{i=1}^{N} \sum_{j=J_{i-1}+1}^{ J_i} t_{N+k}u\left(\sum_{n=I_{k-1}+1}^{ I_k}|h_{nj}|^2 q_n \right) , \end{eqnarray} where $\xi\in(0,1]$ and $Q^{\text C}$ are the PA efficiency and the circuit power consumption of each MTCG, respectively. According to the law of energy conservation \cite{Ng2013Wireless}, we must have \begin{equation}\label{sys1eq7_2} \sum_{i=I_{k-1}+1}^{I_k}\!\! t_{N+k} q_i \!-\!\sum_{i=1}^{ N} \!\sum _{j=J_{i-1}+1}^{J_i} t_{N+k}u\!\left(\!\sum_{n=I_{k-1}+1}^{ I_k}\!|h_{nj}|^2 q_n \!\right)\!\!>\!0, \end{equation} which is the energy loss due to wireless propagation. Based on (\ref{sys1eq5})-(\ref{sys1eq7}), the total energy consumption, $E_{\text {Tot}}$, of the whole system during time $T$ can be expressed as \begin{eqnarray}\label{sys1eq8} \!\!\!\!\!\!E_{\text {Tot}} &&\!\!\!\!\!\!\!\!\!\!=\sum_{i=1}^{N+K} E_i \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!=\sum_{i=1}^{ N}\sum_{j=J_{i-1}+1}^{ J_i}t_i \left(\frac{p_j}{\eta} + P^{\text C}\right) \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \quad+ \sum_{k=1}^{K}\sum_{i=I_{k-1}+1}^{I_k} t_{N+k} \left(\frac{q_i}{\xi} + Q^{\text C}\right) \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \quad- \sum_{k=1}^K \sum_{i=1}^{ N} \sum_{j=J_{i-1}+1}^{ J_i}t_{N+k}u\left(\!\sum_{n=I_{k-1}+1}^{I_k} |h_{nj}|^2 q_n\!\right)\!\!. \end{eqnarray} \subsection{TDMA Strategy} With the TDMA strategy, time $T$ consists of $M+N$ uplink transmission phases for MTCDs and MTCGs, as illustrated in Fig.~\ref{sys5}. All MTCDs transmit data to the corresponding MTCGs in the first $M$ phases with TDMA, and all MTCGs transmit the collected data to the BS in the last $N$ phases with TDMA. Then, we obtain the following transmission time constraint \begin{equation}\label{sys1eq1tdma} \sum_{i =1} ^{M+N}t_i \leq T. \end{equation} \begin{figure} \centering \includegraphics[width=3.35in]{fig4.eps} \caption{Time sharing scheme for TDMA strategy during one uplink transmission period.}\label{sys5} \end{figure} In the $j$-th ($j\leq M$) phase, MTCD $j\in \mathcal J_i$ transmits data to its serving MTCG $i$ with achievable throughput \begin{equation}\label{sys1eq3tdma} r_{ij}=B t_{j}\log_2\left( 1+\frac{|h_{ij}|^2 p_j} { \sigma^2} \right), \quad \forall i \in \mathcal N, j \in \mathcal J_i. \end{equation} In the $(M+i)$-th phase, after having decoded all the messages of its served MTCDs, MTCG $i$ transmits the collected data to the BS with achievable throughput \begin{equation}\label{sys1eq4tdma} r_i=B t_{M+i}\log_2\left( 1+\frac{|h_{i}|^2 q_i} {\sigma^2} \right),\quad \forall i \in \mathcal N. \end{equation} Similar to (\ref{sys1eq5}), the total energy harvested of MTCD $j$ served by MTCG $i$ is \begin{equation}\label{sys1eq5tdma} E_{ij}^{\text H}= \sum_{n=1}^{ N} t_{M+n}u(|h_{nj}|^2 q_n ), \quad \forall i\in\mathcal N, j \in \mathcal J_i. \end{equation} According to (\ref{sys1eq5_2}), it is possible that some MTCDs cannot effectively harvest energy in some slots, due to the fact that the received power is below the receiver sensitivity threshold $P_0$. As in (\ref{sys1eq6}) and (\ref{sys1eq7}), the energy consumption of a transmitter includes both RF transmission power and circuit power \cite{Wu2015Resource}. With allocated transmission time $t_j$, the energy $E_{ij}$ consumed by MTCD $j \in \mathcal J_i$ can be modeled as \begin{equation}\label{sys1eq6tdma} E_{ij}= t_j \left(\frac{p_j}{\eta} + P^{\text C}\right), \quad \forall i\in\mathcal N, j \in \mathcal J_i. \end{equation} With allocated transmission time $t_{M+i}$, the system energy consumption, denoted by $E_{M+i}$, is modeled as \begin{equation}\label{sys1eq7tdma} E_{M+i}= t_{M+i} \left(\frac{q_i}{\xi} + Q^{\text C}\right)-\sum_{n=1}^{N} \sum_{j=J_{n-1}+1}^{ J_n} t_{M+i} u(|h_{ij}|^2 q_i ), \end{equation} where $ \sum_{n=1}^{N} \sum_{j=J_{n-1}+1}^{ J_n} t_{M+i}u(q_i |h_{ij}|^2)$ is the energy harvested by all MTCDs during the transmission time $t_{M+i}$ for MTCG $i$ to transmit data to the BS. According to (\ref{sys1eq5tdma})-(\ref{sys1eq7tdma}), the total energy consumption, $E_{\text {Tot}}$, of the whole system during time $T$ can be expressed as \begin{eqnarray}\label{sys1eq8tdma} E_{\text {Tot}} &&\!\!\!\!\!\!\!\!\!\!=\sum_{i=1}^{ N}\sum_{j=J_{i-1}+1}^{ J_i} E_{ij}+ \sum_{i=M+1}^{M+N} E_i \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!=\sum_{i=1}^{ N}\sum_{j=J_{i-1}+1}^{ J_i}t_j \left(\frac{p_j}{\eta} + P^{\text C}\right)+\sum_{i=1}^{N} t_{M+i} \left(\frac{q_i}{\xi} + Q^{\text C}\right) \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \quad-\sum_{i=1}^{ N} \sum_{j=J_{i-1}+1}^{ J_i}\sum_{n=1}^{ N} t_{M+n}u( |h_{nj}|^2 q_n). \end{eqnarray} \section{Energy Efficient Resource Allocation for NOMA} In this section, we study the resource allocation for an uplink M2M enabled cellular network with NOMA and EH. Specifically, we aim at minimizing the total energy consumption via jointly optimizing power control and time allocation for NOMA. The system energy minimization problem is formulated as \begin{subequations}\label{ee2min1} \begin{align} \!\!\mathop{\min}_{\pmb{p}, \pmb q, \pmb t}\:\; & E_{\text{Tot}} \\ \textrm{s.t.}\qquad \!\!\!\!\!\!\!\! &r_{ij} \geq D_j, \quad\forall i\in \mathcal N, j \in \mathcal J_i\\ &r_i \geq \sum_{j=J_{i-1}+1}^{ J_i} D_j, \quad\forall i \in \mathcal N \\ &E_{ij} \leq E_{ij}^{\text H}, \quad\forall i\in \mathcal N, j \in \mathcal J_i\\ &\sum_{i=1}^{N+K} t_i \leq T \\ &0\leq p_j \leq P_j, 0\leq q_i \leq Q_i, \quad\forall i\in \mathcal N, j \in \mathcal J_i\\ &\pmb t\geq \pmb 0, \end{align} \end{subequations} where $\pmb p=[p_1, \cdots, p_M]^T$, $\pmb q=[q_1, \cdots, q_N]^T$, $\pmb t=[t_{1}, \cdots,$ $ t_{N+K}]^T$, $D_j$ is the payload that MTCD $j$ has to upload within time constraint $T$, $P_{j}$ is the maximal transmission power of MTCD $j$, and $Q_{i}$ is the maximal transmission power of MTCG $i$. It is assumed that all payloads are positive, i.e., $D_j>0$, for all $j$. The objective function (\ref{ee2min1}a) defined in (\ref{sys1eq8}) is the total energy consumption of both MTCDs and MTCGs. Constraints (\ref{ee2min1}b) and (\ref{ee2min1}c) reflect that the minimal required payloads for MTCDs can be uploaded to the BS. The consumed energy of each MTCD should not exceed its harvested energy in time $T$, as stated in (\ref{ee2min1}d). Constraints (\ref{ee2min1}e) reflect that the payloads for all MTCDs are transmitted in time $T$. Note that problem (\ref{ee2min1}) is nonconvex due to nonconvex objective function (\ref{ee2min1}a) and constraints (\ref{ee2min1}b)-(\ref{ee2min1}d). In general, there is no standard algorithm for solving nonconvex optimization problems. In the following, we first find the optimal conditions for problem (\ref{ee2min1}) by exploiting the special structure of the uplink NOMA rate, and then provide an iterative power control and time allocation algorithm. \subsection{Optimal Conditions} By analyzing problem (\ref{ee2min1}), we have the following lemma. \begin{lemma} The optimal solution ($\pmb p^*, \pmb q^*, \pmb t^*$) to problem (\ref{ee2min1}) satisfies \begin{equation}\label{ee2eq1} r_{ij}^*=D_j, \quad \forall i \in \mathcal N, j \in \mathcal J_i. \end{equation} \end{lemma} This observation states that the minimal throughput leads to more energy saving, which is similar to \cite{Yang2017Energy} and is also widely known in the information theory community. Lemma 1 states that the optimal transmit throughput for each MTCD is required minimum. Note that the optimal throughput for each MTCG is not always as its minimum requirement, i.e., constraints (\ref{ee2min1}c) are active at the optimum, since MTCGs should transmit more power to maintain that the harvested energy of each MTCD is no less than the consumed energy. Based on Lemma 1, we further have the following lemma about the optimal transmission power of MTCDs. \begin{lemma} If ($\pmb p^*, \pmb q^*, \pmb t^*$) is the optimal solution to problem (\ref{ee2min1}), we have \begin{eqnarray}\label{ee2eq2} p_{j}^* &&\!\!\!\!\!\!\!\!\!\!= \sum_{l=j+1}^{J_i} \frac{\sigma^2} {|h_{ij}|^2 } {\left( {\text e}^{\frac{a_{l}}{t_i^*}} - 1 \right) \left( {\text e^{\frac{a_{j}}{t_i^*}}} - 1 \right) \text e^{ \frac{b_{jl}}{ t_i^*}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!\quad+ \frac{\sigma^2} {|h_{ij}|^2 } {\left(\text e^{\frac{a_{j}}{t_i^*}} -1\right) } ,\quad \forall i \in \mathcal N, j \in \mathcal J_i, \end{eqnarray} where \begin{equation}\label{ee2eq2_2} a_l={\frac{(\ln 2)D_{l}}{B}}, \quad b_{jl}={\frac{\sum_{s=j+1}^{l-1} (\ln 2) D_{s}}{B}}, \quad \forall i\in \mathcal N, j,l \in \mathcal J_i. \end{equation} Besides, the optimal transmission power $p_j^*$ of MTCD $j\in \mathcal J_i$ is always non-negative and decreases with the transmission time $t_i^*$. \end{lemma} \itshape \textbf{Proof:} \upshape Please refer to Appendix A. \hfill $\Box$ From Lemma 2, large transmission time results in low transmission power. This is reasonable as the minimal payload is limited and large transmission time requires low achievable rate measured in bits/s. It is also revealed from Lemma 2 that the optimal transmission power of MTCD $j \in \mathcal J_i$ served by MTCG $i$ depends only on the variable of the allocated transmission time $t_i$. As a result, the energy $E_{ij}$ consumed by MTCD $j$ in $\mathcal J_i$ is a function of the allocated transmission time $t_i$. Based on (\ref{sys1eq6}) and (\ref{ee2eq2}), we have \begin{eqnarray}\label{ee2eq3} E_{ij}&&\!\!\!\!\!\!\!\!\!\!= \sum_{l=j+1}^{J_i} \frac{\sigma^2 t_i} {\eta |h_{ij}|^2 } {\left( {\text e}^{\frac{a_{l}}{t_i}} - 1 \right) \left( {\text e^{\frac{a_{j}}{t_i}}} - 1 \right) \text e^{ \frac{b_{jl}}{ t_i}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \quad+ \frac{\sigma^2 t_i} {\eta |h_{ij}|^2 } {\left(\text e^{\frac{a_{j}}{t_i}} -1\right) } +t_i P_{\text C}. \end{eqnarray} \begin{theorem} The energy $E_{ij}$ defined in (\ref{ee2eq3}) is convex with respect to (w.r.t.) the transmission time $t_i$. When $P^{\text C}=0$, the energy $E_{ij}$ monotonically decreases with $t_i$. When $P^{\text C}>0$, the energy $E_{ij}$ first decreases with $t_i$ when $0\leq t_i \leq T_{ij}^*$ and then increases with $t_i$ when $t_i>T_{ij}^*$, where $T_{ij}^*$ is the unique zero point of the first-order derivative $\frac {\partial E_{ij}}{\partial t_i}$, i.e., \begin{eqnarray}\label{ee2eq5} \left.\frac {\partial E_{ij}}{\partial t_i} \right|_{t_i=T_{ij}^*} =0. \end{eqnarray} \end{theorem} \itshape \textbf{Proof:} \upshape Please refer to Appendix B. \hfill $\Box$ \begin{figure} \centering \includegraphics[width=3in]{fig5.eps} \caption{The energy $E_{ij}$ versus transmission time $t_i$.}\label{sys3} \end{figure} Fig.~\ref{sys3} exemplifies the energy $E_{ij}$ given in (\ref{ee2eq3}) versus $t_i$. When $P^{\text C}=0$, i.e., the circuit energy consumption of MTCDs is not considered, we come to the same conclusion as in \cite{angelakis2014minimum} and \cite{Chin2015Power} that the consumed energy decreases as the transmission time increases according to Theorem 1. Without considering the circuit power consumption, $R=\log_2(1+\text{SNR})$ and the energy efficiency increases with the decrease of power. Consequently, when $P^{\text C}=0$ and the minimal throughput demand $D_j$ is given, the consumed energy $E_{ij}$ is a decreasing function w.r.t. $t_i$. This fundamentally follows the Shannon's law. When $P^{\text C}>0$, i.e., the circuit energy consumption of MTCDs is taken into account, however, we find from Theorem 1 that the consumed energy first decreases and then increases with the transmission time, which is different from the previous conclusion in \cite{angelakis2014minimum} and \cite{Chin2015Power}. This is because that the total energy contains two parts balancing each other, i.e., the RF transmission energy part which monotonically decreases with the transmission time and the circuit energy part which linearly increases with the transmission time. In the following of this section, we assume that the circuit power consumption of MTCDs and MTCGs is in general positive, i.e., $P^{\text C}>0$ and $Q^{\text C}>0$. \begin{theorem} If $T\leq \max_{\forall i \in \mathcal N}\min_{\forall j \in \mathcal J_i} \{T_{ij}^*\}$, the optimal time allocation $\pmb t^*$ to problem (\ref{ee2min1}) satisfies \begin{equation}\label{ee2eq6} \sum_{i=1}^{N+K}t_i^*=T. \end{equation} If $T \geq T_{\text{Upp}}$, where $T_{\text{Upp}}$ is defined in (\ref{appenCeqn6}), the optimal time allocation $\pmb t^*$ to problem (\ref{ee2min1}) satisfies \begin{equation}\label{ee2eq6_3} \sum_{i=1}^{N+K}t_i^*<T. \end{equation} \end{theorem} \itshape \textbf{Proof:} \upshape Please refer to Appendix C. \hfill $\Box$ From Theorem 2, it is observed that transmitting with the maximal transmission time $T$ is optimal when $T$ is not large. This is because that the reduced energy of RF transmission dominates the additional energy of circuit by increasing transmission time. When the available time $T$ becomes large enough, Theorem~2 states that it is not optimal to transmit with the maximal transmission time $T$. This is due to the fact that the increased energy of circuit power dominates the power consumption while the energy reduction of RF transmission becomes relatively marginal. \subsection{Joint Power Control and Time Allocation Algorithm} Problem (\ref{ee2min1}) has two difficulties: one comes from the non-smooth EH function defined in (\ref{sys1eq5_2}), and the other one is the non-convexity of both objective function (\ref{ee2min1}a) and constraints (\ref{ee2min1}b)-(\ref{ee2min1}d). To deal with the first difficulty, we introduce notation $\mathcal S_{ij}$ as the set of phases during which MTCD $j\in\mathcal J_i$ can effectively harvest energy, i.e., $\mathcal S_{ij}=\{k|\sum_{n=I_{k-1}+1}^{I_k}|h_{nj}|^2q_n>P_0, \forall k \in \mathcal K\}$. With $\mathcal S_{ij}$ in hand, the harvested power of MTCD $j\in\mathcal J_{i}$ can be presented by the smooth function $\bar u(x)$ defined in (\ref{sys1eq5_22}). To deal with the second difficulty, we substitute (\ref{sys1eq3})-(\ref{sys1eq6}), (\ref{sys1eq8}) and (\ref{ee2eq2}) into (\ref{ee2min1}), and the original problem (\ref{ee2min1}) with fixed sets $\mathcal S_{ij}$'s can be equivalently transformed into the following problem: \begin{subequations}\label{ee2min2} \begin{align} \!\!\mathop{\min}_{\pmb q, \pmb t}\; \!& \sum_{i=1}^{ N}\sum_{j=J_{i-1}+1}^{ J_i} \sum_{l=j+1}^{J_i} \frac{\sigma^2 t_i} {\eta |h_{ij}|^2 } {\left( {\text e}^{\frac{a_{l}}{t_i}} - 1 \right) \left( {\text e^{\frac{a_{j}}{t_i}}} - 1 \right) \text e^{ \frac{b_{jl}}{ t_i}}} \nonumber \\ & +\!\sum_{i=1}^{ N}\!\sum_{j=J_{i-1}+1}^{ J_i}\! \frac{\sigma^2 t_i} {\eta |h_{ij}|^2 } {\!\left(\!\text e^{\frac{a_{j}}{t_i}} \!-\!1\!\right)\! } \!+\!\sum_{i=1}^{ N}\!\sum_{j=J_{i-1}+1}^{ J_i}\!t_i P^{\text C} \nonumber\\ &+\sum_{k=1}^K\sum_{i=I_{k-1}+1}^{I_k} t_{N+k} \left(\frac{q_i}{\xi} + Q^{\text C}\right) \nonumber\\ &-\sum_{i=1}^{ N} \sum_{j=J_{i-1}+1}^{ J_i}\sum_{k\in\mathcal S_{ij}}t_{N+k} \bar u \left(\sum_{n=I_{k-1}+1}^{I_k} |h_{nj}|^2 q_n\right) \\ \textrm{s.t.}\qquad \!\!\!\!\!\!\!\!\! &{|h_{i}|^2 q_i}\geq \!\left(\!2^{\frac{\sum_{j\in\mathcal J_i} D_j}{B t_{N+1}}}\!-\!1\!\right) \!\left({ \sum_{l=i+1}^{I_k} |h_{l}|^2 q_l\!+\!\sigma^2}\!\right), \nonumber\\ &\qquad\qquad\forall k \in \mathcal K, i \in \mathcal I_k\\ & \sum_{l=j+1}^{J_i} \frac{\sigma^2} {\eta |h_{ij}|^2 } {t_i \left( {\text e}^{\frac{a_{l}}{t_i}} - 1 \right) \left( {\text e^{\frac{a_{j}}{t_i}}} - 1 \right) \text e^{ \frac{b_{jl}}{ t_i}}} \nonumber\\ &+ \frac{\sigma^2} {\eta |h_{ij}|^2 } {t_i\left(\text e^{\frac{a_{j}}{t_i}} -1\right) } +t_i P^{\text C} \nonumber\\ &\leq \! \sum_{k\in\mathcal S_{ij}}\!t_{N+k} \bar u\!\left(\!\sum_{n=I_{k-1}+1}^{ I_k} \! |h_{nj}|^2 q_n\!\right),\! \quad\forall i\in \mathcal N, j \in \mathcal J_i\\ &\sum_{n=I_{k-1}+1}^{ I_k}\! |h_{nj}|^2 q_n\!\geq\! P_0, \!\!\!\quad \forall i \in \mathcal N, j \in \mathcal J_i, k \in \mathcal S_{ij}\\ &\sum_{i=1}^{N+K} t_i \leq T \\ &\sum_{l=j+1}^{J_i} \frac{\sigma^2} {|h_{ij}|^2 } {\left( {\text e}^{\frac{a_{l}}{t_i}} - 1 \right) \left( {\text e^{\frac{a_{j}}{t_i}}} - 1 \right) \text e^{ \frac{b_{jl}}{ t_i}}} \nonumber\\ & + \frac{\sigma^2} {|h_{ij}|^2 } {\left(\text e^{\frac{a_{j}}{t_i}} -1\right) } \leq P_{j}, \quad\forall i\in \mathcal N, j \in \mathcal J_i\\ & 0\leq q_i \leq Q_i, \quad\forall i\in \mathcal N\\ &\pmb t \geq \pmb 0, \end{align} \end{subequations} where \begin{equation}\label{sys1eq5_22} \bar u(x)= \frac{M(1+\text e^{ab})}{\text{e} ^{ab}+\text{e} ^{-a(x-2b)}} -\frac{M}{\text{e}^{ab}}, \quad \forall x\geq 0. \end{equation} Problem (\ref{ee2min2}) is still nonconvex w.r.t. ($\pmb q, \pmb t$) due to nonconvex objective function (\ref{ee2min2}a) and constraints (\ref{ee2min2}b)-(\ref{ee2min2}c). Before solving problem (\ref{ee2min2}), we have the following theorem. \begin{theorem} Given transmission time $\pmb \tau=[t_{N+1}, \cdots,$ $t_{N+K}]^T$, problem (\ref{ee2min2}) is a convex problem w.r.t. ($\pmb q, \bar {\pmb t}$), where $\bar{\pmb t}=[t_1, \cdots, t_{N}]^T$. Given ($\pmb q, \bar {\pmb t}$), problem (\ref{ee2min2}) is equivalent to a linear problem w.r.t. $\pmb \tau$. \end{theorem} \itshape \textbf{Proof:} \upshape Please refer to Appendix D. \hfill $\Box$ According to Theorem 3, problem (\ref{ee2min2}) with given transmission time $\pmb \tau$ can be effectively solved by using the standard convex optimization method, such as interior point method \cite{boyd2004convex}. Besides, problem (\ref{ee2min2}) with given ($\pmb q, \bar {\pmb t}$) is a linear problem, which can be optimally solved via the simplex method. Based on Theorem 3, we propose an iterative power control and time allocation for NOMA (IPCTA-NOMA) algorithm with low complexity to obtain a suboptimal solution of problem (\ref{ee2min1}). The idea is to iteratively update sets $\mathcal S_{ij}$'s according to the power and time variables obtained in the previous step. \begin{algorithm}[h] \caption{\!\!: Iterative Power Control and Time Allocation for NOMA (IPCTA-NOMA) Algorithm} \begin{algorithmic}[1] \STATE Set $\mathcal S_{ij}^{(0)}=\{k|j\in \cup_{n\in\mathcal I_k}\mathcal J_n, k \in \mathcal K\}$, $\forall i \in \mathcal I$, $j \in \mathcal J_{ij}$, initialize a feasible solution ($\pmb q^{(0)}, {\pmb t}^{(0)}$) to problem (\ref{ee2min2}) with $\mathcal S_{ij}^{(0)}$'s, the tolerance $\theta$, the iteration number $v=0$, and the maximal iteration number $V_{\max}$. \REPEAT \STATE Set $\pmb \tau^*=[t_{N+1}^{(v)}, \cdots, t_{N+K}^{(v)}]^T$. \REPEAT \STATE Obtain the optimal ($ {\pmb q}^{*}, \bar {\pmb t}^{*}$) of convex problem (\ref{ee2min2}) with fixed $\pmb \tau^*$ and sets $\mathcal S_{ij}^{(v)}$. \STATE Obtain the optimal $\pmb \tau^*$ of linear problem (\ref{ee2min2}) with fixed ($ {\pmb q}^{*}, \bar {\pmb t}^{*}$) and sets $\mathcal S_{ij}^{(v)}$. \UNTIL the objective value (\ref{ee2min2}a) with fixed sets $\mathcal S_{ij}^{(v)}$ converges. \STATE Set $v=v+1$. \STATE Denote $\pmb q^{(v)}=\pmb q^*$, $\pmb t^{(v)}=[\bar{\pmb t}^{*T}, \pmb \tau^{*T}]^T$. \STATE Calculate the objective value (\ref{ee2min2}a) with fixed sets $\mathcal S_{ij}^{(v)}$ as $U_{\text {Obj}}^{(v)}=E_{\text {Tot}}({\pmb q}^{(v)}, {\pmb t}^{(v)})$. \STATE Update $\mathcal S_{ij}^{(v)}=\{k|\sum_{n=I_{k-1}+1}^{I_k}|h_{nj}|^2q_n^{(v)}>P_0, \forall k \in \mathcal K\}$, $\forall i \in \mathcal I, j \in \mathcal J_i$. \UNTIL $v \geq 1$ and $|U_{\text {Obj}}^{(v)}-U_{\text {Obj}}^{(v-1)}|/U_{\text {Obj}}^{(v-1)}<\theta$ or $v > V_{\max}$. \end{algorithmic} \end{algorithm} \subsection{Convergence and Complexity Analysis} \begin{theorem} Assuming $V_{\max}\rightarrow\infty$, the sequence (${\pmb q}, {\pmb t}$) generated by the IPCTA-NOMA algorithm converges. \end{theorem} \itshape \textbf{Proof:} \upshape Please refer to Appendix E. \hfill $\Box$ According to the IPCTA-NOMA algorithm, the major complexity lies in solving the convex problem (\ref{ee2min2}) with fixed $\pmb \tau$. Considering that the dimension of the variables in problem (\ref{ee2min2}) with fixed $\pmb \tau$ is $2N$, the complexity of solving problem (\ref{ee2min2}) with fixed $\pmb \tau$ by using the standard interior point method is $\mathcal O(N^3)$ \cite[Pages 487, 569]{boyd2004convex}. As a result, the total complexity of the proposed IPCTA-NOMA algorithm is $\mathcal O(L_{\text{NO}}L_{\text{IT}}N^3)$, where $L_{\text{NO}}$ denotes the number of outer iterations of the IPCTA-NOMA algorithm, and $L_{\text{IT}}$ denotes the number of inner iterations of the IPCTA-NOMA algorithm for iteratively solving nonconvex problem (\ref{ee2min2}) with fixed sets $\mathcal S_{ij}$'s. \section{Energy Efficient Resource Allocation for TDMA} In this section, we study the energy minimization for the M2M enabled cellular network with TDMA. According to (\ref{sys1eq3tdma})-(\ref{sys1eq6tdma}) and (\ref{sys1eq8tdma}), the energy minimization problem can be formulated as \begin{subequations}\label{ee2min1tdma} \begin{align} \!\!\mathop{\min}_{\pmb{p}, \pmb q, \hat{\pmb t}}\; \!& \sum_{i=1}^{ N}\sum_{j=J_{i-1}+1}^{ J_i}t_j \left(\frac{p_j}{\eta} + P^{\text C}\right)+\sum_{i=1}^{N} t_{M+i} \left(\frac{q_i}{\xi} + Q^{\text C}\right) \nonumber\\ &-\sum_{i=1}^{ N} \sum_{j=J_{i-1}+1}^{ J_i}\sum_{n=1}^{ N} t_{M+n}u(|h_{nj}|^2 q_n ) \\ \textrm{s.t.}\qquad \!\!\!\!\!\!\!\!\! &B t_{j}\log_2\left( 1+\frac{|h_{ij}|^2 p_j} { \sigma^2} \right) \geq D_j, \quad\forall i\in \mathcal N, j \in \mathcal J_i\\ &B t_{M+i}\log_2\left( 1+\frac{|h_{i}|^2 q_i} {\sigma^2} \right) \geq \sum_{j=J_{i-1}+1}^{ J_i} D_j, \quad\forall i \in \mathcal N \\ & t_j \left(\frac{p_j}{\eta} + P^{\text C}\right) \leq \sum_{n=1}^{ N} t_{M+n}u( |h_{nj}|^2 q_n), \quad\forall i\in \mathcal N, j \in \mathcal J_i\\ &\sum_{i=1}^{N+1} t_i \leq T \\ &0\leq p_j \leq P_j, 0\leq q_i \leq Q_i, \quad\forall i\in \mathcal N, j \in \mathcal J_i\\ &\hat{\pmb t} \geq \pmb 0, \end{align} \end{subequations} where $\hat{\pmb t}=[t_{1}, \cdots,$ $ t_{M+N}]^T$. Obviously, problem (\ref{ee2min1tdma}) is nonconvex due to nonconvex objective function (\ref{ee2min1tdma}a) and constraints (\ref{ee2min1tdma}b)-(\ref{ee2min1tdma}d). In the following, we first provide the optimal conditions for problem (\ref{ee2min1tdma}), and then we propose a low-complexity algorithm to solve problem (\ref{ee2min1tdma}). \subsection{Optimal Conditions} Similar to Lemma 1, it is also optimal for each MTCD to transmit with the minimal throughput requirement. Accordingly, the following lemma is directly obtained. \begin{lemma} The optimal solution ($\pmb p^*, \pmb q^*, \hat{\pmb t}^*$) to problem (\ref{ee2min1tdma}) satisfies \begin{equation}\label{ee2eq1tdma} B t_{j}^*\log_2\left( 1+\frac{|h_{ij}|^2 p_j^*} { \sigma^2} \right) = D_j, \quad \forall i \in \mathcal N, j \in \mathcal J_i. \end{equation} \end{lemma} According to (\ref{ee2eq1tdma}), the optimal transmission power of MTCD $j$ can be presented as \begin{equation}\label{ee2eq2tdma} p_j=\frac 1 {|h_{ij}|^2}\left(2^{\frac{D_j}{B t_j}}-1 \right), \quad \forall i \in \mathcal N, j \in \mathcal J_i. \end{equation} Substituting (\ref{ee2eq2tdma}) into (\ref{sys1eq6tdma}) yields \begin{equation}\label{ee2eq3tdma} E_{ij}= \frac { t_j} {|h_{ij}|^2 \eta }\left(2^{\frac{D_j}{B t_j}}-1\right) + t_j P^{\text C} , \quad \forall i\in\mathcal N, j \in \mathcal J_i. \end{equation} By analyzing (\ref{ee2eq3tdma}), we can obtain the following theorem. \begin{theorem} The energy $E_{ij}$ defined in (\ref{ee2eq3tdma}) is convex w.r.t. $t_j$. When $P^{\text C}=0$, the energy $E_{ij}$ monotonically decreases with the transmission time $t_j$. When $P^{\text C}>0$, the energy $E_{ij}$ first decreases with $t_j$ when $0\leq t_j \leq T_{ij}^*$ and then increases with $t_j$ when $t_j>T_{ij}^*$, where $T_{ij}^*$ is the unique zero point of the first-order derivative $\frac {\partial E_{ij}}{\partial t_j}$, i.e., \begin{eqnarray}\label{ee2eq5tdma} \left.\frac {\partial E_{ij}}{\partial t_j} \right|_{t_j=T_{ij}^*} =0. \end{eqnarray} \end{theorem} Since Theorem 5 can be proved by checking the first-order derivative $\frac {\partial E_{ij}}{\partial t_j}$ as in Appendix B, the proof of Theorem 4 is omitted. Similar to Theorem 2 for NOMA, we come to the similar conclusion for TDMA that transmitting with the maximal transmission time $T$ is optimal when $T$ is not large, while for large $T$ it is not optimal to transmit with the maximal transmission time $T$. \subsection{Iterative Power Control and Time Allocation Algorithm} Similar to problem (\ref{ee2min1}), problem (\ref{ee2min1tdma}) has two difficulties: one is the non-smooth EH function in (\ref{sys1eq5_2}), and the other is the nonconvex objective function (\ref{ee2min1tdma}a) and constraints (\ref{ee2min1tdma}b)-(\ref{ee2min1tdma}d). To deal with the first difficulty, we introduce notation $\mathcal S_{ij}$ as the set of MTCGs from which MTCD $j\in\mathcal J_i$ can effectively harvest energy, i.e., $\mathcal S_{ij}=\{n||h_{nj}|^2q_n>P_0, \forall n \in \mathcal I\}$. With $\mathcal S_{ij}$ in hand, the harvested power of MTCD $j\in\mathcal J_{i}$ from MTCG $n\in\mathcal S_{ij}$ can be presented by the smooth function $\bar u(x)$ defined in (\ref{sys1eq5_22}). To tackle the second difficulty, we show that problem (\ref{ee2min1tdma}) with fixed sets $\mathcal S_{ij}$'s can be transformed into an equivalent convex problem. \begin{theorem} The original problem in (\ref{ee2min1tdma}) with fixed sets $\mathcal S_{ij}$'s can be equivalently transformed into the following convex problem as \begin{subequations}\label{ee2min2tdma} \begin{align} \!\!\mathop{\min}_{\hat {\pmb p}, \hat{\pmb q}, \hat{\pmb t}}\; \!& \sum_{i=1}^{ N}\sum_{j=J_{i-1}+1}^{ J_i} \left(\frac{\hat p_j}{\eta} +t_j P^{\text C}\right)+\sum_{i=1}^{N} \left(\frac{\hat q_i}{\xi} + t_{M+i}Q^{\text C}\right) \nonumber\\ &-\sum_{i=1}^{ N} \sum_{j=J_{i-1}+1}^{ J_i}\sum_{n\in\mathcal S_{ij}} t_{M+n} \bar u\left(\frac{ |h_{nj}|^2\hat q_n }{t_{M+n}} \right) \\ \textrm{s.t.}\qquad \!\!\!\!\!\!\!\!\! &B t_{j}\log_2\left( 1+\frac{|h_{ij}|^2 \hat p_j} { \sigma^2 t_j} \right) \geq D_j, \quad\forall i\in \mathcal N, j \in \mathcal J_i\\ &B t_{M+i}\log_2\left( 1+\frac{|h_{i}|^2 \hat q_i} {\sigma^2 t_{M+i}} \right) \geq \sum_{j=J_{i-1}+1}^{ J_i} D_j, \quad\forall i \in \mathcal N \\ & \frac{\hat p_j}{\eta} \!+\! t_j P^{\text C} \!\leq\!\! \sum_{n\in\mathcal S_{ij}} \! t_{M+n} \bar u\left(\!\frac{|h_{nj}|^2 \hat q_n }{t_{M+n}}\!\right), \!\!\quad\forall i\!\in\! \mathcal N, j \!\in\! \mathcal J_i\\ & {|h_{nj}|^2 \hat q_n } > P_0 {t_{M+n}}, \quad\forall i\in \mathcal N, j \in \mathcal J_i, n\in\mathcal S_{ij}\\ &\sum_{i=1}^{N+1} t_i \leq T \\ &0\leq \hat p_j \leq P_j t_j, \quad\forall i\in \mathcal N, j \in \mathcal J_i \\ &0\leq \hat q_i \leq Q_i t_{M+i}, \quad\forall i\in \mathcal N\\ &\hat {\pmb t} \geq \pmb 0, \end{align} \end{subequations} where $\hat{\pmb p}=[\hat p_1, \cdots, \hat p_M]^T$ and $\hat{\pmb q}=[\hat q_1, \cdots, \hat q_N]^T$. \end{theorem} \itshape \textbf{Proof:} \upshape Please refer to Appendix F. \hfill $\Box$ Based on Theorem 6, we propose an iterative power control and time allocation for TDMA (IPCTA-TDMA) algorithm with low complexity to obtain a suboptimal solution of problem (\ref{ee2min1tdma}). The idea is to iteratively update sets $\mathcal S_{ij}$'s according to the power and time variables obtained in the previous step. \begin{algorithm}[h] \caption{\!\!: Iterative Power Control and Time Allocation for TDMA (IPCTA-TDMA) Algorithm} \begin{algorithmic}[1] \STATE Set $\mathcal S_{ij}^{(0)}=\{i\}$, $\forall i \in \mathcal I$, $j \in \mathcal J_{ij}$, the tolerance $\theta$, the iteration number $v=0$, and the maximal iteration number $V_{\max}$. \REPEAT \STATE Obtain the optimal ($\hat {\pmb p}^{(v)}, \hat {\pmb q}^{(v)}, \hat {\pmb t}^{(v)}$) of convex problem (\ref{ee2min2tdma}) with fixed sets $\mathcal S_{ij}^{(v)}$. \STATE Calculate the objective value (\ref{ee2min2tdma}a) with fixed sets $\mathcal S_{ij}^{(v)}$ as $U_{\text {Obj}}^{(v)}=E_{\text {Tot}}(\hat {\pmb p}^{(v)}, \hat {\pmb q}^{(v)}, \hat {\pmb t}^{(v)})$. \STATE Set $v=v+1$. \STATE Update $\mathcal S_{ij}^{(v)}=\left\{n\left|\frac{|h_{nj}|^2 \hat q_n^{(v-1)}}{t_{M+n}^{(v-1)}}>P_0\right., \forall n \in \mathcal I\right\}$, $\forall i \in \mathcal I, j \in \mathcal J_i$. \UNTIL $v \geq 2$ and $|U_{\text {Obj}}^{(v)}-U_{\text {Obj}}^{(v-1)}|/U_{\text {Obj}}^{(v-1)}<\theta$ or $v > V_{\max}$. \STATE Output ${\pmb p}^{*}=\hat {\pmb p}^{(v)}, \hat {\pmb t}^{*}=\hat {\pmb t}^{(v)}, q_{i}^* =\hat q_i^{(v)}/ t_{M+i}^{(v)}, \forall i \in \mathcal I$. \end{algorithmic} \end{algorithm} \subsection{Convergence and Complexity Analysis} \begin{theorem} Assuming $V_{\max}\rightarrow\infty$, the sequence ($\hat {\pmb p}, \hat {\pmb q}, \hat {\pmb t}$) generated by the IPCTA algorithm converges. \end{theorem} Theorem 7 can be proved by using the same method as in Appendix E. The proof of Theorem 7 is thus omitted. According to the IPCTA-TDMA algorithm, the major complexity lies in solving the convex problem (\ref{ee2min2tdma}). Considering that the dimension of the variables in problem (\ref{ee2min2tdma}) is $2(M+N)$, the complexity of solving problem (\ref{ee2min2tdma}) by using the standard interior point method is $\mathcal O((M+N)^3)$ \cite[Pages 487, 569]{boyd2004convex}. As a result, the total complexity of the proposed IPCTA-TDMA algorithm is $\mathcal O(L_{\text{TD}}(M+N)^3)$, where $L_{\text{TD}}$ denotes the number of iterations of the IPCTA-TDMA algorithm. \section{Numerical Results} In this section, we evaluate the proposed schemes through simulations. There are 40 MTCDs uniformly distributed with a BS in the center. We adopt the ``data-centric'' clustering technique in \cite{6649207} for cluster formation of MTCDs, the number of MTCGs is set as 12, and the maximal number of MTCDs associated to one MTCG is 4. For NOMA, all MTCGs are classified into 6 clusters based on the strong-weak scheme \cite{Mohammad17uplinkNOMA,7273963}. The path loss model is $128.1+37.6\log_{10} d$ ($d$ is in km) and the standard deviation of shadow fading is $4$ dB \cite{access2010further}. The noise power $\sigma^2=-104$ dBm, and the bandwidth of the system is $B=18$ KHz. For the non-linear EH model in (\ref{sys1eq5_2}), we set $M=24$ mW, $a=1500$ and $b=0.0014$ according to \cite{7264986}, which are obtained by curve fitting from the measurement data in \cite{6259492}. The receiver sensitivity threshold $P_0$ is set as $0.1$ mW. The PA efficiencies of each MTCD and MTCG are set to $\eta=\xi=0.9$, and the circuit power of each MTCG is $Q^{\text C}=500$ mW as in \cite{Wu2016Energy}. We assume equal throughput demand for all MTCDs, i.e., $D_1=\cdots=D_M=D$, and equal maximal transmission power for each MTCD or MTCG, i.e., $P_1=\cdots=P_M=P$, and $Q_1=\cdots=Q_N=Q$. Unless otherwise specified, parameters are set as $P=5$ mW, $P^{\text C}=0.5$ mW, $Q=1$ W, $D=10$ Kbits, and $T=5$ s. Fig.~\ref{fig5} depicts, for instance, $E_{11}$ in (\ref{ee2eq3}) consumed by MTCD 1 served by MTCG 1 versus the transmission time $t_1$ for NOMA. It is observed that $E_{11}$ monotonically decreases with $t_1$ when $P^C=0$. For the case with $P^C=5$ mW or $P^C=10$ mW, $E_{11}$ first decreases and then increases with $t_1$. Both observations validate our theoretical findings in Theorem 1. \begin{figure} \centering \includegraphics[width=3in]{fig6.eps} \vspace{-1em} \caption{Energy $E_{11}$ consumed by MTCD 1 versus the transmission time $t_1$ for NOMA strategy.}\label{fig5} \vspace{-1em} \end{figure} The convergence behaviors of IPCTA-NOMA and IPCTA-TDMA are illustrated in Fig.~\ref{fig12}. From this figure, the total energy of both algorithms monotonically decreases, which confirms the convergence analysis in Section III-C and IV-C. It can be seen that both IPCTA-NOMA and IPCTA-TDMA converge rapidly. \begin{figure} \centering \includegraphics[width=3in]{fig7.eps} \vspace{-1em} \caption{Convergence behaviors of IPCTA-NOMA and IPATC-TDMA.}\label{fig12} \vspace{-1em} \end{figure} In Fig.~\ref{fig10}, we illustrate the total energy consumption versus the circuit power of each MTCD. According to Fig.~\ref{fig10}, the total energy of NOMA outperforms TDMA when the circuit power of each MTCD is low, i.e., $P^C\leq4$ mW in the test case. At low circuit power regime, the total energy consumption of the network mainly lies in the RF transmission power of MTCDs and the energy consumed by MTCGs to charge the MTCDs through EH. For the NOMA strategy, MTCDs served by the same MTCG can simultaneously upload data and the MTCG decodes the messages according to NOMA detections, which requires lower RF transmission power of MTCDs than the TDMA strategy. Thus, the total energy of NOMA is less than TDMA for low circuit power of each MTCD. From Fig.~\ref{fig10}, we can find that the total energy of TDMA outperforms NOMA when the circuit power of each MTCD becomes high, i.e., $P^C\geq5$ mW in our tests. At high circuit power regime, the total energy consumption of the network mainly lies in the circuit power of MTCDs and the energy consumed by MTCGs to charge the MTCDs through EH. For the NOMA strategy, the transmission time of each MTCD with NOMA is always longer than that with TDMA, which leads to higher circuit power consumption of MTCDs than the TDMA strategy. As a result, TDMA enjoys better energy efficiency than NOMA for high circuit power of each MTCD. \begin{figure} \centering \includegraphics[width=3in]{fig8.eps} \vspace{-1em} \caption{Total energy versus the circuit power of each MTCD.}\label{fig10} \vspace{-1em} \end{figure} Fig.~\ref{fig6} illustrates the total energy versus the maximal transmission power of each MTCG. It is observed that the total energy decreases with the increase of maximal transmission power of each MTCG for both NOMA and TDMA. This is because that the increment of maximal transmission power of each MTCG allows the MTCG to transmit with large transmission power, which leads to short EH time of MTCDs to harvest enough energy and low total energy consumption of the network. Moreover, it is found that the total energy of NOMA is more sensitive to the maximal transmission power of each MTCG than that of TDMA for high circuit power case as $P^{\text C}=5$ mW for each MTCD. The reason is that MTCD with low channel gain receives intra-cluster interference due to NOMA and the energy consumption is hence especially large for low maximal transmission power of each MTCG and high circuit power of each MTCD. The total energy versus the maximal transmission power of each MTCD is shown in Fig.~\ref{fig9}. It is observed that the total energy decreases with growing maximal transmission power of each MTCD for both NOMA and TDMA. This is due to the fact that a larger maximal transmission power of each MTCD ensures MTCDs can transmit with more power, and the required payload can be uploaded in a shorter time, which results in low circuit power consumption and low energy consumption. It can be found that the total energy of TDMA converges faster than that of NOMA as the maximal transmission power of each MTCD increases. This is because that the MTCDs served by the same MTCG simultaneously transmit data for NOMA, and the required transmission power of each MTCD for NOMA is always larger than that of MTCD for TDMA. \begin{figure} \centering \includegraphics[width=3in]{fig9.eps} \vspace{-1em} \caption{Total energy versus the maximal transmission power of each MTCG.}\label{fig6} \vspace{-1em} \end{figure} \begin{figure} \centering \includegraphics[width=3in]{fig10.eps} \vspace{-1em} \caption{Total energy versus the maximal transmission power of each MTCD.}\label{fig9} \vspace{-1em} \end{figure} Finally, in Fig.~\ref{fig7}, we illustrate the total energy versus the required payload of each MTCD. The figure shows that the total energy increases with the required payload of each MTCD. This is due to the fact that large payload of each MTCD requires large energy consumption of MTCDs and MTCGs, which leads to high energy consumption of the network. \begin{figure} \centering \includegraphics[width=3in]{fig11.eps} \vspace{-1em} \caption{Total energy versus the required payload of each MTCD.}\label{fig7} \vspace{-1em} \end{figure} \section{Conclusions} This paper compares the total energy consumption between NOMA and TDMA strategies in uplink M2M communications with EH. We formulate the total energy minimization problem subject to minimal throughput constraints, maximal transmission power constraints and energy causality constraints, with the circuit power consumption taken into account. By applying the conditions that it is optimal to transmit with the minimal throughput for each MTCD, we transform the original problem for NOMA strategy into an equivalent problem, which is suboptimally solved through an iterative algorithm. By using a proper variable transformation, we transform the nonconvex problem for TDMA into an equivalent problem, which can be effectively solved. Through simulations, either NOMA strategy or TDMA strategy may be preferred depending on different circuit power regimes of MTCDs. At low circuit power regime of MTCDs, NOMA consumes less energy, while TDMA is preferred at high circuit power regime of MTCDs since the energy consumption for NOMA increases significantly as the circuit power of MTCDs increases. \appendices \section{Proof of Lemma 2} \setcounter{equation}{0} \renewcommand{\theequation}{\thesection.\arabic{equation}} By inserting $r_{ij}=D_j$ into (\ref{sys1eq3}) from Lemma 1, we have \begin{equation}\label{appenAeq1} 2^{\frac{D_{j}}{Bt_i}} \sum_{l=j+1}^{J_i} |h_{il}|^2p_l + \sigma^2 \left(2^{\frac{D_{j}}{Bt_i}} -1\right) = \sum_{l=j}^{J_i} |h_{il}|^2 p_l, \end{equation} for $j=J_{i-1}+1, \cdots, J_i$. To solve those $J_i-J_{i-1}$ equations, we first define \begin{equation}\label{appenAeq2} u_{j}=\sum_{l=j}^{J_i} |h_{il}|^2 p_{l}, \quad \forall j\in\mathcal J_i. \end{equation} Applying (\ref{appenAeq2}) into (\ref{appenAeq1}) yields \begin{equation}\label{appenAeq2_1} {u_{j}}={2^{\frac{D_{j}}{Bt_i}}} u_{{j+1}}+ {\sigma^2 } \left(2^{\frac{D_{j}}{Bt_i}} -1\right), \quad \forall j\in \mathcal J_i . \end{equation} Denote $\pmb u_i=[u_{J_{i-1}+1}, \cdots, u_{J_i}]^T$, \begin{equation}\label{appenAeq2_2_0} \pmb v_i=\left[{\sigma^2 } \left(2^{\frac{D_{J_{i-1}+1}}{Bt_i}} -1\right), \cdots, {\sigma^2 } \left(2^{\frac{D_{J_i}}{Bt_i}} -1\right)\right]^T, \end{equation} and \begin{equation}\label{appenAeq2_2_2} \pmb W_i=\begin{bmatrix} 0&{2^{\frac{D_{J_{i-1}+1}}{Bt_i}}}\\ &0 &\!\!\!\!\!\!\!\!{2^{\frac{D_{J_{i-1}+2}}{Bt_i}}} & \\ &&\ddots&\ddots\\ & && 0&{2^{\frac{D_{J_{i}-1}}{Bt_i}}}\\ &&&& 0 \end{bmatrix}. \end{equation} Equations in (\ref{appenAeq2_1}) can be rewritten as \begin{equation}\label{appenAeq2_2_5} (\pmb E - \pmb W_i) \pmb u_i=\pmb v_i, \end{equation} where $\pmb E$ is an identity matrix of size $(J_i-J_{i-1})\times(J_i-J_{i-1})$. From (\ref{appenAeq2_2_5}), we have \begin{equation}\label{appenAeq2_2_6} \pmb u_i=(\pmb E - \pmb W_i)^{-1}\pmb v_i. \end{equation} Before obtaining the inverse matrix $(\pmb E - \pmb W_i)^{-1}$, we present the following lemma. \begin{lemma} When $l\in[1,J_i-J_{i-1}-1]$, the $l$-th power of matrix $\pmb W_i$ can be expressed as \begin{equation}\label{appenAeq2_2_7} \pmb W_i^l\!=\!\begin{bmatrix} \pmb 0_{l}^T&\!\!\!\!{2^{\frac{\sum_{s=J_{i-1}+1}^{J_{i-1}+l}\!D_{s}}{Bt_i}}}\!\!\\ &\!\!\!\!0 &\!\!\!\!\!\!\!\!\!\!\!\!{2^{\frac{\sum_{s=J_{i-1}+2}^{J_{i-1}+l+1}\!D_{s}}{Bt_i}}} \!\!& \\ &&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\ddots&\!\!\!\!\!\!\!\!\ddots\\ & && \!\!\!\!0&\:\; {2^{\frac{\sum_{s=J_{i}-l}^{J_{i}-1}\!D_{s}}{Bt_i}}}\!\!\\ &&&& \!\!\!\!\pmb 0_l\!\! \end{bmatrix}, \end{equation} where $\pmb 0_{l}$ denotes a $l\times 1$ vector of zeros. When $l= J_i - J_{i-1}$, $\pmb W_i^l=\pmb 0_{(J_i-J_{i-1})\times(J_i-J_{i-1})}$, where $\pmb 0_{(J_i-J_{i-1})\times(J_i-J_{i-1})}$ is a ${(J_i-J_{i-1})\times(J_i-J_{i-1})}$ matrix of zeros. \end{lemma} \itshape \textbf{Proof:} \upshape Lemma 4 can be proved by the principle of mathematical induction. \emph{Basis}: It can be verified that Lemma 4 is valid for $l=1$. \emph{Induction Hypothesis}: For $l\in[1,J_i-J_{i-1}-2]$, assume that the $l$-th power of matrix $\pmb W_i$ can be expressed as (\ref{appenAeq2_2_7}). \emph{Induction Step}: According to (\ref{appenAeq2_2_7}), we can obtain \begin{eqnarray} &&\!\!\!\!\!\!\!\!\!\!\pmb W_i^{l+1}\! =\!\pmb W_i^{l} \pmb W_i \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! = \begin{bmatrix} \pmb 0_{l}^T&\!\!\!\!{2^{\frac{\sum_{s=J_{i-1}+1}^{J_{i-1}+l}\!D_{s}}{Bt_i}}}\!\!\\ &\!\!\!\!0 &\!\!\!\!\!\!\!\!\!\!\!\!{2^{\frac{\sum_{s=J_{i-1}+2}^{J_{i-1}+l+1}\!D_{s}}{Bt_i}}} \!\!& \\ &&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\ddots&\!\!\!\!\!\!\!\!\ddots\\ & && \!\!\!\!0&\:\; {2^{\frac{\sum_{s=J_{i}-l}^{J_{i}-1}\!D_{s}}{B t_i}}}\!\!\\ &&&& \!\!\!\!\pmb 0_l\!\! \end{bmatrix} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \quad \cdot \begin{bmatrix} 0&{2^{\frac{D_{J_{i-1}+1}}{B t_i}}}\\ &0 &\!\!\!\!\!\!\!\!{2^{\frac{D_{J_{i-1}+2}}{B t_i}}} & \\ &&\ddots&\ddots\\ & && 0&{2^{\frac{D_{J_{i}-1}}{B t_i}}}\\ &&&& 0 \end{bmatrix} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! = \begin{bmatrix} \pmb 0_{l+1}^T&\!\!\!\!{2^{\frac{\sum_{s=J_{i-1}+1}^{J_{i-1}+l+1}\!D_{s}}{B t_i}}}\!\!\\ &\!\!\!\!0 &\!\!\!\!\!\!\!\!\!\!\!\!{2^{\frac{\sum_{s=J_{i-1}+2}^{J_{i-1}+l+2}\!D_{s}}{B t_i}}} \!\!& \\ &&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\ddots&\!\!\!\!\!\!\!\!\ddots\\ & && \!\!\!\!0&\:\;{2^{\frac{\sum_{s=J_{i}-l-1}^{J_{i}-1}\!D_{s}}{B t_i}}}\!\!\\ &&&& \!\!\!\!\pmb 0_{l+1}\!\! \end{bmatrix},\nonumber \end{eqnarray} which verifies that the $(l+1)$-th power of matrix $\pmb W_i$ can also be expressed as (\ref{appenAeq2_2_7}). According to (\ref{appenAeq2_2_7}) and (\ref{appenAeq2_2_2}), it is verified that \begin{equation}\label{appenAeq2_2_8} \pmb W_i^{J_i-J_{i-1}}=\pmb W_i^{J_i-J_{i-1}-1}\pmb W_i=\pmb 0_{(J_i-J_{i-1})\times (J_i-J_{i-1})}. \end{equation} Therefore, Lemma 4 is proved. \hfill $\Box$ Now, it is ready to obtain the inverse matrix $(\pmb E - \pmb W_i)^{-1}$. Since \begin{equation}\label{appenAeq2_2_9} (\pmb E - \pmb W_i)\left(\pmb E +\sum_{l=1}^{{J_i-J_{i-1}-1}} \pmb W_i^l\right)= \pmb E - \pmb W_i^{J_i-J_{i-1}}=\pmb E, \end{equation} we have \begin{equation}\label{appenAeq2_2_9_1} (\pmb E - \pmb W_i)^{-1}=\pmb E +\sum_{l=1}^{{J_i-J_{i-1}-1}} \pmb W_i^l. \end{equation} Substituting (\ref{appenAeq2_2_7}) and (\ref{appenAeq2_2_9_1}) into (\ref{appenAeq2_2_6}) yields \begin{equation}\label{appenAeq2_2_9_2} u_j=\sum_{l=j}^{J_i} {\sigma^2} \left(2^{\frac{D_{l}}{Bt_i}} -1\right) {2^{\frac{\sum_{s=j}^{l-1}D_{s}}{Bt_i}}}, \quad \forall j\in \mathcal J_i, \end{equation} where we define $\prod_{s=j}^{j-1} {2^{\frac{D_{s}}{B t_i}}}=2^0$. From (\ref{appenAeq2_2_9_2}), we can obtain $u_j={\sigma^2}\left({2^{\frac{\sum_{s=j}^{J_i}D_{s}}{Bt_i}}}-1\right)$. Since $\sum_{l=J_i+1}^{J_i}p_{j}=0$, we have \begin{equation}\label{appenAeq2_2} u_{J_i+1}=0. \end{equation} From (\ref{appenAeq2}) and (\ref{appenAeq2_2}), we can obtain the transmission power of MTCD $j$ as \begin{equation}\label{appenAeqpowervalue} p_{ij}=\frac{u_{j}-u_{j+1}} {|h_{ij}|^2}, \quad \forall j \in \mathcal J_i. \end{equation} Applying (\ref{appenAeq2_2_9_2}) and (\ref{appenAeq2_2}) to (\ref{appenAeqpowervalue}), we have \begin{eqnarray}\label{appenAeq3} p_{j} &&\!\!\!\!\!\!\!\!\!\! =\sum_{l=j}^{J_i} \frac{ \sigma^2} {|h_{ij}|^2 } \left(2^{\frac{D_{l}}{B t_i}} -1\right){2^{\frac{\sum_{s=j}^{l-1}D_{s}}{B t_i}}} \nonumber \\ &&\!\!\!\!\!\!\!\!\!\!\quad- \sum_{l=j+1}^{J_i} \frac{ \sigma^2} {|h_{ij}|^2 } \left(2^{\frac{D_{l}}{B t_i}} -1\right) {2^{\frac{\sum_{s=j+1}^{l-1}D_{s}}{B t_i}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! = \!\sum_{l=j+1}^{J_i} \frac{\sigma^2} {|h_{ij}|^2 } \!\left(\!2^{\frac{D_{l}}{B t_i}} \!-\!1\!\right)\! \left(\!{2^{\frac{D_{j}}{B t_i}}}\!-\!1\!\right)\!{2^{\frac{\sum_{s=j+1}^{l-1}D_{s}}{B t_i}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!\quad+ \frac{\sigma^2} {|h_{ij}|^2 } \left(2^{\frac{D_{j}}{B t_i}} -1\right){2^{\frac{\sum_{s=j}^{j-1}D_{s}}{B t_i}}}, \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! = \sum_{l=j+1}^{J_i} \frac{\sigma^2} {|h_{ij}|^2 } \left( {\text e}^{\frac{a_{l}}{t_i}} - 1 \right) \left( {\text e^{\frac{a_{j}}{t_i}}} - 1 \right) {\text e^{ \frac{b_{jl}}{ t_i}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!\quad+ \frac{\sigma^2} {|h_{ij}|^2 } \left(\text e^{\frac{a_{j}}{t_i}} -1\right), \quad \forall j \in \mathcal J_i. \end{eqnarray} where $a_l$ and $b_{jl}$ are defined in (\ref{ee2eq2_2}). Since $\text{e}^{x} -1$ with $x\geq 0$ is non-negative and decreases with $t_i$, $ p_{j}$ is also non-negative and decreases with $t_i$ from (\ref{appenAeq3}). \section{Proof of Theorem 1} \setcounter{equation}{0} \renewcommand{\theequation}{\thesection.\arabic{equation}} To show that the energy $E_{ij}$ is convex w.r.t. $t_i$, we first define function \begin{eqnarray}\label{appenBeq1} f_{ijl}(x)= { ( {\text e}^{ {a_{l}x} } - 1 ) ( {\text e^{ {a_{j}x} }} - 1 ) \text e^{ {b_{jl}x} }}, \quad \forall x\geq 0. \end{eqnarray} Then, the second-order derivative follows \begin{eqnarray}\label{appenBeq2} f_{ijl}''(x)= &&\!\!\!\!\!\!\!\!\!\! (a_{l}^2+2a_{l}b_{jl}) ( {\text e^{ {a_{j}x} }} - 1 )\text e^{ ({a_{l}+b_{jl}})x} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! +2a_{l} a_{j} \text e^{ ({a_{l}+ a_{j}+b_{jl}})x} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! +(a_{j}^2+2a_{j}b_{jl}) ( {\text a^{ {a_{l}x} }} - 1 )\text e^{ ({a_{j}+b_{jl}})x} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! +b_{jl}^2 ( {\text e}^{ {a_{l}x} } - 1 ) ( {\text e^{ {a_{j}x} }} - 1 )\text e^{ b_{jl} x}\geq 0, \end{eqnarray} which indicates that $f_{ijl}(x)$ is convex w.r.t. $x$. According to \cite[Page~89]{boyd2004convex}, the perspective of $u(x)$ is the function $v(x,t)$ defined by $v(x,t)=tu(x/t)$, ${\textbf{dom}}\:v= \{(x,t)|x/t\in {\textbf{dom}}\:u, t>0\}$. If $u(x)$ is a convex function, then so is its perspective function $v(x,t)$ \cite[Page~89]{boyd2004convex}. Since $f_{ijl}(x)$ is convex, the perspective function \begin{equation}\label{appenBeq3} \bar f_{ijl}(x, t_i)=t_i f_{ijl}\left(\frac x {t_i} \right) \end{equation} is convex w.r.t. ($x, t_i$). Thus, function $\bar f_{ijl}(1, t_i)$ is also convex w.r.t. $t_i$. Defining function \begin{eqnarray}\label{appenBeq5_1} g_{ij}(x)= {\text e^{ {a_{j}}x} -1}, \quad \end{eqnarray} which is convex w.r.t. $x$. By using the property of perspective function \cite[Page~89]{boyd2004convex}, we also have that function \begin{equation}\label{appenBeq5_3} \bar g_{ij}(x, t_i)=t_i g_{ij}\left(\frac x {t_i} \right) \end{equation} is convex w.r.t. ($x, t_i$) and function $\bar g_{ij}(1, t_i)$ is accordingly convex w.r.t. $t_i$. Substituting (\ref{appenBeq3}) and (\ref{appenBeq5_3}) into (\ref{ee2eq3}), we can obtain \begin{equation}\label{appenBeq5_5} E_{ij}= \sum_{l=j+1}^{J_i} \frac{\sigma^2} {\eta |h_{ij}|^2 } \bar f_{ijl}(1, t_i) + \frac{\sigma^2} {\eta |h_{ij}|^2 }\bar g_{ij}(1, t_i) + t_i P_{\text C}. \end{equation} Due to the fact that both $\bar f_{ijl}(1, t_i)$ and $\bar g_{ij}(1, t_i)$ are convex, $E_{ij}$ is consequently convex w.r.t. $t_i$ from (\ref{appenBeq5_5}). According to (\ref{ee2eq3}), the first-order derivative of $E_i$ w.r.t. $t_i$ is expressed as \begin{eqnarray}\label{appenBeq5_7} \frac {\partial E_{ij}}{\partial t_i}&&\!\!\!\!\!\!\!\!\!\!= \sum_{l=j+1}^{J_i} \frac{\sigma^2 } {\eta |h_{ij}|^2 } {\left( {\text e}^{\frac{a_{l}}{t_i}} - 1 \right) \left( {\text e^{\frac{a_{j}}{t_i}}} - 1 \right) \text e^{ \frac{b_{jl}}{ t_i}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!\quad - \sum_{l=j+1}^{J_i} \frac{\sigma^2 a_l} {\eta |h_{ij}|^2 t_i} { \left( {\text e^{\frac{a_{j}}{t_i}}} - 1 \right) \text e^{ \frac{a_l+b_{jl}}{ t_i}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \quad- \sum_{l=j+1}^{J_i} \frac{\sigma^2 a_{j}} {\eta |h_{ij}|^2 t_i} { \left( {\text e^{\frac{a_{l}}{t_i}}} - 1 \right) \text e^{ \frac{a_{j}+b_{jl}}{ t_i}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!\quad - \sum_{l=j+1}^{J_i} \frac{\sigma^2 b_{jl}} {\eta |h_{ij}|^2 t_i } {\left( {\text e}^{\frac{a_{l}}{t_i}} - 1 \right) \left( {\text e^{\frac{a_{j}}{t_i}}} - 1 \right) \text e^{ \frac{b_{jl}}{ t_i}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \quad+ \frac{\sigma^2 } {\eta |h_{ij}|^2 } {\left(\text e^{\frac{a_{j}}{t_i}} -1\right) } - \frac{\sigma^2 a_j} {\eta |h_{ij}|^2 t_i} { \text e^{ \frac{a_j}{ t_i}}}+ P_{\text C}. \end{eqnarray} Since $E_{ij}$ is convex w.r.t. $t_i$, function $\frac{\partial E_{ij}}{\partial t_i}$ increases with $t_i$. Because $D_j$ is positive for all $j$, we have $a_{l}>0$, $b_{ijl}>0$ and $c_{ijl}>0$ from (\ref{ee2eq2_2}). To calculate the limit of $\frac{\partial E_{ij}}{\partial t_i}$ at $t_i=0+$, we calculate the following limit, \begin{eqnarray}\label{appenBeq5_8} &&\!\!\!\!\!\!\!\!\!\!\quad\lim_{t_i\rightarrow 0+} \frac{\sigma^2 } {\eta |h_{ij}|^2 } {\left( {\text e}^{\frac{a_{l}}{t_i}} - 1 \right) \left( {\text e^{\frac{a_{j}}{t_i}}} - 1 \right) \text e^{ \frac{b_{jl}}{ t_i}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \qquad- \frac{\sigma^2 a_l} {\eta |h_{ij}|^2 t_i} { \left( {\text e^{\frac{a_{j}}{t_i}}} - 1 \right) \text e^{ \frac{a_l+b_{jl}}{ t_i}}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!=\lim_{x\rightarrow +\infty} \frac{\sigma^2 } {\eta |h_{ij}|^2 } {\left( {\text e}^{ {a_{l}}x} - 1 \right) \left( {\text e^{ {a_{j}}x}} - 1 \right) \text e^{ {b_{jl}}x}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \qquad- \frac{\sigma^2 a_l x} {\eta |h_{ij}|^2 } { \left( {\text e^{ {a_{j}}x}} - 1 \right) \text e^{ ({a_l+b_{jl}})x}} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! =\lim_{x\rightarrow +\infty} \frac{\sigma^2 (1-a_l x)} {\eta |h_{ij}|^2 } {\text e}^{( {a_{l}}+{a_{j}}+b_{jl})x} \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! =-\infty. \end{eqnarray} Thus, when $t_i$ approaches zero in the positive direction, we have \begin{eqnarray}\label{appenBeq5_10} \lim_{t_i \rightarrow 0+}\frac {\partial E_{ij}}{\partial t_i} =-\infty. \end{eqnarray} When $t_i$ approaches positive infinity, the limit of first-order derivative $\frac {\partial E_{ij}}{\partial t_i}$ can be calculated as \begin{eqnarray}\label{appenBeq5_11} \lim_{t_i \rightarrow +\infty}\frac {\partial E_{ij}}{\partial t_i} = P_{\text C}. \end{eqnarray} If $P_{\text C}=0$, then $\frac {\partial E_{ij}}{\partial t_i} \leq 0$ for all $t_i \geq 0$. In this case, the energy $E_{ij}$ always decreases with $t_i$. If $P_{\text C}>0$, we can observe that $\lim_{t_i \rightarrow +\infty}\frac {\partial E_{ij}}{\partial t_i}>0$ from (\ref{appenBeq5_11}). Since $\lim_{t_i \rightarrow 0+}\frac {\partial E_{ij}}{\partial t_i}<0$ and $\frac{\partial E_{ij}}{\partial t_i}$ increases with $t_i$, there exists one unique solution $T_{ij}^*$ satisfying (\ref{ee2eq5}), which can be solved by using the bisection method. In particular, $E_{ij}$ decreases with $0\leq t_i\leq T_{ij}^*$ and increases with $t_{ij} > T_{ij}^*$. \section{Proof of Theorem 2} \setcounter{equation}{0} \renewcommand{\theequation}{\thesection.\arabic{equation}} We first consider the case that $T\leq \max_{\forall i \in \mathcal N}$ $\min_{\forall j \in \mathcal J_i} \{T_{ij}^*\}$. Without loss of generality, we denote $T\leq T_{nm}^*=\max_{\forall i \in \mathcal N}\min_{\forall j \in \mathcal J_i} \{T_{ij}^*\}$. Assume that the optimal solution ($\pmb p^*, \pmb q^*, \pmb t^*$) to problem (\ref{ee2min1}) satisfies $\sum_{i=1}^{N+K}t_i^*<T$. Due to that $t_i^*\geq0$ for all $i$, we can obtain that $t_n^*\leq T_{nm}^*\leq T_{nl}^*$, $\forall l \in \mathcal J_n$. With all other power $p_j^*$, $q_i^*$ and time $t_k^*$ fixed, $j\in\mathcal M\setminus\mathcal J_n$, $i\in \mathcal N$, $k\in\mathcal N\setminus\{n\}$, we increase the time $t_n^*$ to $t_n'=t_n^*+\epsilon$ by an arbitrary amount $(0, T-\sum_{i=1,i\neq n}^N t_i^*]$. Using (\ref{ee2eq2}), the corresponding power $p_l^*$ strictly deceases to $p_l'$, $l\in\mathcal J_n$. According to Theorem 1, the energy $E_{nl}$ decreases with the transmission time $0\leq t_n \leq T_{nl}^*$, $\forall l \in \mathcal J_n$. As a result, with new power-time pair ($p_{J_{n-1}+1}', \cdots, p_{J_n}', t_n'$), the objective function (\ref{ee2min1}a) decreases with all the constraints satisfied. By contradiction, we must have $\sum_{i=1}^{N+K}t_i^*=T$ for the optimal solution. This completes the proof of the first half part of Theorem 2. The last half part of Theorem 2 indicates that transmitting with maximal transmission time is not optimal when $T$ is larger than a threshold $T_{\text{Upp}}$. This can be proved by using the contradiction method. Specifically, assuming that total transmission time of the optimal solution is the maximal transmission time $T$, we can find a special solution with total transmission time less than $T$, which strictly outperforms the optimal solution. To obtain the threshold $T_{\text{Upp}}$, we consider a special solution that satisfies all the constraints of problem (\ref{ee2min1}) except the maximal transmission time constraint (\ref{ee2min1}e). Since $E_{ij}$ is convex w.r.t. $t_i$ according to Theorem 1, the energy $E_{i}=\sum _{j =J_{i-1}+1}^{J_i} E_{ij}$ consumed by all MTCDs in $\mathcal J_i$ served by MTCG $i$ is also convex w.r.t. $t_i$. Based on the proof of Theorem 1, we directly obtain the following lemma. \begin{lemma} The energy $E_i$ consumed by all MTCDs in $\mathcal J_i$ first decreases with the transmission time $t_i$ when $0\leq t_i \leq T_i^*$ and then increases with the transmission time $t_i$ when $t_i>T_i^*$, where $T_i^*$ is the unique zero point of first-order derivative $\frac {\partial E_i}{\partial t_i}$, i.e., \begin{eqnarray}\label{appenCee2eq5} \left.\frac {\partial E_i}{\partial t_i} \right|_{t_i=T_i^*} =\sum _{j =J_{i-1}+1}^{J_i}\left.\frac {\partial E_{ij}}{\partial t_i} \right|_{t_i=T_i^*} =0. \end{eqnarray} \end{lemma} Since Lemma 5 can be proved by checking the first-order derivative $\frac {\partial E_i}{\partial t_i}$ as in Appendix B, the proof of Lemma 5 is omitted here. Set $\tilde t_i=T_i^*$, and $\tilde p_{j}$ can be obtained from (\ref{ee2eq2}) with $\tilde t_i$, $i \in \mathcal N$, $j \in \mathcal J_i$. For the transmission power of the MTCGs, we set $\tilde q_i=Q_i$, $i \in \mathcal N$. Denote $\tilde {\pmb p}=[\tilde p_1, \cdots, \tilde p_M]^T$, $\tilde{\pmb q}=[\tilde q_1, \cdots, \tilde q_N]^T$. With power $(\tilde {\pmb p}, \tilde {\pmb q})$ and time $(\tilde t_i, \cdots, \tilde t_N)$ fixed for now, the energy minimization problem (\ref{ee2min1}) without constraint (\ref{ee2min1}e) becomes \begin{subequations}\label{appenCee2min1} \begin{align} \!\!\mathop{\min}_{\pmb \tau}\; \!& \sum_{k=1}^K \sum_{i=I_{k-1}+1}^{I_k}\! t_{N+k} \!\left(\!\frac{Q_i}{\xi} \!+\! Q^{\text C}\!\right) \nonumber\\ & \!-\!\sum_{k=1}^K\sum_{i=1}^{ N}\! \sum_{j=J_{i-1}+1}^{ J_i} \! t_{N+1} u \left(\!\sum_{n=J_{k-1}+1}^{J_k} \!|h_{nj}|^2 Q_n \!\right) \\ \textrm{s.t.}\qquad \!\!\!\!\!\!\!\!\! &B t_{N+i}\log_2\!\left(\! 1\!+\!\frac{|h_{i}|^2 Q_i} { \sum_{n=i+1}^{I_k} \!|h_{n}|^2 Q_n\!+\!\sigma^2} \!\right)\! \!\geq \! \sum_{j\in\mathcal J_i} \!D_j, \nonumber\\ &\qquad\qquad\forall k \in \mathcal K, i \in \mathcal I_k \\ &\tilde t_i\! \left(\!\frac{\tilde p_j}{\eta} \!+\! P^{\text C}\!\right)\!\leq\! \sum_{k=1}^Kt_{N+k} u\!\left(\!\sum_{n=I_{k-1}+1}^{I_k} \! |h_{nj}|^2 Q_n\!\right), \nonumber\\ &\qquad\qquad \forall i \in \mathcal N, j \in \mathcal J_i \\ &\pmb \tau \geq \pmb 0, \end{align} \end{subequations} where $\pmb \tau=[t_{N+1}, \cdots, t_{N+K}]^T$. Problem (\ref{appenCee2min1}) can be obtained by substituting (\ref{sys1eq8}) into (\ref{ee2min1}a), (\ref{sys1eq4}) into (\ref{ee2min1}b), and (\ref{sys1eq5}) and (\ref{sys1eq6}) into (\ref{ee2min1}c). Problem (\ref{appenCee2min1}) is a linear problem, which can be optimally solved via the simplex method. Denote the optimal solution of problem by $ \pmb \tau^*=[T_{N+1}^*, \cdots, T_{N+K}^*]^T$. Denote $\tilde t_{N+k}=T_{N+k}^*$, $\forall k \in \mathcal K$, and $\tilde {\pmb t}=[\tilde t_{1}, \cdots, \tilde t_{N+K}]^T$. As a result, we obtain a special solution ($\tilde {\pmb p}, \tilde {\pmb q}, \tilde {\pmb t}$) that satisfies all the constraints of problem (\ref{ee2min1}) except the maximal transmission time constraint (\ref{ee2min1}e). With solution ($\tilde {\pmb p}, \tilde {\pmb q}, \tilde {\pmb t}$), the total energy consumption $\tilde E_{\text {Tot}}$ obtained from (\ref{sys1eq8}) can be expressed as \begin{eqnarray}\label{appenCeqn5} \!\!\!\!\!\!\tilde E_{\text {Tot}} &&\!\!\!\!\!\!\!\!\!\!=\sum_{i=1}^{N+K} \tilde E_i \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!=\sum_{i=1}^{ N}\sum_{j=J_{i-1}+1}^{ J_i}\tilde t_i \left(\frac{\tilde p_j}{\eta} + P^{\text C}\right)\nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \quad+\sum_{k=1}^K\sum_{i=I_{k-1}+1}^{I_k}\tilde t_{N+k} \left(\frac{\tilde q_i}{\xi} + Q^{\text C}\right) \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \quad\!-\!\sum_{k=1}^K\sum_{i=1}^{ N}\! \sum_{j=J_{i-1}+1}^{ J_i} \! \tilde t_{N+k} u\!\left(\!\sum_{n=I_{k-1}+1}^{I_k}\! |h_{nj}|^2\tilde q_n\!\right)\!\!, \end{eqnarray} where $\tilde E_i$ is the energy consumed by all MTCDs in $\mathcal J_i$, $i\in \mathcal N$, and $\tilde E_{N+k}$ is the system energy consumption during the ($N+k$)-th phase. Denote an upper bound of maximal transmission time by \begin{equation}\label{appenCeqn6} T_{\text{Upp}}=\max\left\{\sum_{i=1}^{N+K} T_i^*, T_{\text{Amp}} \right\}, \end{equation} where \begin{equation} T_{\text{Amp}}= \frac{\left( 1+\sum_{i=1}^N\beta_i\right)\sum_{k=1}^K\tilde E_{N+k} } {\alpha Q^{\text C}}, \end{equation} with $\alpha$ defined in (\ref{appenCeqn7_5}) and $\beta_i$ defined in (\ref{appenCeqn7_55}). If $T\geq T_{ \text{Upp}}$, we show that optimal solution ($\pmb p^*, \pmb q^*, \pmb t^*$) to problem (\ref{ee2min1}) must satisfy constraint (\ref{ee2eq6_3}), i.e., (\ref{ee2min1}e) is inactive, by contradiction. Assume that \begin{equation}\label{appenCeqn6_2} \sum_{i=1}^{N+K}t_i^*=T. \end{equation} With ($\pmb p^*, \pmb q^*, \pmb t^*$), we denote $E_i^*$ as the energy consumed by all MTCDs in $\mathcal J_i$, $i\in \mathcal N$, $E_{N+k}^*$ as the system energy consumption during the ($N+k$)-th phase, and $E_{\text {Tot}}^*$ as the total energy of the whole system. Thus, we have \begin{eqnarray}\label{appenCeqn7} E_{\text {Tot}}^*&&\!\!\!\!\!\!\!\!\!\!=\sum_{i=1}^{N+K} E_i^* \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \overset{(\text a)}{\geq} \sum_{i=1}^{N+1} \tilde E_i+\sum_{k=1}^K E_{N+k}^* \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \overset{(\text b)}{=} \sum_{i=1}^{N}\! \tilde E_i\! +\! \sum_{k=1}^{K} \! t_{N+k}^* \left( \!\sum_{n=I_{k-1}+1}^{I_k}\frac{ q_n^*}{\xi}\! - \!\sum_{i=1}^{ N} \!\sum _{j=J_{i-1}+1}^{J_i} \right. \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \left. \quad u\!\left(\!\sum_{n=I_{k-1}+1}^{ I_k}\!|h_{nj}|^2 q_n^* \!\right)\!\right) + Q^{\text C} \sum_{k=1}^K t_{N+k}^* (I_k-I_{k-1}) \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \overset{(\text c)}{>} \sum_{i=1}^{N} \tilde E_i+\alpha Q^{\text C} \sum_{k=1}^K t_{N+k}^* \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!\overset{(\text d)}{\geq}\sum_{i=1}^{N+K} \tilde E_i =\tilde E_{\text {Tot}}, \end{eqnarray} where inequality (a) follows from the fact that $E_i$ achieves the minimum when $t_i=T_i^*$ according to Lemma 5, equality (b) holds from (\ref{sys1eq5}) and (\ref{sys1eq7}), and inequality (c) follows from (\ref{sys1eq5}), (\ref{sys1eq7_2}), $\xi\in(0,1]$ and \begin{equation}\label{appenCeqn7_5} \alpha\triangleq\min_{k\in\mathcal K}(I_k-I_{k-1}). \end{equation} To explain procedure (d), we substitute (\ref{sys1eq5}) and (\ref{sys1eq6}) into energy causality constraints (\ref{ee2min1}d) to obtain \begin{equation}\label{appenCeqn7_2} t_i^* \left(\frac{p_j^*}{\eta} + P^{\text C}\right)\leq \sum_{k=1}^K t_{N+k}^* u\left(\sum_{n=I_{k-1}+1}^{I_k} |h_{nj}|^2 q_n\right), \end{equation} for all $i \in \mathcal N, j \in \mathcal J_i$. Considering that $p_j^*\geq 0$ in the left hand side of (\ref{appenCeqn7_2}) and $u(x)$ is a increasing function as well as $q_i^*\leq Q_i$ in the right hand side of (\ref{appenCeqn7_2}), we have \begin{eqnarray}\label{appenCeqn7_5_2} t_i^*&&\!\!\!\!\!\!\!\!\!\! \leq\min_{j\in\mathcal J_i}\left\{\frac{ \sum_{k=1}^Kt_{N+k}^* u\left(\sum_{n=I_{k-1}+1}^{I_k}|h_{nj}|^2 Q_n \right)}{P^C}\right\}, \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! \leq \beta_i\sum_{k=1}^Kt_{N+k}^* , \quad\forall i\in \mathcal N, \end{eqnarray} where \begin{equation}\label{appenCeqn7_55} \beta_i=\min_{j\in\mathcal J_i}\left\{\max_{k\in\mathcal K}\frac{u\left(\sum_{n=I_{k-1}+1}^{I_k}|h_{nj}|^2 Q_n \right)}{P^C}\right\}. \end{equation} Combining (\ref{appenCeqn6_2}) and (\ref{appenCeqn7_5}) yields \begin{equation}\label{appenCeqn7_6} \sum_{k=1}^K t_{N+k}^*\geq \frac{T} {1+\sum_{i=1}^N\beta_i}. \end{equation} Hence, inequality (d) follows from (\ref{appenCeqn6}) and (\ref{appenCeqn7_6}). According to (\ref{appenCee2min1}) and (\ref{appenCeqn6}), solution ($\tilde {\pmb p}, \tilde{\pmb q}, \tilde{\pmb t}$) is a feasible solution to problem (\ref{ee2min1}). From (\ref{appenCeqn7}), the objective value (\ref{ee2min1}a) can be decreased with solution ($\tilde {\pmb p}, \tilde{\pmb q}, \tilde{\pmb t}$), which contradicts that ($\pmb p^*, \pmb q^*, \pmb t^*$) is the optimal solution to problem (\ref{ee2min1}). Hence, the optimal solution to problem (\ref{ee2min1}) must satisfy constraint (\ref{ee2eq6_3}). \section{Proof of Theorem 3} \setcounter{equation}{0} \renewcommand{\theequation}{\thesection.\arabic{equation}} We first show that the feasible set of problem (\ref{ee2min2}) with given $\pmb \tau$ is convex. Obviously, constraints (\ref{ee2min2}b), (\ref{ee2min2}d), (\ref{ee2min2}e) and (\ref{ee2min2}g) and (\ref{ee2min2}h) are all linear w.r.t. ($\pmb q, \bar {\pmb t}$). According to (\ref{appenBeq1}) and (\ref{appenBeq3}), constraints (\ref{ee2min2}c) and (\ref{ee2min2}f) can be, respectively, reformulated as \begin{equation}\label{appenDeq1} \sum_{l=j+1}^{J_i} \bar f_{ijl}(1, t_i) + t_i P^{\text C} \leq \sum_{k\in\mathcal S_{ij}} t_{N+k} \bar u\left(\sum_{n=I_{k-1}+1}^{ I_k } |h_{nj}|^2 q_n \right) \end{equation} for all $i\in \mathcal N, j \in \mathcal J_i$, and \begin{equation}\label{appenDeq2} \sum_{l=j+1}^{J_i} f_{ijl}\left(\frac 1 {t_i} \right) \leq P_j, \quad\forall i\in \mathcal N, j \in \mathcal J_i. \end{equation} Based on (\ref{sys1eq5_22}), we have \begin{equation}\label{appenDeq2_2} \bar u''(x)=\frac{-M a^2 (1+\text e ^{ab}) \text e^{-a(x-b)}}{\text e^{ab} (1+\text{e} ^{-a(x-b)})^3} \leq 0, \quad \forall x\geq 0, \end{equation} which shows that $\bar u(x)$ is concave, and $-\bar u(x)$ is convex. Since $\bar f_{ijl}(1, t_i)$ is convex w.r.t. $t_i$ according to Appendix B, constraints (\ref{appenDeq1}) are convex, i.e., constraints (\ref{ee2min2}c) are convex. Due to that $f_{ijl}(x)$ is convex and non-decreasing, and $\frac 1 x$ is convex, $f_{ijl}\left(\frac 1 {x} \right) $ is also convex according to the composition property of convex functions \cite[Page~84]{boyd2004convex}. Thus, constraints (\ref{appenDeq2}) are convex, i.e., constraints (\ref{ee2min2}f) are convex. We then show that the objective function (\ref{ee2min2}a) is convex. Substituting (\ref{appenBeq3}) and (\ref{appenBeq5_3}) into (\ref{ee2min2}a) yields \begin{eqnarray}\label{appenDeq3} &&\!\!\!\!\!\!\!\!\!\!\quad\sum_{i=1}^N \sum_{j=J_{i-1}+1}^{ J_i} \sum_{l=j+1}^{J_i} \frac{\sigma^2} {\eta |h_{ij}|^2 } \bar f_{ijl}(1, t_i) \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! +\sum_{i=1}^N \sum_{j=J_{i-1}+1}^{ J_i} \frac{\sigma^2} {\eta |h_{ij}|^2 } \bar g_{ij}(1, t_i) \nonumber\\ &&\!\!\!\!\!\!\!\!\!\!+\sum_{i=1}^N \sum_{j=J_{i-1}+1}^{ J_i}t_i P_{\text C} +\sum_{k=1}^K\sum_{i=I_{k-1}+1}^{I_k} t_{N+k} \left(\frac{q_i}{\xi} + Q^{\text C}\right) \nonumber\\ &&\!\!\!\!\!\!\!\!\!\! -\sum_{i=1}^{ N} \sum_{j=J_{i-1}+1}^{ J_i} \sum_{k\in\mathcal S_{ij}} t_{N+k} \bar u\left(\sum_{n=I_{k-1}+1}^{ I_k} |h_{nj}|^2 q_n\right), \end{eqnarray} which is convex w.r.t. ($\pmb q, \bar{\pmb t}$) because $\bar f_{ijl}(1, t_i)$, $\bar g_{ij}(1, t_i)$ and $-\bar u(x)$ are convex according to Appendix B and (\ref{appenDeq2_2}). As a result, problem (\ref{ee2min2}) is convex w.r.t. ($\pmb q, \bar {\pmb t}$). With given ($\pmb q, \bar {\pmb t}$), constraints (\ref{ee2min2}b) can be equivalently transformed into \begin{equation}\label{appendeq5} B t_{N+i}\log_2\left( 1+\frac{|h_{i}|^2 q_i} { \sum_{n=i+1}^{I_k} |h_{n}|^2 q_n+\sigma^2} \right) \geq \sum_{j=J_{i-1}+1}^{ J_i} D_j, \end{equation} which is linear w.r.t. $t_{N+i}$. By replacing constraints (\ref{ee2min2}b) with (\ref{appendeq5}), problem (\ref{ee2min2}) with given ($\pmb q, \bar {\pmb t}$) is a linear problem. \section{Proof of Theorem 4} \setcounter{equation}{0} \renewcommand{\theequation}{\thesection.\arabic{equation}} The proof is established by showing that the total energy value (\ref{ee2min2}a) is non-increasing when sequence ($ {\pmb q}, {\pmb t}$) is updated. According to the IPCTA-NOMA algorithm, we have \begin{equation}\label{convergenceProofIPCTA} U_{\text {Obj}}^{(v-1) =E_{\text {Tot}}({\pmb q}^{(v-1)}, {\pmb t}^{(v-1)}) \geq E_{\text {Tot}}({\pmb q}^{(v)}, {\pmb t}^{(v)}) =U_{\text{Obj}}^{(v)}, \end{equation} where the inequality follows from that (${\pmb q}^{(v)}, {\pmb t}^{(v)}$) is a suboptimally optimal solution of problem (\ref{ee2min2}) with fixed sets $\mathcal S_{ij}^{(v)}$, while ($ {\pmb q}^{(v-1)}, {\pmb t}^{(v-1)}$) is the initial feasible solution of problem (\ref{ee2min2tdma}) with fixed sets $\mathcal S_{ij}^{(v)}$. Furthermore, the total energy value (\ref{ee2min2}a) is always non-negative. Since the total energy value (\ref{ee2min2}a) is non-increasing in each iteration according to (\ref{convergenceProofIPCTA}) and the total energy value (\ref{ee2min2}a) is finitely lower-bounded, the IPCTA-NOMA algorithm must converge. \section{Proof of Theorem 6} \setcounter{equation}{0} \renewcommand{\theequation}{\thesection.\arabic{equation}} We first show that problem (\ref{ee2min1tdma}) can be equivalently transformed into problem (\ref{ee2min2tdma}). We introduce a set of new non-negative variables: \begin{equation}\label{Th5eq1} \hat p_{j} =t_{j} p_{j}, \hat q_i=t_{M+i} q_i, \quad \forall i \in \mathcal N, j \in \mathcal J_i. \end{equation} Substituting (\ref{Th5eq1}) into problem (\ref{ee2min1tdma}), we show that problem (\ref{ee2min1tdma}) is equivalent to problem (\ref{ee2min2tdma}). Then, we show that problem (\ref{ee2min2tdma}) is a convex problem. Since constraints (\ref{ee2min2tdma}e)-(\ref{ee2min2tdma}i) are all linear, we only need to check that objective function (\ref{ee2min2tdma}a) and constraints (\ref{ee2min2tdma}b)-(\ref{ee2min2tdma}d) are convex. Due to the fact that $-\bar u(|h_{nj}|^2 \hat q_n)$ is convex w.r.t. $\hat q_n$ from (\ref{appenDeq2_2}), $-t_{M+n} \bar u\left(\frac{|h_{nj}|^2 \hat q_n}{t_{M+n}}\right)$ is convex w.r.t. ($t_{M+n}, \hat q_n$) according to the property of perspective function \cite[Page~89]{boyd2004convex}. Thus, objective function (\ref{ee2min2tdma}a) is convex. Analogously, constraints (\ref{ee2min2tdma}b)-(\ref{ee2min2tdma}d) can also be shown convex. As a result, problem (\ref{ee2min2tdma}) is a convex problem. \bibliographystyle{IEEEtran}

\section{Introduction} \label{Sect_intro} Most of the currently known nearby ($d$$<$25\,pc) stars were first suspected as solar neighbours because of their high proper motions. The lower limit of high proper motion catalogues has decreased with time, from 500\,mas/yr in the Luyten Half Second \citepads[LHS;][]{1998yCat.1087....0L} and 180\,mas/yr in the New Luyten Two Tens \citepads[NLTT;][]{1995yCat.1098....0L} catalogues to 150\,mas/yr in the L{\'e}pine and Shara proper motion catalogue of Northern stars \citepads[LSPM-North;][]{2005AJ....129.1483L}, and now to 40\,mas/yr in the publicly not yet available SUPERBLINK catalogue \citepads{2017AAS...22915601L}. Only 15\% of about 3,300 stars within 25\,pc in the catalogue of nearby stars \citepads[CNS;][]{1995yCat.5070....0G} have proper motions smaller than 180\,mas/yr, and for only 2\% they lie below 40\,mas/yr. Meanwhile, the 25\,pc sample of the REsearch Consortium On Nearby Stars (RECONS\footnote{}) has been improved with respect to the accuracy of the measured trigonometric parallaxes and contains about 4,000 objects in 3,000 systems \citepads{2015IAUGA..2253773H}. While the 25\,pc census of AFGK-type stars had already been completed by the Hipparcos \citepads{1997ESASP1200.....E} mission that surveyed all these bright stars for their trigonometric parallaxes independent of their proper motions and colours, the fainter white dwarfs (WDs) and M-type dwarfs had to be preselected as targets for time-consuming ground-based parallax programmes by their colours and/or proper motions. According to the RECONS, there was an increase of 11\% for WDs and 25\% for M dwarfs if one considers the immediate solar neighbourhood (the 10\,pc sample) in 2012 compared to its status in 2000. A new unbiased survey that will help to finally complete the stellar (and partly also the substellar) census of the solar neighbourhood independent of proper motion limits is now being carried out by the Gaia mission \citepads{2016A&A...595A...1G}. A first relatively small, still incomplete and magnitude-limited ($V< 11.5$), subset of very accurate parallaxes was provided with the about two million stars in the Tycho-Gaia Astrometric Solution \citepads[TGAS;][]{2016A&A...595A...4L}. The TGAS catalogue is expected to include the majority of the GK-type stars within 25\,pc, whereas parallaxes of the brighter AF-type Solar neighbours and the fainter nearby WDs and the most frequent solar neighbours, the M-type stars, as well as most ML- and some of the T-type brown dwarfs \citepads[see][]{2017MNRAS.469..401S} will only be provided in later Gaia data releases. The largest intermediate parallax survey between Hipparcos and Gaia was provided by \citetads{2016AJ....151..160F}, who used data from the first U.S. Naval Observatory Robotic Astrometric Telescope Catalog \citepads[URAT1;][]{2015AJ....150..101Z}. The full URAT Parallax Catalog \citepads[UPC;][]{2016arXiv160406739F} contains more than 112,000 stars north of $\delta$$=$$-$13$^{\circ}$ available from \citetads{2016yCat.1333....0F} with significant parallax measurements, including 53,500 stars, for which no previous parallaxes were available. Note that the UPC is not biased towards high proper motion objects. The positions of all stars published in the first Gaia data release \citepads[Gaia DR1;][]{2016A&A...595A...2G,2016A&A...595A...4L} were used to improve the proper motion measurements of fainter stars in the 5th United States Naval Observatory CCD Astrograph Catalog\citepads[UCAC5;][]{2017AJ....153..166Z}. The UCAC5 proper motions are of similar accuracy as those of TGAS (1-2\,mas/yr) for stars with $R$ magnitudes between 11 and 15. Out of all 107 million UCAC5 stars, 25 million have proper motion errors smaller than 2\,mas/yr. Therefore, we can search among the typically fainter UCAC5 stars for objects that have a common proper motion (CPM) with a TGAS star. The CPM companions found in UCAC5 can then be assumed to lie at the same distances as their primary stars from TGAS, respectively. The CPM method has been usually applied to high proper motion stars, for which the errors were much smaller than the proper motion values \citepads[e.g.][]{1997yCat.1130....0L,2014AJ....148...60L,2011ASPC..448.1375L}. Very wide binaries with projected physical separations of $\gtrsim$1\,pc were already analysed in TGAS data alone by \citetads{2017AJ....153..257O} and \citetads{2017MNRAS.472..675A}. Interestingly, apparent members of open clusters, which lie at distances from the sun between 45\,pc and 450\,pc, were found at very large separations (up to 15\,pc) from their cluster centres \citepads{2017A&A...601A..19G}, not only by their CPM status but also according to their common parallaxes in TGAS. Within the traditional CNS horizon of 25\,pc from the sun, \citetads{2017MNRAS.465.2849T} have found and analysed only four WDs directly observed in TGAS. In addition they identified nine such nearby TGAS stars, which have wide WD companions that were too faint to be included in TGAS. Compared to the total number of more than 180 currently known WDs with accurate parallaxes within 25\,pc \citepads{2017AJ....154...32S}\footnote{According to more than 20\% of these nearby WDs are components of binary/multiple systems.}, the TGAS numbers of directly and indirectly observed WDs are very small. \citetads{2016MNRAS.462.2295H} list 232 WDs with trigonometric parallaxes or spectroscopic distance estimates in their 25\,pc sample. Despite of the obvious incompleteness of the TGAS 25\,pc sample with respect to low-mass stars and WDs, mostly fainter than the TGAS magnitude limit, and although the UCAC5 does also not go deep enough to detect the coolest WDs at about 25\,pc, we aimed at a CPM search for unknown nearby WD companions using the very accurate proper motions in both TGAS and UCAC5. In particular, we wanted to extend the CPM method to those of the known nearest stars that exhibit relatively small proper motions. \section{Common proper motion of nearby wide binaries} \label{Sect_cpm} \subsection{The TGAS 25\,pc sample} \label{Sect_tgas25} We have selected 973 stars from TGAS whose parallaxes are larger than 40\,mas and cross-identified this TGAS 25\,pc sample with the Two Micron All Sky Survey \citepads[2MASS;][]{2006AJ....131.1163S} after converting the TGAS coordinates with epoch 2015 to epoch 2000 taking into account the TGAS proper motions and using TopCat \citepads{2005ASPC..347...29T}. A search radius of 4\,arcsec was used, as the 2MASS coordinates are given for different epochs between 1997 and 2001. Applying the TGAS proper motions to the 2MASS positions, we then converted the latter also to the epoch 2000. The resulting angular distances between the epoch 2000 positions from both catalogues were small (median of 0.12\,arcsec, with only few values above 0.5\,arcsec and a maximum of 1.1\,arcsec). All 973 stars were matched and provided with 2MASS $JHK_s$ magnitudes (except for one star lacking the $K_s$ magnitude). However, only 871 out of the 973 stars have high-quality 2MASS photometry (quality flags ''AAA''). The number of TGAS stars within 25\,pc corresponds to only a quarter of the already mentioned RECONS 25\,pc sample. However, all TGAS stars within 25\,pc have trigonometric parallax errors of less than 1\,mas (for $\approx$75\% the errors are even $<$0.5\,mas), whereas the corresponding RECONS parallaxes were required to have errors of less than 10\,mas. Hence the TGAS 25\,pc sample is highly incomplete but very accurate with respect to its astrometric measurements. The proper motion errors for TGAS stars within 25\,pc ($\approx$88\% of them are Hipparcos stars) are also much smaller than for other TGAS stars (for $\approx$80\% the errors are $<$0.2\,mas/yr). \subsection{Effect of wide orbital motion on proper motions} \label{Sect_pmo} When we search for CPM objects to known nearby stars with accurately measured distances, we can estimate the expected influence of orbital motion on the proper motion difference between both components in a CPM pair. For simplicity we assumed \textbf{(1)} a circular orbit in the plane of the sky, \textbf{(2)} an orbital radius corresponding to the projected physical separation (in AU) that can be computed as the product of the angular separation between both components (in arcsec) and the assumed common distance (in pc), and \textbf{(3)} a total mass of the system of 1.5 solar masses. With these assumptions, and following Kepler's third law, we can compute the orbital period and velocity and transform the latter to a proper motion, again using the known distance. This proper motion effect due to orbital motion, $pmo$, is the expected maximum proper motion difference between the components of a nearby wide binary. This effect increases with smaller distances and separations. For a wide binary located at 25\,pc from the sun, separations of 3600\,arcsec, 60\,arcsec, and 6\,arcsec lead to $pmo$ of about 1\,mas/yr, 8\,mas/yr, and 25\,mas/yr, respectively. With equal separations, the corresponding $pmo$ are already about 4 times larger at 10\,pc, and about 11 times larger at 5\,pc, respectively. Before we used $pmo$ as a criterion for our CPM search in TGAS$+$UCAC5 data (see Sect.~\ref{Sect_tgasucac5wdcpm}), we investigated the known CPM pairs in the TGAS 25\,pc sample alone. Here we have the advantage that we can find physical pairs of stars from their common distance and small angular separation and then check the agreement in their proper motions. As the TGAS proper motion errors are typically very small ($\ll$2\,mas/yr) and can be neglected, we used in this case the total proper motion differences (Eq.~\ref{Eq_2}), leading also to a clearer presentation in Fig.~\ref{Fig_ratio_dpm_pmo}. However, in our TGAS$+$UCAC5 CPM search (Sect.~\ref{Sect_tgasucac5wdcpm}), we applied the $pmo$ criterion to the individual proper motion components (in RA and DE) taking into account their typically larger (and sometimes rather different) UCAC5 errors. \begin{figure} \centering \includegraphics[width=\hsize]{31753_f1.jpg} \caption{Wide binaries and multiple systems with common parallaxes and relatively small angular separations (see text) in the TGAS 25\,pc sample. Shown are their total proper motions \textbf{(top)} and angular separations \textbf{(bottom)} as a function of the ratio of the absolute value of their total proper motion difference $dpm$ (see Eq.~\ref{Eq_2}) to the estimated effect caused by orbital motion ($pmo$). The dotted line indicates the expected maximal ratio of 1. Systems discussed in the text are labeled in the top panel and marked by thick red and blue open symbols. For comparison we added the values for \object{Proxima} with respect to \object{alpha CenAB} (filled magenta symbols), which are not included in TGAS. The symbol sizes increase with parallax. } \label{Fig_ratio_dpm_pmo} \end{figure} We searched for wide binaries and multiple systems within the TGAS 25\,pc sample with angular separations of up to several degrees. First, we allowed only for parallaxes that agree within their TGAS errors and found 18 CPM pairs with total proper motion differences smaller than $pmo$. Their angular separations ranged between 4-120\,arcsec, their parallaxes between 45-114\,mas, and their total proper motions between 130-900\,mas/yr. Allowing for two times larger proper motion differences, we found three more CPM pairs falling in the same ranges of angular separations, parallaxes, and total proper motions, respectively. These three CPM pairs correspond to well-known systems of G-type stars (\object{GJ 765.1A}/\object{GJ 765.1B}), K dwarfs (\object{GJ649.1}(=\object{CCDM J16548+4722AB)}/\object{GJ649.1C}), and of early-M dwarfs (\object{GJ 319}(=\object{CCDM J08427+0933AB})/\object{GJ 319C}), respectively (thick red open symbols in the central parts of the two panels of Fig.~\ref{Fig_ratio_dpm_pmo}). Finally, as shown in Fig.~\ref{Fig_ratio_dpm_pmo}, we arrived at 24 CPM systems, when we conservatively allowed for 10\% differences in the TGAS parallaxes. Interestingly, all the latter three additional cases fall outside of at least one of the above given ranges and occupy edge regions in the two panels of Fig~\ref{Fig_ratio_dpm_pmo} (objects marked by thick blue open symbols). One of these pairs (\object{GJ 325} = \object{CCDM J08555+7048AB}) consists of two early-M dwarfs with a very small angular separation of 3\,arcsec and a large total proper motion of about 1400\,mas/yr. The second pair (\object{GJ 725A}/\object{GJ 725B}), of slightly cooler mid-M dwarfs, has also a relatively small angular separation of 11\,arcsec, a large parallax of $>$280\,mas, and very large total proper motion of about 2300\,mas/yr. These two pairs fall in the lower and upper left parts of the two panels in Fig.~\ref{Fig_ratio_dpm_pmo}, respectively. The third system, consisting of three TGAS sources, the wide binary \object{GJ 282A}/\object{GJ 282B} and its much wider companion \object{GJ 282C} (shown by one symbol at the right edge of the top and bottom panels of Fig.~ref~\ref{Fig_ratio_dpm_pmo}, respectively), does not stand out by its parallax or proper motion but has an extremely large angular separation of about 3900\,arcsec. This is much larger than the angular separations of all other wide binaries in the TGAS 25\,pc sample. At a distance of about 14\,pc it leads to a projected physical separation of about 56000\,AU, even larger than the about 10000\,AU between \object{alpha CenAB} and \object{Proxima}. The early-M dwarf \object{GJ 282C} was studied with respect to the K-type binary \object{GJ 282A}/\object{GJ 282B} and found to rank among the widest physical companions by \citetads{2009ApJ...706..343P}. As our nearest neighbours, \object{Proxima} and \object{alpha CenAB}, represent another prominent case of an extremely wide but bound system \citepads{2017A&A...598L...7K}, we added the corresponding values in Fig.~\ref{Fig_ratio_dpm_pmo}. These objects are not in TGAS, so we used their data as listed in \citetads{2017A&A...598L...7K}. For [\object{GJ 282A}/\object{GJ 282B}]/\object{GJ 282C} we find ratios between the total proper motion differences and our estimated proper motion effect $pmo$ of about 6. These ratios exceed even those of \object{alpha CenAB}/\object{Proxima}. Such very large ratios may indicate that our simple assumptions for computing $pmo$ are not correct and/or that these systems are in the process of dynamical disintegration \citepads{2009ApJ...706..343P}. \subsection{The role of chance alignments} \label{Sect_chance} The larger the separation between the components of a wide binary candidate and the smaller their assumed CPM, the more likely is a chance alignment of unrelated objects. The proper motion errors do also play an important role in the correct identification of CPM pairs, in particular if they are of the same order as the expected proper motion differences due to orbital motion (Sect.~\ref{Sect_pmo}) and/or if they are not much smaller than the proper motion values. \citetads{2007AJ....133..889L} investigated CPM companions of Hipparcos primaries in the LSPM-North catalogue and excluded chance alignments using an empirial relation between the separation ($sep$), total proper motion difference ($dpm$), and the total proper motion ($pm$): \begin{equation} \label{Eq_1} \frac{sep*dpm}{(pm/pm_{min})^{3.8}}<1, \end{equation} where $sep$ was given in arcsec, whereas $dpm$ and $pm$ were here in arcsec/yr (but are in mas/yr throughout this paper), and $pm_{min}$ was set to 0.15\,arcsec/yr, the lower proper motion limit of the LSPM-North catalogue. The total proper motion difference was computed as: \begin{equation} \label{Eq_2} dpm=\sqrt{(dpmRA^2+dpmDE^2)}. \end{equation} In our search for new WD companions (Sect.~\ref{Sect_tgasucac5wdcpm}), we applied several criteria, starting with a lower proper motion limit, selecting a maximum angular separation (similar to the limit used by \citetads{2007AJ....133..889L}), taking into account the proper motion errors, and the ratio of the proper motion differences to the $pmo$ of each CPM candidate. In our final selection we considered only ratios below 1.5 so that we may have excluded some possible extremely wide binaries, similar to the cases described in Sect.~\ref{Sect_pmo}. For an evaluation of our new WD CPM candidates and known WD CPM companions we checked, whether they fulfil the condition of Eq.~\ref{Eq_1}. \section{Nearby TGAS/UCAC5 WD companion candidates} \label{Sect_tgasucac5wdcpm} \subsection{CPM search criteria} \label{Sect_cpmcriteria} Our cross-matching of the TGAS 25\,pc sample with the UCAC5 with a large search radius of 3600\,arcsec yielded more than 7 million UCAC5 stars, on the average about 7200 field stars per TGAS star. For each field star, we determined the $pmo$ effect that we expected if it were a wide binary companion of the TGAS star (see Sect.~\ref{Sect_pmo}). When we consider the proper motion differences $dpm$ (TGAS-UCAC5) of potential CPM pairs, we must also take into account the proper motion errors $epm$ in the TGAS 25\,pc sample and in UCAC5. Even after excluding UCAC5 stars with large proper motion errors, the TGAS proper motions were still about ten times more accurate than those of the remaining UCAC5 stars. This is due to the fact that the TGAS 25\,pc sample consists mostly of Hipparcos stars, for which the TGAS was much more accurate than for Tycho stars. Only the 120 non-Hipparcos stars in the TGAS 25\,pc sample show a similar peak in the proper motion error distribution as in UCAC5. However, the tail of their error distribution contains only about 10 stars with errors $>$3\,mas/yr with a maximum at about 6\,mas/yr. Therefore, we did not exclude potential CPM primaries based on TGAS proper motion errors. We selected candidate TGAS/UCAC5 CPM pairs with the following criteria: \begin{itemize} \item TGAS and UCAC5 total proper motion $>$60\,mas/yr, \item TGAS-UCAC5 separation $<$1800\,arcsec, \item $epm_{UCAC5}<$6\,mas/yr (in RA and DE components), \item $|dpm|<1.5*(pmo+epm_{TGAS}+epm_{UCAC5})$ (in RA and DE). \end{itemize} When we tried to change the first three limits towards smaller total proper motions, larger separations, and larger UCAC5 proper motion errors, respectively, the number of CPM candidates increased. However, as their total proper motions were generally smaller, it happened that several CPM candidates were found for a given TGAS primary, in particular for very nearby TGAS stars. The additional candidates also showed unrealistic absolute magnitudes (assuming a common distance with the primary) with their given colours in near-infrared and optical to near-infrared colour-magnitude diagrams (CMDs). The factor of 1.5 in the fourth selection criterion was a compromise justified by the fact that several known wide binaries among the nearby CPM pairs in TGAS alone and in TGAS/UCAC5 would have been excluded with a factor of 1.0, whereas factors of 2.0-2.5 led again to more doubtful candidates. After applying the first selection criterion, the number of CPM candidates reduced to about 86,000 ($\approx$1\% of the original number of selected field stars). With the following two criteria, it further reduced to about 22,000 and 11,000, respectively. The fourth and decisive criterion led to only 72 ($\approx$0.001\%) CPM candidates, most of which turned out to be main sequence (MS) stars (see Figs.~\ref{Fig_nircmd} and \ref{Fig_optnircmd}). \subsection{WD companion candidates in CMDs} \label{Sect_wdcmd} After selecting among the 973 stars of the TGAS 25\,pc sample only those 871 with accurate 2MASS photometry, the near-infrared CMD in Fig.~\ref{Fig_nircmd} (black plus signs) shows the MS populated from (a few late-F and) early-G to early-M spectral types and a small WD sequence of four isolated WDs, \object{WD 1142-645} (DQ6.4), \object{WD 1647+591} (DAV4.1), \object{WD 1917-077} (DBQZ4.9), and \object{WD 2117+539} (DA3.6), as already mentioned by \citetads{2017MNRAS.465.2849T} \citepads[spectral types are taken from][]{2016MNRAS.462.2295H}. Figure~\ref{Fig_nircmd} also shows the CPM companions identified in our search (filled red lozenges), where we assume that the parallax of a CPM companion is the same as that of its TGAS primary. With our CPM companion search aiming at WD companions we did not expect to find late-M dwarfs (and brown dwarfs with late-M, L or T spectral types) with $M_J$$>$10, as even the nearest of these objects are too faint to be included in UCAC5. Consequently, the UCAC5 CPM companions only slightly extend the MS towards fainter magnitudes. For comparison we show in Fig.~\ref{Fig_nircmd} that late-M and early-L dwarfs occupy the lower right part of the near-infrared CMD as derived from accurate parallaxes by \citetads{2012ApJS..201...19D}. From nine known WD companions of TGAS stars within 25\,pc \citepads{2017MNRAS.465.2849T}, only three are included in UCAC5 (\object{WD 0433+270}, \object{WD 0751-252}, and \object{WD 2154-512}). All three were found by our CPM selection criteria. They extend the small WD sequence in Fig.~\ref{Fig_nircmd} (red filled lozenges overplotted by blue crosses) to the red. Their angular separations range between 28\,arcsec and 400\,arcsec. The other six WD companions that are not in UCAC5 have smaller angular separations (7-17\,arcsec) from their bright primaries, which may have prevented their detection on the UCAC images, and two of them may also have been too faint ($V$$>$16) to be observed within the UCAC survey. The data of known and candidate nearby ($d<25$\,pc) TGAS/UCAC5 CPM pairs including a WD companion, all found with the selection criteria described in Sect.~\ref{Sect_cpmcriteria}, are listed in the upper part of Table~\ref{Tab_cpm}. Two of the three new WD companion candidates have relatively small angular separations, whereas the third, \object{HD 35650 B}, looks with its large angular separation at approximately the same distance similar to the known very wide CPM companion \object{SCR J0753-2524}. Consequently, in these two pairs the $pmo$ effect is relatively small compared to the proper motion errors. \begin{figure}[t!] \centering \includegraphics[width=\hsize]{31753_f2.jpg} \caption{Near-infrared 2MASS CMD for stars in the TGAS 25\,pc sample with photometric quality 'AAA' in the 2MASS catalogue (black plus signs) and for selected CPM companions (see text) from UCAC5 (filled red lozenges). Their absolute magnitudes $M_J$ are based on TGAS parallaxes. Four known WDs in TGAS and three known WD companions in UCAC5 are marked by blue crosses. Our WD candidates (see Eqs.~\ref{Eq_3} and \ref{Eq_4}) fall below both dashed lines in this figure and in Fig.~\ref{Fig_optnircmd}, respectively. The dotted line shows the dividing line at $J-K_s$$=$$+$0.5 between WDs and cool subdwarfs (sd) found by \citetads{2017AJ....154...32S}. Also shown are ''normal'' (not flagged as atypical, and not known as close binaries) late-M (green open triangles) and early-L dwarfs (magenta open squares) as well as late-M and L subdwarfs (cyan open lozenges; marked by their spectral types) from \citetads{2012ApJS..201...19D} based on their parallaxes with relative errors less than 10\%. Three new WD companion candidates are numbered. All CPM candidates with angular separations $>$900\,arcsec are labelled with question marks. } \label{Fig_nircmd} \end{figure} \begin{figure}[t!] \centering \includegraphics[width=\hsize]{31753_f3.jpg} \caption{Optical (Gaia) to near-infrared (2MASS) CMD with the same stars (and symbols) as in Fig.~\ref{Fig_nircmd}. Only part of the late-M and early-L dwarfs, and of the M/L subdwarfs from \citetads{2012ApJS..201...19D} could be identified in Gaia (see text). As in Fig.~\ref{Fig_nircmd}, the absolute Gaia magnitudes $M_G$ of TGAS stars and UCAC5 CPM candidates are based on TGAS parallaxes, whereas those of the comparison objects are based on the parallaxes from \citetads{2012ApJS..201...19D}. Objects located below the dashed lines in this figure and in Fig.~\ref{Fig_nircmd} were considered as WDs. } \label{Fig_optnircmd} \end{figure} From Fig.~\ref{Fig_nircmd} it is obvious that the WD companion candidates '1' and '2' fit well into the sequence of known WDs, whereas candidate '3' lies a bit off and just right of the WD/subdwarf dividing line described by \citetads{2017AJ....154...32S}. Doubtful candidates marked with '?', which had the largest angular separations ($\approx$1000-1750\,arcsec) but relatively small (60-90\,mas/yr) total proper motions, are in this near-infrared CMD either much redder or brighter than the WD sequence. However, because of their small proper motions and the spectral classification (in SIMBAD) of their apparent primaries as normal MS stars, we can not consider these doubtful CPM companions as nearby subdwarf candidates. With their large separations, it is much more likely that these are chance alignments of distant MS stars (with distances in the range of $\approx$50-500\,pc according to their $G-J$ colour and apparent $G$ magnitudes) with the nearby TGAS stars. In fact, if we compute the ratio according to Eq.~\ref{Eq_1}, replacing the $pm_{min}$ used by \citetads{2007AJ....133..889L} with our lower proper motion limit $pm_{min}$=0.06\,arcsec/yr, we get for these six doubtful CPM pairs values between 1.4 and 5.3, indicating chance alignments. The tangential velocities derived from the proper motions and the estimated larger photometric distances of these rejected CPM companions are mostly in the range of 20-80\,km/s, typical of the Galactic thin disk population. Only for one object, which can be identified as \object{2MASS 06233146-5952448}, located just above the dashed line in Fig.~\ref{Fig_optnircmd}, the resulting tangential velocity was larger ($\approx$150\,km/s) so that this might be a distant (K-type) subdwarf (thick disk object), unrelated to our Solar neighbourhood TGAS sample. For candidate '3', this ratio is also large (2.1) indicating a probable chance alignment. However, its separation is only moderately large (about 465\,arcsec) and it lies much closer to the WD sequence. Therefore, we consider this object as a weak WD companion candidate. Interestingly, its relatively blue colour ($G-J\approx+1.1$) and faint magnitude ($G\approx$15) lead to an alternative photometric distance of $\approx$800\,pc and tangential velocity of $\approx$260\,km/s, if it is unrelated. In this case, this would be a distant G-type subdwarf candidate, possibly belonging to the Galactic halo population. For all other CPM candidates, including known WD companions and the WD companion candidates '1' and '2', the ratio from Eq.~\ref{Eq_1} was always smaller than 0.4, with a mean value of 0.02, confirming their real CPM status. Concerning the location of candidate '3' close to the WD/subdwarf dividing line (dotted line in Fig.~\ref{Fig_nircmd}) we note that the $J-K_s$ colours used in \citetads{2017AJ....154...32S} were not directly measured in but transformed to the 2MASS system. The position of this dividing line is questionable in view of the wide-spread $J-K_s$ colour distribution of the shown M/L subdwarfs and obviously not valid for the latest-type (coolest) known subdwarfs, \object{SSSPM J1013-1356} (sdM9.5), \object{2MASS J16262034+3925190} (sdL4), and \object{2MASS J05325346+8246465} (sdL7). The remaining seven M/L subdwarfs seem to represent better the 60\,pc cool subdwarf sample (presumably dominated by sdK and early-sdM subdwarfs) that was used by \citetads{2017AJ....154...32S} in finding the $J-K_s=+0.5$ dividing line. We considered objects as WD candidates if they were falling below two colour-magnitude limits (shown by the dashed lines in Figs.~\ref{Fig_nircmd} and \ref{Fig_optnircmd}): \begin{equation} \label{Eq_3} M_G>10.0+(2.5*(G-J)), \end{equation} \begin{equation} \label{Eq_4} M_J>11.0+(3.0*(J-K_s)). \end{equation} \begin{table*} \caption{Astrometry and photometry of known and candidate ('1-3') WD companions and their primaries in nearby TGAS/UCAC5 CPM pairs and of additional known WDs with no previous trigonometric parallaxes (in SIMBAD) and one WD candidate ('4') found in the UPC}. \label{Tab_cpm} \centering \fontsize{7pt}{0.90\baselineskip}\selectfont \begin{tabular}{@{}ll@{}rclrc@{}crcl@{}c@{}r@{}} \hline\hline Name & RA\tablefootmark{a} & DE\tablefootmark{a} & $plx_{TGAS}$ & $plx_{UPC}$ & $pm$RA & $pm$DE & $pmo$ & $G$\tablefootmark{a} & $J$\tablefootmark{b} & $J-K_s$\tablefootmark{b} &SpT & Ref \\ (WD name/cand) & & & (Sep, PA)\tablefootmark{a} & & & & ($|dpm|$)\tablefootmark{f} \\ & [degrees] & [degrees] & [mas] & [mas] & [mas/yr] & [mas/yr] & [mas/yr] & [mag] & [mag] & [mag] & \\ & & & ([arcsec,$^{\circ}$]) & & & & ([mas/yr]) \\ \hline \object{GJ 171.2 A} & 069.202096 & $+$27.131577 & 57.2$\pm$0.3 & 65.1$\pm$9.0 & $+$232.1$\pm$0.1\tablefootmark{c} & $-$147.7$\pm$0.1\tablefootmark{c} & 9 & 7.64 & 5.95 & $+$0.71 & K3IVke & 3 \\ \object{GJ 171.2 B} & 069.188131 & $+$27.163694 & (124.0,339) & 80.1$\pm$11.0 & $+$230.8$\pm$3.0\tablefootmark{d} & $-$146.6$\pm$2.9\tablefootmark{d} & (1,1) & 15.53 & 14.60 & $+$0.46 & DA9.0 & 1 \\ (= \object{WD 0433+270}) & & & & & $+$229.7$\pm$6.1\tablefootmark{e} & $-$142.9$\pm$6.0\tablefootmark{e} \\ \object{Ross 429} & 118.543958 & $-$25.302335 & 56.2$\pm$0.3 & - & $-$300.8$\pm$0.1\tablefootmark{c} & $+$201.0$\pm$0.1\tablefootmark{c} & 5 & 8.97 & 7.02 & $+$0.85 & M0 & 4 \\ \object{SCR J0753-2524}\tablefootmark{i} & 118.484536 & $-$25.399551 & (399.8,209) & - & $-$286.4$\pm$5.2\tablefootmark{d} & $+$205.5$\pm$5.3\tablefootmark{d} & (14,4) & 15.93 & 14.75 & $+$0.44 & DA9.8 & 1 \\ (= \object{WD 0751-252}) \\ \object{GJ 841 A} & 329.421448 & $-$51.007760 & 66.1$\pm$0.5 & - & $-$35.3$\pm$0.2\tablefootmark{c} & $-$379.8$\pm$0.2\tablefootmark{c} & 25 & 9.24 & 6.75 & $+$0.87 & M2Ve & 5 \\ \object{GJ 841 B} & 329.409514 & $-$51.010412 & (28.7, 251) & - & $-$46.5$\pm$1.2\tablefootmark{d} & $-$395.5$\pm$1.1\tablefootmark{d} & (11,16) & 14.48 & 14.01 & $+$0.20 & DQ8.3 & 1 \\ (= \object{WD 2154-512}) \\ \object{HD 166435} & 272.339436 & $+$29.951968 & 40.7$\pm$0.2 & 46.1$\pm$8.8 & $+$71.9$\pm$0.1\tablefootmark{c} & $+$61.1$\pm$0.1\tablefootmark{c} & 12 & 6.62 & 5.69 & $+$0.37 & G1IV & 3 \\ \object{HD 166435 B}& 272.331393 & $+$29.956100 & (29.2,301) & 50.4$\pm$3.3 & $+$64.8$\pm$1.2\tablefootmark{d} & $+$72.4$\pm$1.2\tablefootmark{d} & (7,11) & 13.16 & 13.70 & $-$0.16 & DA2.2$\pm$0.2 & 2 \\ (= '1') & & & & & $+$67.4$\pm$2.4\tablefootmark{e} & $+$71.5$\pm$2.4\tablefootmark{e} \\ \object{TYC 3980-1081-1} & 327.909017 & $+$59.294450 & 120.6$\pm$1.0 & 154.7$\pm$12.1 & $-$70.8$\pm$2.5\tablefootmark{c} & $+$76.1$\pm$2.4\tablefootmark{c} & 84 & 9.20 & 6.53 & $+$0.88 & $\approx$M2 & 2 \\ \object{TYC 3980-1081-1 B} & 327.916408 & $+$59.292935 & (14.6,112) & 131.0$\pm$4.4 & $-$87.2$\pm$2.2\tablefootmark{d} & $-$31.2$\pm$2.0\tablefootmark{d} & (16,107) & 14.28 & 12.97\tablefootmark{g} & $+$0.32\tablefootmark{g} & $\approx$DA10? & 2 \\ (= '2') & & & & & $-$79.5$\pm$3.3\tablefootmark{e} & $-$17.3$\pm$3.2\tablefootmark{e} \\ \object{HD 35650} & 081.125938 & $-$38.969890 & 57.4$\pm$0.3 & - & $+$43.1$\pm$0.1\tablefootmark{c} & $-$57.3$\pm$0.1\tablefootmark{c} & 5 & 8.43 & 6.70 & $+$0.78 & K6Vke & 6 \\ \object{HD 35650 B} & 081.252385 & $-$39.053688 & (464.9,131) & - & $+$46.3$\pm$1.9\tablefootmark{d} & $-$48.8$\pm$2.0\tablefootmark{d} & (3,9) & 14.99 & 13.87 & $+$0.53 & $\approx$DA9? & 2 \\ (= '3') \\ \hline \object{GJ 3285} & 066.473476 & $+$12.195784 & - & 58.0$\pm$8.0 & $-$94.6$\pm$6.4\tablefootmark{e} & $-$227.9$\pm$6.4\tablefootmark{e} & - & 15.21 & 14.49 & $+$0.24 & DC8.2 & 1 \\ (= \object{WD 0423+120} & \\ \object{GJ 275.2 B}\tablefootmark{h} & 112.695736 & $+$48.168756 & - & 89.6$\pm$4.2 & $-$216.6$\pm$4.0\tablefootmark{e} & $-$1274.9$\pm$3.8\tablefootmark{e} & - & 14.84 & 13.08 & $+$0.33 & DA10.0+DA10.1 & 1 \\ (= \object{WD 0727+482} & \\ \object{GJ 1098} & 113.379072 & $+$64.156548 & - & 54.3$\pm$8.2 & $+$50.9$\pm$6.8\tablefootmark{e} & $-$260.5$\pm$5.9\tablefootmark{e} & - & 15.91 & 14.81 & $+$0.43 & DA11.1 & 1 \\ (= \object{WD 0728+642} & \\ \object{GJ 4165}\tablefootmark{j} & 312.278706\tablefootmark{e} & $+$37.471154\tablefootmark{e} & - & 63.1$\pm$2.9 & $+$167.1$\pm$2.9\tablefootmark{e} & $+$154.3$\pm$2.9\tablefootmark{e} & - & - & 13.30 & $-$0.13 & DA3.6 & 1 \\ (= \object{WD 2047+372} \\ \object{UPC 72924} & 277.995093 & $+$46.974870 & - & 40.2$\pm$6.6 & $+$61.5$\pm$4.6\tablefootmark{e} & $+$38.4$\pm$4.4\tablefootmark{e} & - & 15.05 & 14.52 & $+$0.14 & $\approx$DA7? & 2 \\ (= '4') \\ \hline \end{tabular} \tablefoot{\fontsize{7pt}{0.90\baselineskip}\selectfont Gaia coordinates are for (J2000, epoch 2015.0) and were rounded to 0.000001 degrees. Although the TGAS data may be more accurate, all parallaxes and their errors were rounded to 0.1\,mas, proper motions and their errors to 0.1\,mas/yr. The absolute values of measured proper motion differences $dpm$ and the estimated effect because of orbital motion $pmo$ were rounded to 1\,mas/yr. Magnitudes and colours were rounded to 0.01\,mag. Spectral types with question marks are only guesses based on $G-J$ and $M_G$ (Figs.~\ref{Fig_optnircmd} and \ref{Fig_optnirupc}). Further notes on the data: \tablefoottext{a}{Gaia DR1,} \tablefoottext{b}{2MASS,} \tablefoottext{c}{TGAS,} \tablefoottext{d}{UCAC5,} \tablefoottext{e}{UPC,} \tablefoottext{f}{two values are given for RA and DE, respectively,} \tablefoottext{g}{$J$ magnitude with very poor goodness-of-fit quality of the profile-fit photometry,} \tablefoottext{h}{An accurate parallax of 87.4$\pm$0.5\,mas measured for this close binary by \citetads{2015ASPC..493..501N},} \tablefoottext{i}{\citetads{2009AJ....137.4547S} measured a parallax of 56.54$\pm$0.95\,mas,} \tablefoottext{j}{This object is not plotted in Figs.~\ref{Fig_nirupc} and \ref{Fig_optnirupc}, since it was not identified in Gaia DR1.} } \tablebib{\fontsize{7pt}{0.90\baselineskip}\selectfont (1) \citetads{2016MNRAS.462.2295H}; (2) this paper; (3) \citetads{2003AJ....126.2048G}; (4) \citetads{2014MNRAS.443.2561G}; (5) \citetads{2006A&A...460..695T}; (6) \citetads{2006AJ....132..161G}. } \end{table*} Contrary to the WD/subdwarf confusion in a near-infrared $M_J/J-K_s$ diagram (Fig.~\ref{Fig_nircmd}) we expect a clearer separation of the MS, subdwarf and WD sequences in an optical to near-infrared CMD. As Gaia provides accurate magnitudes, albeit in a very broad optical band, we studied the TGAS 25\,pc sample and our UCAC5 CPM candidates together with the comparison objects from \citetads{2012ApJS..201...19D} in an $M_G/G-J$ diagram (Fig.~\ref{Fig_optnircmd}). Unfortunately, only three out of ten M/L subdwarfs could be identified in Gaia DR1 (with Gaia magnitudes $16.3<G<17.7$). This low identification rate may be due to the very large proper motions, probably problematic for Gaia DR1, of these ultracool subdwarfs. Among the comparison late-M and early-L dwarfs, we identified 6 out of 58 (only 10\%; $15.8<G<17.7$) and 7 out of 12 (58\%; $18.5<G<20.6$) objects in Gaia DR1, respectively. The very low identification rate of the relatively bright and nearby late-M dwarfs may be caused by their large parallaxes and proper motions and the selection criteria for problematic sources in Gaia DR1. The much fainter early-L dwarfs have typically smaller parallaxes and proper motions so that their identification rate is similar to that found for all L and T dwarfs in Gaia DR1 \citepads[45\%;][]{2017MNRAS.469..401S}. The MS in Fig.~\ref{Fig_optnircmd} shows a large gap at the expected position of early- to mid-M dwarfs. Three labelled late-M subdwarfs form a parallel sequence shifted from the MS by $\approx$0.5\,mag to the blue. Some of the doubtful CPM candidates marked with '?' seem to fall in the same region, but we consider them as chance alignments rather than nearby subdwarf candidates (see above). The WD sequence is also almost parallel with respect to the MS, but shows a large blue-shift of $\approx$3.0\,mag. The new WD candidates '1' and '2' are now located at the blue and red end of the WD sequence, suggesting a hot and a cool WD, respectively. The two L-type subdwarfs, which were falling right in the WD sequence in the near-infrared CMD (Fig.~\ref{Fig_nircmd}) have no Gaia magnitudes yet, but we expect them to fall in the lower right corner of Fig.~\ref{Fig_optnircmd}, well separated from the WDs in $G-J$ colour. Note that all the M/L subdwarfs from \citetads{2012ApJS..201...19D} are fainter than the UCAC5 magnitude limit of $R\approx16$\,mag. The same is true for the MS late-M and early-L dwarfs. As already discussed above, the Gaia magnitude intervals for the M/L subdwarfs and the MS late-M/early-L dwarfs are almost identical, and their UCAC5 magnitudes fall in the same range. With our quality cut for UCAC5 proper motion errors, we effectively reduced the limiting UCAC5 magnitude of the investigated CPM candidates further. The M/L subdwarfs have in addition very large proper motions (between $\approx$600\,mas/yr and $\approx$3500\,mas/yr) so that we were not expecting to find such objects in our CPM search aiming at relatively small proper motions. The region between WDs and cool subdwarfs in Fig.~\ref{Fig_optnircmd} should be empty. Therefore, all CPM candidates falling in this CMD close to the arbitrary drawn dividing line between WDs and subdwarfs have to be taken with caution independent of their possible chance alignment that we evaluated using Eq.~\ref{Eq_1}. \section{UPC parallaxes confirming two WD companions} \label{Sect_upc} The primaries of our WD companion candidates '1' and '2', \object{HD 166435} and \object{TYC 3980-1081-1}, were only recently added to the 25\,pc sample. The original Hipparcos \citepads{1997ESASP1200.....E} parallax of the well-known early-G star \object{HD 166435} placed it at about 25.2\,pc, whereas the revised Hipparcos \citepads{2007A&A...474..653V} results led to a distance of 24.8\,pc, and the TGAS parallax finally fixed it at 24.58\,pc. The Tycho star \object{TYC 3980-1081-1}, was discovered as a close neighbour of the sun by \citetads{2014AJ....148..119F}, who estimated a photometric distance of 5.9\,pc, before a first trigonometric parallax from URAT data was reported by \citetads{2016AJ....151..160F} giving a distance of about 6.5\,pc. The full UPC catalogue, including all significant parallaxes determined from the complete northern hemisphere URAT data, was only recently made available \citepads{2016arXiv160406739F,2016yCat.1333....0F}. The TGAS parallax of \object{TYC 3980-1081-1} corresponds to a distance of 8.29\,pc, still well within the 10\,pc limit. From its absolute magnitude of $M_J=6.94$ we estimate a spectral type of $\approx$M2 for \object{TYC 3980-1081-1} using the relationship between absolute magnitudes and spectral types from \citetads{2005A&A...442..211S}. Interestingly, not only \object{HD 166435} and \object{TYC 3980-1081-1} can be found in the UPC, but also their WD companion candidates '1' and '2', which we selected based on their CPM in TGAS and UCAC5. We call them \object{HD 166435 B} and \object{TYC 3980-1081-1 B}, respectively. Their UPC parallaxes, albeit much less accurate than the TGAS parallaxes, are similarly large (see upper part of Table~\ref{Tab_cpm}) as for the primaries and confirm a common distance, respectively. This is important in particular for our second candidate, \object{TYC 3980-1081-1 B}, which shows a large difference in the DE proper motion component compared to that of \object{TYC 3980-1081-1}. However, the close distance and small separation of this CPM pair lead to a very large $pmo$ value, and the proper motion difference is not much larger than this expected effect from orbital motion. In case of the known wide binary \object{Ross 429}/\object{SCR J0753-2524}, discovered by \citetads{2005AJ....130.1658S}, the proper motion differences are also relatively large with respect to the corresponding $pmo$ value (Table~\ref{Tab_cpm}). Note that with our fourth selection criterion described in Sect.~\ref{Sect_cpmcriteria} we took also into account the proper motion errors, which are relatively large for the UCAC5 proper motion of \object{SCR J0753-2524}. The UPC proper motions of three CPM companions, also given in the upper part of Table~\ref{Tab_cpm}, are less accurate than their UCAC5 proper motions. However, for the two relatively bright WD companions \object{HD 166435 B} and \object{TYC 3980-1081-1 B} the combined parallax and proper motion solution in the UPC led to more similar proper motions compared to their primaries, respectively. For the fainter known companion \object{WD 0433+270} the UPC errors of its parallax and proper motion are much larger, and the UCAC5 proper motion is in better agreement with the TGAS proper motion of the primary than the UPC proper motion. \begin{figure} \centering \includegraphics[width=\hsize]{31753_f4.jpg} \caption{Near-infrared (2MASS) CMD for our selected UPC 25\,pc sample. Absolute magnitudes are based on UPC parallaxes. The start and end values of both axes are the same as in Fig.~\ref{Fig_nircmd}, the dashed line shows our previously chosen limit for the WD region in this CMD (Eq.~\ref{Eq_3}), the dotted line marks again the WD/subdwarf dividing line of \citetads{2017AJ....154...32S}. Stars with photometric quality 'AAA' in 2MASS are plotted as plus signs. Overplotted as red filled squares, blue open squares and black dots are stars without previous parallax measurements (UPC $srcflg=0$), with already published data in \citetads{2016AJ....151..160F} ($srcflg=3$), and in the literature ($srcflg=1,4,5$), respectively. All WD candidates, defined by us as falling below the dashed lines in this figure and in Fig.~\ref{Fig_optnirupc}, are shown as open lozenges. The new WD candidates '1' and '2' (already found in our TGAS CPM search in UCAC5), and '4' are marked. A close binary cool WD is also labelled.} \label{Fig_nirupc} \end{figure} \begin{figure} \centering \includegraphics[width=\hsize]{31753_f5.jpg} \caption{Optical (Gaia) to near-infrared (2MASS) CMD for our selected UPC 25\,pc sample. As in Fig.~\ref{Fig_nirupc}, absolute magnitudes are based on UPC parallaxes. The start and end values of both axes are the same as in Fig.~\ref{Fig_optnircmd}, the dashed line shows again our limit for the WD region in this CMD (Eq.~\ref{Eq_4}). Symbols are the same as in Fig.~\ref{Fig_nirupc}.} \label{Fig_optnirupc} \end{figure} \section{Additional WD candidates from the UPC} \label{Sect_addupc} \smallskip \citetads{2016AJ....151..160F} have already studied stars falling in a 25\,pc sample according to significant parallax measurements in the UPC. They mentioned many new nearby stars with small proper motions ($<$200\,mas/yr) but did not pay special attention to potential WDs among their new neighbour candidates. In their notes on individual systems they discussed the nearest new discoveries and found many of them suspicious with respect to blended or elongated images and/or large scatter in the combined fit for parallax and proper motion. These suspicious nearest candidates have parallax and proper motion errors of typically 10-25\,mas and 10-20\,mas/yr, respectively. From $\approx$112,000 stars in the full UPC available at the CDS \citepads{2016yCat.1333....0F}, $\approx$5,200 have parallaxes $>$40\,mas, and for $\approx$3,700 of them the parallax errors are smaller than 10\,mas. As the UPC covers only about half of the sky, this number appears very high in comparison with the already mentioned RECONS all-sky 25\,pc census, and the new UPC neighbours need to be taken with caution. To evaluate potential WDs in near-infrared and optical to near-inrared CMDs, similar to those shown in Figs~\ref{Fig_nircmd} and \ref{Fig_optnircmd}, we cross-matched the UPC with 2MASS and Gaia DR1. We found $\approx$3,700 UPC stars within 25\,pc that have both 2MASS and Gaia counterparts. As expected from the two different parts of the UPC \citepads[stars with or without previously known parallaxes; see][]{2016arXiv160406739F,2016yCat.1333....0F} the error distribution for the UPC parallaxes rises to a peak at 10\,mas before it abruptly turns down to a long tail reaching to about 60\,mas. The proper motion errors show a more symmetric distribution with a maximum at about 4\,mas/yr, a smooth decrease to 10\,mas/yr, and also a long tail continuing to about 40\,mas/yr. For our final UPC 25\,pc sample, consisting of $\approx$1,600 stars, we applied the following quality criteria concerning the parallax and proper motion errors, $eplx$ and $epm$, respectively, and also used a lower limit for the total proper motion: \begin{itemize} \item $eplx_{UPC}<$10\,mas, \item $epm_{UPC}<$10\,mas/yr (in RA and DE components), \item UPC total proper motion $>$60\,mas/yr. \end{itemize} The resulting CMDs are shown in Figs.~\ref{Fig_nirupc} and \ref{Fig_optnirupc}. As in Figs.~\ref{Fig_nircmd} and \ref{Fig_optnircmd}, we plot stars with reliable 2MASS photometry as plus signs, but overplot all stars independent of their 2MASS quality with different symbols corresponding to their source flags ($srcflg$) given in the UPC. Stars with $srcflg=0$, without previously published parallaxes in SIMBAD or other catalogues, are plotted as filled red squares (except for \object{GJ 4165} lacking a $G$ magnitude) and are listed in the lower part of Table~\ref{Tab_cpm} if not already listed in the upper part. Because of the deeper limiting magnitude of URAT compared to UCAC5, the MS is now mainly populated with M dwarfs that were lacking in TGAS and TGAS/UCAC5 CPM data. With $M_J$=6-11, these are M0-M7 dwarfs according to \citetads{2005A&A...442..211S}. As expected from the lower accuracy of the UPC parallaxes in comparison to TGAS, the MS and the WD sequence in Figs.~\ref{Fig_nirupc} and \ref{Fig_optnirupc} do not look as narrow as in Figs.~\ref{Fig_nircmd} and \ref{Fig_optnircmd}. In particular, many of the new UPC neighbours ($srcflg=0$ and $srcflg=3$), are located left of the MS, where only few subdwarfs are expected. However, all our WD candidates defined by Eqs.~\ref{Eq_3} and \ref{Eq_4} are well separated from the cloud of suspicious subdwarfs left of the MS. The only WD located close to the corresponding dashed line in Fig.~\ref{Fig_optnircmd}, \object{WD 0727+482}, is a close binary, expected to be brighter. We also note that none of the 34 known WDs and 3 WD candidates ('1', '2', and '4') in Fig.~\ref{Fig_nirupc} falls right of the WD/subdwarf dividing line of \citetads{2017AJ....154...32S}. The new candidate '4' is a previously anonymous field star, \object{UPC 72924}. When we tried to allow for larger parallax and proper motion errors in the UPC 25\,pc sample, e.g. 12\,mas and 12\,mas/yr, or 15\,mas and 15\,mas/yr, respectively, a few additional known WDs were recovered, They were all falling in the WD regions below the dashed lines in both CMDs, whereas new WD candidates were not found. On the other hand, a lower proper motion limit of 40\,mas/yr led to many more suspicious objects falling closer to the dashed lines in both CMDs but only one more WD candidate that we nevertheless considered as unreliable because of its relatively large image elongation and large scatter of the post-fit residuals given in the UPC. Note that, unless we did not search for CPM candidates in the UPC alone or with respect to TGAS, one of the known WDs with a UPC parallax in Table~\ref{Tab_cpm}, \object{WD 0727+482}, being itself a close binary, is a member of a wide multiple system. It includes \object{GJ 275.2 A} that has a similarly large UPC parallax and proper motion of 101.2$\pm$4.8\,mas and $(-192.5\pm4.4, -1268.4\pm4.1)$\,mas/yr, respectively. \section{One spectroscopically confirmed WD companion} \label{Sect_specwd} Our WD candidate \object{HD 166435 B} was already included in our earlier large programme concentrating on the spectroscopic classification of nearby cool WD and subdwarf candidates. Within this programme it was one of the targets with the smallest proper motion, at that time selected from an earlier UCAC version. Large numbers of comparison objects with known spectral types were also observed. The low-resolution spectroscopic observations were carried out in 2008 (August 8/9) using the focal reducer and faint object spectrograph CAFOS mounted at the 2.2m telescope at Calar Alto, Spain. We used the grism B~200 giving a wavelength coverage from about 3500\,{\AA} to 7000\,{\AA} and a dispersion of 4.7\,{\AA} per pixel. All spectra were reduced with standard routines from the ESO MIDAS data reduction package. Fig.~\ref{Fig_wdspec} shows the normalised spectrum of \object{HD 166435 B} together with those of two comparison objects observed during the same observing run. In this comparison \object{HD 166435 B} appears very similar to the DA2.4 and DA2.8 white dwarfs and could be classified as DA2.0 from its bluer continuum. The very blue continuum of \object{HD 166435 B} is consistent with its magnitudes and the resulting negative $FUV-NUV$ colour index measured by the Galaxy Evolution Explorer \citepads[GALEX;][]{2011Ap&SS.335..161B}, $FUV=12.068\pm0.003$ and $NUV=12.735\pm0.004$, compared to $FUV=12.604\pm0.007$ and $NUV=12.688\pm0.005$ and a zero color index for \object{WD 2149+021} (\object{WD 2032+248} was not measured by GALEX). The distances of \object{HD 166435 B} and \object{WD 2149+021} are very similar with $\approx$24.6\,pc (TGAS) and $\approx$24.5\,pc \citepads{2015ApJS..219...19L}, respectively. We have also measured the equivalent widths of the H$\beta$ and H$\alpha$ lines, relative to their well-defined continuum in the three spectra of Fig.~\ref{Fig_wdspec}. Whereas our results for H$\alpha$ support a slightly earlier classification of \object{HD 166435 B} (DA1.9) compared to \object{WD 2032+248} (DA2.1) and \object{WD 2149+021} (DA2.5), the equivalent widths of H$\beta$, measured with higher signal-to-noise, hint at DA2.4, DA2.2, and DA2.7, respectively. We assign a preliminary spectral type of DA2.2$\pm$0.2 to \object{HD 166435 B}. \begin{figure} \centering \includegraphics[width=\hsize]{31753_f6.jpg} \caption{Low resolution Calar Alto 2.2m/CAFOS spectrum of candidate '1' (= \object{HD 166435 B}), marked on Fig.~\ref{Fig_nircmd} and Fig.~\ref{Fig_optnircmd}, overplotted by the spectra of two known WDs, \object{WD 2032+248} (DA2.4; red line) and \object{WD 2149+021} (DA2.8; green line) \citepads{2011ApJ...743..138G}, observed with the same instrument setup. All spectra are normalised at 4600\,{\AA}. } \label{Fig_wdspec} \end{figure} For the other two WD companion candidates, \object{TYC 3980-1081-1 B} and \object{HD 35650 B}, we only assume cool DA types from their location in the optical to near-infrared CMD (Fig.~\ref{Fig_optnircmd}) in comparison to the known WDs \object{GJ 171.2 B} and \object{SCR J0753-2524} (Table~\ref{Tab_cpm}). For the WD candidate \object{UPC 72924} we assume a spectral type of $\approx$DA7 from comparing its $G-J$ colour with those of known WDs found in our study of the UPC 25\,pc sample. \section{Conclusions} \label{Sect_concl} \begin{enumerate} \item Compared to only four known WDs in TGAS and three known WD companions in UCAC5, our three new WD companion candidates represent an increase of $\approx$43\%. \item Two of our new WD companion candidates are confirmed by parallax measurements in the UPC. One, \object{HD 166435 B}, orbits a G-type star at a distance of $\approx$24.6\,pc with a projected physical separation of $\approx$700\,AU, the other, \object{TYC 3980-1081-1 B}, an early-M dwarf at only $\approx$8.3\,pc distance with a projected physical separation of $\approx$120\,AU, respectively. The third WD companion candidate, \object{HD 35650 B}, has a very large projected physical separation from its primary of $\approx$8100\,AU, similar to that of the known wide companion \object{SCR J0753-2524}. However, we can not exclude the possibility that \object{HD 35650 B} is only a chance alignment and represents in fact a distant halo star. \item Our additional search for nearby WDs using parallaxes from the UPC confirmed our WD selection criteria in two CMDs already applied for our CPM candidates and led to one more WD candidate, \object{UPC 72924}, at $\approx$25\,pc. \item Considering the estimated effect $pmo$ of orbital motion on the proper motion difference in CPM pairs, we were able to recover the faint components of known very wide CPM systems (\object{GJ 282AB}/\object{GJ 282C} and \object{Ross 429}/\object{SCR J0753-2524}). On the other hand, this allowed us to discover \object{TYC 3980-1081-1 B} as a cool WD companion candidate with relatively small angular separation from a very nearby star, where the $pmo$ effect is very large. \item \object{HD 166435 B} was spectroscopically confirmed as a relatively hot WD, as already suspected from its $G-J=-0.54$ and large negative GALEX colour index $FUV-NUV\approx-0.67$, whereas \object{TYC 3980-1081-1 B} is with $G-J=+1.31$ a cool WD candidate (possibly $\approx$DA10). Compared to the 25\,pc WD sample of \citetads{2016MNRAS.462.2295H}, \object{HD 166435 B} ranges as $\approx$DA2.2 among the hottest nearby DA WDs and is an important addition of a 'Sirius-like' system (a WD with a luminous primary of spectral type K or earlier) in the volume between 20\,pc and 25\,pc. \item The discovery of \object{TYC 3980-1081-1 B}, if confirmed by follow-up spectroscopy, shows that previous assumptions on the completeness of the WD sample within 13\,pc may be not correct. This concerns in particular cool WDs with small proper motions. \item The identification of relatively close companions to very bright stars, such as \object{HD 166435 B} and \object{TYC 3980-1081-1 B} has always been a challenge with photographic all-sky surveys. We took advantage of the better resolution of CCD images in UCAC5 and Gaia. Gaia's next data releases will certainly help identifying more close WD companions of nearby stars. On the other hand, \citetads{2015ApJS..219...19L} found eight {\it single} WDs as probable new members of the 25\,pc sample in their photographic (SUPERBLINK) survey of the Northern sky with a lower proper motion limit of 40\,mas/yr. A similar number of missing nearby WDs can be expected in the Southern sky, and additional WDs with even smaller proper motions and trigonometric distances $<$25\,pc may still be discovered in forthcoming Gaia all-sky catalogues. \end{enumerate} \begin{acknowledgements} We have extensively used the SIMBAD database and the VizieR catalogue access tool operated at the CDS/Strasbourg. The spectroscopic confirmation of one of our WD candidates was part of a larger observing campaign carried out with the 2.2\,m telescope of the German-Spanish Astronomical Centre at Calar Alto, Spain. Part of the observations were carried out in service mode. We would like to thank the Calar Alto staff for their kind support and for their help with the observations. We thank Axel Schwope for providing equivalent widths of Balmer lines in DA WDs. It is a pleasure to thank Benjamin Braun and Peter Grodzewitz for their help with a first look at the TGAS and TGAS/UCAC5 nearby subsamples during their internships at AIP in February and May 2017, respectively. We would like to thank the referee for many valuable comments and suggestions that helped us improve the paper. \end{acknowledgements} \bibliographystyle{aa}

\section{Introduction} Fluctuations and correlations of conserved charges provide diagnostic tools for the nature of strongly interacting matter described by Quantum Chromodynamics (QCD). The first-principles calculations by lattice QCD (LQCD) have provided not only equation of state at the physical quark masses but also the fluctuations and correlations of the net-baryon, net-electric charge, and net-strangeness \cite{}. The fluctuations and correlations can also be measured in heavy-ion collisions to identify the state of created matter. In particular, non-Gaussian (higher-order) fluctuations have been expected to probe critical properties of the system, such as the QCD critical point in the beam-energy scan program at RHIC and a remnant of the $O(4)$ criticality at small baryon density \cite{ejiri06:_hadron_fluct_at_qcd_phase_trans,friman11:_fluct_as_probe_of_qcd}. For a physical interpretation of the thermodynamic properties, the hadron resonance gas (HRG) model \cite{Statmodelreview_QGP3,} has been used as a reference. The equation of state and fluctuations have been well described by the model below the chiral crossover temperature $T = 154\pm 9$ MeV \cite{karsch11:_probin_freez_out_condit_in,bazavov12:_chiral_and_decon_aspec_of_qcd_trans}, and the experimental data of fluctuations have been analyzed in terms of a gas of hadronic states \cite{garg13:_conser,}. At vanishing baryon density, the critical behavior due to the chiral phase transition appears at the sixth order of net-baryon fluctuations \cite{friman11:_fluct_as_probe_of_qcd}. This is because the singular contribution to the free energy is suppressed in the lower-order fluctuations even in the vicinity of the expected second order transition. Therefore, one expects that the lower order fluctuations and correlations may be well described by the hadronic degrees of freedom, and they would provide a reliable baseline for exploring the critical behavior \cite{ejiri06:_hadron_fluct_at_qcd_phase_trans}. According to the recent LQCD calculations, however, some correlations between conserved charges cannot be explained by a conventional HRG model. Particularly interesting quantities are those involving net-baryon number, baryon-charge (BQ) and baryon-strangeness (BS) correlations. Since mesons do not contribute to these quantities, one may gain access to a role of baryonic degrees of freedom and interactions which are not visible in the equation of state and other meson-dominated quantities due to their heavy masses \cite{}. In the HRG model the interaction of hadrons is replaced by resonances, and in the first approximation their widths can be neglected. Then, the partition function of the interacting hadronic system can be written as a mixture of free gases of all stable and resonant hadrons \cite{Dashen1969}. The validity of the vanishing-width approximation for the fluctuations has been recently examined based on the S-matrix formalism for $K$--$\pi$ \cite{} and $\pi$--$N$--$\Delta$ \cite{} systems. It has been found that an explicit treatment of the width can have a substantial effect on the fluctuations. On the other hand, at finite temperature and density one expects changes in the spectral property of hadrons. In particular, a search for those modifications because of the partial restoration of the QCD chiral symmetry has been one of the central subjects in heavy-ion collisions \cite{}. Since the vacuum masses are used in the HRG model, the observed agreement of the equation of state in the HRG and LQCD suggests no need for a strong mass reduction in the dominant degrees of freedom such as light mesons. Although chiral symmetry predicts substantial medium modifications of low-lying mesons \cite{Hatsuda93,Hohler2014}, their reliable estimates in LQCD have been limited to the screening masses \cite{}. Recently, masses of the non-strange and strange baryons with positive and negative parity have been extracted from the temporal correlation functions by the FASTSUM collaboration~\cite{}. The negative-parity states clearly show downward mass shifts, whereas the positive-parity baryons stay insensitive to temperature. The obtained spectra follow an expectation from the parity doublet picture of the chiral symmetry, indicating that the masses of a negative- and positive-parity partners tend to degenerate when approaching the chiral crossover. The above medium modification was used to possibly explain a missing contribution in the correlations between the conserved charges \cite{Aarts_SQM2017}. In the following we investigate the fluctuations and correlations of the net-baryon with net-charge and net-strangeness on the basis of the HRG model implementing in-medium mass modifications. We employ temperature dependent masses of the negative-parity octet and decouplet baryons from the lattice QCD and from a chiral effective model with parity doubling \cite{sasaki17:_parity}. We show that the strong downward mass shift in hyperons can accidentally reproduce some correlation ratios, however it also tends to overshoot the individual fluctuation and correlation observables. This indicates that to quantify fluctuation and correlation of conserved charges in the presence of chiral symmetry restoration it is not sufficient to implement in-medium hadron masses in the statistical sum of the hadron resonance gas. \section{Baryon masses in parity doublet picture} \subsection{Lattice QCD} \begin{table*}[!t] \caption{Assignment of negative parity states in in-medium HRG} \label{} \begin{tabular}[t]{l|cccccccc}\hline P$^-$ State & $N$ & $\Lambda$ & $\Sigma$ & $\Xi$ & $\Delta$ & $\Sigma^*$ & $\Xi^*$ & $\Omega$ \\\hline Mass { LGT} \cite{} & 1779 & 1899 & 1823 & 1917 & 2138 & 2131 & 2164 & 2193 \\ $M_-^i(T_c)$ [MeV]& 1254 & 1172 & 1329 & 1295 & 1405 & 1398 & 1426 & 1383 \\ $b_i$ & 0.338 & 0.369 & 0.257 & 0.275 & 0.312 & 0.257 & 0.246 & 0.213 \\ \hline Assignment A & $1535$ & $1405$ & $1750$ & $1690$ & $1700$ & $1670$ & $1820$ & 2250 \\ Assignment B & $1535$ & $1670$ & $1750$ & $1950$ & 1700 & $1940$ & $1820$ & $2250$ \\ \hline Mass [MeV] & 1535 & 1790 & 1880 & 2090 & 1710 & 1930 & 2150 & 2380 \\ Assignment C & 1535 & $1800$ & $1880$ & $2120$ & $1700$ & $1940$ & $2250$ & $2380$ \\ \hline \end{tabular} \end{table*} The FASTSUM collaboration presented the temperature dependent masses of $N$, $\Delta$, and $\Omega$ states, extracted from imaginary time correlators of the corresponding interpolating operators in $N_f = 2 + 1$ lattice simulations ~\cite{}. Their calculations were performed for heavier light-quark mass than the physical one; $m_\pi=384$ MeV, while the strange quark is set to the physical one. Thus, the mass of baryons except $\Omega$ are heavier than the physical ones. This yields the nucleon with positive parity $N_+$ has $m_{N_+}=1158$ MeV (939 MeV in PDG), while the positive parity $\Omega_+$ has $m_{\Omega_+}=1661$ MeV (1672 MeV in PDG). The results of other octet and decuplet states have been shown in \cite{Aarts_SQM2017}. Here we use the masses at $T/T_c=0.24$, $0.76$, $0.84$, and $0.95$. Note that owing to the heavier pion mass, the crossover temperature $T_c=185$ MeV, which was determined from the renormalized Polyakov loop, is also higher than the physical one, $T_c = 154$ MeV. \begin{figure}[!h] \centering \includegraphics[width=0.6\columnwidth]{Mass_N-Delta.eps} \caption{Temperature dependence of the mass of $N_-$ (a) and $\Delta_-$ (b) from lattice-motivated parameterization Eq.~\eqref{} (lines) and from a chiral effective model (squares).} \label{} \end{figure} In Table \ref{}, we list the masses of the octet and decuplet negative-parity states at $T=0.24T_c$. They are well parameterized, for the given the state $i$, by~\cite{} \begin{align} M_-^i(T) &= M_-^i(T=0)\omega(T,b_i) + M_-^i(T_c)(1-\omega(T,b_i))\label{ \omega(T,b_i)&= \tanh[(1-T/T_c)/b_i] / \tanh(1/b_i)\label{} \end{align} where $M_-^i(T=0)$ is fixed to be the value obtained at $T=0.24T_c$ and $M_-^i(T_c)$ and $b_i$ are the fitting parameters fixed for each channel. The data are given as a function of $T/T_c$; the value of $T_c$ does not affect the fit but later it is set to the physical value 154~MeV. The parameter $b_i$ corresponds to the width of the chiral crossover. The results of the fits are shown in Table \ref{}. \begin{figure*}[!tb] \centering \includegraphics[width=0.7\textwidth]{Mass_Hyperon_AB.eps} \caption{Temperature dependence of octet (upper row) and decuplet (lower row) hyperons from lattice-motivated parameterization Eq.~\eqref{}. } \label{} \end{figure*} To correct the unphysical effect from the heavy up and down quarks, $M_-^i(T=0)$ is set to its PDG mass in the following. The value of $M_-(T_c)$ needs to be corrected as well, and we shall re-scale it by multiplying the factor $M_+^{\text{PDG}}(T=0)/M_+^\text{lattice}(T=0)$ with $M_-(T_c)$. It is not clear how to assign the negative-parity states to the observed ones because of unknown quantum numbers of some of the candidates. We follow the suggestion in \cite{Aarts_SQM2017} and further adopt two different assignments, A and B. The set A takes the lighter state for $\Lambda_-$, $\Xi_-$, and $\Sigma^*_-$ while the set B does the heavier ones. The obtained temperature-dependent masses are displayed in Figs.~\ref{} and \ref{}. The masses of negative-parity states start to drop around $T/T_c \simeq 0.6$, then approach those of their positive-parity partners. In comparing the non-strange baryons with the hyperons, one finds the correction of $M(T_c)$ significantly affects the baryons with more light quarks, as expected. \subsection{Chiral model} The parity-doubled baryons can be modeled in an effective chiral approach~\cite{}, where a chiral-invariant mass is naturally introduced. In \cite{sasaki17:_parity}, the mass relations to the light quark $\sigma_q$ and strange quark $\sigma_s$ condensates are given for the octet and decuplet states as~\footnote{ In~\cite{sasaki17:_parity} the same $m_0$ was assumed common to the octet and decuplet baryons. Here we lift this constraint and introduce two independent masses. Their splitting can be deduced from the spin-spin interaction. } \begin{eqnarray} m_{N_\pm}&=& \left(a_N \mp b_N\right)3\sigma_q + m_{N0}\,, {\hskip 1.6cm~~~~~~~~~~m_{\Delta_\pm} = \left(a_\Delta \mp b_\Delta\right)3\sigma_q + m_{\Delta 0}\,} \nonumber\\ m_{\Sigma_\pm} &=& \left(a_N \mp b_N\right)\left(2\sigma_q + \sqrt{2}\sigma_s\right) {}+ m_{N0} + m_1\,, ~~m_{\Sigma^\ast_\pm} = \left(a_\Delta \mp b_\Delta\right)\left(2\sigma_q + \sqrt{2}\sigma_s\right) {}+ m_{\Delta 0} + m_s\, \nonumber\\ m_{\Lambda_\pm} &=& \left(a_N \mp b_N\right)\left(2\sigma_q + \sqrt{2}\sigma_s\right) {}+ m_{N0} + m_3\,, ~~ m_{\Xi^\ast_\pm}= \left(a_\Delta \mp b_\Delta\right)\left(\sigma_q + 2\sqrt{2}\sigma_s\right) {}+ m_{\Delta 0} + 2m_s\, \nonumber\\ m_{\Xi_\pm} &=& \left(a_N \mp b_N\right)\left(\sigma_q + 2\sqrt{2}\sigma_s\right) {}+ m_{N0} + m_2\,, ~~m_{\Omega_\pm} = \left(a_\Delta \mp b_\Delta\right)3\sqrt{2}\sigma_s {}+ m_{\Delta 0} + 3m_s\,. \end{eqnarray} The parameters $m_{1,2,3}$ are related via the Gell-Mann--Okubo relation as \begin{equation} m_1 = 2m_2 - 3m_3\,. \end{equation} The parameters in the above expressions are determined at zero temperature as in Table~\ref{paraB}. \begin{table*} \begin{center} \begin{tabular*}{12cm}{@{\extracolsep{\fill}}ccccccc} \hline $a_N$ & $b_N$ & $m_1$ [GeV] & $m_2$ [GeV] & $a_\Delta$ & $b_\Delta$ & $m_s$ [GeV] \\ \hline $1.22$ & $1.08$ & $0.248$ & $0.367$ & $1.16$ & $0.862$ & $0.139$ \\ \hline \end{tabular*} \end{center} \caption{ Set of parameters in the baryon-mass relations. We set $m_{N0} = 0.9$ GeV and $m_{\Delta 0} = 1.15$ GeV. } \label{paraB} \end{table*} The masses at $T=0$ are also summarized in Table \ref{}. Thermal modifications of the baryon masses are driven by the quark condensates. Following \cite{sasaki17:_parity}, we shall use the in-medium condensates measured in LQCD by the HotQCD collaboration~\cite{bazavov12:_chiral_and_decon_aspec_of_qcd_trans}. One readily finds that some of the hyperon states have larger masses than those expected from the lattice baryon-spectrum in~\cite{}. Thus, we assign these hyperons to the PDG states as in the bottom line of Table \ref{} and refer to set C. For a comparison with LQCD, we also apply the set C to the lattice-motivated parametrization \eqref{}. The non-strange baryons exhibit the thermal behavior as in Fig.~\ref{}. To some extent, their trends are similar to the results by the FASTSUM collaboration, whereas the parametrization~(\ref{}) leads to a sizable difference in the mass dropping near $T_c$. This is particularly evident when the re-scaling of $M_-(T_c)$ is made. Figure \ref{} displays the mass of the hyperons in the assignment C. The most distinct difference from the non-strange baryons lies in the mass of the hyperons near $T_c$; the model calculation shows a rather moderate downward-shift, while the lattice QCD results lead to a strong mass drop. Comparing $\Lambda$ ($\Sigma$), $\Xi$, and $\Omega$ states between the model and the lattice-inspired scaling, one clearly finds that the difference becomes more significant as the baryon contains more strangeness. This may be attributed to the difference in the masses between the light and strange quarks ; The FASTSUM setup with a large pion mass $m_\pi=384$ MeV and the physical kaon mass is rather close to the flavor SU$_f$(3) limit, while the model calculation done with the physical pion mass is dominated by SU$_f$(2). Therefore, it will be intriguing to see whether the strong mass reduction of $\Omega$ and $\Xi$ would still persist in the lattice simulations with a lighter pion. \begin{figure*}[!t] \centering \includegraphics[width=0.7\textwidth]{Mass_Hyperon_C.eps} \caption{Comparison of hyperon masses between the lattice-motivated parametrization (lines) and the chiral effective model (squares) in the assignment C. As in Fig.~\ref{}, the octet and decuplet hyperons are shown in upper and lower panels, respectively. } \label{} \end{figure*} \section{Fluctuations and correlations from in-medium HRG} \label{} We employ the hadron resonance gas model to explore the fluctuations and correlations in the hot hadronic matter. In this work we entirely rely on the zero-width approximation; all the resonances are treated as point-like particles and their spectral functions are given by the delta function. Consequently, the thermodynamic pressure at temperature $T$ and chemical potentials ${\mu_j}$ $(j=B,Q,S)$ is given by the summation of ideal gas pressure of each particle species as \begin{equation} p^{\text{HRG}}(T,\mu_B,\mu_Q,\mu_S) = \sum_{i=\text{hadrons}}p^{\text{ideal}}(T,\mu_i;m_i) \end{equation} where $\mu_i = B_i \mu_B + Q_i \mu_Q + S_i \mu_S$ is the chemical potential of particle $i$ with baryon number $B_i$, electric charge $Q_i$, and strangeness $S_i$ and the ideal gas pressure with degeneracy $d$ reads \begin{equation} p^{\text{ideal}}(T,\mu;m) = \mp \frac{d T}{2\pi^2} \int_{0}^{\infty}\!\! dp p^2 \ln[1 \mp e^{-(\sqrt{p^2+m^2} - \mu)/T}]. \end{equation} The signs are negative for mesons and positive for baryons. The fluctuations and the correlations at vanishing chemical potentials can be expressed as the generalized susceptibilities \begin{equation} \chi_{ijk}^{BQS} \equiv \frac{\partial^{i+j+k}p^{\text{HRG}}/T^4}{\partial^i(\mu_B/T) \partial^j(\mu_Q/T) \partial^k(\mu_S/T)} \end{equation} Although heavy particles ($m \gg T$) are thermally suppressed, the existence of missing states are expected likely in the strange sectors \cite{PhysRevLett.113.072001,Lo2015}. In fact, the assignment C in Table~\ref{} contains several 1-star states whose spins are undetermined in the latest PDG \cite{pdg2016}. For those state we shall assume the same spin with its partner. For consistency we also include other unconfirmed state; we have in total 28 nucleons, 22 $\Delta$ baryons, 18 $\Lambda$ baryons, 21 $\Sigma$ baryons, 8 $\Xi$ baryons, 3 $\Omega$ baryons, and their isospin multiplets. We also include a deuteron, triton, $^3$He and $^4$He as they are considered to be thermal ingredients in QCD thermodynamics and heavy ion collisions \cite{}. In this work we focus on $\chi^{B}_2$, $\chi^{BQ}_{11}$, and $\chi^{BS}_{11}$ at $\mu_B=\mu_Q=\mu_S=0$. Since $\chi_{ijk}^{BQS} \propto (B/T)^{i}(Q/T)^j (S/T)^k$, mesons do not contribute to these susceptibilities. Thus they are good measures of in-medium effects in the baryon sector on the QCD thermodynamics. In general, particle masses in a medium can depend not only on temperature but also on chemical potentials. We neglect such an intrinsic chemical-potential dependence in the baryon masses in the present calculations. This is well justified at small chemical potential since the quark condensates are not much affected. \begin{figure*}[!t] \centering \includegraphics[width=0.99\textwidth]{chiBB-BQ-BS.eps} \caption{Fluctuations and correlations from the HRG model with and without the mass shifts. From left to right, (a) Baryon number susceptibility, (b) Baryon-charge correlation, and (c) Baryon-strangeness correlation are shown together with the LQCD results from HotQCD \cite{bazavov12:_fluct_and_correl_of_net,} and Budapest-Wuppertal Collaboration \cite{}} \label{} \end{figure*} We display $\chi_2^B$, $\chi_{11}^{BQ}$, and $-\chi_{11}^{BS}$ in Fig.~\ref{}. In the baryon number susceptibility $\chi_2^B$, the HRG model agrees with the LQCD results below $T_c$. \footnote{The present HRG leads to a slightly larger $\chi_2^B$ than the conventional one because of including the 1-star states and the multi-baryons.} With the in-medium mass shift, a decrease in the negative-parity baryon masses lead to an enhancement of the susceptibility. One sees a moderate enhancement in all the three scenarios, Lattice (A) (B) and (C), and it becomes stronger when $M(T_c)$ is corrected, since the correction further reduces the masses just below $T_c$. On the other hand, the different assignments of the parity-partners lead to a minor change. The results with the masses from the chiral effective model follows the same trend, but the amount of the enhancement is much smaller than the $\tanh$-parameterization, owing to the weaker downward mass shifts in the hyperon sectors (see Fig.~\ref{}.) In any case, all the results with mass reduction overshoot the lattice data of $\chi_2^B$. The $\chi_{11}^{BQ}$ shows a different tendency in Fig.~\ref{; the results of Lattice (A)-(C) without the correction of $M(T_c)$ do not exhibit any enhancement from the HRG, even below the ``Model'' result. In the $\chi^{BQ}_{11}$, $I=1$ states do not contribute because neutral particles do not contribute and the positively-charged states cancel with the negatively-charged states. Charged $N$ states and doubly charged $\Delta^{++}$ states contribute positively, while negatively-charged $\Xi_-^-$, $\Xi_-^{*-}$ and $\Omega_-^-$ states contribute oppositely. Without a strong mass-shift of these multi-strange states, the $\chi^{BQ}_{11}$ becomes larger due to the smallness of the hyperon contribution suppressed by the Boltzmann factor, as in the ``Model'' case. With the strong mass shift of the hyperons, they tend to cancel a part of the positive contribution, and suppress the $\chi_{11}^{BQ}$. The strong enhancement with the corrected $M(T_c)$ clearly signals that the effect of the correction is smaller for $\Xi$ states and larger for $N$ and $\Delta$ states, as seen in Figs.~\ref{}. The behavior of $-\chi_{11}^{BS}$ in Fig.~\ref{ can be understood in a similar way to the $\chi_2^B$; the ``Model'' result does not differ much from the HRG, because of a weaker mass reduction. Owing to the different trend in the hyperon sectors, the curves deviate more sensitively depending on the fate of the masses. The observed differences can be more pronounced by taking the ratio between the susceptibilities. In Fig.~\ref{, we display the $\chi_{11}^{BQ}/\chi_2^B$ together with the LQCD data. The results from Lattice (A) and (B) reproduce the one shown in \cite{Aarts_SQM2017}; they follows the trend seen in the HotQCD data. As explained above, however, the individual susceptibilities cannot be reproduced. The coincidence comes from the reduction of $\chi_{11}^{BQ}$ owing to the strong mass reductions in the charge asymmetric states. The ``Model'' case shows the opposite trend that the ratio is above the HRG, due to the much milder mass-reduction in the hyperon sectors. Therefore, the data cannot be explained solely by the mass reduction. Also shown in Fig.~\ref{ is the $-\chi_{11}^{BQ}/\chi_{11}^{BS}$. Again, while the sets (A) and (B) follow the trend of the HotQCD data, the ``Model'' shows the opposite. This supports the observation that naive insertion of the in-medium masses of parity doublers into the HRG results in fluctuation observables which are inconsistent with LGT data. Nevertheless, our results do not rule out the manifestation of parity doublers. Several studies \cite{} have revealed that the conventional treatment of the particles in the HRG is not sufficient and a proper inclusion of the width improves the thermodynamics. Although those analysis have been limited to the $S-$matrix approaches where interactions are incorporated via the two-body scattering phase shift in the vacuum, a consistent treatment of the resonance widths at finite temperature and densities should clarify the consequences of the partial restoration of the chiral symmetry in the baryonic sector of correlations and fluctuations. \begin{figure*}[!t] \centering \includegraphics[width=0.4\columnwidth]{chiBQchiBB_Tdepmass.eps} \includegraphics[width=0.4\columnwidth]{chiBQchiBS_Tdepmass.eps} \caption{Ratio of the baryon-charge correlation to the baryon number fluctuation (a) and the baryon-strangeness correlation (b). The legends are the same as Fig.~\ref{}.} \label{} \end{figure*} \section{Concluding remarks} \label{} We have studied the fluctuations and correlations of conserved charges by making use of the hadron resonance gas model with in-medium masses of the low-lying parity-doubled baryons. Motivated by the recent LQCD result on the baryon masses at finite temperature, we have examined the consequences of several scenarios with in-medium masses for the negative-parity octet and decouplet baryons, adopting a lattice-motivated parametrization and that from a chiral effective theory. We have computed the baryon number fluctuation $\chi_2^B$, baryon-charge correlation $\chi_{11}^{BQ}$, and baryon-strangeness correlation $\chi_{11}^{BS}$. The results have been understood on the basis of the presence or absence of the strong mass reduction in the hyperon sectors. We have pointed out that the reproduction of the LQCD results of the ratios $\chi_{11}^{BQ}/\chi_2^B$ and $-\chi_{11}^{BQ}/\chi_{11}^{BS}$ is just accidental when the in-medium masses are naively introduced in the conventional HRG approach. Since the strong mass shifts of the hyperons observed in LQCD can be regarded as a consequence of the approximate SU$_f$(3) nature due to a heavy pion mass, it is a critical issue to be examined whether the hyperon mass-shift persists for the physical pion mass. Although our results clarify the interplay of the different particle species in the behavior of the baryon-charge and baryon-strangeness correlations, we emphasize that the treatment of the resonances in the conventional HRG model is insufficient to quantify correlation and fluctuations of conserved charges. This indicates, that in order to correctly account for the influence of the chiral symmetry restoration on the fluctuation observables, a consistent framework of in-medium effects beyond hadron mass shifts is required. \subsection*{Acknowledgments} We acknowledge stimulating discussions with G\'{a}bor Alm\'{a}si and Bengt Friman. This work has been partly supported by the Polish Science Center NCN under Maestro grant 2013/10/A/ST2/00106 and RIKEN iTHES project. \input{ref_inmediumHRG.bbl} \end{document} \bibitem{} H.~T.~Ding, F.~Karsch and S.~Mukherjee, Int.\ J.\ Mod.\ Phys.\ E {\bf 24}, no. 10, 1530007 (2015)

\section{Appendix: Cyclic order of unbounded closed connected sets}\label{} Suppose that $\mathcal{A}$ is any pairwise disjoint collection of unbounded, closed, connected subsets of $\C$ none of which separates the plane. Then there is a natural \emph{cyclic order} on $A$. Recall that a cyclic order is a ternary relation $A\prec B \prec C$ that is cyclic, asymmetric, transitive and total. In our case, the relation $A\prec B \prec C$ means that $B$ lies between $A$ and $C$ in positive orientation. More precisely, let $R$ be sufficiently large to ensure that each of the three sets contains a point of modulus at most $R$. Then $\C\setminus (\overline D_R(0) \cup C)$ has exactly one connected component $W$ from which $\infty$ is accessible. Since a point is accessible from a domain by a curve if and only if it is accessible by a closed set (see e.g.\ \cite[Lemma~A.1]{Re08}), it follows that $A$ and $B$ tend to infinity within $W$. Let $U$ be a connected component of $\exp^{-1}(W)$; observe that $U$ lies in the right half-plane $H\defeq \{a+ib\colon a>\log R\}$. Furthermore, $\partial U\cap H$ has an upper and a lower boundary; for a fixed $a>\log R$, the upper boundary separates $U$ from all points of the form $a+ib$ with $b$ sufficiently large and positive, and the lower boundary separates $U$ from all points of the form $a+ib$ with $b$ sufficiently large and negative. Set $\alpha \defeq \exp^{-1}(A)\cap U$ and $\beta \defeq \exp^{-1}(B)\cap U$. Then we say that $A\prec B \prec C$ if $\beta$ separates all sufficiently large points of $\alpha$ from the upper boundary of $U$, or equivalently if $\alpha$ separates all sufficiently large points of $\beta$ from the lower boundary. Otherwise, $B\prec A \prec C$. It can be checked that this does indeed define a cyclic order on $\mathcal{A}$. Suppose that $\tilde{\mathcal{A}}$ is a second collection as above, wher every element of $\tilde{\mathcal{A}}$ is contained in an element of $\mathcal{A}$ and every element of $\mathcal{A}$ contains exactly one element of $\tilde{\mathcal{A}}$. Then the cyclic order on $\tilde{\mathcal{A}}$ coincides with the corresponding order on $\mathcal{A}$. \begin{remark} There are some subtleties to the assumptions and the construction, compared with the case of arcs to infinity which has been previously considered in the complex dynamics literature. Firstly, note that the assumption that the sets in $\mathcal{A}$ are closed is crucial. Indeed, consider the case of a Knaster bucket-handle continuum $X$, whose terminal point (that is, the initial point of the half-ray running through all of the endpoints of the complementary intervals of the ternary Cantor set) has been placed at $\infty$, and consider the collection of path-connected components of this set. Each such component is unbounded and connected, but since each component accumulates everywhere upon $X$, there is no sensible circular order among them. Secondly, let $C\in\mathcal{A}$ as above. Then it is possible for $\C\setminus (\overline D_R(0)\cup C)$ to have more than one connected component. Indeed, consider the case where $C$ is the union of $[0,\infty)$ together with another ray, which starts within $D_R(0)$ and accumulates, in the manner of the $\sin(1/x)$-curve, on the interval $[R+1,\infty)$. With some more care, an example can be constructed where this occurs for infinitely many $R$; furthermore, dreadlocks with this property can arise as Julia continua by the results of \cite{arclike}. However, as noted above, $\infty$ can be accessible from only one of these components, and hence only this component is relevant to the definition of circular order. \end{remark} We can use this observation to define cyclic order also for pairwise disjoint collections of \emph{open} unbounded domains, each of which contains exactly one homotopy class of curves to infinity. (Simply replace each domain by a representative in the mentioned homotpy class.) Furthermore, suppose that $U$ and $\tilde{U}$ are unbounded domains in $\C$, that $\phi\colon U\to \tilde{U}$ is a conformal isomorphism. Also suppose that $\mathcal{A}$ and $\tilde{\mathcal{A}}$ are collections as above, whose elements are contained in $U$ and $\tilde{U}$, respectively, that $\phi$ maps every element of $\mathcal{A}$ to an element of $\tilde{\mathcal{A}}$, and that all elements of $\tilde{\mathcal{A}}$ arise in this manner. Then the action of $\phi$ on $\mathcal{A}$ preserves cyclic order. Finally, suppose that $\tilde{A}$ is a pairwise disjoint collection of closed, connected sets in $\C^*=\C\setminus\{0\}$, and that the closure of element of $A$ contains both $0$ and $\infty$. Then the cyclic order on $\tilde{A}$ at $\infty$ and the cyclic order on $\tilde{A}$ at $0$ (defined in the analogous manner) coincide. \section{Appendix: Unbounded postsingular sets}\label{} As mentioned in the introduction, the Douady-Hubbard landing theorem no longer holds for polynomials with escaping singular values. It is still true that every repelling (or parabolic) periodic point is accessible from the basin of infinity, and even by an external ray (understood appropriately in the case where the ray passes through critical points). Compare \cite{ELv,LP}. However, it is possible for the set of external rays landing at such a point to be uncountable, and for none of these rays to be periodic; compare \cite[Appendix~C]{goldbergmilnor} and \cite{LP}. Let us now briefly discuss the case of transcendental entire functions $f$ with unbounded postsingular set $\P(f)$. When $f\notin \mathcal{B}$, the structure of the escaping set may change dramatically within a given parameter space (compare \cite[Appendix~B]{rempesixsmith}), and hence it is not clear whether questions concerning the landing of rays or dreadlocks are even meaningful in this setting. Let us hence restrict to the case of $f\in\mathcal{B}$. First suppose that $f$ has an escaping singular value. In addition to the above-mentioned behaviour that occurs already for polynomials, it is also possible for a repelling periodic point to not be accessible from the escaping set at all (by hairs or dreadlocks). Indeed, this is the case for the fixed point of the exponential map $z\mapsto e^z$ having imaginary part between $0$ and $\pi$, and shows that the question of landing behaviour at periodic points becomes considerably more subtle when $\P(f)$ is unbounded. However, consider now the full family of exponential maps, $f_a\colon z\mapsto e^z+a$. Suppose that the singular value $a$ has an unbounded orbit but does not belong to the escaping set. Then $f_a$ is criniferous. In \cite{Re06}, it is shown that that all periodic hairs of $f_a$ land. Conversely, every periodic point, with the exception of at most one periodic orbit, is the landing point of a periodic hair. The exceptional orbit cannot be parabolic, but it is an open question whether it can be repelling. However, it is shown in \cite{Re06} that a plausible conjecture about parameter space of exponential maps (the ``no ghost limbs conjecture'') would imply that this is not the case. Hence it is plausible that the Douady-Hubbard landing theorem remains valid for exponential maps as above, which raises the question whether the main theorem of our paper may also have an extension for functions $f\in\mathcal{B}$ with unbounded but non-escaping singular orbits. However, it appears that fundamentally new approaches would be required to resolve this question. Indeed, the proofs in \cite{Re06} use sophisticated results on the structure of exponential parameter space, which rely in an essential way on the one-dimensionality of the family. Already for cosine maps $z\mapsto ae^z+ be^{-z}$ with at least one unbounded critical orbit, the landing of periodic rays remains open in general, to say nothing of the accessibility of repelling periodic points. \section{Introduction} Let $p\colon\C\to\C$ be a polynomial. The \emph{filled-in Julia set} $K(p)$ consists of those points $z\in\C$ whose orbits remain bounded under repeated application of $p$. In their study of the dynamics of complex polynomials and the Mandelbrot set \cite{DH84}, Douady and Hubbard introduced the notion of \emph{external rays}, which can be characterised as the gradient lines of the Green's function on the basin of attraction of infinity, $\C\setminus K(p)$. Such a ray is said to \emph{land} if it has a unique accumulation point in $K(p)$. \emph{Periodic} external rays are of particular importance, due to the following result. \begin{douadyhubbardlandingtheorem} Let $p$ be a polynomial whose \emph{post-critical set} \begin{equation}\label{} \mathcal{P}(p) \defeq \overline{\bigcup_{c\colon p'(c)=0} \{p^n(c)\colon n\geq 1\}} \end{equation} is bounded. (Equivalently, assume that $K(p)$ is connected.) Then every periodic external ray of $p$ lands at a repelling or parabolic periodic point, and conversely every repelling or parabolic periodic point of $p$ is the landing point of at least one and at most finitely many periodic external rays. \end{douadyhubbardlandingtheorem} \begin{remark} The first half of the theorem, concerning the landing of periodic external rays, can be found in \cite[Expos\'e~VIII.II, Proposition 2]{DH84}. The second half, which is more difficult, is due to Douady; the first published proofs are in \cite{ELv,Hu}. \end{remark} This theorem has been a cornerstone of the study of polynomial dynamics ever since. In particular, it forms the basis of the ``puzzle techniques'' that were pioneered by Yoccoz, Branner and Hubbard, and continue to lead to fundamental new results; see \cite{Hu,roeschyin,avilalyubichshen}. In the study of rational functions and transcendental entire functions, there is no immediate analogue of the basin of infinity, and this is one of the reasons that the study of these classes has presented greater challenges than that of polynomials. Nonetheless, in both settings there are cases where analogues of the above-mentioned puzzle techniques have been employed with considerable success. We refer to \cite{Ro08,Be15} for two such examples. Our goal is to extend the Douady-Hubbard landing theorem to the case of a transcendental entire function $f$. In this setting, the role that critical values play in polynomial dynamics is taken by the larger set $S(f)$ of \emph{singular values} of $f$. These are those points not having a neighbourhood in which all branches of $f^{-1}$ are defined and holomorphic. Analogously to~\eqref{}, the \emph{postsingular set} of $f$ is defined as \begin{equation}\label{} \P(f) \defeq \overline{\bigcup_{s\in S(f)} \{f^n(s)\colon n\geq 0\}}. \end{equation} For transcendental maps, $\infty$ is an essential singularity, rather than a super-attracting fixed point. Hence the definition of external rays for polynomials as gradient lines of a Green's function has no natural analogue. Nonetheless, it has long been known that the \emph{escaping set} \[ I(f) \defeq \{z\in\C\colon f^n(z)\to\infty\} \] often contains curves to infinity; indeed, in some cases this was already noticed by Fatou~\cite{fatou26}. It was the work of Devaney and his collaborators (see e.g.\ \cite{DK,DT86}) that really began the study of these \emph{hairs} or \emph{dynamic rays} in the 1980s, particularly for functions in the \emph{exponential family}, \begin{equation}\label{} f_a\colon z\mapsto e^z+a. \end{equation} Devaney, Goldberg and Hubbard were probably the first to suggest that such hairs could serve as analogues of external rays of polynomials; compare \cite{dgh,BD}. Subsequently, Schleicher and Zimmer \cite{expescaping} and Schleicher and Rottenfu{\ss}er \cite{cosescaping} proved that, for the families of exponential maps~\eqref{} and of \emph{cosine maps} $z\mapsto ae^z+ be^{-z}$, respectively, the \emph{entire} escaping set $I(f)$ consists of hairs. On the other hand, in \cite{R3S} it is shown that there is a transcendental entire function $f$ for which $I(f)$, and indeed the Julia set $J(f)$, contains no arcs. Hence there are no curves in $I(f)$, of any kind, landing at any of the repelling periodic points of $f$. (Recall that repelling periodic points are dense in the Julia set of any transcendental entire function.) Furthermore, the postsingular set $\P(f)$ of this function is bounded. Indeed, $S(f)$ is a compact subset of the immediate basin of attraction of a single attracting fixed point. \subsection*{Dreadlocks} In view of the preceding example, we will develop a novel approach to the landing problem that removes the focus on hairs altogether, by connecting repelling periodic points to infinity using not necessarily curves, but more general sets of escaping points. More precisely, we introduce a notion of \emph{dreadlocks} for postsingularly bounded entire functions. These are certain unbounded connected sets of escaping points generalising the concept of hairs. (See Section~\ref{} for formal definitions.) The set of dreadlocks has a natural combinatorial structure, and in tame cases, all dreadlocks are in fact hairs. In general, however, dreadlocks can be topologically much more complicated. Indeed, it follows from \cite{arclike} that the closure of a dreadlock may be a hereditrarily indecomposable continuum.\footnote{% More precisely, \cite{arclike} shows that the \emph{Julia continuum} of a disjoint-type entire function may have this property, Every such Julia continuum is the closure of a dreadlock in our sense; compare Lemma~\ref{} and Remark~\ref{}.} With this terminology, we are able to prove the following generalisation of the Douady-Hubbard landing theorem for postsingularly bounded entire functions: Every periodic dreadlock lands, and every repelling or parabolic periodic point is the landing point of at least one and at most finitely many periodic dreadlocks (Theorem~\ref{}). In particular, {without requiring the definitions of Section~\ref{}, we can state the following result.} \begin{thm}[Landing at periodic points]\label{} Let $f$ be a transcendental entire function such that $\P(f)$ is bounded, and let $\zeta$ be a repelling or parabolic periodic point. Then there is a connected and unbounded set $A\subset I(f)$ and a period $p$ with the following properties. \begin{enumerate}[(a)] \item $\overline{A} = A\cup \{\zeta\}$;\label{} \item $f^p(A)=A$ and $f^j(A)\cap A = \emptyset$ for $1\leq j < p$;\label{} \item for every $\eps>0$, $f^n$ tends to $\infty$ uniformly on $\{z\in A\colon \lvert z-\zeta\rvert \geq \eps\}$.\label{} \end{enumerate} If $\tilde{\zeta}$ is a second repelling or parabolic periodic point and $\tilde{A}$ is as above, then $A\cap\tilde{A}=\emptyset$. \end{thm} This result offers the possibility of developing puzzle-type arguments for all post\-singularly bounded entire functions, and of using the powerful techniques of symbolic dynamics to study the behaviour of non-escaping points. As mentioned above, this is the reason why the structure of polynomial Julia sets is so well understood. Theorem~\ref{} opens up large classes of entire transcendental functions to the same type of analysis. \subsection*{Existence and landing of periodic hairs} In many interesting cases periodic dreadlocks will in fact be periodic hairs. That is, the connected set $A$ in Theorem~\ref{} will be an arc connecting $\zeta$ to $\infty$. In particular, this will be the case for functions satisyfing the following property, which states that the escaping set consists entirely of hairs. \begin{defn}[{{Criniferous}} functions] We say that an entire function $f$ is \emph{{{criniferous}}} if the following holds for every $z\in I(f)$: For all sufficiently large $n$ there is an arc $\gamma_n$ connecting $f^n(z)$ to $\infty$, in such a way that $\gamma_{n+1} = f(\gamma_n)$ and $\min_{z\in \gamma_n} \lvert z\rvert\to\infty$ as $n\to \infty$. \end{defn} The counterexample from \cite{R3S} mentioned above shows that entire functions, even those with bounded postsingular sets, need not be {{criniferous}}. However, the same article also establishes {{criniferousness}} for a large and natural class of functions, as follows. The \emph{Eremenko-Lyubich class} $\mathcal{B}$ consists of those transcendental entire function for which $S(f)$ is bounded, and hence compact. (If $\P(f)$ is bounded, then $f\in \mathcal{B}$ by definition.) It is proved in \cite{R3S} that $f$ is {{criniferous}} whenever $f\in\mathcal{B}$ and $f$ has \emph{finite order of growth}, i.e., \[ \log \log \lvert f(z)\rvert= O(\log \lvert z \rvert).\] Furthermore, any finite composition of functions with these properties is also {{criniferous}}. To discuss periodic hairs, let us use the following definition from~\cite{Re08}. \begin{defn}[Periodic hairs]\label{} An \emph{invariant hair} of a transcendental entire function $f$ is a continuous and injective curve $\gamma\colon \R\to I(f)$ such that $f(\gamma(t)) = \gamma(t+1)$ for all $t$ and $\lim_{t\to+\infty} \lvert f(t)\rvert = \infty$. A \emph{periodic hair} is a curve that is an invariant hair for some iterate $f^n$ of $f$. Such a hair \emph{lands} if the limit $z_0 = \lim_{t\to -\infty} \gamma(t)$ exists; this limit is called the \emph{landing point} (sometimes also \emph{endpoint}) of the hair $\gamma$. \end{defn} With this terminology, Theorem~\ref{} takes the following form for {{criniferous}} functions. \begin{thm}[Landing theorem for periodic hairs]\label{} Let $f$ be a transcendental entire function such that the postsingular set $\P(f)$ is bounded. Then every periodic ray of $f$ lands at a repelling or parabolic periodic point. If, in addition, $f$ is {{criniferous}}, then conversely every repelling or parabolic periodic point of $p$ is the landing point of at least one and at most finitely many periodic hairs. \end{thm} The first part of the theorem, concerning landing behaviour of periodic rays, is not new. It was proved for exponential maps in~\cite{SZ03}, and later in full generality by the second author~\cite[Corollary~B.4]{Re08}; see also~\cite{De}. The proof uses similar ideas as in the polynomial case, namely expansion properties for the hyperbolic metric, although there are also some additional ingredients. On the other hand, the usual proofs for accessibility of repelling and parabolic periodic points in the polynomial case \cite{ELv,Hu,Pr94} strongly rely on the presence of the open basin of attraction of infinity, and thus break down completely in the transcendental setting. Nonetheless, there has been some previous work in this direction. Under the additional dynamical assumption that $f$ is \emph{geometrically finite}, the theorem was proved by Mihaljevi\'c-Brandt \cite{Mi10}. Furthermore, the first author and Lyubich \cite{BL14} proved Theorem~\ref{} when $f$ belongs to the exponential family~\eqref{}. For exponential maps, boundedness of the postsingular set is a strong dynamical condition (though weaker than geometrical finiteness), as it implies \emph{non-recurrence} of the singular value $a$. However, the non-recurrence property is not used in any essential way in \cite{BL14}, and the ideas used there form one of the ingredients in our proofs of Theorems~\ref{} and~\ref{}. During the preparation of this manuscript, Dierk Schleicher informed us that he has an alternative approach to Theorem~\ref{}, using ideas from~\cite{SZ03}. \subsection*{Hyperbolic sets} As in \cite{BL14}, our results apply not only to repelling (and parabolic) periodic points, but also to \emph{hyperbolic sets}; for polynomials this was first proved in \cite{Pr94}. Recall that a compact, forward-invariant set $K \subset \C$ is called \emph{hyperbolic} if for some $k\in\N$ and $\eta>1$ we have $\lvert (f^k)'(z)\rvert > \eta$ for all $z\in K$. If $\P(f)$ is bounded and $K$ is such a hyperbolic set, then we prove that every point of $K$ is ``accessible'' from the escaping set, via a dreadlock (see Theorem~\ref{}). If $f\in \mathcal{B}_{\operatorname{hairs}}$, these dreadlocks are indeed hairs (arcs in the escaping set connecting a finite endpoint in $K$ to infinity). Again, we can state the following result without requiring the terminology of dreadlocks. \begin{thm}[Landing at hyperbolic sets]\label{} Let $f$ be a transcendental entire function such that $\P(f)$ is bounded, and let $K$ be a hyperbolic set of $f$. Then there is a collection $\mathcal{A}$ of pairwise disjoint, connected and unbounded sets $A\subset I(f)$ with the following properties. \begin{enumerate}[(a)] \item For every $A\in\mathcal{A}$, there is $z_0= z_0(A)\in K$ such that $\overline{A} = A\cup \{z_0(A)\}$, and the function $\mathcal{A}\to K; A\mapsto z_0(A)$ is surjective;\label{} \item $f(A)\in \mathcal{A}$ for all $A\in\mathcal{A}$; \label{} \item for every $\eps>0$, $f^n$ tends to $\infty$ uniformly on $\{z\in \bigcup \mathcal{A}\colon \dist(z, K) \geq \eps\}$;\label{} \item If $z_0(A)$ is periodic of period $p$, then $f^{kp}(A)=A$ for some $k\geq 1$.\label{} \end{enumerate} If $f$ is {{criniferous}}, then every $A\in\mathcal{A}$ is an arc connecting $z_0(A)\in K$ to $\infty$. \end{thm} This generalisation is of particular relevance in the case where $\P(f)$ itself is a hyperbolic set, which is often the case for non-recurrent entire functions (see \cite{RvS}). Hence, in this case, each singular value can itself be connected to infinity by a dreadlock, which in turn allows one to study the Julia set via symbolic dynamics rather closely. {For example, in \cite{Be15}, the existence of a ray landing at the omitted value is exploited to prove strong \emph{rigidity} properties of non-recurrent parameters in this family, extending previous work \cite{Be11} in the postsingularly finite case.} To conclude the introduction, we remark on the case where the postsingular set $\P(f)$ is unbounded. If $f$ is a polynomial, then any unbounded orbit in fact escapes to infinity. For polynomials with escaping singular orbits, the Douady-Hubbard landing theorem no longer holds. Indeed, it is possible that a repelling periodic point is the landing point of uncountably many external rays, none of which are periodic. Compare \cite{LP}. For transcendental entire functions, it is possible for singular orbits to be unbounded without converging to infinity. It is conceivable that, for $f\in\mathcal{B}$ with all singular orbits nonescaping, a version of the landing theorem holds. However, even for exponential maps this is not known (but see \cite{Re06} for a partial result), and it appears that significant new ideas would be required to approach it. See Section~\ref{} for some further discussion. \subsection*{Structure of the paper} {Section~\ref{} gives an overview of expansivity properties for functions in class $\mathcal{B}$ without the assumption of bounded postsingular set. It also defines the concept of external addresses, and gives sufficient conditions on such addresses to be realised by certain unbounded connected sets of points. {Much of this material is} already contained in the literature, e.g.\ in \cite{EL92}, \cite{eremenkoproperty}, \cite{Re08}, \cite{Re09}, but using slightly different formulations and assumptions. Hence we give a systematic and unified account here.} {From Section~\ref{} onward, we restrict to functions with bounded postsingular sets, beginning by discussing hyperbolic expansion estimates for such maps, and introducing the important combinatorial notion of \emph{fundamental tails}. With these preparations, Section~\ref{} introduces dreadlocks for functions with bounded postsingular set, studies their main topological and combinatorial properties, and also shows that the escaping set consists of dreadlocks. This section is a systematic development, investigation and extension of ideas introduced in~\cite{eremenkoproperty}. In Section~\ref{}, we discuss the relation between dreadlocks and hairs.} {Section~\ref{} introduces accumulation sets of dreadlocks at bounded-addresses, and gives different characterisations of when a dreadlock \emph{lands}. This section contains a {crucial} innovation, which is {central} to the proofs of our main theorems: We shall see that, rather than having to contend with the potentially extremely complicated topological structure of dreadlocks, we can instead study their landing properties by considering the accumulation sets of appropriate preimages of a single base point. Having introduced this terminology, we are ready to state our main theorems concerning dreadlocks in Section~\ref{}, and formally deriving Theorems~\ref{} and~\ref{} from these.} {The} three following sections are dedicated to proving the main theorems from this paper. Section~\ref{Periodic dreadlocks land} establishes the landing of periodic dreadlocks, in Section~\ref{ at hyperbolic sets} we show accessibility of hyperbolic sets and repelling periodic orbits, and finally Section~\ref{} is dedicated to the proof of accessibility of parabolic points. We remark that one can take an alternative, less natural but more direct, approach to establishing our theorems, bypassing most of the material in Sections~\ref{} and \ref{}. In particular, this material is not required if one is interested primarily in the case of hairs. Readers interested in such a short-cut are referred to Remark~\ref{}. {To round off the paper, Section~\ref{} discusses bounds on the number of rays landing together at a given point in a hyperbolic set.} {We also include two appendices. The first, Section~\ref{}, gives some details concerning the cyclic order at infinity of unbounded connected sets, which are used in some of our arguments. The second, Section~\ref{}, discusses open questions about landing theorems for entire functions with unbounded postsingular sets.} \subsection*{Notation and preliminaries} We write $\C$ for the complex plane and $\hat{\C}$ for the Riemann sphere. For a set $A\subset\C$ we denote by $\overline{A}$ its closure in $\C$ (occasionally by $\operatorname{cl}(A)$), while we denote by $\hat{A}$ its closure in $\Ch$. The Euclidean disk of radius $R$ around a point $z$ is denoted by $D_R(z)$; the unit disk is $\D\defeq D_1(0)$. If $D$ is any Euclidean disc, we also write $r(D)$ for the radius of $D$. We denote Euclidean distance and diameter by $\dist$ and $\diam$, respectively. If $U\subset\C$ is an open set omitting more than two points, then we denote hyperbolic distance on $U$ by $\dist_U$, and similarly $\diam_U$ for hyperbolic diameter. We also denote the density of the hyperbolic metric of $U$ at a point $z$ by $\rho_U(z)$. That is, the hyperbolic metric is given by $\rho_U(z)\lvert {\rm d} z\rvert$. \subsection*{Acknowledgements} We are extremely grateful to Dave Sixsmith for an extraordinarily careful reading of our manuscript and many suggestions that considerably improved the presentation of the paper. \section{Landing at parabolic points}\label{} We now complete the proof of our analogue of the Douady-Hubbard landing theorem, Theorem~\ref{}, by showing that parabolic periodic points are also accessible by dreadlocks. \begin{thm}[Parabolic points are accessible by dreadlocks]\label{} Let $f\in\mathcal{B}$ with bounded postsingular set, and let $z_0$ be a parabolic periodic point. Then there is a periodic dreadlock of $f$ that lands at $z_0$. \end{thm} Let $f$ be as in the statement of the theorem. By passing to an iterate, we may assume that $f'(z_0)=1$. So $z_0$ is a multiple fixed point of $f$, say of multiplicity $m+1$ for $f$. Then there are $m$ unit vectors $v_1\ldots v_n$, called \emph{repelling directions} at $z_0$. Any backward orbit of $f$ converging to $z_0$ must asymptotically converge to $z_0$ along one of these directions; see \cite[Lemma 10.1]{Mi}. Similarly, there are $n$ attracting directions $w_n$ such that any forward orbit $(f^n(z))_{n=0}^{\infty}$ converging to $z_0$ must converge to $z_0$ along one of these attracting directions $w_n$. Let $U$ be a small simply connected neighborhood of $z_0$ on which $f$ is univalent, and let $\psi\colon f(U)\to U$ be the branch of $f^{-1}$ that fixes $z_0$. A \emph{petal} for an attracting (resp. repelling) direction $w$ (resp. ${v}$) is an open set $P\subset U$ containing $z_0$ on its boundary, such that \begin{enumerate} \item $f(P)\subset P$ (resp. $\psi(P)\subset P$ ); \item an orbit $z\rightarrow f(z)\rightarrow\ldots$ (resp. $z\rightarrow \psi(z)\rightarrow\ldots$) is eventually absorbed by $P$ if and only if it converges to $z_0$ from the direction $w$ (resp. $v$). \end{enumerate} Petals for a given repelling or attracting direction are far from unique. For each repelling direction $v$, we can choose a repelling petal $P_v$ for $v$ which is simply connected, and such that $\overline{\psi(P_v)} \subset P_v\cup\{z_0\}$ and $\psi^n|_{P_v}\to z_0$ uniformly on $P_v$. Similarly, for each attracting direction $w$ we choose a simply connected attracting petal $P_w$ such that $f^n\to z_0$ uniformly on $P_w$. We furthermore require that the union of these $n$ attracting and $n$ repelling petals forms a punctured neighborhood of $z_0$ (see Definition 10.6 and Theorem 10.7 in \cite{Mi} and the subsequent discussion). \begin{defn}[Landing of dreadlocks along a repelling direction] Let $\zeta\in \WW$, let $G_{{\underline{s}}}$ be a periodic dreadlock of $f$, and let $v$ be a repelling direction at $z_0$. We say that $G_{{\underline{s}}}$ \emph{lands at $z_0$ along $v$} if the backwards orbit $(\zeta_n({\underline{s}}))_{n=1}^{\infty}$ converges to $z_0$ along the direction $v$. \end{defn} We remark that it is not difficult to see that this is equivalent to requiring that $V_{n}({\underline{s}})\subset P_v$ for all sufficiently large $n$, where $V$ is as in Lemma~\ref{}. In particular, the definition is independent of the choice of the base point $\zeta$. The following establishes Theorem~\ref{}. \begin{prop}[Accessibility along repelling directions]\label{Parabolic accessibility}\label{} Let $v$ be a repelling direction of $f$ at $z_0$. Then there is at least one periodic dreadlock landing at~$z_0$ along~$v$. \end{prop} \begin{proof} Let $\zeta\in\WW$ and let $V$ be as in Lemma \ref{}, with $\mathcal{F}$ once again the finite collection of fundamental domains whose closure intersects $\overline{D}$. Since $\overline{V}\subset \C\setminus \P(f)$ and $z_0\in\P(f)\subset D$, we may assume that the repelling petals $P_v$ and attracting petals $P_w$ chosen above all have closures disjoint from $V\cup \WW$. Let us define $B_i\defeq \psi^i(P_v)$ for $i\geq 0$. Let $\AA$ be the union of the attracting petals $P_w$. Since the union of attracting and repelling petals is a punctured neighbourhood of $z_0$, all points of $\partial B_0$ that are sufficiently close to $z_0$ must lie in $\AA\cup\{z_0\}$. So $\partial B_0 \setminus (\AA\cup \{z_0\})$ is a compact set disjoint from $\overline{B_1}$, and \[ \eps \defeq \dist\bigl( (\partial B_0)\setminus (\AA \cup \{z_0\}), \overline{B_1} \bigr) > 0.\] Since $B_1$ intersects $J(f)$, there is an $N_1$ such that $f^{-n}(\zeta)\cap B_1\neq \emptyset$ for $n\geq N_1$. In particular, there exists some finite external address of length $n$ such that $\tau_n({\underline{s}})\cap B_1\neq \emptyset$. By Lemma~\ref{}, there is $N\geq N_1$ such that, for all $n\geq N$ and all infinite external addresses ${\underline{s}}$ with $V_n({\underline{s}})\cap \overline{B_0}\neq \emptyset$, $\diam V_n({\underline{s}}) < \eps$ whenever $n\geq N$. Observe that $V_n({\underline{s}})\cap \AA=\emptyset$ by our choice of petals. In particular, if $n\geq N$ and $V_n({\underline{s}})\cap B_1\neq\emptyset$, then $V_n({\underline{s}})\cap \partial B_0 = \emptyset$, and hence $V_n({\underline{s}})\subset B_0$. As in the proof of Theorem~\ref{}, for $n\geq N$ we define $\S_n$ to consist of those finite external addresses of length $n$ for which $\tau_n({\underline{s}})$ intersects $B_{n-N}$. The remainder of the proof then proceeds analogously. \end{proof} In the case that all periodic dreadlocks are hairs (for example, if $f$ is criniferous), our Proposition~\ref{Parabolic accessibility} is a corollary of the Main Theorem in \cite{BF15} (since the hypothesis that periodic rays land is implied by assuming bounded postsingular set), with a completely different proof. We remark that it is plausible that the results of \cite{BF15} can also be extended to non-criniferous functions, using dreadlocks instead of hairs. \begin{proof}[Proof of Theorem~\ref{}] That every periodic dreadlock lands at a repelling or parabolic point was proved in Theorem~\ref{}. Let $z_0$ be a repelling or parabolic point. If $z_0$ is repelling, then the orbit of $z_0$ is a hyperbolic set, and it follows from Theorem~\ref{} that $z_0$ is the landing point of a periodic dreadlock. If $z_0$ is a parabolic point, then this fact follows from Theorem~\ref{}. By Lemma~\ref{ for all} there are only finitely many dreadlocks landing at $z_0$ and they are all periodic of the same period. \end{proof} \section{Dreadlocks landing together at a hyperbolic set}\label{} Recall from Theorem~\ref{} that, for a repelling periodic point $z_0$ of a postsingularly bounded function $f$, the number of dreadlocks landing at $z_0$ is finite. In the polynomial case, this holds also for every point $z_0$ in a hyperbolic set $K$ of $f$. It is plausible that this remains true also in the transcendental entire case. For postsingularly exponential map, the claim is proved in \cite[Proposition~4.5]{BL14}, where it is proved that the number of rays in question is even \emph{uniformly} bounded (depending on $K$). However, the proof uses the fact that postsingularly bounded exponential maps are non-recurrent, and hence the postsingular set is itself a hyperbolic set. Here we shall be content with proving that the number of dreadlocks of a postsingularly bounded function $f$ landing at a given point of a hyperbolic set is (pointwise) finite, in the important special case where $f$ belongs to the \emph{Speiser class}; i.e.,\ the set of singular values $S(f)$ is finite. \begin{thm}[Finitely many dreadlocks landing together]\label{} Let $f$ be a postsingularly bounded entire function with finitely many singular values, and suppose that ${\underline{s}}$ is a bounded external address such that $G_{{\underline{s}}}$ lands at a point $z_0$ that is neither periodic nor preperiodic. Then the number of bounded-address dreadlocks $G_{\tilde{{\underline{s}}}}$ landing at $z_0$ is finite. \end{thm} Recall that, by Lemma~\ref{ for all}, the conclusion also holds when $z_0$ is a (pre-)periodic point that is the landing point of at least one (pre-)periodic dreadlock. Since $f$ is assumed to be postsingularly bounded, it follows by Theorem~\ref{} that the theorem holds for all periodic points $z_0\in J(f)$, with the possible exception of Cremer periodic points and their preimages. \begin{remark} Similarly as discussed in~Remark~\ref{}, the assumptions that $f$ is postsingularly bounded and that $\tilde{{\underline{s}}}$ is a bounded address are made only to ensure that we can speak about dreadlocks and their landing properties, in the context of the definitions made in this theorem. In particular, if $S(f)$ is finite (not necessarily with bounded postsingular set), and ${\underline{s}}^1$ is the bounded address of a hair landing at $z_0$, then the number of hairs landing at $z_0$ is finite, and all of them have bounded periodic addresses. \end{remark} \begin{cor}[Finiteness of ray portraits at hyperbolic sets] Let $f$ be a postsingularly bounded entire function with finitely many singular values. If $K$ is a hyperbolic set for $f$, then every point $z_0\in K$ is the landing point of at least one and at most finitely many dreadlocks at bounded addresses. \end{cor} \begin{proof} If $z_0$ is not periodic, then the claim follows from Theorem~\ref{} and Theorem~\ref{}. Otherwise, it follows from Theorem~\ref{} and Lemma~\ref{ for all}. \end{proof} We now fix a postsingularly bounded entire function $f$ with $\# S(f)<\infty$ for the remainder of the section. The key property that we need to establish in the proof of Theorem~\ref{} is that the addresses of dreadlocks landing at $z_0$ are uniformly bounded, in the sense that they all take their entries from a common finite family of fundamental domains. This is the content of the following lemma. \begin{lem}\label{} Let $\mathcal{F}_1$ be a finite collection of fundamental domains for $f$. Then there exists another finite collection $\mathcal{F}_2\supset \mathcal{F}_1$ of fundamental domains such that the following holds. Suppose that $ {\underline{s}}^1$ takes only entries from $\mathcal{F}_1$ and that $G_{{\underline{s}}^1}$ lands at a point $z_0\in\C$. If ${\underline{s}}^2$ is bounded and $G_{{\underline{s}}^2}$ also lands at $z_0$, then all entries of ${\underline{s}}^2$ belong to $\mathcal{F}_2$. \end{lem} Let us suppose for a moment that the function $f$ is {{criniferous}}. Then the idea of the proof of Lemma~\ref{} can be described as follows. If $G_{{\underline{s}}^1}$ and $G_{{\underline{s}}^2}$ land together at a point $z_0$, the rays $G_{\sigma({\underline{s}}^1)}$ and $G_{\sigma({\underline{s}}^2)}$ also land together at $f(z_0)$, by continuity of $f$. Since we can obtain the ray pair $G_{{\underline{s}}^1}\cup \{z_0\}\cup G_{{\underline{s}}^2}$ by analytic continuation along $G_{\sigma({\underline{s}}^1)}\cup \{f(z_0)\} \cup G_{\sigma({\underline{s}}^2)}$, the first entry of ${\underline{s}}^2$ depends (essentially) only on the first entry of ${\underline{s}}^1$ and the homotopy class of the latter curve in $\C\setminus S(f)$. As different ray pairs are disjoint, there are only finitely many possible such homotopy classes, and the claim follows. In order to make this argument precise in the general case, i.e.\ where the dreadlocks are not necessarily hairs, we should clarify what we mean by ``homotopy classes''. Let us fix the postsingularly bounded function $f$ with finite singular set for the remainder of the section. Let $\Gamma$ be the class of continuous curves $\gamma\colon \R\to \C\setminus S(f)$ that tend to infinity within $\WW$ in both directions. In a slight abuse of notation, we shall say that $\gamma_1$ and $\gamma_2$ are homotopic (relative $S(f)$ and $\delta$) if they are homotopic in $\C\setminus (S(f)\cup \tilde{\delta})$ (relative to their endpoints at infinity) for some infinite piece $\tilde{\delta}$ of the curve $\delta$. Similarly, let $\tilde{\Gamma}$ denote the set of curves connecting a finite endpoint $z_0\in\C$ (possibly belonging $S(f)$) to infinity within $\C\setminus S(f)$, again tending to infinity within $\WW$. Then we similarly define homotopy classes for curves in $\tilde{\Gamma}$ having the same endpoint. With this notation, we can introduce the following convenient notion for bounded-address dreadlocks. \begin{lemdef}[Homotopy of dreadlocks] Let ${\underline{s}}$ be a bounded-address dreadlock, and suppose that $G_{{\underline{s}}}$ lands at a point $z_0\in\C$. Let $\zeta\in\WW$. Then for any $n_0\geq 1$, there is an arc $\gamma_{{\underline{s}}}\in\tilde{\Gamma}$ connecting $z_0$ to $\infty$, parameterised as $\gamma_{{\underline{s}}}\colon \R\to \C$, with the following properties. \begin{enumerate} \item $\gamma(t)\to \infty$ as $t\to+\infty$, and $\gamma(t)\to z_0$ as $t\to-\infty$. \item $\gamma\bigl([-n_0,\infty)\bigr)\subset \tail_{n_0}({\underline{s}})$, and $\gamma\bigl([-n,1-n)\bigr)\subset \tail_n({\underline{s}})$ for $n\geq n_0$. \item $\gamma(-n) = \zeta_n({\underline{s}})$ for $n\geq n_0$. \end{enumerate} The \emph{homotopy class} of $G_{{\underline{s}}}$ is defined to be the homotopy class of $\gamma_{{\underline{s}}}$ relative $S(f)$ and $\delta$, in the above sense. This homotopy class is independent of the choice of $\zeta$ and $\gamma_{{\underline{s}}}$. Furthermore, $f(\gamma_{{\underline{s}}})$ is homotopic to $\gamma_{\sigma({\underline{s}})}$. If finitely many different dreadlocks as above are given (not necessarily all with the same landing point), then the corresponding curves can be chosen to be pairwise disjoint. If two such dreadlocks $G_{{\underline{s}}^1}$ and $G_{{\underline{s}}^2}$ land at the same point $z_0\notin S(f)$, we refer to them as a \emph{dreadlock pair}, and the homotopy class of the dreadlock pair is the homotopy class of the curve from $\infty$ to $\infty$ given by $\gamma_{{\underline{s}}^1} \cup \{z_0\} \cup \gamma_{{\underline{s}}^2}$ (oriented by passing first through $\gamma_{{\underline{s}}^1}$ from $\infty$ to $z_0$), relative $S(f)$ and $\delta$. \end{lemdef} \begin{proof} All claims are easily deduced from the definitions; we leave the details to the reader. \end{proof} \begin{lem}[Finitely many homotopy classes]\label{} There are only finitely many different homotopy classes of dreadlock pairs not landing at singular values. Similarly, for any $z_0\in\C$, there are only finitely many homotopy classes of dreadlocks landing at $z_0$. \end{lem} \begin{proof} The curves representing the homotopy class of two different dreadlock pairs are disjoint, except for the endpoints at infinity, and possibly a single additional point (if the dreadlock pairs land at the same point). Also recall that neither curve self-intersects. It follows that, if both curves wind around the same collection of singular values in positive orientation, and both either surround or do not surround an infinite piece of $\delta$, they represent the same homotopy class. As there are only finitely many singular values, the set of homotopy classes is finite. The second claim follows in the same manner. \end{proof} The following is immediate from the homotopy lifiting property. \begin{obs}[Connecting fundamental domains]\label{} Let $\gamma\in \Gamma$. Suppose that $F$ is a fundamental domain, and let $\tilde{\gamma}\colon (-\infty,\infty)\to\C\setminus f^{-1}(S(f))$ be the unique lift of $\gamma$ under $f$ such that $\tilde{\gamma}(-t)\in F$ for all sufficiently large $t$. Then there is a fundamental domain $\tilde{F}$ such that $\tilde{\gamma}(t)\in\tilde{F}$ for large $t$, and $\tilde{F}$ depends only on $F$ and the homotopy class of $\gamma$ (in the above sense). Similarly, let $\gamma\in\tilde{\Gamma}$ connect a finite point $z_0\in\C$ to $\infty$. If $F$ is a fundamental domain, and $\tilde{\gamma}$ is the lift of $\gamma$ under $f$ that tends to infinity within $F$, then the finite endpoint $w_0$ of $\tilde{\gamma}$ depends only on $F$ and the homotopy class of $\gamma$. \end{obs} \begin{proof}[Proof of Lemma~\ref{}] Let $\tilde{\mathcal{F}}_2$ consist of all domains $\tilde{F}$ as in Observation~\ref{}, where $F$ ranges over the finitely many elements of $\mathcal{F}_1$, and the homotopy class of $\gamma$ ranges over the finitely many homotopy classes of dreadlock pairs of $f$. Now suppose that $G_{{\underline{s}}^1}$ and $G_{{\underline{s}}^2}$ form a dreadlock pair, with $F^1_0\in \mathcal{F}_1$. Let $z_0$ be the common landing point of the two dreadlocks. If $f(z_0)\notin S(f)$, then it follow from Observation\ref{} (applied to the curve $\gamma_{\sigma({\underline{s}}^1)}\cup \{f(z_0)\} \cup \gamma_{\sigma({\underline{s}}^2)}$ that $F^2_0\in \tilde{\mathcal{F}}_2$. On the other hand, suppose that $s=f(z_0)\in S(f)$. Then, by Observation~\ref{}, $z_0$ depends only on the homotopy class of $\gamma_{\sigma({\underline{s}}^1)}$, and the entry $F^1_0$. Hence, for each singular value $s$, there are only finitely many possible preimages $z_0$ that can arise as landing points of dreadlocks whose first entry is in $\mathcal{F}_1$. Consider such $z_0$, and the $\gamma_{\sigma({\underline{s}}^2)}\in\tilde{\Gamma}$ connecting $f(z_0)$ to $\infty$. Then $\gamma$ has $d$ different lifts starting at $z_0$, where $d$ is the local degree of $f$ at $z_0$, tending to infinity within fundamental domains $\tilde{F}_1,\dots,\tilde{F}_d$. This collection of fundamental domains depends only on the homotopy class of $\gamma$ by Observation~\ref{}. In particular, there is a collection $\mathcal{F}(z_0)$ of at most $m\cdot d$ fundamental domains, where $m$ is the (finite) number of homotopy classes of dreadlocks connecting $s$ to $\infty$, such that $F^2_0\in \mathcal{F}(z_0)$ whenever ${\underline{s}}^2$ is as above. Recall that there are only finitely many singular values $s$, and for each of these only finitely many preimages $z_0$ as above. Thus we can add the finitely many sets $\mathcal{F}(z_0)$ to $\tilde{\mathcal{F}}_2$ to obtain a set $\mathcal{F}_2$ with the desired property. \end{proof} \begin{proof}[Proof of Theorem~\ref{}] By passing to a forward iterate, we may assume that the forward orbit of $z_0$ does not contain a critical point. Let $\mathcal{F}_1$ be the set of fundamental domains occurring in ${\underline{s}}$, and let $\mathcal{F}_2$ be the set whose existence is guaranteed by Theorem~\ref{}; say $\mathcal{F}_2 = \{ F^0, F^1 , \dots , F^{m-1}\}$, where we assume that \[ F^0 \prec F^1 \prec \dots \prec F^{m-1} \prec F^0 \] with respect to the cyclical order at infinity. Let $X$ be the set of points on the unit circle $S^1 = \R/\Z$ having an $(m+1)$-ary expansion that contains only the entries $0,\dots,m-1$. Via the $(m+1)$-ary expansion, this set is order-isomorphic to $\{0,\dots,m-1\}^\N$, which in turn is clearly order-isomorphic to $\mathcal{F}_2^{\N}$. Let $\phi\colon \mathcal{F}_2^{\N}\to X$ be this order-isomorphism; then $\phi$ conjugates the shift on $\mathcal{F}_2^{\N}$ to the $(m+1)$-tupling map on $X$. Suppose that $T_0$ is a collection of $p\geq 1$ bounded external addresses that land at $z_0$; we claim that $p\leq m+1$. Indeed, for $j\geq 0$, define $T_j\defeq \sigma^j(T_0)$. Then all dreadlocks at addresses in $T_j$ land at $f^j(z_0)$. Since $z_0$ is not pre-periodic and its orbit does not pass through any critical points, the $T_j$ are pairwise disjoint, and $\sigma\colon T_j\to T_{j+1}$ is an order-preserving bijection for all $j$. Furthermore, the $T_j$ are pairwise \emph{unlinked}. That is, if $j\neq \tilde{j}$, then all elements of $T_j$ lie between the same two adjacent elements of $T_{\tilde{j}}$ with respect to circular order. This means that the set $\phi(T_0)$ is a \emph{wandering $p$-gon} for the $(m+1)$-tupling map on $S^1$. By a result of Kiwi~\cite{Kiwi02}, $p\leq m+1$ as claimed. \end{proof} \begin{remark} It seems likely that one can also directly prove the absence of wandering $d+2$-gons for maps with at most $d$ singular values. (Compare \cite{alhabib-rempe15} for the proof of the case $d=1$, i.e.\ the \emph{no wandering triangles theorem} for exponential maps.) This would imply that the number of dreadlocks in Theorem~\ref{} is always bounded by $d+1$ (assuming that $z_0$ is not pre-critical). \end{remark} \section{Unbounded sets of escaping points}\label{} {In this section, we briefly review basic properties of the dynamics of a function $f\in\mathcal{B}$, and review the definition of external addresses for such maps. Then we state a theorem (Theorem~\ref{}) about the existence of unbounded connected sets for these addresses, and devote the rest of the section to the proof thereof. These sets will provide the basis of the ``dreadlocks'' that are introduced (for postsingularly bounded entire functions) in the next section.} {Throughout this section, fix a function $f\in\mathcal{B}$. Recall that this implies that $S(f)$ is bounded. For now, we do \emph{not} assume that $\PC(f)$ is also bounded. Let us} begin by reviewing the method of partitioning the locus where $f$ is large into (topological) half-strips known as \emph{fundamental domains}. (Compare e.g.\ \cite[Section~2]{Re08} or \cite[Section~2]{guenterthesis}.) For this construction, we fix a Euclidean disk $D$ around the origin containing $S(f)$. The connected components of $f^{-1}(\C\setminus \overline{D})$ are called the \emph{tracts} of~$f$. If $T$ is a tract, then $f\colon T \to \C\setminus \overline{D}$ is a universal covering map; in particular, $T$ is unbounded and simply-connected. In fact (applying the same argument to a slightly smaller disc than $D$), $T$ is a Jordan domain in $\Ch$ whose boundary passes through infinity, and $f$ is a universal covering $f\colon \overline{T} \to \C\setminus D$ on the closure of $T$ (in $\C$). Now let $\delta$ be {an arc} connecting a point of $\overline{D}$ to infinity in the complement of the closure of the tracts. (It is easy to see that such an arc always exists.) We define \begin{equation} \WW \defeq \C\setminus( \overline{D}\cup\delta). \end{equation} Then the connected components of $f^{-1}(\WW)$ are called the \emph{fundamental domains} of $f$; see Figure~\ref{}. \begin{figure} \def\svgwidth{.65\textwidth \begingroup% \makeatletter% \providecommand\color[2][]{% \renewcommand\color[2][]{}% }% \providecommand\transparent[1]{% \renewcommand\transparent[1]{}% }% \providecommand\rotatebox[2]{#2}% \ifx\svgwidth\undefined% \setlength{\unitlength}{705.78776855bp}% \ifx\svgscale\undefined% \relax% \else% \setlength{\unitlength}{\unitlength * \real{\svgscale}}% \fi% \else% \setlength{\unitlength}{\svgwidth}% \fi% \global\let\svgwidth\undefined% \global\let\svgscale\undefined% \makeatother% \begin{picture}(1,0.73807518)% \put(0,0){\includegraphics[width=\unitlength]{TractsLasse2.pdf}}% \put(0.35,0.35){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\small{$D\supset S(f)$}}}% \put(-0.0016161,0.70263163){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\small{$W_0=\C\setminus(\overline{D}\cup\delta)$}}}% \put(0.1213254,0.32438186){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\small{$\delta$}}}% \put(0.13562029,0.0512865){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\small{$T_1$}}}% \put(0.73980271,0.58009572){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\small{$T_2$}}}% \put(0.73860633,0.47002682){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\small{$F_i$}}}% \end{picture}% \endgroup% \caption{\label{ The dynamical plane of a function with two tracts. Fundamental domains obtained by taking preimages of $\delta$ are shown in one of the two tracts.} \end{figure} We remark that there are only finitely many fundamental domains that intersect a given compact set, due to the following simple fact. \begin{lem}[Preimage components intersecting a compact set]\label{} Let $f\colon X \to Y$ be a holomorphic function between Riemann surfaces $X$ and $Y$. Furthermore, let $U\subset Y$ be an open set whose boundary (in $Y$) is locally connected; i.e.\ every point of the boundary of $U$ in $Y$ has arbitrarily small connected relative neighbourhoods in $Y$. Then for any compact set $K\subset X$, only finitely many connected components of $f^{-1}(U)$ intersect $K$. \end{lem} \begin{remark} {The condition that $\partial U$ is locally connected is necessary: Let $X=Y=\C$, $f=\exp$, and let $U$ be a simply-connected domain in the unit disc that spirals in towards the unit circle. (I.e., any branch of the argument on $U$ tends to infinity as $\lvert z\rvert\to 1$ in $U$.) Then infinitely many components of $f^{-1}(U)$ intersect the closed unit disc.} \end{remark} \begin{proof} Let $\mathcal{C}$ be the set of connected components of $f^{-1}(U)$; we claim that every point $z\in f^{-1}(\overline{U})$ has a neighbourhood $V_z$ that intersects only finitely many elements of $\mathcal{C}$. Indeed, $z$ has a neighbourhood $V_1$ that is topologically mapped as by $z\mapsto z^d$, where $d$ is the local degree of $f$ at $z$. Furthermore, the assumption of local connectivity implies that there is a finite collection of connected open sets {$W_1,\dots,W_n\subset f(V_1)\cap U$} such that $\{f(z)\}\cup W_1\cup \dots \cup W_n$ is a neighbourhood of $f(z)$ in $\{f(z)\}\cup U$. Hence if $V_z\subset V_1$ is a sufficiently small disc around $z$, then any point in $f^{-1}(U)\cap V_z$ maps into some $W_j$. As $f^{-1}(W_j)$ has $d$ connected components in $V_1$, and each of these is contained in an element of $\mathcal{C}$, it follows that $V_z$ intersects at most $dn$ elements of $\mathcal{C}$. So the compact set $K\cap f^{-1}(\overline{U})$ has an open cover by sets $V_z$, each of which intersects only finitely many elements of $\mathcal{C}$. The claim follows by taking a finite subcover. \end{proof} In particular, there are only finitely many fundamental domains $F$ whose closure intersects the disc $\overline{D}$. When this does not occur for any $F$, the function $f$ is dynamically particularly simple; more precisely, it is of \emph{disjoint type} (hyperbolic with connected Fatou set). For a detailed study of the topological dynamics of such functions, see~\cite{arclike}. (Compare also the discussion of disjoint-type addresses in Remark~\ref{}.) For any fundamental domain $F$, let $\unbdd{F}$ denote the unbounded connected component of $F\setminus \overline{D}$. \subsection*{Expansion properties and relative cylindrical distance} It is known that functions in $\mathcal{B}$ are expanding near infinity. More precisely, if $f\in\mathcal{B}$, then the \emph{cylindrical derivative} of $f$ is large whenever $f(z)$ is large \cite[Lemma~1]{EL92}. That is, \begin{equation}\label{} \lVert {\rm D} f(z)\rVert_{\operatorname{cyl}} \defeq \left\lvert f'(z) \cdot \frac{z}{f(z)}\right\rvert \to \infty \qquad \text{as}\qquad \lvert f(z)\rvert \to \infty. \end{equation} In view of~\eqref{}, we may assume that the radius $r(D)$ is chosen sufficiently large to ensure that \begin{equation}\label{} \lVert {\rm D} f(z)\rVert_{\operatorname{cyl}} \geq 2 \end{equation} whenever $f(z)\notin D$. Observe that this implies, in particular, that $f(0)\in D$. These assumptions will remain in place for the remainder of the paper. A number of results in the literature are phrased not for the function $f$ directly, but in terms of a \emph{logarithmic transform} $\mathcal{L}$ of $f$. (See e.g.\ \cite{EL92} or \cite{eremenkoproperty}.) Such a transform can be obtained using the change of variable $z = \rho \cdot \exp(\zeta)$, where $\rho=r(D)$ is the radius of $D$. I.e., there is a $2\pi i$-periodic function $\mathcal{L}$ defined by \[ \rho\cdot \exp(\mathcal{L}(\zeta)) = f(\rho\cdot \exp(\zeta)), \] defined whenever the right-hand side (i.e., $f(z)$) belongs to $\C\setminus\overline{D}$. Dynamical properties of $\mathcal{L}$ easily translate to properties of $f$ on the set of points whose orbits remain outside $\overline{D}$ forever. Our assumptions on $D$ imply that the function $\mathcal{L}$ is \emph{normalised} in the sense of \cite{eremenkoproperty,R3S}. We occasionally cite results from other articles that are phrased in this language, but never use the logarithmic transform $\mathcal{L}$ directly in this article. Instead, we use the following terminology, which is inspired by this change of coordinates. See Figure~\ref{}. \begin{defn}[Relative cylindrical distance]\label{} If $z,w\in\WW$, we define the \emph{relative cylindrical distance} $\dist_{\cyl}^{\WW}(z,w)$ to be the shortest cylindrical length of a curve $\gamma$ from $z$ to $w$ that is {homotopic, in $\C\setminus\overline{D}$, to a curve in $\WW$}. Equivalently, if $U$ is a connected component of $\exp^{-1}(\WW)$ and $\zeta$, $\omega$ are the logarithms of $z$ and $w$ that belong to $U$, then \[ \dist_{\operatorname{cyl}}^{\WW}(z,w) = \lvert \zeta - \omega \rvert. \] We similarly define the distance between two subsets of $\WW$, and the diameter $\diam_{\cyl}^{\WW}$ with respect to this metric. \end{defn} \begin{figure} \subfloat[Definition~\ref{}] \fbox \def\svgwidth{.47\textwidth \input{W0-illustration.pdf_tex}\label{ }}\hfill \subfloat[Lemma~\ref{}] \fbox \def\svgwidth{.47\textwidth \input{W0-crosscuts.pdf_tex}\label{ }} \caption{\label{fig:W0 In~\protect\subref{}, the cylindrical distance between $z$ and $w$ is less than $\pi$. However, any curve connecting $z$ and $w$ in $\WW$ takes more than two full turns around $D$, and hence the relative distance in the sense of Definition~\ref{} is greater than $4\pi$. The second image, in~\protect\subref{} illustrates Lemma~\ref{} in this case; here $\Gamma$ is the union of three cross-cuts of $\WW$; the domain $\WW^{\Gamma}$ is shown shaded in grey.} \end{figure} Since $f$ is expanding with respect to the cylindrical metric on $\WW$ by~\eqref{}, we have \begin{equation}\label{} \dist_{\cyl}^{\WW}(f(z),f(w)) \geq 2\cdot \dist_{\cyl}^{\WW}(z,w) \end{equation} whenever $z$ and $w$ both belong to $\unbdd{F}$ for some fundamental domain $F$. Let us also note the following fact for future reference. \begin{obs}\label{} Let $\zeta_0\in \WW$. Then $\lvert z\rvert \to \infty$ in $\WW$ if and only if $\dist_{\cyl}^{\WW}(z,\zeta_0)\to\infty$. In other words, if $(z_n)_{n=0}^{\infty}$ is a sequence of points in $\WW$, then $\lvert z_n\rvert \to\infty$ if and only if $\dist_{\cyl}^{\WW}(z_n, \zeta_0)\to\infty$. \end{obs} \subsection*{External addresses and symbolic dynamics} The reason for introducing fundamental domains is that they can be used to assign symbolic dynamics to points whose orbit stays sufficiently large, and in particular to escaping points. \begin{defn}[External addresses] Let $f\in\mathcal{B}$, and let fundamental domains be defined as above. An \emph{(infinite) external address} is a sequence ${\underline{s}}= F_0 F_1 F_2 \dots$ of fundamental domains of $f$. The address ${\underline{s}}$ is \emph{bounded} if the set of fundamental domains occurring in ${\underline{s}}$ is finite; it is \emph{periodic} if there is $k$ such that $F_{n+k}=F_n$ for all $n\geq 0$. Let ${\underline{s}}=F_0 F_1 F_2\ldots$ be an external address, and recall that $\unbdd{F}_n$ denotes the unbounded connected component of $F_n\setminus\overline{D}$. Then we define \[J^0_{{\underline{s}}}(f):=\{z\in\C\colon f^n(z)\in\unbdd{F_n} \text{ for all $n\geq 0$}\}. \] \end{defn} \begin{remark} For the purpose of this paper, we shall often use ``address'' synonymously with ``external address''. \end{remark} The main goal of this section is to prove the following. \begin{thm}[Realisation of addresses]\label{} Let ${\underline{s}}$ be an external address. \begin{enumerate}[(a)] \item Suppose that $J^0_{{\underline{s}}}$ contains some point $z_0$. Then $J^0_{{\underline{s}}}$ also contains a {closed} unbounded connected set $X$ on which the iterates of $f$ tend to infinity uniformly. Moreover, $\dist_{\cyl}^{\WW}(z_0, X)\leq 2\pi$. \label{} \item If $X_1$ and $X_2$ are unbounded, closed, connected subsets of $J^0_{{\underline{s}}}$ with $X_1\not\subset X_2$, then $X_2\subset X_1$ and $f^n|_{X_2}\to\infty$ uniformly. \label{} \item If ${\underline{s}}$ is bounded, then $J^0_{{\underline{s}}}\neq\emptyset$. {Furthermore, there exists $R>r(D)$, depending on the finite collection of fundamental domains occurring in ${\underline{s}}$ but not otherwise on ${\underline{s}}$, such that the set $X$ in~\ref{} can be chosen to contain a point of modulus $R$.}\label{} \item\label{ Conversely, if $\mathcal{F}$ is a finite collection of fundamental domains, then there is $R>0$ such that the iterates of $f$ tend to infinity uniformly on the closed set \begin{equation}\label{} \bigcup_{{\underline{s}}} J_{{\underline{s}}}^0\setminus D_R(0) = \left\{z\in \C\colon \lvert z \rvert \geq R \text{ and } f^n(z)\in {\bigcup_{F\in\mathcal{F}} \overline{F}} \text{ for all $n\geq 0$} \right\}. \end{equation} \end{enumerate} \end{thm} This collection of results is not entirely new. Claims~\ref{} and~\ref{} are essentially Proposition~3.2 and Corollary~3.4 of \cite{eremenkoproperty}. Part~\ref{} is \cite[Lemma~2.1]{Re08}. Claim~\ref{} is proved in \cite[Theorem~2.4]{Re08} for fixed addresses, and the proof extends directly to the case of arbitrary bounded addresses. Alternatively, the fact that $J_{{\underline{s}}}\neq\emptyset$ for bounded addresses ${\underline{s}}$ (and even for certain unbounded ${\underline{s}}$, see Proposition~\ref{} below) is proved for disjoint-type functions in \cite[Corollary~B' on p.~405]{BK07}. Compare also~\cite[Proposition~2.10]{arclike}. The claim then follows for general functions in the class $\mathcal{B}$ by the results of~\cite{Re09}. Since the papers in question all use slightly different notation, we give a self-contained and unified account of Theorem~\ref{} in the remainder of the section. We begin with a simple property of the set $J_{{\underline{s}}}^0$, which is essentially \cite[Theorem~1]{EL92}. \begin{lem}[One-dimensionality of $J_{{\underline{s}}}^0$]\label{lem:J} Let ${\underline{s}}=F_0 F_1\ldots$ be an external address. Then~$\overline{J_{{\underline{s}}}^0}$ is a subset of $J(f)$, has empty interior and does not separate the plane. \end{lem} \begin{proof} For each $n\geq 0$, $U_n\defeq \C\setminus \operatorname{cl}(\unbdd{F}_n)$ is connected. Suppose, by contradiction, that $V_0$ was a connected component of $\C\setminus \partial J_{{\underline{s}}}^0$ that does not contain $U_0$. Then $V_0\subset \unbdd{F}_0$, and $f\colon V_0 \to f(V_0)$ is a conformal isomorphism. It follows inductively that $f^n(V_0)$ is contained in a connected component of $\C\setminus \partial f^n( J_{{\underline{s}}}^0)$ that does not contain $U_n$, and hence $f^n(V_0)\subset \unbdd{F}_n$ for all $n\geq 0$. On the other hand, fix $z_0\in V_0$, and set $z_n\defeq f^n(z_0)$. {It follows from the above that $f^n\colon V_0\to f^n(V_0)$ is univalent for all $n$. By~\eqref{} and the definition of the cylindrical derivative, we have $\lvert (f^n)'(z_0) \rvert / \lvert f^n(z_0)\rvert \to\infty$ as $n\to\infty$. By Koebe's $1/4$-theorem it follows that $0\in f^n(V_0)$ for sufficiently large $n$. This is a contradiction and proves that $\overline{J_{{\underline{s}}}^0}$ has empty interior and does not separate the plane.} Let $z\in \overline{J_{{\underline{s}}}^0}$. Suppose first that $\dist(f^n(z),D)\to\infty$ as $n\geq 0$. If $z$ belonged to the Fatou set of $f$, then it would follow from equicontinuity that there is a neighbourhood $V$ of $z$ such that $f^n(V)\cap\overline{D}=\emptyset$ for all $n\geq 0$. But then $V\subset J_{{\underline{s}}}^0$, and this contradicts the result we have just proved. So $z\in J(f)$. Otherwise, there is a sequence $n_k$ such that $f^{n_k}(z)\not\to\infty$. Then the spherical derivative of $f^{n_k}$ at $z$ is comparable to the corresponding cylindrical derivative. By~\eqref{}, the latter tends to infinity as $k\to\infty$. Thus the family of iterates of $f$ is not normal at $z$ by Marty's theorem, and again $z\in J(f)$. \end{proof} \subsection*{A separation lemma} We now formulate a key technical lemma, which is crucial for our unified proof of Theorem~\ref{}. It is in essence a combination and reformulation of Lemmas~3.1 and~3.3 of \cite{eremenkoproperty}; we give the proof for completeness. \begin{lem}[Cross-cuts of $\WW$]\label{} Let $z_0\in\WW$. Then there exists a union $\Gamma\ni z_0$ of cross-cuts of $\WW$ such that $\dist_{\cyl}^{\WW}(z_0,\zeta)\leq 2\pi$ for all $\zeta\in\Gamma$ and such that the unbounded connected component ${\WW}^{\Gamma}$ of $\WW\setminus \Gamma$ has the following property. If $R>0$ such that $z_0$ belongs to the unbounded connected component of $\WW\setminus \overline{D_R(0)}$, then $\WW^{\Gamma}$ is also disjoint from $\overline{D_R(0)}$. Here $\Gamma$ can be chosen to consist only of arcs of the circle of radius $\lvert z_0\rvert$ centred at $0$. Moreover, suppose that $A\subset \WW$ is any unbounded connected set with $z_0\in A$. Then, for all $z\in {\WW^{\Gamma}}$, $\dist_{\cyl}^{\WW}(z,A) \leq 2\pi$. \end{lem} \begin{remark}[Remark~1] It may appear at first as though the statement is obvious, since any round circle centred at $0$ has cylindrical length $2\pi$. However, the diameter of the intersection of such a circle with $\WW$ may be arbitrarily large when measured with respect to $\dist_{\cyl}^{\WW}$. See~Figure~\ref{fig:W0}. \end{remark} \begin{remark}[Remark~2] {Clearly $\delta$ can be chosen to be piecewise analytic, in which case the number of cross-cuts in $\Gamma$ is necessarily finite. However, we do not require this.} \end{remark} \begin{proof} Let $U$ be a connected component of $\exp^{-1}(\WW)$; then $\exp\colon U\to \WW$ is a conformal isomorphism. Let $\zeta_0$ be the unique point of $\exp^{-1}(z_0)$ in $U$. Let $I$ be the vertical segment $\zeta_0 + i\cdot [-2\pi , 2\pi]$. Observe that the endpoints of $I$ are elements of $\exp^{-1}(z_0)$ different from $\zeta_0$, and hence do not belong to $U$. We set $X \defeq I\cap U$; then $X$ is a collection of cross-cuts of $U$. Set $\Gamma \defeq \exp(X)$; we will prove the claims of the lemma by considering the unbounded connected component $V_1$ of $U\setminus I$. Since $\exp(V_1)= \WW^{\Gamma}$, we must show the following: \begin{enumerate}[(a)] \item {$\re \zeta > \log R$ for all $\zeta\in V_1$, for $R$ as in the statement of the lemma} \label{} \item if $A\subset U$ is an unbounded connected set with $\zeta_0\in A$, then $\dist(z, A)\leq 2\pi$ for all $z\in U$.\label{} \end{enumerate} We prove~\ref{} first. Set $A^+ \defeq \overline{A} + 2\pi i$ and $A^- \defeq \overline{A} - 2\pi i$. Then $A^-,A^+\subset\C\setminus U$, and both $A^-$ and $A^+$ intersect $I$. Let $V_2$ be the connected component of $\C\setminus (A^+\cup A^- \cup I)$ containing $V_1$. Since $A$ tends to infinity within $V_1$, it follows easily that for all $\zeta \in V_2$, the segment $\zeta + i(0, 4\pi]$ meets $A^+$ (and similarly $\zeta - i(0,4\pi]$ intersects $A^-$). It follows that $\zeta + i[-2\pi , 2\pi]$ intersects $\overline{A}$, and in particular~\ref{} holds as required. Furthermore, by assumption, $\zeta_0$ can be connected to infinity by a curve $\gamma_0$ not containing any points at real parts at most $\log R$. Letting $A=\gamma_0$ be such a curve, we conclude that every point $\zeta \in V_2$ has the same real part as some point in $A$, and hence~\ref{} follows. \end{proof} \subsection*{Existence of unbounded sets of escaping points} With these preparations, we are now able to prove the existence of unbounded connected subsets of $J_{{\underline{s}}}^0$ under very general hypotheses. \begin{thm}[Unbounded subsets of $J_{{\underline{s}}}^0$]\label{} Let ${\underline{s}} = F_0 F_1 F_2 \dots$ be an external address. Suppose that $(z_n)_{n=0}^{\infty}$ is a sequence of points such that each $z_n\in \unbdd{F}_n$ for all $n \geq 0$. Suppose furthermore that there is $C>0$ such that $\dist_{\cyl}^{\WW}(z_n, f|_{F_n}^{-1}(z_{n+1}))\leq C$ for all $n\geq 0$ for which $(f|_{F_n})^{-1}(z_{n+1})\notin\overline{D}$. Then there is a closed, unbounded and connected set $X\subset J_{{\underline{s}}}^0$ such that {$\dist_{\cyl}^{\WW}(z_0, X)\leq 2\max(2\pi, C)$}. \end{thm} \begin{proof} For $n\geq 0$, let $\Gamma_n$ be the union of cross-cuts from Lemma~\ref{}, applied to $z_n$. Let $A_n^k$ be a sequence of unbounded closed connected sets, defined for $n, k \geq 0$ as follows. Let $A_n^0$ be the closure of the unbounded connected component of $F_n\setminus \Gamma_n$. For $k\geq 0$, let $A_{n}^{k+1}$ be the closure of the unbounded connected component of $(f|_{F_n})^{-1}(A_{n+1}^k)\setminus \Gamma_n$. Observe that $A_n^{k+1}\subset A_n^k$ for all $n$ and $k$. \begin{claim} {$\dist_{\cyl}^{\WW}(A_n^k , z_n) \leq 2\max(2\pi, C)$ for all $n,k\geq 0$.} \end{claim} \begin{subproof} This is true by definition for $k=0$. Suppose that $k\geq 0$ is such that the claim holds for all $n\geq 0$. Let $n\geq 0$, and set $B\defeq (f|_{F_n})^{-1}(A_{n+1}^k)$. If $B\cap \Gamma_n\neq\emptyset$, then $A_n^{k+1}$ intersects $\Gamma_n$ and the claim is immediate from the properties of $\Gamma_n$. Otherwise, $A_n^{k+1}=B$. {Suppose first that $(f|_{F_n})^{-1}(z_{n+1}))\notin\overline{D}$. Then \[ \dist_{\cyl}^{\WW}( A_n^{k+1} , z_n) \leq \dist_{\cyl}^{\WW}( A_n^{k+1} , (f|_{F_n})^{-1}(z_{n+1})) + C \leq \max(\pi , C ) + C \leq 2\max( \pi , C) \] by~\eqref{}. Now suppose that $(f|_{F_n})^{-1}(z_{n+1}))\in\overline{D}$. By the inductive hypothesis, and since $\Gamma_n$ separates $\partial D$ from $B$, we can connect $B$ to $\Gamma_n$ by a curve $\gamma$ with $\diam_{\cyl}^{\WW}(\gamma)\leq \max(2\pi,C)$. Since $\diam_{\cyl}^{\WW}(\Gamma_n)\leq 2\pi$, the claim follows.} \end{subproof} Now set $X_1 \defeq \bigcap_{k\geq 0} A_0^k$. Then $X_1 \cup \{\infty\}$ is compact and connected as a countable intersection of compact connected sets. Moreover, $X_1$ contains a point $\zeta$ with $\dist_{\cyl}^{\WW}( \zeta , z_0)\leq 2\max(\pi , C)$. If $X$ is the connected component of $X_1$ containing $\zeta$, then $X$ is unbounded by the boundary bumping theorem \cite[Theorem~5.6]{continuumtheory}, and the proof is complete. \end{proof} In particular, we obtain the following partial results towards Theorem~\ref{}. \begin{cor}[Realised addresses have unbounded sets]\label{} Suppose that ${\underline{s}}$ is an external address, and that there is a point $z_0\in J_{{\underline{s}}}^0$. Then $J^0_{{\underline{s}}}$ contains an unbounded closed connected set $X$, and $\dist_{\cyl}^{\WW}(z_0, X)\leq 2\pi$. \end{cor} \begin{proof} Set $z_n \defeq f^n(z_0)$, and apply Theorem~\ref{}. \end{proof} \begin{cor}[Bounded addresses are realised]\label{} Suppose that $\mathcal{F}$ is a finite collection of fundamental domains. Then there is $R>0$ with the following property. If ${\underline{s}}$ is an external address whose entries are all in $\mathcal{F}$, then $J^0_{{\underline{s}}}$ contains an unbounded connected set $X$ which contains a point of modulus at most $R$. \end{cor} \begin{proof} Pick a base-point $\zeta_0\in\WW$. For $F\in\mathcal{F}$, let $\zeta_F$ be the preimage of $\zeta_0$ in $F$. Then there is a constant $C$ such that $\dist_{\cyl}^{\WW}(\zeta_0, \zeta_F)\leq C$ for all $F\in\mathcal{F}$ (when defined). If ${\underline{s}}$ is an external address whose entries are all in $\mathcal{F}$, we can set $z_n\defeq \zeta_0$ for all $n\geq 0$, and apply Theorem~\ref{}. We obtain an unbounded connected set $X$ with $\dist_{\cyl}^{\WW}(\zeta_0, X)\leq 2\max(\pi,C)$. If $R$ is sufficiently large (depending only on $C$), then $X$ contains a point of modulus at most $R$. \end{proof} \subsection*{Uniform escape to infinity} To complete the proof of Theorem~\ref{}, we consider the question of uniform escape to infinity on an unbounded connected subset of $J^0_{{\underline{s}}}$. \begin{lem}[Uniform escape on unbounded connected sets]\label{} Let ${\underline{s}} = F_0 F_1 F_2 \dots$ be an external address, and suppose that $X\subset J^0_{{\underline{s}}}$ is unbounded and connected. Furthermore, assume that there is a sequence $(z_n)_{n=0}^{\infty}$ such that $z_n\in \unbdd{F}_n$, and such that $\dist_{\cyl}^{\WW}(z_n, f^n(X))\to\infty$. Then $f^n|_{\overline{X}}\to\infty$ uniformly. \end{lem} \begin{proof} Let $\zeta_0\in \WW$ be any base point. We may assume that $\dist_{\cyl}^{\WW}(z_n, \zeta_0)\to\infty$. Indeed, set $\eta_n\defeq \dist_{\cyl}^{\WW}(z_n, f^n(X))$ and let $(\xi_n)_{n=0}^{\infty}$ be any sequence in $\WW$ with \[ \dist_{\cyl}^{\WW}(\xi_m , \zeta_0)\to \infty \qquad\text{and}\qquad \dist_{\cyl}^{\WW}( \xi_m , \zeta_0) \leq \eta_n/3. \] Now define \[ \tilde{z}_n \defeq \begin{cases} z_n &\text{if } \dist_{\cyl}^{\WW}(z_n, \zeta_0) > \eta_n/3 \\ \xi_n &\text{otherwise}.\end{cases}\] Clearly $\dist(\tilde{z}_n, \zeta_0)\to\infty$ and $\dist_{\cyl}^{\WW}(\tilde{z}_n, f^n(X)) \geq \eta_n/3\to\infty$, as desired. Let $\Gamma_n$ be the union of cross-cuts associated to $z_n$ by Lemma~\ref{}. Then, for sufficiently large $n$, $f^n(\overline{X})$ is disjoint from $\Gamma_n$, and hence belongs to $\WW^{\Gamma_n}$. Since $\lvert z_n\rvert \to\infty$ by Observation~\ref{}, it follows that $f^n(\overline{X})\to\infty$ uniformly, as required. \end{proof} \begin{proof}[Proof of Theorem~\ref{}] We first prove~\ref{}, so let $X_1, X_2\subset J_{{\underline{s}}}^0$ be closed, unbounded and connected with $X_1\not\subset X_2$. Let $z_0\in X_1\setminus X_2$ and set $z_n\defeq f^n(z_0)$. Then $\dist(z_n , f^n(X_2))\to\infty$ by~\eqref{}. In particular, $f^n|_{X_2}\to\infty$ uniformly by Lemma~\ref{}. Let $n_0$ be sufficiently large that $\dist_{\cyl}^{\WW}(z_n, f^n(X_2)) > 2\pi$ for $n\geq n_0$. Let $\Gamma_n$ be the union of crosscuts associated to $z_n$ by Lemma~\ref{}. Then $f^n(X_2)\subset \WW^{\Gamma_n}$ for $n\geq n_0$. Since $f^n(X_1)$ connects $z_n$ to $\infty$, we have $\dist(f^n(X_1), f^n(\zeta))\leq 2\pi$ for all $\zeta\in X_2$ and all $n\geq 0$. Again by~\eqref{}, we have $\dist(X_1, \zeta)=0$, and hence $\zeta\in X_1$, as required. Now let us prove~\ref{}, so suppose that $z_0\in J_{{\underline{s}}}^0$. By Corollary~\ref{}, there is an unbounded closed connected set $X\subset J_{{\underline{s}}}^0$ with $\dist_{\cyl}^{\WW}(X,z_0)\leq 2\pi$. We may assume that $z_0\notin X$; indeed, otherwise we may replace $X$ by a connected component of $X\setminus D_{\eps}(z_0)$, where $\eps$ is sufficiently small. Now set $z_n\defeq f^n(z_0)$. By~\eqref{}, we have $\dist_{\cyl}^{\WW}(f^n(X), z_n)\to\infty$, and hence it follows from Lemma~\ref{} that $f^n|_X\to\infty$ uniformly. This completes the proof of~\ref{} of Theorem~\ref{}. Part~\ref{} follows directly from Corollary~\ref{}. To prove~\ref{}, {observe first that equality of the sets in~\eqref{} holds as soon as $R$ is sufficiently large. Indeed, suppose that $F,F'\in\mathcal{F}$ and that $z\in\overline{F}$ maps to some point in $\overline{F'}$ of modulus at least $R$. Since $f(z)\in \WW$, we must in fact have $z\in F$, and additionally $z\in\unbdd{F}$ if $R$ is sufficiently large.} Now let $\zeta_0$ and $(\zeta_F)_{F\in\mathcal{F}}$ be defined as in the proof of Corollary~\ref{}. Let $R$ be sufficiently large such that any point $z\in\WW$ of modulus at least $R$ has $\dist_{\cyl}^{\WW}(\zeta_0, z)\geq 3\max(C,2\pi)$. {Let $z$ be a point whose orbit is contained in $\bigcup_{F\in\mathcal{F}} \overline{F}$, and furthermore $\lvert z\rvert \geq R$. Then} it follows that \[ \dist_{\cyl}^{\WW}(\zeta_0, f(z)) \geq 2\cdot \dist_{\cyl}^{\WW}(\zeta_F, z) \geq 2\cdot ( \max(C,2\pi) + \dist_{\cyl}^{\WW}(\zeta_0 , z)) \geq \frac{4}{3} \dist_{\cyl}^{\WW}(\zeta_0, z). \] It follows by induction that $\dist_{\cyl}^{\WW}(\zeta_0, f^n(z))\to\infty$ uniformly in $n$, and the claim follows. \end{proof} \subsection*{Exponentially bounded addresses} As noted above, the results on bounded addresses can be generalised to certain \emph{unbounded} addresses. While we do not require this fact for this paper, we shall record it for future reference. \begin{defn}[Exponentially bounded addresses]\label{} Let $\zeta_0\in \WW$ be an arbitrary base point. For any fundamental domain $F$, let $\zeta_F$ be the unique preimage of $\zeta_0$ in $F$. We say that an infinite external address ${\underline{s}}=F_0F_1\ldots$ is \emph{exponentially bounded} if there exists a positive real number $T$ with the following property. For all $n\geq 0$, if $\zeta_{F_n}\in \unbdd{F}_n$, then \begin{equation} \dist_{\cyl}^{\WW}(\zeta_0 , \zeta_{F_n}) \leq \exp^n(T). \end{equation} \end{defn} \begin{remark} If $\xi$ is another base-point, then it follows from~\ref{} that $\dist_{\cyl}^{\WW}( \zeta_F , \xi_F)$ is uniformly bounded (where defined) for all fundamental domains $F$. Thus it follows that the definition of exponentially bounded addresses is independent of the choice of base point $\zeta_0$. Exponentially bounded addresses were defined previously for exponential maps \cite{expescaping} and cosine maps \cite{cosescaping}; it is easy to see that in these cases the definition agrees with ours. For these families, the class of exponentially bounded addresses ${\underline{s}}$ agrees precisely with those for which $J_{{\underline{s}}}^0\neq\emptyset$. This is no longer true for general $f$, even when $f$ has finite order of growth; see~\cite{simonannalasse}. In \cite[Corollary~B']{BK07}, it is shown that $J_{{\underline{s}}}^0\neq\emptyset$ for a certain class of addresses, namely those whose orbits remain within finitely many tracts, and such that the ``index'' of the corresponding fundamental domains within each tract does not grow faster than an iterated exponential. It is easy to see that such addresses are exponentially bounded in our sense, but the converse is not the case. (For example, our definition allows for addresses taking values in fundamental domains that lie in infinitely many different tracts.) \end{remark} We now show that~\ref{} and~\ref{} of Theorem~\ref{} can be extended to exponentially bounded addresses as follows. \begin{prop}[Exponentially bounded addresses are realised]\label{} If ${\underline{s}}$ is exponentially bounded, then $J_{{\underline{s}}}^0\neq\emptyset$. More precisely, fix a base point $\zeta_0\in \WW$ and $T>0$, and denote by $\S(\zeta_0,T)$ the set of all addresses which are exponentially bounded for this choice of $\zeta_0$ and $T$. Then there is a number $R>0$, depending only on $\zeta_0$ and on $T$, with the following property. If ${\underline{s}}\in\S(\zeta_0,T)$, then $J_{{\underline{s}}}^0$ contains an unbounded connected set $X$ on which the iterates tend to infinity uniformly, and which contains a point of modulus $R$. Conversely, the iterates of $f$ tend to infinity uniformly on \begin{equation}\label{} \bigcup_{{\underline{s}}\in\S(\zeta_0,T)} J_{{\underline{s}}}^0\setminus D_R(0). \end{equation} \end{prop} \begin{proof} We shall use an expansion estimate \cite[Lemma~3.1]{R3S}, which is stronger than~\eqref{} at large distances. (Compare also~\cite[Lemma~3.3]{BK07}.) We will use this estimate in the following form, which follows easily from the version stated in \cite{R3S}: There are constants $C_1,C_2>0$ with the following property. If $F$ is a fundamental domain and $\zeta_1,\zeta_2\in \unbdd{F}$ with $\dist_{\cyl}^{\WW}(\zeta_1,\zeta_2)\geq C_1$, then \begin{equation}\label{} \dist_{\cyl}^{\WW}(f(\zeta_1), f(\zeta_2)) \geq \exp(C_2 \cdot \dist_{\cyl}^{\WW}(\zeta_1, \zeta_2)). \end{equation} Now fix $\zeta_0$ and $T>0$, and write $\mathcal{S}\defeq \mathcal{S}(\zeta_0,T)$. We may assume without loss of generality that $\zeta_0$ is chosen sufficiently large to ensure that $\zeta_F\in \unbdd{F}$ for every fundamental domain $F$. (To this end, we may need to increase $T$, but only by a finite amount according to the remark following Definition~\ref{}.) Define $E(t)\defeq \exp(C_2 \cdot t)$. By basic properties of exponential growth, there is $\tilde{R}>0$ such that the following hold for all $x\geq \tilde{R}/3$. \begin{align} \label{ &> 3E(x/2) > E(x/2) > 2x, \qquad\text{and} \\ \label{ &\geq E^n(\tilde{R}/3) > \exp^n(T) + 2\pi. \end{align} We choose $R$ sufficiently large to ensure that $\dist_{\cyl}^{\WW}(z,\zeta_0) > \tilde{R}$ whenever $\lvert z \rvert \geq R$. Suppose that ${\underline{s}}\in\mathcal{S}$, and let $z\in J_{{\underline{s}}}^0$ such that $t\defeq \dist_{\cyl}^{\WW}(z,\zeta_0) \geq \tilde{R}$. We claim that \begin{equation}\label{} \dist_{\cyl}^{\WW}( f^n(z) , \zeta_0) \geq 3E^n(t/3). \end{equation} Indeed, this is trivial for $n=0$, and if the claim holds for $n$, then \[ \dist_{\cyl}^{\WW}(f^n(z) , \zeta_{F_n}) \geq \dist_{\cyl}^{\WW}(f^n(z) , \zeta_0) - \exp^n(T) \geq 2E^n(t/3). \] Hence, by~\eqref{} and~\ref{}, \[ \dist_{\cyl}^{\WW}(f^{n+1}(z), \zeta_0) = \dist_{\cyl}^{\WW}(f^{n+1}(z), f(\zeta_{F_n})) \geq E(2E^n(t/3)) \geq 3E^{n+1}(t/3). \] The claim follows by induction. In particular, \[ \dist_{\cyl}^{\WW}(f^n(z),\zeta_0) \geq 3E^n(\tilde{R}/3) \geq \exp^n(T) \] Since this applies to all points in the union~\eqref{}, the second claim of the proposition follows. Now let us prove the first. Set $t\defeq \tilde{R}/3$. For $n\geq 0$, choose $z_n\in \unbdd{F}_n$ with $\dist_{\cyl}^{\WW}(z_n , \zeta_0) = E^n(t)$. Such a point exists because $\dist(\zeta_{F_n} , \zeta_0) < E^n(t)$, and $\zeta_{F_n}$ can be connected to infinity within $\unbdd{F}_n$. For $n\geq 0$, let $\Gamma_n$ be the union of cross-cuts from Lemma~\ref{}, applied to $z_n$, and let $A_n^k$, for $n,k\geq 0$, be defined precisely as in the proof of Theorem~\ref{}. \begin{claim} $\dist_{\cyl}^{\WW}(A_n^k , \zeta_0) \leq 3E^n(t)$ for all $n,k\geq 0$. \end{claim} \begin{subproof} If $A_n^k\cap \Gamma_n \neq \emptyset$, then the claim follows by choice of $z_n$. In particular, this is always the case for $k=0$. So suppose that $k\geq 0$ is such that the claim holds for all $n\geq 0$, and that $n$ is such that $A_n^{k+1}\cap \Gamma_n = \emptyset$. Then, by definition, $f(A_n^{k+1}) = A_{n+1}^k$. By the inductive hypothesis, the latter contains a point $w$ with $\dist_{\cyl}^{\WW}(w, \zeta_0) \leq 3E^{n+1}(t)$. Set $\tilde{w} \defeq (f|_{F_n})^{-1}(w) \in A_{n+1}^k$. Then, by~\eqref{} and~\eqref{}, \[ \dist_{\cyl}^{\WW}( \tilde{w} , \zeta_{F_n}) \leq E^{-1}(3 E^{n+1}(t)) \leq 2E^n(t). \] So \[ \dist_{\cyl}^{\WW}(\tilde{w} , \zeta_0) \leq \dist_{\cyl}^{\WW}(\tilde{w} , \zeta_{F_n}) + \dist_{\cyl}^{\WW}(\zeta_{F_n} , \zeta_0) \leq 2E^n(t) + \exp^n(T) \leq 3E^n(t) \] by~\eqref{}, as claimed. \end{subproof} In particular, the set $\bigcap_{k\geq 0} A_0^k$ has a connected component $X$ containing a point with $\dist(A_n^k , \zeta_0) \leq 3t = \tilde{R}$, and hence of modulus at most $R$. Since $X$ is connected and unbounded, it also contains a point of modulus exactly $R$. The proof is complete. \end{proof} \section{Hyperbolic expansion and fundamental tails}\label{} For the remainder of the article, we shall specialise to the case where our transcendental entire function $f$ has bounded postsingular set. This section collects some fundamental preliminary material concerning the functions. We begin by noting a global expansion property away from the postsingular set, and then proceed to introduce the notation of \emph{fundamental tails}, which will later be used to study the structure of the set of escaping points. \begin{prop}[Hyperbolic expansion]\label{} For every transcendental entire function $f$, $\# \P(f) \geq 2$. Now suppose that $\P(f)$ is bounded, let $\P$ be a compact forward-invariant set with $\P(f)\subset \P$, and let $\Omega$ be the unbounded connected component of $\C\setminus\P$. If $z\in V\defeq f^{-1}(\Omega)\subset \Omega$, then $f$ is strictly expanding at $z$ in the hyperbolic metric of $\Omega$. Moreover, this expansion factor tends to infinity as $z\to\infty$ in $V$. That is, \begin{equation}\label{} \lVert{\rm D} f(z)\rVert_{\Omega} \defeq \lvert f'(z)\rvert \cdot \frac{\rho_{\Omega}(f(z))}{\rho_{\Omega}(z)} > 1, \end{equation} and \begin{equation}\label{} \lVert {\rm D} f(z)\rVert_{\Omega} \to\infty \qquad\text{as}\qquad \lvert z\rvert \to \infty. \end{equation} (Recall that $\rho_{\Omega}$ denotes the density of the hyperbolic metric on $\Omega$.) In particular, for every $\eps>0$ there is $\Lambda>1$ such that $\lVert {\rm D} f(z)\rVert_{\Omega}\geq \Lambda$ whenever $z\in V$ with $\dist(z, \P)\geq \eps$. \end{prop} This result was proved, for a hyperbolic entire function $f$ and a certain choice of $\P$, in \cite[Lemma 5.1]{Re09}. The same proof goes through whenever $\P$ intersects the unbounded connected component of $\C\setminus S(f)$. This can always be ensured by adding a periodic orbit to $\P$ that intersects this component. (Compare also \cite[Section~4.2]{De}.) This is sufficient for all our purposes. However, for completeness and future reference, we shall prove the result for general $\P$, with a slightly simpler proof than that given in \cite{Re09}. To do so, we use the following simple fact. \begin{lem}[Preimages in annuli]\label{} Let $f$ be an entire transcendental function which is bounded on an unbounded connected set. Let $z_1,z_2\in\C$. Then, for all $C>1$, and all sufficiently large $R$, $f^{-1}(\{z_1,z_2\})$ contains a point of modulus between $R/C$ and $C\cdot R$. \end{lem} \begin{proof} Let $C>1$ and set $A\defeq \{z\in\C\colon 1/C < \lvert z \rvert <C\}$. Suppose by contradiction that $R_n\to\infty$ is a sequence such that the functions \[ g_n\colon A\to\C; \qquad z\mapsto f(R_n\cdot z) \] omit both {$z_1$ and $z_2$}. By Montel's theorem, this sequence of functions is normal, and hence converges locally uniformly, possibly after restricting to a subsequence. By assumption, $\limsup \min_{|z|=1} g_n(z) <\infty$, and hence the limit function is holomorphic. But this implies that $f$ remains bounded on the circle of radius $R_n$ as $n\to\infty$. Hence $f$ is bounded by the maximum principle and hence constant by Liouville's theorem, a contradiction. \end{proof} \begin{proof}[Proof of Proposition~\ref{}] The fact that $\# \P(f) \geq 2$ is well-known: Otherwise, $f$ would be a self-covering of a punctured plane, and hence conformally conjugate to $z\mapsto z^d$ for some $d$. However, $f$ is transcendental. So $\Omega$ is indeed a hyperbolic domain. Since $f\colon V\to\Omega$ is a covering map, and hence a local isometry with respect to the hyperbolic metrics of $V$ and $\Omega$, $\rho_V(z)=\rho_\Omega(f(z))\lvert f'(z)\rvert$ for all $z\in V$. By forward-invariance of $\P$, we have $V\subset \Omega$. Since $\# \mathcal{P} \geq 2$, and by Picard's theorem, $\C\setminus V = \C\setminus f^{-1}(\P)$ contains points of arbitrarily large modulus. Hence $V\subsetneq \Omega$. By Pick's theorem, it follows that \[ \lVert {\rm D} f(z)\rVert_{\Omega} = \lvert f'(z)\rvert \cdot \frac{\rho_\Omega(f(z))}{\rho_{\Omega}(z)} = \frac{\rho_V(z)}{\rho_{\Omega}(z)} > 1 \] for all $z\in V$. This establishes the first claim. Furthermore, since $f\in\mathcal{B}$, there is an unbounded connected set on which $f$ is bounded. (E.g.,\ the boundary of one of the tracts of $f$.) {By Lemma~\ref{}, there is a sequence $(c_n)_{n\geq 0}$ in $\Omega\setminus V$ such that $c_n\to\infty$ and $\lvert c_{n+1}/c_n\rvert \leq 2$. (We can even ensure $\lvert c_{n+1}/c_n\rvert \to 1$ by letting the constant $C$ in Lemma~\ref{} tend to $1$, but do not require this here.)} Hence $1/\rho_V(z) = O(\lvert z\rvert)$ as $z\to\infty$ in $V$. (See \cite[Lemma 2.1]{Re09} and compare also \cite[Proposition~3.4]{wandering} and \cite{mindaquotients}.) On the other hand, $\rho_{\Omega}(z) = O(1/(\lvert z\rvert \log \lvert z \rvert))$ as $z\to\infty$. The claim follows. \end{proof} \begin{cor}[Sets remaining in $\Omega$]\label{} Let $\Omega, \P$ be as in Proposition~\ref{}, and suppose that {$U\subset \Omega$} is open with $f^n(U)\subset \Omega$ for all $n$. Then $\dist(f^n(z),\P)\to 0$ locally uniformly for $z\in U$. In particular, if $z_0\in\Omega$ with $f^n(z_0)\in\Omega$ for all $n$, and $\limsup \dist(f^n(z_0),\P)>0$, then $z_0\in J(f)$. \end{cor} \begin{proof} It follows from Lemma~\ref{} that $\dist_{\Omega}(z, f^{-1}(\P))$ remains bounded as $z\to\infty$ in $\Omega$. Hence $\dist_{\Omega}(z,f^{-1}(\P))\to \infty$ only if $\dist(z,\P)\to 0$. Suppose that $z_0$ is a point such that $f^n(z_0)\in \Omega$ for all $n$, and that additionally there is $\eps>0$ and a sequence $n_k\to\infty$ such that $\dist(f^{n_k}(z_0),\P)\geq \eps$ for all $k$. By the above, $\dist_{\Omega}(f^{n_k}(z_0), f^{-1}(\P))$ remains bounded as $k\to\infty$. Connecting $f^{n_k}(z_0)$ to a point of $f^{-1}(\P)$ by a curve of bounded length in $\Omega$, and pulling back, it follows that there is a sequence of points $w_k \in f^{-(n_k+1)}(\P) \subset f^{-(n_k+1)}(\C\setminus \Omega) \subset \C\setminus U $ converging to $z_0$. Since $U$ is open, it follows that $z_0\notin U$. This proves the first claim. Suppose $z_0$ is as in the final claim, and $\dist(f^{n_k}(z),\P)\geq \eps$ for some sequence $n_k\to\infty$ and some $\eps>0$. If $z\in F(f)$, then by equicontinuity of the iterates of $f$, there would be a neighbourhood $U$ of $z$ such that $\dist(f^{n_k}(z))\geq \eps/2$ for all $n_k$ and all $z\in U$. By backward invariance of $\Omega$, we would have $f^n(U)\subset\Omega$ for all $n$, and this contradicts the first part of the Corollary. \end{proof} \begin{remark} The result can also be proved by appealing to the classification of Fatou components. \end{remark} \subsection*{Fundamental tails} We now return to the setting of Section~\ref{}, where we used a disc $D$ centred at the origin and a curve $\delta$ to define fundamental domains for a function $f\in\mathcal{B}$, as preimages of the domain $\WW = \C\setminus (\overline{D}\cup \delta)$. Since we now assume additionally that $f$ has bounded postsingular set, we may assume that $D$ is chosen sufficiently large to ensure that $\P(f)\subset D$. In fact, it will be convenient to allow for some extra flexibility, and fix any forward-invariant compact set $\P$ containing $\P(f)$, and assume that $\P\subset D$. We also denote the unbounded connected component of $\C\setminus\P(f)$ by $\Omega$. In everything that follows, we can more generally let $\Omega$ be as in Proposition~\ref{}; i.e.\ the unbounded connected component of $\C\setminus \P$ where $\P$ is a forward-invariant compact set containing $\P(f)$. However, we shall not require this extra generality. By~\eqref{}, we may assume that $D$ is chosen sufficiently large to ensure that \begin{equation}\label{} \lVert {\rm D} f(z)\rVert_{\Omega} \geq 2 \end{equation} whenever $f(z)\notin D$ (in addition to the assumptions already made in Section~\ref{}). The concepts that follow depend a priori on the choice of fundamental domains, i.e.\ on the choice of the initial configuration consisting of $D$ and $\delta$. However, it turns out that this choice is not essential. {(Compare Observation~\ref{}.)} Recall that the postsingular set $\P(f)$ is contained in $D$, and the image of any fundamental domain is contained in $\C\setminus \overline{D}$. Hence, for any fundamental domain $F$, its closure $\overline{F}$ does not intersect the postsingular set. It follows that any connected component $\tail$ of $f^{-(n-1)}(F)$, for any $n\geq 1$, is a Jordan domain on the Riemann sphere, whose boundary contains $\infty$, and whose closure is mapped homeomorphically onto $\overline{F}$ by $f^{n-1}$. Moreover, $f^k(z)\to\infty$ as $z\to\infty$ in $\tail$, for all $k\leq n$. \begin{defn}[Fundamental tails] Let $n\geq 1$. A connected component $\tail$ of $f^{-n}(\WW)$ is called a \emph{fundamental tail of level $n$}. In particular, the fundamental tails of level $1$ are precisely the fundamental domains of $f$. \end{defn} \begin{prop}[Facts about fundamental tails]\label{} Fundamental tails of the same level are disjoint. Moreover, for every $N$ and any compact set $K$, there are at most finitely many fundamental tails of level at most $N$ that meet $K$. If $\tail$ is a fundamental tail of level $n>1$, then $f(\tail)$ is a fundamental tail of level $n-1$. \end{prop} \begin{proof} Recall that the fundamental tails of level $N$ are precisely the connected components of $f^{-N}(\WW)$. Hence they are pairwise disjoint, and the second claim follows from Lemma~\ref{}. The final claim holds by definition. \end{proof} \begin{lem}[Fundamental domain associated to fundamental tail]\label{} Let $\tail$ be a fundamental tail of level $n$. Then there is a unique fundamental domain $F$ such that $F\cap \tail$ is unbounded. In fact, if $n>1$, then $F$ contains all sufficiently large points of $\overline{\tail}$. \end{lem} \begin{proof} We proceed by induction. The claim is trivial for $n=1$. Now suppose that $n>1$. By induction, there is a unique fundamental domain $F_1$ whose intersection with the fundamental tail $f(\tail)$, of level $n-1$, is unbounded. Let $A_1$ be the unbounded connected component of $\overline{F_1}\setminus D$, and let $A_2$ be the unbounded connected component of $A_1\cap \overline{f(\tail)}\setminus \overline{D}$. Then $f(\tail)\setminus A_2$ is bounded. Indeed, this is clear if $n=2$, and otherwise follows from the inductive hypothesis. Recall that $f\colon \overline{\tail} \to \overline{f(\tail)}$ is a homeomorphism. Since $A_2\subset \WW$, there is a unique fundamental domain $F$ containing $A\defeq f^{-1}(A_2)\cap \overline{\tail}$, and $\overline{\tail}\setminus A\subset \overline{\tail}\setminus F$ is bounded, as claimed. \end{proof} It follows from the above that we can associate natural symbolic sequences to fundamental tails. \begin{defn}[Addresses of fundamental tails] Let $\tail$ be a fundamental tail of level $n$, and denote by $F_k(\tail)$ the unique fundamental domain whose intersection with the fundamental tail $f^k(\tail)$ is unbounded. We call the finite sequence ${\underline{s}}= F_0(\tail) F_1(\tail) \dots F_{n-1}(\tail)$, of length $n$, the \emph{(finite) external address} of~$\tail$. \end{defn} Conversely, we can construct a fundamental tail having an arbitrary prescribed (finite) address by taking repeated pull-backs along the correct branches. Recall that a sequence ${\underline{s}}^1$, say of length $n$, is a \emph{prefix} of another sequence ${\underline{s}}^2$ of length $m\geq n$ if the first $n$ entries of ${\underline{s}}^2$ coincide with those of ${\underline{s}}^1$. \begin{deflem}[Tails at a given address]\label{} Let ${\underline{s}}= F_0 F_1 \dots$ be a finite or infinite sequence of fundamental domains having length at least $n\geq 1$. Then there is a unique fundamental tail $\tail$ of level $n$ having address $\tilde{{\underline{s}}} \defeq F_0 F_1 \dots F_{n-1}$. We denote this fundamental tail by $\tail_n({\underline{s}})$. We also define the inverse branches \[ f^{-n}_{{\underline{s}}} \defeq (f^n|_{\tail_n({\underline{s}})})^{-1}\colon \WW \to \tail_n({\underline{s}}). \] \end{deflem} \begin{proof} The proof is by induction on the level $n$ of the tails. For $n=1$ and for every fundamental domain $F_0$, the fundamental tail of level 1 and address $F_0$ is the fundamental domain $F_0$ itself, which is unique and has address $F_0$ by definition. Now let ${\underline{s}}$ be in the claim and let $\tail=\tail_{n-1}(\sigma({\underline{s}}))$ be the fundamental tail of level $n-1$ and address $F_1\ldots F_{n-1}$. This tail exists and is unique by the inductive hypothesis. Let $R$ be sufficiently large, and let ${\tail^{1}}$ be the unique unbounded connected component of $\tail\setminus \overline{D_R(0)}$. By Lemma~\ref{}, ${\tail}^{1}$ is contained in the fundamental domain $F_1=F_0(\tail)$ for sufficiently large $R$. Hence, if we additionally assume that $R$ is greater than the radius of $D$, we have $\tail^{1}\subset f(F_0) = \WW$. So there is a unique connected component $\tilde{\tail}^{1}$ of $f^{-1}(\tail^{1})$ contained in $F_0$, and a unique connected component $\tilde{\tail}$ of $f^{-1}(\tail)$ containing $\tilde{\tail}^{1}$. Then $F_0(\tilde{\tail})=F_0$; i.e.,\ $F_0$ is the initial entry in the address of the fundamental tail $\tilde{\tail}$, which hence has address ${\underline{s}}$. \end{proof} The following are immediate consequences of the preceding results and definitions. \begin{obs}\label{ tails} Let $\tail$ be a fundamental tail of level $n$, and let ${\underline{s}}$ be the address of $\tail$. Then the address of $f(\tail)$ is $\sigma({\underline{s}})$, where $\sigma$ denotes the shift map. Suppose that $\tail^1$ and $\tail^2$ are fundamental tails of levels $n_1$ and $n_2$, with $n_1 \geq n_2$. Let ${\underline{s}}^1$ and ${\underline{s}}^2$ be the addresses of $\tail^1$ and $\tail^2$, respectively. Then $\tail^1 \cap \tail^2$ is unbounded if and only if ${\underline{s}}^1$ is a prefix of ${\underline{s}}^2$. In this case, if addditionally $n_1>n_2$, all sufficiently large points of $\overline{\tail^1}$ lie in $\tail^2$. \end{obs} \section{Dreadlocks}\label{} Maintaining the same notation as in the previous section, we now define and study the central objects of this article: dreadlocks. Recall that $\P(f)$ is bounded; the main goal of this section is to show that, under this assumption, each of the sets $J_{{\underline{s}}}^0$ defined in Section \ref{} can be consistently extended to a larger~-- and, in a certain sense, maximal~-- set $J_{{\underline{s}}}$. The intersection of $J_{{\underline{s}}}$ with the escaping set $I(f)$ forms the \emph{dreadlock} $G_{{\underline{s}}}$ at address ${\underline{s}}$. As we shall see, each dreadlock is an unbounded connected set of escaping points, the escaping set can be written as the union of dreadlocks, and via their external addresses the collection of dreadlocks is endowed with a natural combinatorial structure. Furthermore, the definition of dreadlocks does not depend on the initial choices made in the construction of fundamental domains. Together these facts indicate that dreadlocks can indeed be considered a natural generalisation of ``hairs'' or ``rays''. For a fundamental domain $F$, recall that $\unbdd{F}$ denotes the unbounded connected component of $F\setminus\overline{D}$. We extend this definition to fundamental tails as follows. \begin{defn}[Unbounded parts of tails] Let $\tau$ be a fundamental tail of level $n$. We define $\unbdd{\tail}$ to be the unbounded connected component of $\tail\setminus f^{-(n-1)}(\overline{D})$. In other words, if $F=f^{n-1}(\tail)$, then $\unbdd{\tail}$ is the component of $f^{-(n-1)}(\unbdd{F})$ contained in $\tail$. \end{defn} Observe that, if ${\underline{s}}$ is an external address and $n\geq 2$, then $\unbdd{\tail}_n({\underline{s}})$ is precisely the unbounded connected component of $\tail_n({\underline{s}})\cap \tail_{n-1}({\underline{s}})$. \begin{defn}[Dreadlocks]\label{} Let ${\underline{s}}$ be an (infinite) external address ${\underline{s}}$. We say that a point $z\in \C$ \emph{has external address ${\underline{s}}$} if $z\in\unbdd{\tail}_{n+1}({\underline{s}})$ for all sufficiently large~$n$. We denote the set of all points $z\in \C$ having external address ${\underline{s}}$ by $J_{{\underline{s}}}$. The \emph{dreadlock} $G_{{\underline{s}}}$ is defined to be $G_{{\underline{s}}}\defeq J_{{\underline{s}}}\cap I(f)$. \end{defn} \begin{remark} If $z\in J_{{\underline{s}}}$, then $f^n(z)$ belongs to $\unbdd{F}_n$ for all sufficiently large $n$. In particular, $z\in J(f)$ by Lemma~\ref{lem:J}. Observe, however, that $J_{{\underline{s}}}$ is not closed in general. Note that other notions of points having external address ${\underline{s}}$ appear in the literature; for example, in~\cite{eremenkoproperty}, $z$ is said to have address ${\underline{s}}=F_0 F_1 \dots$ if $f^n(z)\in F_n$ for all $n\geq 0$. The advantage of the above definition, in the context of postsingularly bounded functions, is that we shall see that every escaping point has an external address (Corollary~\ref{}), and that this address is in a certain sense independent of the initial choice of fundamental domains (Observation~\ref{}). \end{remark} \begin{lem}[Properties of addresses and dreadlocks]\label{} Suppose that ${\underline{s}}$ is an external address and $z\in J_{{\underline{s}}}$. Then \begin{enumerate}[(a)] \item $z\notin J_{\tilde{{\underline{s}}}}$ for ${\underline{s}}\neq \tilde{{\underline{s}}}$. \item The point $f(z)$ has address $\sigma({\underline{s}})$. \item If $w\in f^{-1}(z)$, then $w$ has an address in $\sigma^{-1}({\underline{s}})$, and every such address is realised by exactly one element of $f^{-1}(z)$. \item The restriction $f\colon J_{{\underline{s}}} \to J_{\sigma({\underline{s}})}$ is a continuous bijection. \end{enumerate} \end{lem} \begin{proof} The first claim is trivial since fundamental tails of a given level are disjoint. Now let $\tail_n\defeq \tail_n({\underline{s}})$, and write ${\underline{s}} = F_0 F_1 F_2 \dots$. Recall that $z$ belongs to $\unbdd{\tail}_n$ for all sufficiently large $n$; say for $n\geq n_0$. {Let $n\geq \max(1,n_0-1)$. Observe that $f({\unbdd{\tail}}_{n+1}) = \unbdd{\tail}_n(\sigma({\underline{s}}))$, and $f\colon \unbdd{\tail}_{n+1} \to \unbdd{\tail}_n(\sigma({\underline{s}}))$ is a conformal isomorphism. Hence $f(z)\in \unbdd{\tail}_n(\sigma({\underline{s}}))$ as required.} {Now let $w\in f^{-1}(z)$. For $n \geq n_0$, we have $z = f(w) \in \unbdd{\tail}_n$. In particular, for $n\geq n_0+1$, there is a fundamental tail $\tailtilde_{n}$ such that $f(\tailtilde_{n}) = \tail_{n-1}({\underline{s}})$ and $w\in\unbdd{\tailtilde}_n$. Recall that $\unbdd{\tail}_n({\underline{s}})\subset \tail_{n-1}({\underline{s}})$, and that the intersection $\unbdd{\tailtilde}_{n+1} \cap \tailtilde_n$ is non-empty (since it contains~$w$). Hence $\unbdd{\tailtilde}_{n+1}\subset \tailtilde_n$ for all $n\geq n_0+1$. That is, $\tailtilde_{n+1}$ tends to infinity within $\tailtilde_n$. So if $F$ is the fundamental domain whose intersection with $\tailtilde_{n_0+1}$ is unbounded, then $\tailtilde_n$ tends to infinity in $F$ for $n\geq n_0+1$. It follows that $\tailtilde_n = {\tail}_n(F{\underline{s}})$. In particular, $w$ has address $F{\underline{s}}\in \sigma^{-1}({\underline{s}})$.} Conversely, let $\tilde{{\underline{s}}} \in \sigma^{-1}({\underline{s}})$; that is, $\tilde{{\underline{s}}} = F {\underline{s}}$ for some fundamental domain $F$. Then the fundamental tail $\tailtilde_{n_0+1}$ of level $n_0+1$ associated to the address $\tilde{{\underline{s}}} \defeq F{\underline{s}}$ is a component of $f^{-1}(\tail_{n_0})$. Hence there is a unique element $w$ of $f^{-1}(z)$ in $\tailtilde_{n_0+1}$, and $w$ has address $\tilde{{\underline{s}}}$ as above. The final claim follows from the previous two. \end{proof} The following proposition establishes a connection between the sets $J_{{\underline{s}}}$ defined in Definition~\ref{} and the sets $J_{{\underline{s}}}^0$ studied in Section~\ref{}. \begin{prop}[$J_{{\underline{s}}}$ and $J_{{\underline{s}}}^0$]\label{} Let ${\underline{s}}$ be an (infinite) external address. If $z\in J_{{\underline{s}}}^0$, then $z\in \unbdd{\tail}_n({\underline{s}})$ for all $n\geq 1$. In particular, $J_{{\underline{s}}}^0\subset J_{{\underline{s}}}$. \end{prop} \begin{proof} This is essentially the content of the second paragraph of \cite[Proof of Theorem~1.1]{eremenkoproperty}. Since that proof is somewhat concise, and we are using a different terminology, we provide the details. For $n\geq 1$, let $\tailtilde_n$ denote the fundamental tail of level $n$ containing $z$. By assumption, $f^{n-1}(z)\in \unbdd{F}_{n-1}$, and hence $z\in \unbdd{\tailtilde}_n$. We claim that $\unbdd{\tailtilde}_{n+1} \subset \tailtilde_n$; that is, $\tailtilde_{n+1}$ tends to infinity within $\tailtilde_n$. Indeed, $\unbdd{\tailtilde}_{n+1}$ is a connected component of $f^{-n}(\unbdd{F}_n)$ and $\tailtilde_n$ is a connected component of $f^{-n}(\WW)$. Since $\unbdd{F}_n\subset \WW$ and $\unbdd{\tailtilde}_{n+1}\cap \tailtilde_n \neq\emptyset$, the claim follows. Inductively, $\tailtilde_n$ tends to infinity within $F_0=\tailtilde_1$ for each $n\geq 0$. Applying this fact to $f^k(z)$, we see likewise that $f^k(\tailtilde_n)$ tends to infinity within $F_k$ for $n\geq k$. Thus, for any $n\geq 1$, we conclude that $\tailtilde_n =\tail_n({\underline{s}})$, and hence $z\in \unbdd{\tailtilde}_n = \unbdd{\tail}_n({\underline{s}})$, as desired. \end{proof} \begin{cor}[Escaping points are organised in dreadlocks]\label{} \mbox{} \begin{enumerate}[(a)] \item A point $z$ has an external address if and only if there is $n$ and an external address ${\underline{s}}$ such that $f^n(z)\in J_{\sigma^n({\underline{s}})}^0$.\label{} \item\label{} There is a number $R>0$ with the following property. If $z\in J(f)$ is such that $\lvert f^n(z)\rvert \geq R$ for all sufficiently large $n$, then $z$ has an external address ${\underline{s}}$. \item Every escaping point $z\in I(f)$ has an external address ${\underline{s}}$, and hence belongs to a dreadlock $G_{{\underline{s}}}$.\label{} \end{enumerate} \end{cor} \begin{proof} The ``only if'' direction of~\ref{} is immediate from Definition~\ref{}. On the other hand, if $z$ has the stated property, then $f^{n}(z)$ has address ${\underline{s}}$ by Proposition~\ref{}, and hence $z$ also has an external address by Lemma~\ref{}. To prove~\ref{}, let $\mathcal{F}$ be the set of fundamental domains that intersect $D$. {Recall that $F\setminus \unbdd{F}$ is bounded for all $F\in\mathcal{F}$, and that $\mathcal{F}$ is finite by Lemma~\ref{}. Now} fix \[ R > \sup\left\{ \lvert z\rvert \colon z\in \overline{D}\cup \bigcup_{F\in\mathcal{F}} F\setminus\unbdd{F}\right\}. \] Suppose that $\lvert f^n(z)\rvert \geq R$ for all $n\geq n_0$. Then, for all $n\geq n_0$, $f^n(z)$ belongs to some fundamental domain $F$, and by choice of $R$, it must belong to $\unbdd{F}$. Now $z$ has an address by~\ref{}. Finally,~\ref{} is an immediate consequence of~\ref{}. \end{proof} \begin{rmk} Under the hypotheses of Proposition~\ref{}, $f^n(z)$ belongs to $\unbdd{F}_n$ for all $n$. However, it may well be that $f^n(z)$ belongs to a \emph{bounded} component of $\tail_k({\underline{s}})\setminus \overline{D}$ for all sufficiently large $k$ and similarly for all points on the forward orbit of $z$. Indeed, typically when $I(f)$ does not consist of hairs, exactly this is the case for many escaping points, so Proposition~\ref{} is not at all trivial, and uses the boundedness of the postsingular set in an essential manner. No analogue thereof is currently known for functions with unbounded postsingular set, and in particular in this setting there is no canonical way of associating external addresses to arbitrary escaping points as in Corollary~\ref{}. This is a major challenge in showing the unboundedness of components of the escaping set. (Compare Corollary~\ref{} below.) \end{rmk} \subsection*{Connectedness of dreadlocks and uniform escape to infinity} Thanks to Proposition~\ref{}, we can use the results of Section~\ref{} to study dreadlocks. We use the following definition from \cite{arclike}. \begin{defn}[Escaping composants] \label{} Let $z\in I(f)$. Then the \emph{escaping composant} $\mu(z)=\mu(z,f)$ is the union of all connected sets $A$ containing $z$ on which the iterates of $f$ tend to infinity uniformly. \end{defn} \begin{remark} While $\mu(z)$ is a union of sets on which the iterates tend to infinity uniformly, typically $f^n$ will not tend to infinity uniformly on $\mu(z)$. \end{remark} \begin{lem}[Dreadlocks and uniform escape]\label{} Let ${\underline{s}}$ be an external address, and suppose that $z\in G_{{\underline{s}}}$. Let $A\ni z$ be a connected set {and suppose that there is $N\geq 0$ such that $f^n(A)\cap D=\emptyset$ for $n\geq N$. (In particular, this is the case when the iterates of $f$ tend to infinity uniformly on $A$.)} Then there is $n_0\geq N$ such that $A\subset \unbdd{\tail}_n({\underline{s}})$ for $n\geq n_0$. In particular, $\mu(z)\subset G_{{\underline{s}}}$. \end{lem} \begin{proof} {Let $n_0\geq N+1$ be such that $z\in \unbdd{\tail}_n({\underline{s}})$ for $n\geq n_0$. Let $n\geq n_0$. Then \[ A \subset f^{-n}(\WW) \cap f^{-(n-1)}(\WW). \] Now $\unbdd{\tail}_n({\underline{s}})$ is a connected component of the set on the right-hand set, {$z\in A\cap \unbdd{\tail}_n({\underline{s}})$}, and $A$ is connected. Hence $A\subset \unbdd{\tail}_n({\underline{s}})$ for $n\geq n_0$, and all points in $A$ have address ${\underline{s}}$.} \end{proof} We remark that a dreadlock $G_{{\underline{s}}}$ may contain uncountably many different {escaping composants}. Moreover, it is possible for $\mu(z)$ to consist of a single point. (See \cite[Theorem~1.3]{arclike}.) However, by the results of Section~\ref{}, $G_{{\underline{s}}}$ contains a distinguished uniformly {escaping composant}, namely the one consisting of those points for which the set $A$ can be taken to be unbounded. \begin{deflem}[The core of a dreadlock]\label{} Let ${\underline{s}}$ be an infinite external address. Let $\mathcal{X}$ be the collection of all closed, unbounded, connected sets {$X\subset G_{{\underline{s}}}$} on which the iterates of $f$ tend to infinity uniformly. Then no element of $\mathcal{X}$ separates the plane. Furthermore, $\mathcal{X}$ is linearly ordered by inclusion; i.e., if $X_1,X_2\in \mathcal{X}$ then $X_1\subset X_2$ or $X_2\subset X_1$. In particular, if $J_{{\underline{s}}}\neq \emptyset$, then the union $\mu_{{\underline{s}}}\defeq \bigcup \mathcal{X} \subset G_{{\underline{s}}}\neq\emptyset$ satisfies $\mu_{\underline{s}}=\mu(z)$ for all $z\in \mu_{{\underline{s}}}$. We call $\mu_{{\underline{s}}}$ the \emph{core} of the dreadlock $G_{{\underline{s}}}$. \end{deflem} \begin{proof} If $X\in \mathcal{X}$, then there is $n_0\geq 1$ such that $X\subset \unbdd{\tail}_n({\underline{s}})$ for $n\geq n_0$. In particular, $f^n(X)\subset J_{\sigma^n({\underline{s}})}^0$ for $n\geq n_0$. Let $n\geq n_0$. Since $f^n\colon \tau_n({\underline{s}})\to \WW$ is a conformal isomorphism, and $X$ is unbounded, it follows that $f^n(X)$ is unbounded for all $n$. By Lemma~\ref{lem:J} we know that $f^n(X)$, and hence $X$, does not separate the plane. Furthermore, if $X_1,X_2\in \mathcal{X}$, then we can choose $n_0$ sufficiently large so that the above holds for both sets. It follows from part~\ref{} of Theorem~\ref{} that one of $f^{n_0}(X_1)$ and $f^{n_0}(X_2)$ is contained in the other, and the same holds for $X_1$ and $X_2$. By part~\ref{} of Theorem~\ref{}, together with Proposition~\ref{}, the set $\mu_{{\underline{s}}}$ (as defined in the statement of the lemma) is non-empty. By the fact we just proved, $\mu_{{\underline{s}}}$ is a nested union of connected sets, and hence itself connected. By definition, if $z\in \mu_{{\underline{s}}}$, then $\mu_{{\underline{s}}}\subset \mu(z)$. On the other hand, $\mu(z)\subset \mu_{{\underline{s}}}$ by Lemma~\ref{}. \end{proof} \begin{prop}[Dreadlocks are connected]\label{} Let ${\underline{s}}$ be an external address with $J_{{\underline{s}}}\neq\emptyset$. Then $\mu_{{\underline{s}}}$ is dense in $J_{{\underline{s}}}$ and $G_{{\underline{s}}}$. In particular, both of these sets are connected and unbounded. \end{prop} \begin{proof} Let $z\in J_{{\underline{s}}}$. {By Corollary~\ref{},} there is $n_0$ such that $f^n(z)\in J_{\sigma^n({\underline{s}})}^0$ for $n\geq n_0$. Let $X_n$ be the unbounded connected subset of $\mu_{\sigma^n({\underline{s}})}$ whose existence is guaranteed by Theorem~\ref{}. Recall that $X_n$ can be connected to $f^n(z)$ by a curve $\gamma_n\subset\C\setminus \overline{D}$ of cylindrical length at most $2\pi$ such that $\gamma_n$ is homotopic to a curve in $\WW$. In particular, the pullback $\tilde{\gamma}_n$ of $\gamma_n$ along the orbit of $z$ connects $z$ to the set $\tilde{X}_n \defeq f^{-n}_{{\underline{s}}}(X_n)\subset \mu({\underline{s}})$. The density $\rho_{\Omega}$ of the hyperbolic metric on $\Omega$ tends to zero like $1/\lvert z\rvert \lvert \log z\rvert$. (Recall that $\Omega$ is the unbounded connected component of $\C\setminus\P(f)$.) Therefore the hyperbolic length $\ell_{\Omega}(\gamma_n)$ is also uniformly bounded, independently of $n$. It follows by Proposition~\ref{} that $\ell_{\Omega}(\tilde{\gamma}_n)\to 0$ as $n\to \infty$. Hence $\dist_{\Omega}(z , \tilde{X}_n) \to 0$, and $z\in \overline{\mu_{{\underline{s}}}}$, as claimed. \end{proof} We note that the above result (and its proof) is essentially a reformulation of the main argument in the proof of the main theorem of \cite{eremenkoproperty}, which {we recover as} follows. \begin{cor}[Unbounded sets of escaping points]\label{} If $f$ is a postsingularly bounded entire function, then every connected component of the escaping set $I(f)$ is unbounded. \end{cor} \begin{proof} Let $z\in I(f)$. By Corollary~\ref{}, $z\in G_{{\underline{s}}}$ for some external address ${\underline{s}}$, and by Proposition~\ref{}, $G_{{\underline{s}}}$ is an unbounded connected subset of $I(f)$. \end{proof} \subsection*{Independence of dreadlocks from the construction} Note that the definition of addresses, and hence of dreadlocks, depends a priori on the choice of fundamental domains, and hence on the domain $\WW$ (i.e., on the choice of the disc $D$ and the curve $\delta$). However, if dreadlocks are to be considered canonical objects associated to $f$, then this dependence should not essential. We shall briefly discuss here why this is indeed the case. Suppose that $\tilde{W}_0$ is a different choice of base domain, and that $\tilde{\tail}$ is a fundamental tail of level at least $2$, using this alternative initial configuration. Then it is easy to see that there is a unique fundamental domain $F$ for the original domain $\WW$ such that all sufficiently large points of $\tilde{\tail}$ lie in $F$. Indeed, if $z\in \tilde{\tail}$ is sufficiently large, then $f(z), f^2(z)\notin \overline{D}$, and in particular $f(z)\notin \delta$. So $f(z)\in \WW$, and the claim is established. In particular, if $\tilde{{\underline{s}}}$ is an external address with respect to $\tilde{W}_0$, then we can associate to $\tilde{{\underline{s}}}$ an address ${\underline{s}}= F_0 F_1 \dots$ with respect to $\WW$. {Here for each $n\geq 0$, $F_n$ is the fundamental domain associated, in the above manner, to the fundamental tails $\tilde{\tail}_k(\sigma^n(\tilde{{\underline{s}}}))$, for $k\geq 2$. (Observe that $F_n$ is independent of $k$.)} It is easy to see that, in turn, $\tilde{{\underline{s}}}$ can be obtained from ${\underline{s}}$ using the same procedure in the opposite direction, so that the correspondence ${\underline{s}}\mapsto \tilde{{\underline{s}}}$ is a bijection between external addresses defined using the original configuration $\WW$, and those defined using the modified one. {The following observation is an easy consequence of this construction.} \begin{obs}[Dreadlocks are independent of the initial configuration]\label{} Let ${\underline{s}}$ and $\tilde{{\underline{s}}}$ correspond to each other as in the above construction; let $G_{{\underline{s}}}$ be the dreadlock obtained by using preimages of $\WW$ in the construction, and let $\tilde{G}_{\tilde{{\underline{s}}}}$ be the corresponding dreadlock according to the choice $\tilde{W}_0$. Then $G_{{\underline{s}}} = \tilde{G}_{\tilde{{\underline{s}}}}$. In other words, the collection $\mathcal{G} = \{G_{{\underline{s}}}\}$ of subsets of $J(f)$ is independent of the initial choice of $\WW$. \end{obs} We remark that the same is not true for the set $J_{{\underline{s}}}$. Indeed, suppose that $z$ is a non-escaping point that has an external address for one choice of $\WW$. Then we can choose a different initial configuration $\tilde{W}_0$, where the initial disc $\tilde{D}$ is chosen sufficiently large to ensure that $z$ enters $\tilde{D}$ infinitely many times. Clearly $z$ does not have an external address with respect to this configuration. In similar fashion, we see that dreadlocks are preserved under iteration. Indeed, let $n\geq 1$. Suppose that $\tilde{D}$ is a disc centred around zero such that $f^{n-1}(D)\subset \tilde{D}$, and that $\tilde{\delta}\subset\delta$ connects $\tilde{D}$ to $\infty$ outside of $\tilde{D}$. Then $\tilde{W}_0\defeq \C\setminus(\operatorname{cl}(\tilde{D})\cup \tilde{\delta})$ is a valid initial configuration for $f^n$. Every fundamental domain of $f^n$ (with respect to $\tilde{W}_0$) is contained in a unique fundamental tail of level $n$ for $f$. Conversely, every fundamental tail of $f$ contains exactly one fundamental domain of $f^n$. Hence there is a natural bijection between the fundamental domains of $f^n$ and finite external addresses of length $n$ for $f$, and thus between external addresses of $f^n$ and those of $f$. In particular, we obtain the following. \begin{obs}[Dreadlocks and iterates]\label{} Every dreadlock of $f$ is a dreadlock of $f^n$, for $n\geq 1$, and vice versa. \end{obs} \subsection*{Cyclic order of addresses and dreadlocks} There is a natural \emph{cyclic order} on the set of fundamental tails of any given level, and in particular on the set of fundamental domains of $f$: if $A, B, C$ are fundamental tails, then $A \prec B \prec C$ means that $B$ tends to infinity between $A$ and $C$ in positive orientation. (See Section~\ref{} for background on the cyclic order near infinity.) We can also define a cyclic order on the collection of dreadlocks, by choosing for each a closed, connected, unbounded set on which the iterates tend to infinity, as in Proposition~\ref{}, and considering the cyclic order of these sets. Recall that the function $f$ acts in a natural way on dreadlocks, and maps fundamental tails of level $n+1$ to fundamental tails of level $n$. By the remark above, this action locally preserves cyclic order, in the following sense. Let $A\prec B \prec C$ be either dreadlocks or fundamental tails of some fixed level $n>1$. If the addresses of $A$, $B$ and $C$ all have the same initial entry, then $f(A)\prec f(B)\prec f(C)$. We can also define a ``lexicographical'' cyclic order on (finite or infinite) external addresses. To do so, we use the curve $\delta$ to convert the cylic order on fundamental domains to a linear order in the usual sense, setting $F < \tilde{F}$ if and only if $\delta \prec F \prec \tilde{F}$. This linear order gives rise to a lexicographic order $<$ on external addresses in the usual sense. The cyclic order on addresses is then defined by ${\underline{s}}^1 \prec {\underline{s}}^2 \prec {\underline{s}}^3$ if and only if $s^1<s^2<s^3$, $s^2<s^3<s^1$, or $s^3<s^1<s^2$. It follows from what was written above that this cyclic order on addresses agrees with the cyclic order of the associated fundamental tails or dreadlocks. \subsection*{Disjoint-type addresses} To conclude this section, we discuss a particularly well-behaved type of dreadlock. Note that, in general, the points in the closure of a dreadlock $G_{{\underline{s}}}$ need not belong to $J_{{\underline{s}}}$. Indeed, this is the case for those dreadlocks of greatest interest to us, namely those accumulating on a periodic point whose orbit does not lie outside of $\overline{D}$. We show now that this cannot occur when the address ${\underline{s}}$ (or a forward iterate thereof) contains only fundamental domains that do not intersect $\overline{D}$. \begin{lem}[(Eventually) disjoint-type addresses]\label{} Let ${\underline{s}} = F_0 F_1 \dots$ be an external address, and suppose that there is $n_0\geq 0$ such that $\overline{F_n}$ does not intersect $\partial D$ for $n \geq n_0$. Then $\overline{G_{{\underline{s}}}} = J_{{\underline{s}}} = f_{{\underline{s}}}^{-n}( J_{\sigma^{n}({\underline{s}})}^0)\subset \tail_n({\underline{s}})$ for $n\geq n_0$. \end{lem} \begin{proof} By assumption, we have $\overline{F_{n}}\subset W_0$, and in particular $\unbdd{F_n}=F_n$, for $n \geq n_0$. Thus $\overline{\tail_{n+1}({\underline{s}})}\subset \tail_n({\underline{s}})$ for $n \geq n_0$ (where we use the convention that $\tail_0({\underline{s}})= W_0$ for convenience). In particular, $\unbdd{\tail}_{n+1}({\underline{s}}) = \tail_{n+1}({\underline{s}})$, and \[ J_{{\underline{s}}} = \bigcup_{k> 0} \bigcap_{j\geq k} \unbdd{\tail}_j({\underline{s}}) = \bigcap_{j> n} \tail_j({\underline{s}}) = \{ z\in \tail_{n}({\underline{s}})\colon f^j(z)\in F_j\text{ for $j\geq n$}\} = f_{{\underline{s}}}^{-n}( J_{\sigma^n({\underline{s}})}^0) \] by definition. Furthermore, \[ \overline{G_{{\underline{s}}}} = \overline{J_{{\underline{s}}}} = \bigcap_{n > n_0} \overline{\tail_n({\underline{s}})} \subset \bigcap_{n > n_0} \tau_{n+1}({\underline{s}}) = J_{{\underline{s}}} \] (where the first equality holds by Proposition~\ref{}). This completes the proof. \end{proof} \begin{rmk}[Disjoint-type addresses]\label{} If $n_0=0$, then we say that ${\underline{s}}$ is \emph{of disjoint type} (since $\tau_n({\underline{s}})$ and $\tau_{n'}({\underline{s}})$ have disjoint boundaries for $n\neq n'$). For a \emph{disjoint-type} function (i.e., one that is hyperbolic with connected Fatou set), all addresses are of disjoint-type, given a suitable choice of fundamental domains. In this case, every component of the Julia set $J(f)$ is one of the sets $J_{{\underline{s}}}= J_{{\underline{s}}}^0$, and the closure $J_{{\underline{s}}} \cup \{\infty\}$ in $\Ch$ is called a \emph{Julia continuum}. Compare \cite{arclike}. Suppose that, additionally, $G_{{\underline{s}}}$ is a hair (in the sense of Definition~\ref{} below). Then it follows, using \cite[Corollary~5.6]{arclike}, that either $G_{{\underline{s}}} = J_{{\underline{s}}}$, and $J_{{\underline{s}}}$ is an arc to $\infty$ on which the iterates of $f$ tend to infinity uniformly, or $J_{{\underline{s}}}\setminus G_{{\underline{s}}}$ contains a single non-escaping point $z_0$; here $G_{{\underline{s}}}$ is an arc connecting $z_0$ to $\infty$. (For functions satisfying a head-start condition, this follows already from the results of~\cite{R3S}, without the need for the results of~\cite{arclike}.) In fact, it can be shown that, whenever ${\underline{s}}$ is of disjoint type, the set $J_{{\underline{s}}}$ is homeomorphic to the component of the Julia set of a suitable disjoint-type function, with escaping points corresponding to escaping points. Hence the above observation remains true for disjoint-type addresses, even if $f$ itself is not of disjoint type. So we may think of dreadlocks at disjoint-type addresses as always ``landing''. When ${\underline{s}}$ is bounded, we show below that this is true in a precise sense (see Proposition~\ref{}). However, when $G_{\underline{s}}$ is not a hair, it may happen in general that $J_{{\underline{s}}}$ contains a dense or uncountable set of non-escaping points, even when $f$ and hence all addresses are of disjoint type \cite[Theorem~1.8]{arclike}. In this article, we only consider landing properties for dreadlocks at bounded external addresses, so these subtleties will not become relevant. \end{rmk} \section{Hairs and dreadlocks}\label{} We shall now discuss the relationship between hairs and dreadlocks. We begin by making the notion of hairs precise for postsingularly bounded functions. \begin{defn}[Hairs]\label{} A dreadlock $G_{{\underline{s}}}$ is a hair {if one of the following holds}. \begin{enumerate}[(a)] \item $G_{{\underline{s}}}$ is an arc connecting a finite endpoint to infinity, and $f^n|_{G_{{\underline{s}}}}\to\infty$ uniformly.\label{} \item There is a continuous bijection $\gamma\colon (0,\infty) \to G_{{\underline{s}}}$ such that $\gamma(t)\to\infty$ as $t\to \infty$.\label{} \end{enumerate} \end{defn} \begin{remark}[Remark 1] There are a number of different definitions of ``hairs'' or ``rays'' for entire functions in the literature; see e.g.\ {\cite[Definition~2.2]{R3S} or \cite[p.~740]{devaneyhairs}.} These definitions typically allow the possibility that a dreadlock contains a hair as a proper subset. For example, it follows from~\cite{arclike} that there is an entire function $f$ with bounded postsingular set (and indeed of disjoint type), having a bounded-address dreadlock $G_{{\underline{s}}}$ with the following properties. The set $\overline{G_{{\underline{s}}}}\setminus G_{{\underline{s}}}$ consists of a single point $z_0$ with bounded orbit, and $G_{{\underline{s}}}\cup\{z_0,\infty\}$ is homeomorphic to a $\sin(1/x)$ continuum, with the starting point of the accumulating curve corresponding to $\infty$, and one of the endpoints of the limiting interval situated at $z_0$. According to the definitions cited above, the accumulating curve itself would be considered a ray, but this ray does not include all points of $G_{{\underline{s}}}$. (In particular, $G_{{\underline{s}}}$ is not a hair in the sense of Definition~\ref{}.) Here, we shall leave open the question of what the ``correct'' notion of a hair should be, and in particular whether such a hair should be allowed to be a proper subset of a dreadlock. Instead, we only discuss the case where the entire dreadlock $G_{{\underline{s}}}$ is a hair; for such hairs we shall see in Lemma~\ref{} that our definition coincides with those mentioned above. Moreover, for criniferous functions~-- and in particular for the class of functions for which hairs are constructed in \cite{R3S}~-- we see below that every dreadlock is both a hair in the sense of Definition~\ref{} and a dynamic ray in the sense of~\cite[Definition~2.2]{R3S}, so this issue does not arise there at all. \end{remark} \begin{remark}[Remark 2] There exist examples of {postsingularly bounded} entire functions having a dreadlock which is an arc connecting a finite endpoint to infinity, but which are not a hair in the above sense, {since the iterates on this arc do not tend to infinity uniformly}. \end{remark} \begin{remark}[Remark 3] The bijection in~\ref{} will not be a homeomorphism, in general, since it may be that e.g.\ the hair accumulates everywhere upon itself. \end{remark} \begin{remark}[Remark 4] Our terminology means that a dreadlock may {just consist of a single hair}, and that any dreadlock contains at most one hair. \end{remark} \begin{lem}[Properties of hairs]\label{} Suppose that $G_{{\underline{s}}}$ is a hair. If case~\ref{} of Definition~\ref{} holds, then $f^n_{\gamma([t,\infty))}\to\infty$ uniformly for all $t>0$. Moreover, if ${\underline{s}}$ is bounded and $G_{{\underline{s}}}$ is a hair, then $G_{{\underline{s}}}$ satisfies~\ref{} of Definition~\ref{}. \end{lem} \begin{proof} Suppose~\ref{} holds, and let $t>0$. Then $\gamma_1\defeq \gamma([t,\infty))$ is a closed and unbounded proper subset of $J_{{\underline{s}}}$. By Definitinion~\ref{}, there is a closed unbounded set $A\subset {\mu}_{{\underline{s}}}$ on which the iterates of $f$ tend to infinity uniformly, and such that $A\not\subset \gamma_1$. It follows that $\gamma_1\subset A$, and the claim is proved. Moreover, suppose ${\underline{s}}$ is bounded and let $R>0$ be as in Corollary~\ref{}. Then, for all $n\geq 0$, the dreadlock $f^n(G_{{\underline{s}}}) = G_{\sigma^n({\underline{s}})}$ contains a point of modulus at most $R$. Hence the iterates of $f$ do not tend to infinity uniformly on $G_{{\underline{s}}}$. \end{proof} \begin{prop}[Characterisation of hairs]\label{} A dreadlock $G_{{\underline{s}}}$ of $f$ is a hair if and only if, for every $z\in G_{{\underline{s}}}$, there is an arc connecting $z$ to $\infty$ on which $f^n\to\infty$ uniformly. \end{prop} \begin{proof} If $G_{{\underline{s}}}$ is a hair, then the stated condition holds by definition (in case~\ref{}) or by Lemma~\ref{} (in case~\ref{}). So now suppose that every point of $z\in G_{{\underline{s}}}$ can be connected to infinity by an arc $\gamma_z$ on which the iterates tend to infinity uniformly. So $\gamma_z\subset \mu(z) \subset \mu_{{\underline{s}}}\subset G_{{\underline{s}}}$ {by Lemma~\ref{}. By Lemma~\ref{}, the arcs $\gamma_z$ are linearly ordered by inclusion, and the arc $\gamma_z$ is unique.} Let $(x_n)_{n=0}^{\infty}$ be a countable dense subset of $G_{{\underline{s}}}$. Set $n_0=0$ and define $n_{k+1}$ inductively as the minimal value of $n$ for which $x_n$ does not lie on the arc $\gamma_{x_{n_k}}$. {(If no such $n$ exists, then $G_{{\underline{s}}}=\gamma_{x_{n_k}}$ satisfies Definition~\ref{} and we are done.)} We set $y_k \defeq x_{n_k}$. Then the union of the arcs $\gamma_{y_k}$ is a single continuous injective curve $\gamma\colon (0,\infty)\to G_{{\underline{s}}}$, which can be parameterised such that $\gamma([1/k,\infty)) = \gamma_{y_k}$ for all $k$. First suppose that $\gamma(t)$ has a limit $z_0\in G_{{\underline{s}}}$ as $t\to 0$. Then {$\gamma_{z_0} = \gamma\cup \{z_0\} =\overline{\gamma} = G_{{\underline{s}}}$, since $\gamma$ is dense in $G_{{\underline{s}}}$. Hence $G_{{\underline{s}}}$ satisfies Definition~\ref{}.} Otherwise, $\gamma\bigl((0,\infty)\bigr)\not\subset \gamma_z$ for all $z\in G_{{\underline{s}}}$. Hence we see that $\gamma_z\subset \gamma_{y_k}$ for sufficiently large $k$; it follows that $\gamma$ is surjective, and Definition~\ref{} holds. \end{proof} \begin{cor}[Dreadlocks and {{criniferous}} functions] A transcendental entire function $f$ with bounded postsingular set is {{criniferous}} if and only if every dreadlock is a hair. \end{cor} \begin{proof} Recall that every escaping point of $f$ belongs to some dreadlock. Hence the claim is immediate from {Proposition~\ref{}}. \end{proof} Observe that we now have two apparently different definitions of periodic hairs, namely Definition~\ref{}, and the case of a periodic dreadlock (i.e., a dreadlock $G_{{\underline{s}}}$ at a periodic address~${\underline{s}}$) that is a hair. We conclude the section by showing that the two coincide. \begin{prop}[Periodic hairs]\label{} Every periodic hair in the sense of Definition~\ref{} is a periodic dreadlock, and this dreadlock is a hair in the sense of Definition~\ref{}. Conversely, any periodic dreadlock that is a hair has a parameterisation as a periodic hair in the sense of Definition~\ref{}. \end{prop} \begin{proof} Suppose that $\gamma$ is a periodic hair, say of period $p$. Then, for every $t\in\R$, the iterates $f^n$ tend to infinity uniformly on $\gamma\bigl([t,\infty)\bigr)$. Hence, by Lemma~\ref{}, $\gamma$ is contained in a dreadlock $G_{{\underline{s}}}$, whose address ${\underline{s}}$ is necessarily periodic. Let us show that $G_{\underline{s}}\subset\gamma$ and hence that $G_{\underline{s}}=\gamma$ is a hair in the sense of definition~\ref{}. Set \[ R\defeq \exp(2\pi)\cdot \max_{t\in [0,1]} \lvert \gamma(t)\rvert, \] and let $t_0>1$ be maximal such that $\lvert \gamma(t_0)\rvert = R$. Let $z\in G_{{\underline{s}}}$, and set $z_n\defeq f^{pn}(z)$ for $n\geq 0$. Then there is $n_0$ such that $z_n$ belongs to the unbounded connected component of $\WW\setminus \overline{D_R(0)}$ for $n\geq n_0$. By Lemma~\ref{}, for every $n\geq n_0$, there is a point $\zeta_n\in \gamma\bigl([t_0,\infty)\bigr)$ with $\dist_{\cyl}^{\WW}( z_n , \zeta_n) \leq 2\pi$; say $\zeta_n = \gamma(T_n)$, with $T_n\geq t_0$. Set $t_n\defeq T_n - (n-n_0)$ and $\xi_n \defeq \gamma(t_n)$. By expansion of $f$, it follows that $\xi_n \to z_{n_0}$. Furthermore, $\dist_{\cyl}^{\WW}( f^{pk}(\xi_n) , z_k) \leq 2\pi$ for $0\leq k \leq n-n_0$, and hence \[ \lvert \gamma(t_n + k) \rvert = \lvert f^{pk}(\xi_n) \rvert > \max_{t\in [0,1]} \lvert \gamma(t)\rvert \] for all $k\geq 0$. Thus $t_n> 1$, and hence $z_{n_0}\in \gamma\bigl([1,\infty)\bigr)$. In particular, $z\in \gamma$; hence $G_{{\underline{s}}} = \gamma$ as desired. Now suppose that ${\underline{s}}$ is periodic and that $G_{{\underline{s}}}$ is a hair. Let $\tilde{\gamma}\colon (0,\infty) \to G_{{\underline{s}}}$ be a parameterisation of $G_{{\underline{s}}}$ as in part~\ref{} of Definition~\ref{}. Consider the point $z_0\defeq \tilde{\gamma}(1)$, and its image $z_1 = f(z_0)$, then $z_1 = \tilde{\gamma}(t)$ for some $t$. It follows easily that $t>1$, and that the piece $\tilde{\gamma}\bigl((1,t)\bigr)$ is disjoint from its forward and backward images in $G_{{\underline{s}}}$. We may reparameterise $\tilde{\gamma}$ to a curve $\gamma\colon\R\to G_{{\underline{s}}}$ such that $\gamma([0,1])$ corresponds to $\tilde{\gamma}([1,t])$, and such that $\gamma(t+1) = f(\gamma(t))$ for all $t$. Then $\gamma$ is a periodic hair in the sense of definition~\ref{}. \end{proof} \section{Accumulation sets and landing properties of bounded-address dreadlocks}\label{} We now study the accumulation behaviour of dreadlocks at bounded external address. In the case where the dreadlock in question is a hair, it is clear that the ``accumulation set'' of this dreadlock should be the set of limit points of the curve $\gamma(t)$ as $t\to 0$, where $\gamma$ is the curve from Definition~\ref{}. For periodic hairs, one can see easily that this is equivalent to fixing a base point on this hair, considering its successive preimages along $\gamma$, and studying the accumulation set of this sequence. This motivates the following definition. \begin{defn}[Accumulation sets and landing of dreadlocks]\label{} Let ${\underline{s}}$ be a bounded external address, and let $\zeta\in \WW$. For $n\geq 1$, set $\zeta_n \defeq \zeta_n({\underline{s}}) \defeq f_{{\underline{s}}}^{-n}(\zeta)\in \tau_n({\underline{s}})$. (Recall from Definition~\ref{} that $f_{{\underline{s}}}^{-n}=(f^n|_{\tail_n({\underline{s}})})^{-1}$.) Then the \emph{accumulation set} of the dreadlock $G_{{\underline{s}}}$ is defined to be the accumulation set (in $\Ch$) of the sequence $(\zeta_n)_{n=1}^{\infty}$, and denoted by $\Lambda(G_{{\underline{s}}})\defeq \Lambda({\underline{s}})\defeq \Lambda({\underline{s}},\zeta)$. The dreadlock $G_{\underline{s}}$ is said to \emph{land} at a point $z_0$ if and only if $\Lambda({\underline{s}})=\{z_0\}$. \end{defn} Note that this is an abstract definition of accumulation sets; a priori this definition uses only the notion of fundamental tails, and does not require the definition of dreadlocks or their properties. It may not be clear a priori that this is a natural notion, but the results of this section will show that this is indeed the case. We begin by verifying that $\Lambda({\underline{s}})=\Lambda({\underline{s}},\zeta)$ is indeed independent of $\zeta$. This follows from the following well-known fact about the shrinking of univalent preimages. (Compare also~\cite[Proposition 3]{Lyu83}.) Recall that $\Omega$ is the unique unbounded component of $\C\setminus P(f)$. \begin{lem}[Euclidean shrinking]\label{} Suppose that $V\Subset \Omega$ is a bounded Jordan domain. Then, for any $\eps>0$ and any compact set $K\subset\C$, there exists $N_\eps$ with the following property. For $n\geq N_\eps$, every connected component of $f^{-n}(V)$ that meets $K$ has Euclidean diameter at most $\eps$. \end{lem} \begin{proof} Suppose, by contradiction, that there is a sequence $(V_{k})_{k=0}^{\infty}$ of $n_k$-th preimages of $V$, with $n_k\to\infty$, with $V_k\cap K\neq \emptyset$ and $\inf_k \diam(V_k)>0$. Let $\tilde{V}\Subset \Omega$ be a slightly larger Jordan domain than $V$ with $\overline{V}\subset\tilde{V}$, and let $\tilde{V}_{k}$ be the component of $f^{-n_k}(\tilde{V})$ containing $V_{n_k}$. Then $f^{n_k}\colon \tilde{V}_{k}\to \tilde{V}$ is a covering map and hence a conformal isomorphism, whose inverse $\phi_k\defeq (f|_{\tilde{V}_k})^{-1} \colon \tilde{V}\to \tilde{V}_k$ maps $V_k$ to $V$. By assumption, there is a sequence $(z_k)_{k=0}^{\infty}$ with $z_k\in K$. By Koebe's distortion theorem, $\tilde{V}_k$ contains a round disc around $z_k$ whose diameter is comparable to that of $V_k$. By assumption, the latter is bounded from below. Hence, if $D$ is a sufficiently small disc centred at a limit point of the sequence $(z_k)$, then $D$ is contained in infinitely many $\tilde{V}_{n_k}$. It follows that $f^n(D)\subset \Omega$ for all $n\geq 0$, and $f^{n}(D)\subset \tilde{V}$ for infinitely many $n$. This contradicts Corollary~\ref{}. \end{proof} \begin{cor}[Spherical shrinking]\label{} Suppose that $V\Subset \Omega$ is a bounded Jordan domain. Then, for any $\eps>0$ there exists $N$ such that for $n\geq N$ every connected component of $f^{-n}(V)$ has spherical diameter at most $\eps$. \end{cor} \begin{proof} Let $R>0$ be so large that the complement of $K\defeq \overline{B(0,R)}$ has spherical diameter less than $\eps$. Since the spherical and Euclidean metrics are comparable on any compact subset of the plane, there is $\eps_1$ such that any set of Euclidean diameter at most $\eps_1$ that intersect $K$ has spherical diameter at most $\eps$. Let $N=N_{\eps_1}$ be as in Lemma~\ref{}, let $n\geq N$, and let $X$ be a connected component of $f^{-n}(V)$. Then either $X\cap K=\emptyset$, and hence $X$ has spherical diameter less than $\eps$, or $X\cap K\neq\emptyset$ and Lemma~\ref{} applies. In the latter case, $\diam X\leq \eps_1$, and hence $X$ also has spherical diameter at most $\eps$ by choice of $\eps_1$. \end{proof} \begin{cor}[Accumulation sets are well-defined] Let ${\underline{s}}$ be a bounded external address. Then the accumulation set $\Lambda({\underline{s}})$ is independent of the base point $\zeta$. That is, $\Lambda({\underline{s}},\zeta^1) = \Lambda({\underline{s}},\zeta^2)$ for any two choices of $\zeta^1,\zeta^2\in\WW$.\label{} \end{cor} \begin{proof} Let $V$ be a bounded Jordan domain in $W_0$ containing both $\zeta^1$ and $\zeta^2$. For $j\in\{1,2\}$ and $n\geq 1$, write $\zeta_j^n \defeq f^{-n}_{{\underline{s}}}(\zeta^j) \in {\tail}_n\defeq {\tail}_n({\underline{s}})$. Since $f^{n}\colon \tau_n \to W_0$ is univalent, for any $n$ the points $\zeta^1_n$ and $\zeta^2_n$ belong to the same connected component of $f^{-n}(V)$. By Corollary~\ref{}, the spherical diameter of this component tends to zero as $n\to\infty$. Hence the sequences $(\zeta^1_n)_{n=1}^{\infty}$ and $(\zeta^2_n)_{n=1}^{\infty}$ have the same accumulation set in $\Ch$. \end{proof} The following establishes a number of fundamental properties of accumulation sets. \begin{prop}[Properties of accumulation sets]\label{} Let ${\underline{s}}$ be a bounded external address, and $G_{{\underline{s}}}$ the dreadlock at address ${\underline{s}}$. \begin{enumerate}[(a)] \item\label{} The accumulation set $\Lambda({\underline{s}})$ is a nonempty connected subset of $\Ch$. \item\label{ {The closure $\hat{G_{{\underline{s}}}}$ of $G_{{\underline{s}}}$ in $\Ch$ is given precisely by \[\hat{G}_{{\underline{s}}} = G_{{\underline{s}}} \cup \Lambda({\underline{s}}) \cup \{\infty\} . \] \item If $J_{{\underline{s}}}\setminus G_{{\underline{s}}} \neq\emptyset$, then this set has a unique element $z_0$, and $G_{{\underline{s}}}$ lands at $z_0$, which has bounded orbit.} \label{} \item {If ${\underline{s}}$ is of disjoint type, then $G_{{\underline{s}}}$ lands at a point $z_0\in J_{{\underline{s}}}$ having bounded orbit.}\label{} \item {Let $U\subset \Ch$ be open with $U\cap \hat{G}_{{\underline{s}}}\neq\emptyset$.} {If $\hat{U}\cap\Lambda({\underline{s}})=\emptyset$, then $f^n\to\infty$ uniformly on $\hat{U}\cap \overline{G_{{\underline{s}}}}$, and in particular $\hat{U}\cap \overline{G_{{\underline{s}}}}= \hat{U}\cap J_{{\underline{s}}} = \hat{U}\cap G_{{\underline{s}}}$. On the other hand, if $U\cap\Lambda({\underline{s}})\neq\emptyset$, then $f^n$ does not tend to infinity uniformly on $U\cap \mu_{{\underline{s}}}\subset U\cap G_{{\underline{s}}}$.}\label{} \end{enumerate} \end{prop} Before proving these facts, we observe a few consequences. Firstly, we see that we can characterise the accumulation set purely in terms of $G_{{\underline{s}}}$ as a subset of the escaping set, justifying the notation $\Lambda(G_{{\underline{s}}})=\Lambda({{\underline{s}}})$. \begin{cor}\label{} Let $G_{{\underline{s}}}$ be a dreadlock at a bounded address ${\underline{s}}$. Then the accumulation set of $G_{{\underline{s}}}$ consists precisely of those points in $\hat{G}_{{\underline{s}}}$ having a neighbourhood $U$ such that $f^n$ does not tend to infinity uniformly on $U\cap G_{{\underline{s}}}$. In particular, $G_{{\underline{s}}}$ lands at a point $z_0\in\Ch$ if and only if {\(\hat{G}_{{\underline{s}}} = G_{{\underline{s}}}\cup\{z_0,\infty\}\)} and furthermore, for every neighbourhood $U$ of $z_0$ in $\Ch$, the iterates of $f$ tend to infinity uniformly on $G_{{\underline{s}}}\setminus U$. \end{cor} \begin{proof}[Proof of Corollary~\ref{}, using Proposition~\ref{}] The first claim follows directly from~\ref{} of Proposition~\ref{}. The second claim follows from the first (together with the fact that the accumulation set of $G_{{\underline{s}}}$ is nonempty). \end{proof} {In particular, it follows that for hairs our definition of the accumulation set agrees with the usual one. Recall from Lemma~\ref{} that, if a bounded-address dreadlock is a hair, it must satisfy part~\ref{} of Definition~\ref{}.} \begin{cor}[Accumulation sets of hairs]\label{} Suppose that a dreadlock $G_{{\underline{s}}}$ at bounded address is a hair, and let $\gamma\colon (0,\infty)\to G_{{\underline{s}}}$ be the continuous bijection from Definition~\ref{}. Then $\Lambda(G_{{\underline{s}}})$ is precisely the accumulation set of $\gamma(t)$ as $t\to 0$. \end{cor} \begin{proof}[Proof of Corollary~\ref{}, using Proposition~\ref{}] {Let $\Lambda(\gamma)\subset\Ch$ denote the set of accumulation points of $\gamma(t)$ as $t\to 0$. Recall that the iterates of $f$ tend to infinity uniformly on $\gamma([t,\infty))$ for all $t>0$, by Lemma~\ref{}. Thus we see from part~\ref{} of Proposition~\ref{} (or from Corollary~\ref{}) that $\Lambda(G_{{\underline{s}}}) \subset \Lambda(\gamma)$.} {It remains to prove the opposite inclusion, $\Lambda(\gamma)\subset \Lambda(G_{{\underline{s}}})$. If $\gamma$ lands, i.e., when $\#\Lambda(\gamma) = 1$, this follows from the first inclusion and the fact that $\Lambda(G_{{\underline{s}}})\neq\emptyset$. So we may assume that $\Lambda(\gamma)$ is a nondegenerate continuum.} {Since $\Lambda(\gamma)\subset \hat{G}_{{\underline{s}}}$, we have \[ \tilde{\Lambda} \defeq \Lambda(\gamma)\setminus (G_{{\underline{s}}} \cup \{\infty\}) \subset \Lambda(G_{{\underline{s}}}) \] by~\ref{} of Proposition~\ref{}. As $\Lambda(G_{{\underline{s}}})$ is closed, it thus suffices to prove that $\tilde{\Lambda}$ is dense in $\Lambda(\gamma)$.} {This is clear if $\Lambda(\gamma)\cap G_{{\underline{s}}}=\emptyset$. Otherwise, let $t_0>0$ be such that $\gamma(t_0)\in \Lambda(\gamma)$. We claim that $\gamma\bigl((0,t_0]\bigr)\subset \Lambda(\gamma)$.} {This follows from a well-known argument that we sketch as follows; compare~\cite[Lemma~5.1]{topescapingnew} for the case of exponential maps. For every $t>0$, there are pieces of other hairs of $f$ accumulating uniformly from above and below on $\gamma\bigl([t,\infty)\bigr)$. (For example, this follows from \cite[Proposition~7.1]{arclike} via \cite[Theorem~1.1]{Re09}.) Thus, in order to accumulate on $\gamma(t_0)$, the curve $\gamma$ must also accumulate on $\gamma([t,t_0])$; letting $t\to 0$, the claim is established.} {Let $t_0\in (0,\infty]$ be the supremum over all possible choices of $t_0$, and let $(t_n)_{n=0}^{\infty}$ be a decreasing sequence with $t_n\to 0$. Then $A_n \defeq \gamma([t_n,t_0])$ is a compact and nowhere dense subset of $\Lambda(\gamma)$ for all $n$. By Baire's theorem, \[ \tilde{\Lambda} = \Lambda(\gamma)\setminus \bigcup A_n \] is dense in $\Lambda(\gamma)$, as claimed. This completes the proof.} \end{proof} \begin{remark}[Remark 1] {The proof shows that a hair $G_{{\underline{s}}}$ either lands, or otherwise a generic point in $\Lambda(G_{{\underline{s}}})$ belongs to $\C\setminus G_{{\underline{s}}}$. It is plausible that this is true without the assumption that $G_{{\underline{s}}}$ is a hair, with a similar proof. This would simplify the characterisation in Corollary~\ref{} as follows: A dreadlock $G_{{\underline{s}}}$ at a bounded address ${\underline{s}}$ lands at a finite point $z_0\in\C$ if and only if $\overline{G_{{\underline{s}}}}\setminus G_{{\underline{s}}} = \{z_0\}$.} \end{remark} \begin{remark}[Remark 2] For periodic addresses, or for addresses satisfying a head-start condition as in \cite{R3S}, the proof of Corollary~\ref{} is considerably simpler. Indeed, in this setting it is easy to see directly that the iterates of $f$ do not tend to infinity uniformly on any neighbourhood of any point of $\Lambda(\gamma)$. \end{remark} Finally, we observe that \emph{most} periodic rays land and \emph{most} periodic points are landing points. Compare also \cite{BK07,BF15,Be16}. \begin{cor}[Most periodic dreadlocks land] Let $p\geq 1$. Then, for all but finitely many periodic addresses ${\underline{s}}$ of period $p$, the dreadlock $G_{{\underline{s}}}$ lands at a periodic point $z\in J_{{\underline{s}}}$. (In particular, $z$ has period $p$.) Conversely, for all but finitely many periodic points $z$ of period $p$, the point $z$ is repelling, there is a periodic dreadlock $G_{{\underline{s}}}$ that lands at $z$, and $z\in J_{{\underline{s}}}$. \end{cor} \begin{proof} By passing to an iterate, we may assume that $p=1$. Only finitely many fundamental domains $F$ intersect $\overline{D}$, and hence all but finitely many fixed addresses of $f$ are of disjoint type. Hence the first claim follows from Proposition~\ref{}. Similarly, all but finitely many fixed points of $f$ are contained in $\WW$, and hence have a fixed external address. So the second claim is a consequence of Proposition~\ref{} \end{proof} The remainder of the section is dedicated to establishing Proposition~\ref{}. We shall do so by applying Corollary~\ref{} to a suitable large Jordan domain $V$, depending on the collection of fundamental domains involved. The following technical lemma collects the properties that we require of $V$. \begin{lem}[Domains for bounded-address dreadlocks]\label{} Let $\mathcal{F}$ be any finite collection of fundamental domains of $f$, let $\zeta\in W_0$ and let $R>0$. Then there is a Jordan domain $V\Subset \Omega$ with the following properties. \begin{enumerate}[(i)] \item $\zeta\in V$. \item For all $F\in\mathcal{F}$, the unique preimage $\zeta_F$ of $\zeta$ in $F$ also belongs to $V$. \item For all $F\in\mathcal{F}$, there is a connected component $A_F$ of $V\cap \overline{F}$ containing $\zeta_F$ as well as all points of $\overline{F}$ having modulus at most $R$.\label{item:AF} \item If $U$ is the connected component of $V\setminus \overline{D}$ {containing $\zeta$}, then $U\cap A_F$ intersects the unbounded connected component {$\unbdd{F}$} of $F\setminus \overline{D}$, for all $F$. \end{enumerate} \end{lem} \begin{proof} See Figure~\ref{fig:V}. Write $\mathcal{F} = \{F^1,\dots F^n\}$. Let $\gamma$ be a Jordan curve in $\C\setminus\overline{D}$ that intersects the arc $\delta$ in exactly one point. We may assume that $\gamma$ is chosen so large that $\gamma$ surrounds $\zeta$, $f(\overline{D})$ and the set $\{f(z)\colon \lvert z\rvert \leq R\}$. For each $i$, let $\gamma_i = f^{-1}(\gamma)\cap F^i$ and $\zeta_i \defeq \zeta_{F^i}$. Then $\gamma_i$ is a cross-cut of $F^i$ that separates $\zeta_{i}$, $F^i\cap\overline{D}$ and all points of modulus at most $R$ in $F^i$ from $\infty$ in $F^i$. In particular, $\gamma^i$ belongs to the unbounded connected component of $F^i\setminus\overline{D}$. Let $X^i$ be the closure of the bounded component of $F^i\setminus \gamma_i$, and let $X$ be the union of these components. Observe that different $X^i$ may intersect when the corresponding fundamental domains are adjacent. In this case, the corresponding two $X^i$ intersect precisely in a preimage component of the piece of $\delta$ that connects $\overline{D}$ to $\gamma$. Let $X^1,\dots,X^k$ be the $k\leq n$ connected components of $X$. The set $X$ is a disjoint union of $k$ closed Jordan domains in $\Omega\setminus\delta$. (Recall that the closure of a fundamental domain does not meet $\delta$.) Let $\tilde{X}$ be obtained from $X$ by adding, for each $i\leq k$, an arc $\beta_i$ in $W_0\setminus X$ joining $\zeta$ to a point of $\gamma^i$. We may assume that two different $\beta^i$ intersect only at $\zeta$. Then $\tilde{X}$ is a compact and full set, and we may let $V$ be a Jordan domain containing $\tilde{X}$. The point $\zeta$ and all $\zeta_i$ belong to $V$ by construction. Moreover, each $X^i$ is connected, and hence belongs to a single connected component of $V\cap \overline{F^i}$. Finally, the union of all $\beta^i$ and all $\gamma^i$ is connected by construction, is contained in $\C\setminus \overline{D}$ and intersects the unbounded connected component of $F^i\setminus\overline{D}$ for all $i$. This completes the construction. \end{proof} \begin{figure} \def\svgwidth{.65\textwidth \begingroup% \makeatletter% \providecommand\color[2][]{% \errmessage{(Inkscape) Color is used for the text in Inkscape, but the package 'color.sty' is not loaded}% \renewcommand\color[2][]{}% }% \providecommand\transparent[1]{% \errmessage{(Inkscape) Transparency is used (non-zero) for the text in Inkscape, but the package 'transparent.sty' is not loaded}% \renewcommand\transparent[1]{}% }% \providecommand\rotatebox[2]{#2}% \ifx\svgwidth\undefined% \setlength{\unitlength}{512.05929974bp}% \ifx\svgscale\undefined% \relax% \else% \setlength{\unitlength}{\unitlength * \real{\svgscale}}% \fi% \else% \setlength{\unitlength}{\svgwidth}% \fi% \global\let\svgwidth\undefined% \global\let\svgscale\undefined% \makeatother% \begin{picture}(1,0.82915159)% \put(0,0){\includegraphics[width=\unitlength]{FigureV3.pdf}}% \put(0.19342145,0.41931805){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$D$}}}% \put(0.47550362,0.635){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$F_1$}}}% \put(0.47988486,0.4593828){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$F_2$}}}% \put(0.47871883,0.10){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$F_3$}}}% \put(0.57943319,0.31950585){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$\zeta$}}}% \put(0.38458639,0.19473898){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$\zeta_{F_3}$}}}% \put(0.68394288,0.45005447){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$\zeta_{F_2}$}}}% \put(0.69443728,0.70425243){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$\zeta_{F_1}$}}}% \put(0.81570605,0.75789049){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$\gamma_1$}}}% \put(0.82037022,0.4523865){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$\gamma_2$}}}% \put(0.58716108,0.09557654){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$\gamma_3$}}}% \put(0.13034575,0.67533117){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$D_R$}}}% \put(0.66793204,0.17402376){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$\beta_3$}}}% \put(0.85757137,0.37685538){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$\beta_2$}}}% \put(0.8015041,0.59452831){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$\beta_1$}}}% \put(0.42387446,0.71985521){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$V$}}}% \put(0.45,0.06){\color[rgb]{0,0,0}\makebox(0,0)[lb]{\smash{$F_4$}}}% \end{picture}% \endgroup% \caption{\label{fig:V A domain $V$ as described in Lemma~\ref{}. The domain $V$ is shaded.} \end{figure} The domain $V$ from Lemma~\ref{} allows us to study the accumulation sets of dreadlocks. The following lemma is crucial not only in our study of accumulation sets, but also for the proofs of our main theorems. The key idea is that, in order to investigate the landing of a given dreadlock $G_{{\underline{s}}}$, we can study a certain chain of simply-connected domains (obtained as iterated preimages of the domain $V$ from Lemma~\ref{}) whose diameters shrink to zero. \begin{lem}[Preimage domains]\label{} Let $\mathcal{F}$ be a finite collection of fundamental domains of $f$, {and assume that $\mathcal{F}$ contains every fundamental domain $F$ with $\overline{F}\cap \overline{D}\neq\emptyset$.} Let $\zeta$, $R$ and $V$ be as in Lemma~\ref{}. If $R$ was chosen sufficiently large (depending only on $\mathcal{F}$), then the following holds. Let ${\underline{s}}=F_0 F_1 \dots $ be any external address. For $n\geq 1$, set $\zeta_n \defeq f^{-n}_{{\underline{s}}}(\zeta)\in \tail_n\defeq \tail_n({\underline{s}})$. Also let $V_n= V_{n}({\underline{s}})$ be the unique component of $f^{-n}(V)$ containing $\zeta_n$. Then the following hold for all $n\geq 1$. \begin{enumerate}[(1)] \item The spherical diameter of $V_n$ tends to $0$ as $n\to\infty$. In particular, if $(\omega_n)_{n=0}^{\infty}$ is any sequence with $\omega_n\in V_n$ for all sufficiently large $n$, then $\Lambda({\underline{s}})$ coincides with the set $\Lambda_\omega$ of accumulation points of the sequence $(\omega_n)$. \label{} \item $\tail_{n+1}({\underline{s}}) \subset \unbdd{\tail}_{n+1}({\underline{s}})\cup V_n \subset \tail_n({\underline{s}})\cup V_n$ and in fact $\overline{\tail_{n+1}({\underline{s}})} \subset \tail_n({\underline{s}})\cup V_n$.\label{} \end{enumerate} {If $F_k\in\mathcal{F}$ for all $k\geq0$, then additionally:} \begin{enumerate}[(1),resume] \item $V_n \cap V_{n+1}\neq \emptyset$. More precisely, $\zeta_{n+1}\in V_n$, and $f^{n}(\zeta_{n+1})=\zeta_{F_{n+1}}\in V$. \label{} \item $V_n\cap \mu_{{\underline{s}}}\neq \emptyset$. \label{} \item $f^n\to\infty$ uniformly on {$G_{{\underline{s}}}\setminus \bigcup_{j\geq n} V_j$. (In particular, this set is closed and agrees with $J_{{\underline{s}}}\setminus \bigcup_{j\geq n} V_j$.) In fact, if $z$ belongs to this set, then $\lvert f^m(z)\rvert \geq \rho_m$ for $m\geq n$, where the sequence $(\rho_m)$ depends only on $\mathcal{F}$ and $n$.}\label{} \end{enumerate} \end{lem} \begin{proof} By Corollary~\ref{} the spherical diameter of $V_n$ tends to $0$ as $n\to\infty$. In particular, the set $\Lambda_{\omega}$ in~\ref{} is independent of the choice of the sequence $(\omega_n)$. Also recall that $\zeta\in V$ by assumption; if we choose $\omega_n = \zeta_n$, then $\Lambda({\underline{s}})=\Lambda({\underline{s}},\zeta) = \Lambda_\omega$. This proves~\ref{}. Part~\ref{} is trivial if $\overline{F_n}\subset \WW$, as then $\tau_{n+1}=\unbdd{\tau}_{n+1}$ and $\overline{\tau_{n+1}}\subset \tau_n$ by definition. Otherwise, $F_n\in\mathcal{F}$, $F_n\subset \unbdd{F}_n \cup A_{F_n}$ and $\overline{F_n}\subset \WW \cup A_{F_n}$, where $A_{F_n}\subset V$ is the connected set from Lemma~\ref{}. Recall that $\tau_{n+1}$ is the connected component of $f^{-n}(F_n)$ containing $\zeta_n$, that $\tau_n$ is the connected component of $f^{-n}(\WW)$ containing $\zeta_n$, and that $V_n$ contains the connected component of $f^{-n}(A_{F_n})$ containing $\zeta_n$. This proves~\ref{}. Now assume that all fundamental domains $F_n$ occuring in ${\underline{s}}$ belong to $\mathcal{F}$. Let $n\geq 0$. To prove~\ref{}, recall from Lemma~\ref{} that we can connect $\zeta$ to a point $x$ of $\unbdd{F}_{n+1}$ by a curve $\gamma_1\subset V\setminus\overline{D}$, and then furthermore connect this point $x$ to $\zeta_{F_{n+1}}$ by a curve $\gamma_2\subset V\cap F_{n+1}$. The curve $\tilde{\gamma}_1\defeq f_{{\underline{s}}}^{-n}(\gamma_1)\subset \tau_n\cap V_n$ connects $\zeta_n$ to the point \[ \tilde{x} \defeq f^{-n}_{{\underline{s}}}(x)\in f^{-n}_{{\underline{s}}}(\unbdd{F}_n)\subset \tau_{n+1}({\underline{s}}). \] Now let $\tilde{\gamma_2}$ be the connected component of $f^{-n}(\gamma_2)$ containing $\tilde{x}$. This curve connects $\tilde{x}$ to a point $\tilde{\zeta}_{n+1}$ of $f^{-(n+1)}(\zeta)$. Since $\tilde{\gamma_2}\subset \tau_{n+1}({\underline{s}})\cap V_n$, we must have $\zeta_{n+1} = f_{{\underline{s}}}^{-(n+1)}(\zeta) = \tilde{\zeta}_{n+1}\in V_n$. This proves~\ref{}. For $n\geq 0$, let $X_n\subset \mu_{\sigma^n({\underline{s}})}\subset G_{\sigma^n}({\underline{s}})$ be a closed unbounded connected set as in Theorem~\ref{}. By Theorem~\ref{}, if $R$ is large enough (depending only on $\mathcal{F}$), then $X_n$ can be chosen to contain a point of radius at most $R$. In particular, $X_n$ intersects the set $A_{F_n}$ from Lemma~\ref{}. So if $\tilde{X_n}\subset \mu_{{\underline{s}}}$ is the connected component of $f^{-n}(X_n)$ contained in $\tau_{n+1}$, then $\tilde{X_n}\cap V_n\neq \emptyset$. This proves~\ref{}. Finally, suppose that $z\in A_n \defeq G_{{\underline{s}}}\setminus \bigcup_{j\geq n} V_j$. Since $z\in G_{{\underline{s}}}$, there is $m_0$ such that $f^n(z)\in \tail_m$ for $m\geq m_0$. By assumption and~\ref{}, we can in fact take $m_0 = n$. So, for $m> n$, we have $f^n(z) \in F_n$ and $\lvert f^n(z)\rvert > R$. So $f^n(A_n)$ belongs to the set from Theorem~\ref{}, on which the iterates tend to infinity uniformly (assuming $R$ is chosen sufficiently large). The same holds for $f^n(\overline{A_n})$; in particular, this set is contained in~$G_{\sigma^n({\underline{s}})}$. \end{proof} \begin{proof}[Proof of Proposition~\ref{}] Let $\mathcal{F}$ be a finite collection of fundamental domains containing all fundamental domains occurring in ${\underline{s}}$, and all fundamental domains whose closure intersects $\overline{D}$. Let $\zeta$, $V$ and $V_n$ be as in Lemma~\ref{}. {Here we assume that $R$ is chosen at least as large as the numbers from Theorem~\ref{} and~\ref{}.} By~\ref{}, \[ \Lambda({\underline{s}}) = \bigcap_{N\geq 1} \operatorname{cl}_{\Ch}\left(\bigcup_{n\geq N} V_n\right). \] Each of the sets in the intersection on the right is compact and connected by~\ref{} of Lemma~\ref{}; claim~\ref{} follows. We next prove the technical claim~\ref{}, from which we then deduce the remaining facts. First let $U$ be a neighbourhood of some point $z_0\in \Lambda({\underline{s}})$. By~\ref{}, there are infinitely many $n$ such that $V_n\subset U$. By definition, all points in $V_n$ map to the bounded set $V$ after $n$ iterations, and $V_n$ contains a point of $\mu_{{\underline{s}}}$ by~\ref{}. Hence $f^n$ does not tend to infinity uniformly on $\mu_{{\underline{s}}}\cup U$. On the other hand, suppose that $U$ intersects $\hat{G}_{{\underline{s}}}$, and hence $G_{{\underline{s}}}$, but $\hat{U}\cap \Lambda({\underline{s}})=\emptyset$. Then there is $N\geq 1$ such that \[ \hat{U} \cap \overline\bigcup_{n\geq N} V_n = \emptyset. \] So $f^n$ tends to infinity uniformly on $\hat{U}\cap \overline{G_{{\underline{s}}}} = \hat{U}\cap G_{{\underline{s}}}$ by~\ref{}. This completes the proof of~\ref{}. Now we can establish the remaining claims in Proposition~\ref{}. Using~\ref{} and~\ref{}, it follows that every point of $\Lambda({\underline{s}})$ is in $\hat{G}_{{\underline{s}}}$. Conversely, by~\ref{}, every point of $\overline{G_{{\underline{s}}}}\setminus \Lambda({\underline{s}})$ belongs to $G_{{\underline{s}}}$. This proves~\ref{}. Next suppose that $z_0\in J_{{\underline{s}}} \setminus G_{{\underline{s}}}$; that is, $z_0$ has address ${\underline{s}}$ but is not escaping. {There is $n_0$ such that $f^n(z_0)\in J_{\sigma^n(z)}^0$ for $n\geq n_0$. Hence \(\lvert f^n(z_0)\rvert < R \) by Theorem~\ref{}. So $f^n(z_0)\in V$ for $n\geq n_0$, and thus $z_0\in V_n$.} By~\ref{}, this proves $\Lambda({\underline{s}})= \{z_0\}$, as claimed. It remains to prove~\ref{}. Recall from Lemma~\ref{} that, if ${\underline{s}}$ is of disjoint type, then $J_{\sigma^n({\underline{s}})} = \overline{G_{\sigma^n({\underline{s}})}}\subset F_n$ for all $n\geq 0$. {Let $X$ be the set from~\ref{}. Then there is $n_0$ such that $\lvert f^n(z)\rvert > R$ for all $n\geq n_0$ and all $z\in X$. Set \[ \tilde{R}\defeq \max\{ \lvert f^j(z)\rvert \colon \lvert z\rvert \leq R\text{ and } j \leq n_0\}. \] By Theorem~\ref{}, for all $n\geq 0$ there is $\zeta_n\in J_{{\underline{s}}}$ such that $\lvert f^n(\zeta_n)\rvert \leq R$. We claim that $\lvert f^j(\zeta_n)\rvert \leq \tilde{R}$ for all $n\geq n_0$ and all $j\leq n$. Indeed, let $j$ be minimal such that $\lvert f^j(\zeta_n)\rvert >R$ (if no such $j$ exists, there is nothing to prove). Then $f^j(\zeta_n)\in X$, and hence we must have $j> n-n_0$, and the claim follows by the definition of $\tilde{R}$.} Let $z_0\in J_{{\underline{s}}}$ be a limit point of the sequence $(\zeta_n)$; then all points on the orbit of $z_0$ have modulus at most $\tilde{R}$. The claim now follows from~\ref{}. \end{proof} \begin{rmk}[Coding tree] Fix $\mathcal{F}$, $\zeta$ and $V$ as in Lemmas~\ref{} and~\ref{}. For each $F\in\mathcal{F}$, we can choose an arc $\gamma_F$ connecting $\zeta$ to the point $\zeta_F$, and we may assume that these arcs are disjoint except at $\zeta$. For each $\zeta_{F_0}$ and each arc $\gamma_{F_1}$, there is a component of $f^{-1}(\gamma_{F_1})$ connecting $\zeta_{F_0}$ to some point $\zeta_{F_0 F_1}$ of $f^{-2}(\zeta)$. By Lemma~\ref{}, this is precisely the point contained in the fundamental tail at address $F_0 F_1$. Continuing inductively, we obtain an infinite tree with root $\zeta$, whose vertices of depth $n>0$ are the elements of $f^{-(n)}(\zeta)$ contained in fundamental domains of level $n-1$ whose addresses contain only entries from $\mathcal{F}$, and whose edges are all components of $f^{-n}(\gamma_F)$ for some $F\in\mathcal{F}$. Recall that the spherical length of these edges tends to zero as $n\to\infty$. This tree can be considered to be an analogue of the \emph{geometric coding tree} used by Przytycki \cite{Pr94} in the case of rational functions. We see that, for each address ${\underline{s}}$ whose entries are drawn from $\mathcal{F}$, the accumulation set of the dreadlock $G_{{\underline{s}}}$ coincides precisely with the accumulation set of a branch of this coding tree. However, we will not use this language in the following. \end{rmk} \section{Landing theorems for dreadlocks}\label{} With the notation from Section 2 and Section 3~-- in particular, the definition of dreadlocks in Definition~\ref{} and their accumulation sets in Definition~\ref{ we can now state the main result of our paper. Recall that a dreadlock $G_{{\underline{s}}}$ is \emph{periodic} if the address ${\underline{s}}$ is periodic under the shift map. Equivalently, $G_{{\underline{s}}}$ is periodic under the action of $f$ as a subset of $\C$. \begin{thm}[Douady-Hubbard landing theorem for dreadlocks]\label{} Let $f$ be a transcendental entire function with bounded post-singular set. Then every periodic dreadlock of $f$ lands at a repelling or parabolic periodic point, and conversely every repelling or parabolic periodic point of $f$ is the landing point of at least one and at most finitely many periodic dreadlocks. \end{thm} We also obtain a corresponding theorem about landing properties at more general hyperbolic sets. \begin{thm}[Landing at hyperbolic sets]\label{} Let $f$ be a transcendental entire function with bounded postsingular set. Moreover, suppose that $K$ is a hyperbolic set for $f$. Then every point $z\in K$ is the landing point of a dreadlock at some bounded external address ${\underline{s}}$; if $z$ is periodic, then so is ${\underline{s}}$. \end{thm} In fact, we establish the following more precise version of Theorem~\ref{}. Note that the space of external addresses is equipped with the product topology; this is the same as the order topology arising from the cyclic order on addresses. \begin{thm}[Accessibility of points in hyperbolic sets]\label{} Let $K$ be a hyperbolic set of a postsingularly bounded entire function. Then there exists a compact and forward-invariant set $\mathcal{S}_K$ of bounded addresses of $f$, with the following properties. \begin{enumerate}[(a)] \item\label{ For every ${\underline{s}}\in \mathcal{S}_K$, the dreadlock at address ${\underline{s}}$ lands at a point $z_0({\underline{s}})\in K$. \item\label{ The function $\mathcal{S}_K\to K; {\underline{s}}\mapsto z_0({\underline{s}})$ is surjective and continuous. (In particular, every point of $K$ is the landing point of a dreadlock). \item\label{ If $z_0$ is periodic, then all bounded-address dreadlocks landing at $z_0$ are periodic with the same period, and the number of such dreadlocks is finite. \item\label{} The dreadlocks at addresses in $\S_K$ land uniformly at $K$, in the following sense. Let $\zeta\in\WW$ and let $\eps>0$. Then there is $n_0$ such that $\dist(\zeta_n({\underline{s}}),K)\leq \eps$ for all $n\geq n_0$ and all ${\underline{s}}\in \S_K$. (Recall from Definition~\ref{} that $\zeta_n({\underline{s}}) = f_{{\underline{s}}}^{-n}(\zeta)$.) \end{enumerate} \end{thm} The remainder of the paper will be dedicated to the proofs of these theorems. Let us first show that they indeed imply the theorems stated in the introduction. \begin{proof}[Proof of Theorem~\ref{}, using Theorem~\ref{}] Let $\zeta$ be a repelling or parabolic periodic point. By Theorem~\ref{}, $\zeta$ is the landing point of a periodic dreadlock $G_{{\underline{s}}}$. Let us set $A \defeq G_{{\underline{s}}}$, and let $p$ be the period of {${\underline{s}}$. Since $G_{\underline{s}}$ lands at $\zeta$, we see from Corollary~\ref{} that $\overline{A}=A\cup\{\zeta\}$ and that, for any neighborhood $U$ of $\zeta$, $f^n\to\infty$ uniformly on $A \setminus U$. This proves that $A$ satisfies properties~\ref{} and~\ref{} of Theorem~\ref{}. Since $p$ is the period of ${\underline{s}}$, we have $f^p(A)=G_{\sigma^p({\underline{s}})}=A$ and $f^j(A)\cap A=G_{\sigma^j({\underline{s}})}\cap G_{{\underline{s}}} = \emptyset$ if $0<j< p.$} \end{proof} \begin{proof}[Proof of Theorem~\ref{}, using Theorem~\ref{}] Recall that every periodic hair is a periodic dreadlock, and for criniferous functions, every periodic dreadlock is a periodic hair. Moreover, {by Corollary~\ref{}} such a dreadlock lands if and only if the corresponding hair lands in the sense of Definition~\ref{}. Thus Theorem~\ref{} follows immediately from Theorem~\ref{}. \end{proof} \begin{proof}[Proof of Theorem~\ref{}, using Theorem~\ref{}] Define $\mathcal{A}\defeq \{G_{{\underline{s}}}\colon {\underline{s}}\in \mathcal{S}_K\}$, where $\mathcal{S}_K$ is as in Theorem~\ref{}. Properties~\ref{} and~\ref{} in Theorem~\ref{} follow immediately, as in the proof of Theorem~\ref{}. Furthermore, if $f$ is {{criniferous}}, then every element of $\mathcal{A}$ is an arc connecting its landing point $\zeta(A)$ to $\infty$, by Corollary~\ref{}. Thus it remains to establish~\ref{}. Observe that, since $\mathcal{S}_K$ is compact and forward-invariant, only finitely many different fundamental domains can occur within the addresses in $\mathcal{S}_K$. Let $\mathcal{F}$ be the collection of all these fundamental domains, as well as of all fundamental domains whose closure intersects $\overline{D}$. Let $V$ be as in Lemma~\ref{}, and let $n_0$ be as in part~\ref{} of Theorem~\ref{}; for $\eps/2$. By Lemma~\ref{}, we can find $n_1\geq n_0$ such that $\diam V_n({\underline{s}}) < \eps/2$ for $n\geq n_1$. Thus $\dist(z, K)<\eps$ for all $z\in V_n({\underline{s}})$. Let $z\in G_{{\underline{s}}}$ for some ${\underline{s}}\in\mathcal{S}_K$, and suppose that $\dist(z,K)\geq \eps$. Then, by the above $z \in G_{{\underline{s}}}\setminus \bigcup_{j\geq n_1}V_j({\underline{s}})$. By~\ref{} of Lemma~\ref{}, the set of points with this property escape to infinity at a rate that depends only on $\mathcal{F}$ and on $n_0$ (and hence only on $\eps$). This completes the proof. \end{proof} \begin{rmk}[Shortcut to the landing theorems]\label{} Observe that Definition~\ref{}, concerning landing and accumulation properties, requires only the definition of the fundamental tails $\tau_n({\underline{s}})$ associated to an address ${\underline{s}}$, rather than any properties of the dreadlock $G_{{\underline{s}}}$ itself. Moreover, a key point in our proofs of Theorems~\ref{} and~\ref{} is that we can ignore the fine structure of dreadlocks and indeed use only the sets $V_n$ from Lemma~\ref{}. We only require properties~\ref{}, \ref{} and~\ref{} of this lemma, all of which are likewise independent of the construction and analysis of dreadlocks in Sections~\ref{} and~\ref{}. Hence it would be possible to prove these theorems without requiring any results from these sections, and only the elementary parts of Section~\ref{}. Furthermore, while we used properties of dreadlocks above to deduce the statements of Theorems~\ref{} and~\ref{}, many of these can be established more easily \emph{a posteriori} for the dreadlocks under consideration, using the additional information that these dreadlocks land. Nonetheless, the general material concerning dreadlocks is crucial to the interpretation of our results as a natural extension of the classical Douady-Hubbard landing theorem for functions without hairs. For criniferous functions, we require only Corollary~\ref{} in order to deduce Theorem~\ref{} from Theorem~\ref{}, and to deduce Theorem~\ref{} for criniferous functions from Theorem~\ref{}. Recall that, for periodic hairs and general hairs of functions satisfying a head-start condition, the proof of Corollary~\ref{} becomes particularly simple. \end{rmk} \section{Periodic dreadlocks land}\label{Periodic dreadlocks land} We now prove the first half of Theorem~\ref{}, which we restate here as follows. \begin{thm}[Landing of periodic dreadlocks]\label{} Let $f\colon\C\to\C$ be a transcendental entire function with bounded postsingular set. Then every periodic dreadlock of $f$ lands at a repelling or parabolic periodic point of $f$, where the period of the landing point divides that of the dreadlock. \end{thm} \begin{proof}[Proof of Theorem~\ref{}] Let $G_{\underline{s}}$ be a periodic dreadlock. Recall from Observation~\ref{} that any dreadlock of $f$ is also a dreadlock of any iterate of $f$ (and vice versa). Hence, it is no loss of generality to assume that ${\underline{s}}$ is fixed by $\sigma$; i.e., ${\underline{s}}= F F F F \dots$, where $F$ is a fundamental domain of $f$. Fix $\zeta$ and $V$ as in Lemmas~\ref{} and \ref{}. Let $\tail_n$ be the fundamental tail of level $n$ associated to ${\underline{s}}$, and let $\zeta_n$ be the unique element of $f^{-n}(\zeta)$ in $\tail_n$. Recall that $\zeta_1 = \zeta_F \in V$, and let $\gamma\colon [0,1]\to V$ be a rectifiable curve with $\gamma(0)=\zeta_1$ and $\gamma(1)=\zeta$. Then, by Lemma~\ref{}, for every $n\geq 1$ there is a component $\gamma_n$ of $f^{-n}(\gamma)$ connecting $\zeta_n$ and $\zeta_{n+1}$. We can hence combine these to a continuous curve $\gamma\colon [-\infty,1]\to\C$, with $f(\gamma(t))=\gamma(t+1)$ for $t\leq 0$, where $\gamma|_{[0,1]} = \gamma$, and $\gamma|_{[-n,1-n]} = \gamma_n$. By Lemma~\ref{}, the spherical diameter of $\gamma_n$ tends to zero as $n\to\infty$. Therefore the set $\Lambda(\gamma)$ of accumulation points of $\gamma(t)$ as $t\to -\infty$ is precisely the accumulation set $\Lambda(G_{\underline{s}})=\Lambda({\underline{s}},\zeta)$ of $G_{\underline{s}}$ in the sense of Definition~\ref{}. \begin{claim} As $t\to -\infty$, $\gamma(t)$ converges to a fixed point of $f$. \end{claim} \begin{subproof} Recall that the spherical distance between $\zeta_n$ and $\zeta_{n+1}$ tends to zero as $n\to \infty$, and $f(\zeta_{n+1})=\zeta_n$. Hence, by continuity, any point of $\Lambda(\gamma) = \Lambda({\underline{s}},\zeta)$ is a fixed point of $f$. Since the set of fixed points is discrete in $\C$, and since $\Lambda(\gamma)$ is connected, the latter set is a singleton, whose sole element is either a fixed point or $\infty$. We must exclude the second possibility. So suppose, by contradiction, that $\gamma(t)\to\infty$ as $t\to-\infty$. Recall that $\Omega$ is the unbounded connected component of $\C\setminus \P(f)$. Let $\rho_n\defeq \ell_{\Omega}(\gamma_n)$ be the hyperbolic length of $\gamma_n$ in $\Omega$. By Proposition~\ref{} and by assumption, we have \[ \rho_{n+1} \leq \rho_n/2 \] for all sufficiently large $n$. It follows that the hyperbolic length of $\gamma$ is bounded. As the hyperbolic metric on $\Omega$ is complete, this contradicts our assumption that $\gamma$ tends to $\infty$. \end{subproof} The classical snail lemma of Douady and Sullivan (see \cite[Expos\'e VIII, Proposition 2, p. 60]{DH84} or \cite[Lemma 16.2]{Mi}) shows that the limiting fixed point of $f$ is either repelling or parabolic with multiplier $1$, and the proof is complete. \end{proof} {\begin{remark} In fact, the above claim follows already from~\cite[Theorems~B.1 and~B.2]{Re08}. Here Theorem~B.1 is a hyperbolic expansion argument going back to the proof by Douady and Hubbard~\cite[Expos\'e VIII]{DH84} of the first half of the landing theorem for polynomials. On the other hand, the proof of \cite[Theorem~B.2]{Re08}, which shows that $\gamma(t)$ cannot converge to infinity as $t\to-\infty$, used the notion of ``extendability'', which was developed for more general purposes in \cite{Re08}. For the reader's convenience, we gave the complete and much simpler proof above, in the spirit of Deniz~\cite{De}. \end{remark}} {It is possible that the landing point in Theorem~\ref{} is also the landing point of other dreadlocks. However, as we now observe, there can only be finitely many of these, and they all need to be periodic. The idea of the proof is very similar to the polynomial case \cite[Lemma~18.12]{Mi}, but we need to take into account the non-compactness of the space of addresses. Compare also~\cite[Lemma~3.2]{expcombinatorics} for the case of exponential maps.} \begin{lem}\label{ for all} Let $f$ be a transcendental entire function with bounded postsingular set, and let $z$ be the landing point of a dreadlock with periodic address. Then the number of bounded-address dreadlocks landing at $z$ is finite, and their addresses are all periodic of the same period. \end{lem} \begin{proof} {Let ${\underline{s}}^0$ be the address of the periodic dreadlock landing at $z$; we may assume that its period $p$ is minimal with this property. Passing to an iterate, we may then assume that $p=1$. Let $\mathcal{S}_z$ be the set of bounded external addresses ${\underline{s}}$ for which $G_{{\underline{s}}}$ lands at $z$. \begin{claim} There is a finite collection $\mathcal{F}$ of fundamental domains with the following property. Every address ${\underline{s}}\in\mathcal{S}_z$ contains some element of $\mathcal{F}$ infinitely many times. \end{claim} \begin{subproof} Let $\mathcal{F}$ consist of all fundamental domains $F$ such that either $\overline{D}\cap \overline{F}\neq\emptyset$, or $z\in F$. Clearly $\mathcal{F}$ is finite. Suppose that ${\underline{s}} = F_0 F_1 F_2 \dots$ is such that $F_n\notin \overline{F}$ for $n\geq n_0$. Then $\sigma^{n_0}({\underline{s}})$ is of disjoint type, and $z\notin F_{n_0} \supset f^{n_0}(\overline{G_{{\underline{s}}}})$. In particular, ${\underline{s}}$ does not land at $z$. So any address in $\mathcal{S}_z$ contains infinitely many entries from $\mathcal{F}$; since the latter set is finite, at least one of these is itself repeated infinitely many times. \end{subproof} Since $f$ maps a neighborhood of $z$ diffeomorphically into another neighborhood of $z$, $f$ preserves the cyclic order of these $G_{{\underline{s}}}$ at $z$. As remarked at the end of Section~\ref{}, this implies that $f$ also preserves the cyclic order of these dreadlocks at $\infty$. In other words, the shift map $\sigma\colon \mathcal{S}_z \to \mathcal{S}_z$ is injective and preserves the cyclic order of addresses on $\mathcal{S}_z$. Recall that ${\underline{s}}^0$ is a fixed address, say ${\underline{s}}^0 = F^0 F^0 F^0 \dots$. Hence $\sigma$ also preserves the (linear) order $<$, where ${\underline{s}} < \tilde{{\underline{s}}}$ means that ${\underline{s}}^0\prec {\underline{s}} \prec \tilde{{\underline{s}}}$ in the cyclic ordering. So if ${\underline{s}} = F_0 F_1 F_2\dots$ is an element of $\mathcal{S}_z$, then $(\sigma^n({\underline{s}}))_{n=0}^{\infty}$ is monotone. If $F\in\mathcal{F}$ is the domain from the claim above, then we clearly must have $F_n = F$ for all sufficiently large $n$. By injectivity of $\sigma$, we conclude that ${\underline{s}} = F F F \dots$ is itself a fixed address. As $\mathcal{F}$ is finite, the proof is complete.} \end{proof} \begin{rmk}\label{} Recall that we defined dreadlocks only for maps with bounded postsingular sets, and landing of dreadlocks only for dreadlocks at bounded addresses. This allows us to state the lemma in the above simple form, which is all that will be required for the purpose of our main results. However, observe that the proof of the lemma is purely combinatorial, and does not utilise either assumption in an essential manner. In particular, let $f\in\mathcal{B}$ be arbitrary (not necessarily with bounded postsingular set), and let $z_0\in J(f)$ be a fixed point of $f$. Suppose that $\mathcal{S}_z$ is a forward-invariant set of external addresses ${\underline{s}}$ of $f$, and that for every ${\underline{s}}\in\mathcal{S}$ there is an unbounded connected set $A_{{\underline{s}}}$ with the following properties: \begin{itemize} \item $z_0\in\overline{A_{{\underline{s}}}}$ for all ${\underline{s}}\in \mathcal{S}_z$; \item $\overline{A_{{\underline{s}}}}\cap A_{\tilde{{\underline{s}}}}=\emptyset$ for ${\underline{s}}\neq \tilde{{\underline{s}}}$; \item $\overline{A_{{\underline{s}}}}$ does not separate the plane; \item $f(A_{{\underline{s}}}) = A_{\sigma({\underline{s}})}$. \end{itemize} If $\mathcal{S}$ contains a periodic element, then it follows as above that $\mathcal{S}_z$ is finite and contains only periodic addresses. In particular, for any $f\in\mathcal{B}$, the landing point of a periodic hair cannot be the landing point of a non-periodic hair. \end{rmk} \section{Landing at hyperbolic sets}\label{ at hyperbolic sets} \begin{proof}[Proof of Theorem~\ref{}] Let $K$ be a hyperbolic set of $f$. Replacing $f$ by a sufficiently high iterate, there is a neighbourhood $U$ of $K$ such that $\lvert f'(z)\rvert \geq 2$ for all $z\in U$. (It is easy to see that proving the theorem for an iterate of $f$ also establishes it for $f$ itself; we leave the details to the reader.) We may additionally assume that the disc $D$ in the definition of fundamental domains is chosen so large that $K\subset D$. Set $\delta \defeq \dist(K, \partial U)$. For $z\in K$, define $B_0(z)\defeq B(z,\delta)$, and let $B_n(z)$ denote the connected component of $f^{-n}(B_0(f^n(z)))$ containing $z$. Then $\overline{B_{n+1}(z)}\subset B_n(z)$ for all $n$, and $f\colon B_{n+1}(z)\to B_{n}(f(z))$ is a conformal isomorphism. For $n\geq 0$, define \[ U_n \defeq \bigcup_{z\in K} B_n(z). \] Then $\overline{U_{n+1}}\subset U_n\subset U$ and $f(U_{n+1})=U_n$ for all $n$. Clearly $K=\bigcap_{n\geq 0} \overline{U_n} = \bigcap_{n\geq 0} U_n$. Fix $\zeta\in W_0$. By the blowing-up property of the Julia set (see e.g.\ \cite[Lemma 2.2]{Ba84}), and compactness of $K$, there is some $N_1$ with the following property. If $n\geq N_1$ and $z\in K$, then $f^{-n}(\zeta)\cap B_0(z)\neq\emptyset$. In particular, for each $z\in K$ there is a finite external address of length $n$ such that $\zeta_n({\underline{s}})\in B_0(z)$, and hence $\tail_n({\underline{s}})\cap B_0(z)\neq \emptyset$. Define \[ \eps \defeq \frac{\delta}{2} \leq \min_{z\in K} \dist(\partial B_0(z) , \overline{B_1(z)}). \] Now let $V$ be as in Lemma~\ref{} for the finite collection $\mathcal{F}$ of fundamental domains that intersect $D$. Similarly as in Lemma~\ref{}, if ${\underline{s}}$ is an infinite external address or a finite external address of length at least $n$, we define $V_{n}({\underline{s}})$ to be the unique component of $f^{-n}(V)$ containing $\zeta_n({\underline{s}})\defeq f^{-n}_{{\underline{s}}}(\zeta)\in \tail_n({\underline{s}})$. By Lemma~\ref{}, there exists $N\geq N_1$ with the following property. Suppose that $n\geq N$ and that ${\underline{s}}$ is a finite or infinite external address of length $n\geq N$ with $V_n({\underline{s}})\cap \overline{U_0} \neq \emptyset$. Then \begin{equation} \diam V_{n}({\underline{s}})< \eps. \end{equation} For $n\geq N$ and $z\in K$, we now define $\S_n(z)$ to be the set of finite external addresses ${\underline{s}}$ of length $n$ for which the tail $\tau_{n}({\underline{s}})$ intersects $B_{n-N}(z)$. Observe that \[ \S_n\defeq \bigcup_{z\in K} \S_n(z)\] is finite for every $n$ by Lemma~\ref{}. We also define $\S_K$ to be the set of infinite addresses ${\underline{s}}$ such that every prefix of length $n\geq N$ of ${\underline{s}}$ is an element of $\S_n$. In order to show that this set has the properties asserted in Theorem~\ref{}, we investigate the sets $\S_n(z)$ more closely. \begin{claim}[Claim 1] The following hold for all $n\geq N$ and all $z\in K$. \begin{enumerate}[(i)] \item $\S_N(z)\neq \emptyset$.\label{item:B1} \item The shift map $\sigma$ maps $\S_{n+1}(z)$ onto $\S_n(f(z))$.\label{ \item $\S_n(z)\neq\emptyset$.\label{} \item Suppose that ${\underline{s}}=F_0 F_1 \dots F_{n}\in \S_{n+1}(z)$. Let $\pi_n({\underline{s}})$ be the prefix of length $n$ of ${\underline{s}}$; i.e., $\pi_n({\underline{s}}) = F_0 F_1 \dots F_{n-1}$. Then $\pi_n({\underline{s}})\in \S_n(z)$.\label{} \item In particular, $\pi_N(\sigma^j({\underline{s}}))\in \S_N\bigl(f^j(z)\bigr)$ for $j=0,\dots,n-N$.\label{} \end{enumerate} \end{claim} \begin{subproof} The first claim holds by choice of $N_1$. Part~\ref{} is immediate from the fact that $f\colon B_{n+1-N}(z)\to B_{n-N}(f(z))$ is a conformal isomorphism. Claim~\ref{} follows from these first two claims by induction. Now let us prove~\ref{}. So suppose that ${\underline{s}}\in \S_{n+1}(z)$, and let $w\in \tail_{n+1}({\underline{s}})\cap B_{n+1-N}(z)$. Recall from Lemma~\ref{} that $\tail_{n+1}({\underline{s}})\subset \unbdd{\tail}_{n+1}({\underline{s}}) \cup V_{n}({\underline{s}})$. First consider the case where $w\in \unbdd{\tail}_{n+1}({\underline{s}})$. Then ${\underline{s}}\in \S_{n}(z)$ since $w\in \unbdd{\tail}_{n+1}({\underline{s}})\cap B_{n+1-N}(z)\subset \tail_{n}({\underline{s}}) \cap B_{n-N}(z)$. Now suppose that $w\in V_{n}({\underline{s}})$. By definition, \begin{equation}\label{} \zeta_n({\underline{s}}) \in V_{n}({\underline{s}})\cap \tail_{n}({\underline{s}}).\end{equation} Set $\tilde{{\underline{s}}} \defeq\sigma^{n-N}({\underline{s}})$, $\tilde{z}\defeq f^{n-N}(z)$ and $\tilde{w}\defeq f^{n-N}(w)$. Then $\tilde{w}\in V_{N}(\tilde{{\underline{s}}})\cap B_1(\tilde{z})$. By definition of $\eps$ and choice of $N$, it follows that $V_N(\tilde{{\underline{s}}})\subset B_0(\tilde{z})$. Now $B_{n-N}(z)$ is mapped conformally to $B_0(\tilde{z})$ by $f^{n-N}$. Since $V_n({\underline{s}})$ is a connected component of $f^{N-n}(V_N(\tilde{{\underline{s}}}))$ and intersects $B_{n-N}(z)$, we see that $B_{n-N}(z)\supset V_{n}({\underline{s}})\ni \zeta_n({\underline{s}})$. By~\eqref{}, we see that $\tail_{n}({\underline{s}})\cap B_{n-N}(z)\neq \emptyset$. This proves~\ref{}). The final claim~\ref{} follows by induction from~\ref{} and~\ref{}. \end{subproof} \begin{claim}[Claim 2] There is a finite collection $\tilde{\mathcal{F}}\supset\mathcal{F}$ of fundamental domains such that all entries of addresses in $\bigcup_{n\geq N} \S_n$ are in $\tilde{\mathcal{F}}$. \end{claim} \begin{subproof} Let $\tilde{\mathcal{F}}$ be obtained by adding to $\mathcal{F}$ all fundamental domains appearing in the finitely many external addresses in $\S_N$. By part~\ref{} of Claim 1, it follows that all entries of ${\underline{s}}\in \S_n$ belong to $\tilde{\mathcal{F}}$, for all $n\geq N$. \end{subproof Applying Lemma~\ref{} again, this time to the collection $\tilde{\mathcal{F}}$, we obtain a simply-connected domain $\tilde{V}$ that we can use to study the accumulation sets of the addresses in $\S_K$. \begin{claim}[Claim~3] For every $k\geq 0$ there is an $n_0\geq N$ with the following property. If $z\in K$, $n\geq n_0$ and ${\underline{s}}\in \S_n(z)$, then $\tilde{V}_n({\underline{s}})\subset B_k(z)$. \end{claim} \begin{subproof} By assumption, $K\subset D$. In particular, there exists $M>0$ such that no fundamental tail of level $N$ intersects the neighbourhood $U_M$ of $K$. Let $n\geq N+M$, let $z\in K$, and let ${\underline{s}}\in \S_n(z)$. Then \[ \tau_n({\underline{s}}) \subset \tau_N({\underline{s}}) \cup \bigcup_{\tilde{n}=N}^{n-1} V_{\tilde{n}}({\underline{s}}) \] by Lemma~\ref{}. Since $\tau_n({\underline{s}})$ intersects $B_{n-N}(z)\subset U_M$ by definition of $\S_n$, while $\tau_N$ is disjoint from $U_M$, there is $\tilde{n}\in \{N,\dots,n-1\}$ such that $V_{\tilde{n}}({\underline{s}}) \cap B_{n-N}(z)\neq \emptyset$. Let $\tilde{n}$ be maximal with this property. We claim that $\tilde{n}\geq n-M$. This is trivial when $n=N+M$. Now suppose, inductively, that $\tilde{V}_{\tilde{n}}(\sigma({\underline{s}}))\cap B_{n-1-N}(f(z))\neq \emptyset$, with $\tilde{n}\geq n-M$. Now $f$ maps $\tilde{V}_{\tilde{n}+1}({\underline{s}})$ conformally to $\tilde{V}_{\tilde{n}}(\sigma({\underline{s}}))$, and $B_{n-N}(z)$ to $B_{n-1-N}(f(z))$. Hence $\tilde{V}_{\tilde{n}+1}({\underline{s}}) \cap B_{n-N}(z)\neq \emptyset$, and $\tilde{n}+1 \geq n+1-M$, as required. Let $k\geq 0$. If $n_1\geq N$ is sufficiently large, then by Lemma~\ref{}, \[ \diam(\tilde{V}_n({\underline{s}})) < \eps_k \defeq \frac{\min_{z\in K} \dist(B_{k+1}(z),\partial B_k(z))}{M+1} \] whenever $n \geq n_1$, ${\underline{s}}\in \S_{n}(z)$ for some $z\in K$, and $\tilde{V}_n({\underline{s}})\cap U_0\neq \emptyset$. Set \[ n_0\defeq \max(n_1 + M , k+N+1) \] and suppose that $z$, $n$ and ${\underline{s}}$ are as in the statement of Claim~3. Let $\tilde{n}$ be as above; i.e., $\tilde{n}$ is maximal such that $\tilde{V}_{\tilde{n}}({\underline{s}})\cap B_{n-N}(z)\neq \emptyset$. Recall that $\tilde{n}\geq n-M\geq n_1$. Since $n-N\geq n_0 - N \geq k+1$, we see that $\tilde{V}_{\tilde{n}}({\underline{s}})$ intersects $B_{k+1}(z)$. It follows inductively for $j=\tilde{n}, \tilde{n}+1, \dots , n$ that, for all $\zeta \in \tilde{V}_j({\underline{s}})$, \[ \dist(\zeta , B_{k+1}(z)) < (j+1-\tilde{n})\cdot \eps_k \leq \dist(B_{k+1}(z) , \partial B_k(z)). \] (In the inductive step, we use that $\tilde{V}_{j}({\underline{s}})\cap \tilde{V}_{j-1}({\underline{s}})\neq\emptyset$ by Claim~2 and Lemma~\ref{}, and that $j\geq n_1$.) In particular, $\tilde{V}_n({\underline{s}})\subset B_k(z)$, as required. \end{subproof} \begin{claim}[Claim~4] Let $n_0$ be as in Claim~3, for $k=1$. Then $\sigma\colon \S_{n+1}(z) \to \S_n(z)$ is a bijection for all $n\geq n_0$ and $z\in K$. \end{claim} \begin{subproof} By Claim~1~\ref{}, it remains to show that $\sigma\colon \S_{n+1}(z)\to \S_n(z)$ is injective. Let $z\in K$ and let ${\underline{s}}^1,{\underline{s}}^2\in \S_{n+1}(z)$ with $\sigma({\underline{s}}^1) =\sigma({\underline{s}}^2) \eqdef \tilde{{\underline{s}}}$. Then, for $j=1,2$, we know that $\tilde{V}_{n+1}({\underline{s}}^j)\subset B_1(z)$ by choice of $n_0$, and $\tilde{V}_n(\tilde{{\underline{s}}}) = f(\tilde{V}_{n+1}({\underline{s}}^j)) \subset B_0(f(z))$. Since $f\colon U_1(z)\to U_0(f(z))$ is univalent, it follows that $\tilde{V}_{n+1}({\underline{s}}^1)=\tilde{V}_{n+1}({\underline{s}}^2)$, and hence ${\underline{s}}^1={\underline{s}}^2$, as required. \end{subproof} Now consider the directed graph $G$ whose vertices are the elements of \[ \mathcal{V}(G) \defeq \bigcup_{n\geq N} \S_n, \] and which contains an edge from $\pi_n({\underline{s}})$ to ${\underline{s}}$ for every ${\underline{s}} \in \S_{n+1}$. For $z\in K$, let $G_z$ be the induced subgraph of $G$ whose vertices are the elements of $\bigcup_{n\geq N} \S_n(z)$. By~\ref{} and~\ref{} of Claim~1, we can apply K\"onig's lemma, and $G_z$ contains an infinite path for every $z\in K$. Recall that $\S_K$ is the set of infinite external addresses ${\underline{s}}$ such that $\pi_n({\underline{s}})\in \S_n$ for all $n\geq N$. If ${\underline{s}}\in \S_K$, the sequence $(\pi_n({\underline{s}}))_{n=N}^{\infty}$ forms an infinite path in $G$. Conversely, every infinite path in $G$ determines an associated address ${\underline{s}}\in\S_K$. For $z\in K$, denote by $\S_z$ the set of all ${\underline{s}}\in \S_K$ with $\pi_n(z)\in \S_n(z)$ for all $n\geq N$. By the above, $\S_z\neq\emptyset$ for all $z\in K$. The set $\S_K$ is shift-invariant by part~\ref{} of Claim~1. Furthermore, $\S_K$ is contained in the compact set of addresses all of whose entries are taken from $\tilde{\mathcal{F}}$; we need to show that $\S_K$ is itself compact. Suppose that $({\underline{s}}^k)_{k=0}^{\infty}$ is a sequence of addresses in $\S_K$ converging to some address ${\underline{s}}$. Then the prefixes $\pi_{n+1}({\underline{s}})$ and $\pi_{n+1}({\underline{s}}^k)$ agree for all sufficiently large $k$, and in particular $\pi_n({\underline{s}})$ and $\pi_{n+1}({\underline{s}})$ are two vertices of $G$ connected by an edge. It follows that ${\underline{s}}$ is indeed represented by an infinite path in $G$, and hence ${\underline{s}}\in \S_K$ as required. To prove claim~\ref{} of Theorem~\ref{}, we must show that $G_{{\underline{s}}}$ lands at a point $z_0({\underline{s}})\in K$ for all ${\underline{s}}\in\S_K$. By Claim~3, there is some $n_0$ such that, for all $n\geq n_0$ and all ${\underline{s}}\in \S_K$, there is $z\in K$ such that $\tilde{V}_{n}({\underline{s}})\subset B_0(z)$. In particular, $\diam \tilde{V}_{n_0}({\underline{s}}) \leq 2\delta$, and by expansion of $f$ on $U$, we conclude that \[ \diam \tilde{V}_{n}({\underline{s}}) \leq 2^{n-n_0+1} \cdot \delta. \] In particular, $\zeta_n({\underline{s}})$ is a Cauchy sequence, and hence convergent. So $G_{{\underline{s}}}$ lands at a point $z_0({\underline{s}})$ with \begin{equation}\label{} \dist(z_0({\underline{s}}) , \tilde{V}_n({\underline{s}})) \leq 2^{n-n_0}\cdot \delta, \end{equation} for all $n\geq n_0$. It is clear from Claim~3 that the landing point $z_0({\underline{s}})$ belongs to $K$. Furthermore, if $z\in K$ and ${\underline{s}}\in \S_z$, then $z_0({\underline{s}})=z$ by Claim~3. As $\S_z\neq\emptyset$ for all $z\in K$, this shows that the function ${\underline{s}}\mapsto z_0({\underline{s}})$ is surjective. To prove continuity, suppose that ${\underline{s}},\tilde{{\underline{s}}}\in \S_K$ agree in the first $n\geq n_0$ entries. Then $\tilde{V}_n({\underline{s}}) = \tilde{V}_n(\tilde{{\underline{s}}})$, and hence \[ \dist( z_0({\underline{s}}) , z_0(\tilde{{\underline{s}}})) \leq 3 \cdot 2^{n-n_0}\cdot \delta \] by~\eqref{}. This completes the proof of~\ref{}. Part~\ref{} follows directly from Claim~3. It remains to establish~\ref{}. So let $z\in K$ be a periodic point of period $p$. By Lemma~\ref{ for all} and~\ref{}, it is enough to show that $\S_z$ contains a periodic address. Let $n_0$ be as in Claim~4; by increasing $n_0$ if necessary, we can assume that $p$ divides $n_0$. Let $\psi\colon \S_{n_0}(z) \to \S_{2n_0}(z)$ be the inverse of $\sigma^{n_0}|_{\S_{2n_0}(z)}$. Since $\S_{n_0}(z)$ is finite, the function $\phi \defeq \pi_{n_0}\circ \psi$ has a periodic element; say $\phi^{k_0}({\underline{s}}) ={\underline{s}}$. For $k\geq 0$, let ${\underline{s}}^k$ be the unique preimage of ${\underline{s}}$ under $\sigma^{k n_0}$ in $\S_{(k+1)n_0}(z)$. We claim that \[ \pi_{n_0}({\underline{s}}^{k+1}) = \phi(\pi_{n_0}({\underline{s}}^k)) \] for all $k\geq 0$. This is true for $k=0$ by definition. If $k>0$, we have \[ \pi_{2n_0}({\underline{s}}^{k+1}) = \psi(\sigma^{n_0}(\pi_{2n_0}({\underline{s}}^{k+1}))) = \psi( \pi_{n_0}(\sigma^{n_0}({\underline{s}}^{k+1}))) = \psi(\pi_{n_0}({\underline{s}}^k)). \] Hence \[ \pi_{n_0}({\underline{s}}^{k+1}) =\pi_{n_0} ( \psi(\pi_{n_0}({\underline{s}}^k))) = \phi(\pi_{n_0}({\underline{s}}^k)). \] So $\pi_{n_0}({\underline{s}}^k) = \phi^k({\underline{s}})$ for all $k\geq 0$. Hence ${\underline{s}}^k$ can be written as a concatenation \[ {\underline{s}}^k = \phi^k({\underline{s}}) {\underline{s}}^{k_1}= \dots = \phi^k({\underline{s}}) \phi^{k-1}({\underline{s}}) \dots \phi^1({\underline{s}}) {\underline{s}}. \] Since ${\underline{s}}$ is periodic under $\phi$, of period $k_0$, we conclude that ${\underline{s}}^k = \pi_{(k+1)n_0}({\underline{s}}^{k+k_0})$. Hence there is an infinite path in $G_z$ passing through the vertices ${\underline{s}}^{j\cdot k_0}$, $j\geq 0$. The associated address is the periodic sequence $(\phi^{k_0-1}({\underline{s}}) \dots \phi^1({\underline{s}}) {\underline{s}})^{\infty}$, and the proof is complete. \end{proof} We note the following corollary, which proves the accessibility of certain singular values. For definitions, we refer to~\cite{RvS}. \begin{cor} Let $f$ be a postsingularly bounded transcendental entire function, and let $v\in J(f)$ be a non-recurrent singular value for $f$ whose forward orbit does not pass through any critical points. Suppose that the $\omega$-limit set of $v$ does not contain parabolic points, and does not intersect the $\omega$-limit set of a recurrent critical point or of a singular value contained in a wandering domain. Then there is a bounded-address dreadlock of $f$ that lands at $v$. \end{cor \begin{proof} By \cite[Theorem 1.2]{RvS}, if the postsingular set is bounded then any forward invariant compact subset of the Julia set is hyperbolic provided it does not contain parabolic points, critical points, or it intersects the $\omega$-limit set of a critical point or of a singular value contained in wandering domains. Hence $P(a):=\overline{\bigcup_n f^n(a)}$ is hyperbolic and every point in $P(a)$ is the landing point of a dreadlock. \end{proof}

\section{Introduction} \IEEEPARstart{T}{he} rapid development of wireless sensor networks (WSNs) in the last decade is in large part responsible for the recent upsurge in interest in WSN-based distributed tracking. A typical decentralized WSN consists of a number of spatially distributed, interconnected sensors that have independent sensing and calculation capabilities and (only) communicate with the neighbors for information sharing, namely peer-to-peer (P2P) communication. In particular, due to the appealing fault-tolerance and scalability to large networks, consensus-based distributed algorithms have gained immense popularity in the sensor networks community. In the consensus-oriented distributed filtering setup, each sensor operates an independent filter while sharing information with its neighbors iteratively to ameliorate each other's estimation with the goal of converging to the same estimate over the entire network. As a result, local estimation that are made based on both local observation and the information disseminated from neighbors are similar to each other (or in other words, the sensors asymptotically reach a consensus), which are better as compared to the independent estimation without network cooperation \cite{Liu07,Akselrod12,Mohammadi15}. In this paper, we consider the scenario with a time-varying, unknown, number of targets, which are synchronously observed by all sensors in the presence of false and missing observations. Great interest has been seen for extending the theory of average consensus \cite{Xiao04,Olfati-Saber07} for which the item being estimated may be the arithmetic average (considered as the default manner \cite{Olfati-Saber07}, akin to the linear opinion pool \cite{Heskes98} ) or the geometric average \cite{Olfati-Saber06} (akin to the logarithmic opinion pool \cite{Heskes98} ) of the initial values. With regard to the type of information disseminated, three main categories of protocols exist; we note that there are protocols such as diffusion \cite{sayed14,Dedecius17} that may belong to either. Our approach falls into the last category: \begin{enumerate} \item \textit{Measurement/Likelihood}. Disseminating raw measurement can be practically preventable in communication. Instead, the likelihood function, as a compact representation of the measurement information, is a promising alternative \cite{Hlinka12, Hlinka13, Battistelli14b}. However, in multi-sensor multi-target cases, computationally cumbersome measurements-to-targets association or enumeration \cite{Mahler10} is typically required. Moreover, it is nontrivial to fuse raw measurements reported at sensors of different profiles including detection probabilities, clutter rates, etc. To date, measurement/likelihood consensus is mainly limited to the single target case. \item \textit{Estimate/Track}. This involves running tracking algorithms on each sensor separately, yielding a set of tracks to be associated between sensors based on their proximities and then be fused, namely track-to-track fusion \cite{Mori12, Noack14}. When tracks are distant in the state space, this may work well, e.g., \cite{Beaudeau15, Zhu17} otherwise it suffers from the fragility and high computational cost for maintaining a large number of hypotheses. \item \textit{Posterior/Intensity}. This involves disseminating and fusing the multi-target posterior \cite{Meyer14, Meyer17} or the density of its statistical moments between sensors. In particular, the probability hypothesis density (PHD) that is the first order moment of the random target-state set, has been developed as a powerful alternative to the full posterior for time series recursion \cite{Mahler03, Vo06}. In this case, the key is to disseminate and fuse PHDs. \end{enumerate} As the state of the art, the geometric average for PHDs/multi-target densities is referred to as the Kullback-Leibler average (KLA) \cite{Battistelli13, Battistelli14a, Battistelli14b}, also known as the geometric mean density (GMD) or the exponential mixture density (EMD) \cite{Uney11, Uney13, Bailey12}. The fusion approach is known as generalized covariance intersection (GCI) \cite{Mahler00, Clark10, Mahler12} which extends the Chernoff fusion/covariance intersection \cite{Uhlmann95,Julier97} to multi-target densities. In spite of its success in certain scenarios, deficiencies of GCI, most of which have already been recognized in the literature, are noted in this paper. However, it is not our intention to revise or improve these geometric average consensus approaches. Instead, we propose novel arithmetic average consensus approaches for PHD fusion, which are closely connected to the so-called covariance union (CU) \cite{Uhlmann03,Julier05,Bochardt06,Reece10} and arithmetic mean density (AMD) \cite{Bailey12}. In short, there are two distinguishable features with our approaches: \begin{enumerate} \item Only the significant components of local PHDs, which are more likely target signals rather than false alarms, are disseminated between neighbors, while the insignificant components that are more suspected to be false alarms will not be involved in either the P2P communication or the consensus fusion. As such, the signal-to-noise ratio (SNR) can be positively enhanced at local sensors. This significantly differs from existing consensus approaches where the (complete) consensus is conditioned on \textit{all} the information available in the network. The consensus in our approach is made based on a part of the information of local posteriors, termed \textit{partial consensus}. % \item Only closely distributed components, which are more likely corresponding to the same target, are fused, in either of two conservative manners: averaging and merging, based on union calculation rather than intersection. The resulting consensus remains defined in the default \textit{arithmetic average} sense rather in the KLA sense, which demonstrates particular advantages in dealing with the false and missing observations which are inevitably involved in realistic tracking. \end{enumerate} A preliminary part of the merging-for-fusion protocol has been presented in our conference paper \cite{Li17merging_Conf}, in which, however, we did not analyze its connection to AMD/CU, nor did we provide any conservativeness analysis. The merging scheme adopted initially is a standard one \cite{Salmond90}, which as found in this paper can be optimized in the covariance-fusion part. These have now been completed in this article. In addition, we present much more technical extension and new results, including a new, communicatively much cheaper, partial consensus protocol. Therefore, this paper serves as a significant revision and extension to \cite{Li17merging_Conf}. The remainder of this paper is organized as follows. The basics of GM-PHD, conservative fusion and GCI are given in Section \ref{}, followed by the motivation and key idea of our approach in section \ref{}. The proposed distributed GM-PHD fusion protocol is detailed in Section \ref{}. Simulations are given in Section \ref{} for comparing our approaches with the GCI. In particular, the weakness of the GCI is note . We conclude in Section \ref{}. \section{Background and Preliminary} \label{} \subsection{RFS and GM-PHD} We consider an unknown and time-varying number $M_k$ of targets with random states $\mathbf{x}_k^{(n)}$ in the state space $\chi \subseteq \mathbb{R}^d$, $n = 1,2,\cdots,M_k$. The collection of target states at time $k$ can be modeled by a \emph{random finite set} (RFS) $X_k=\left\{{\mathbf{x}_{k,1},\cdots,\mathbf{x}_{k,M_k}}\right\}$ with random cardinality $M_k =|X_k|$. The cardinality distribution $\rho(n)$ of $X_k$ is a discrete probability mass function of $M_k$, i.e., $\rho(n) \triangleq \mathrm{Pr}[M_k \!=\! n]$. A RFS variable $X$ is a random variable that takes values as unordered finite sets and is uniquely specified by its cardinality distribution $\rho(n)$ and a family of symmetric joint distributions $p_n (\mathbf{x}_1,\mathbf{x}_2,\cdots,\mathbf{x}_n)$ that characterize the distribution of its elements over the state space, conditioned on the set cardinality $n$. Here, a joint distribution function $p_n (\mathbf{x}_1,\mathbf{x}_2,\cdots,\mathbf{x}_n )$ is said to be symmetric if its value remains unchanged for all of the $n!$ possible permutations of its variables. The probability density function (PDF) $f(X)$ of a RFS variable $X$ is given as $f(\{\mathbf{x}_1,\mathbf{x}_2,\cdots, \mathbf{x}_n \})=n! \rho(n) p_n (\mathbf{x}_1,\mathbf{x}_2,\cdots, \mathbf{x}_n )$. Instead of propagating the full multi-target density which has been considered computationally intractable, the PHD filter propagates its first order statistical moment \cite{Mahler03}. For a multi-target RFS variable $X$ with the PDF $f(X)$, its first order moment PHD $D_S(\mathbf{x})$ in a region $S \subseteq \chi$ is given as: \begin{equation}\label{} D_S(\mathbf{x})=\int_S \delta_X(\mathbf{x})f(X)\delta X \hspace{0.5mm}, \end{equation} where $\delta_X(\mathbf{x}) \triangleq \sum_{w\in X}\delta_w(\mathbf{x})$ which is used to convert the finite set $X=\{\mathbf{x}_1,\mathbf{x}_2,\cdots \}$ into vectors since the first order moment is defined in the single-target vector space, $\delta_\mathbf{y}(\mathbf{x})$ denotes the generalized Kronecker delta function, and the RFS integral in the region $S \subseteq \chi$ is defined as: \begin{align}\label{} & \int_S f(X)\delta X\nonumber \\ & \triangleq f(\emptyset) + \sum_{n=1}^\infty \int_{S^n} \frac{f(\{\mathbf{x}_1,\mathbf{x}_2,\cdots, \mathbf{x}_n\})}{n!} d\mathbf{x}_1d\mathbf{x}_2\cdots d\mathbf{x}_n \hspace{0.5mm}. \end{align} Straightforwardly, the GM approximation of the whole PHD at filtering time $k$ can be written as: \begin{equation}\label{} D_{k}(\mathbf{x})= \sum_{i=1}^{J_k} w_k^{(i)} \mathcal{N}(\mathbf{x};\mathbf{m}_k^{(i)},\mathbf{P}_k^{(i)}) \hspace{0.5mm}, \end{equation} where $\mathcal{N}(\mathbf{x;m,P})$ denotes a Gaussian component (GC) with mean $\mathbf{m}$ and covariance $\mathbf{P}$, $J_k$ is the number of GCs in total, and $w_k^{(i)}$ is the weight of $i$th GC. The PHD is uniquely defined by the property that its integral in any region gives the expected number of targets in that region. Therefore, the expected number of targets can be approximated by the weight sum $W_k$ of all GCs, i.e., \begin{equation}\label{eq:4} W_k = \sum_{i=1}^{J_k} w_k^{(i)} \hspace{0.5mm}. \end{equation} In this paper, we assume each local sensor running a GM-PHD filter, e.g., as given in \cite{Vo06} and that they are synchronous. Our work focuses on the posterior GM dissemination and fusion between neighboring sensors. \subsection{Conservative Fusion and Mixture Reduction} We consider now a sensor network where all the sensors observe the same set of targets but their measurements are conditionally independent. The errors of estimates yielded at local sensors are correlated with each other where the correlation is due to the common prior/noise in the models and common information shared between sensors, etc. Despite any a priori information on the cross-covariance \cite{Chang97,Chong04, Gao16}, it is typically intractable, if not impossible, to quantify the correlation between sensors, which is at least time-varying due to the P2P communication and prevents optimal fusion (e.g., in the sense of minimum mean square error); see also \cite{Bakr17}. Therefore, a pseudo-optimal, ``conservative", fusion rule is resorted to for avoiding underestimating the actual squared estimate errors. The benefit to do so includes better fault tolerance and robustness \cite{Uhlmann03,Julier05}. To be more precise, we have the following definition on the notion of ``conservative'', as used in \cite{Uhlmann95,Uhlmann03,Bochardt06,Ajgl14}: \textit{Definition 1 (conservativeness)}. An estimate pair ($\mathbf{\hat{x}},\mathbf{P}$) of the real state $\mathbf{x}$ (a random vector), consisting of a vector estimate mean $\mathbf{\hat{x}}$ with an associated error covariance matrix $\mathbf{P}$, is called conservative when $\mathbf{P}$ is no less than the actual error covariance of the estimate, i.e., $\mathbf{P} - \mathrm{E} [(\mathbf{x}-\mathbf{\hat{x}})(\mathbf{x}-\mathbf{\hat{x}})^\mathrm{T}]$ is positive semi-definite. With the associated covariance matrix being ``conservative'', the estimate pair is also called ``consistent'' \cite{Uhlmann03,Reece10}. However, a consistent estimator is traditionally an estimator that converges in probability to the quantity being estimated as the sample size grows. To avoid confusion, we shall only use the terminologies of ``conservative'' or ``conservativeness''. \begin{lemma} \label{lemma1} A sufficient condition for the fused estimate pair ($\mathbf{\hat{x}},\mathbf{P}$), due to fusing a set of estimate pairs $(\mathbf{\hat{x}}_i,\mathbf{P}_i), i \in \mathcal{I} =\{1,2,\cdots\}$, in which at least one pair is unbiased and conservative, to be conservative is that \begin{equation} \label{} \mathbf{P} \geq \mathbf{P}_i + (\mathbf{\hat{x}}-\mathbf{\hat{x}}_i)(\mathbf{\hat{x}}-\mathbf{\hat{x}}_i)^\mathrm{T}, \hspace{1mm} \forall i \in \mathcal{I} \hspace{0.5mm}. \end{equation} \end{lemma} \begin{proof} Without lose of generality, supposing $(\mathbf{\hat{x}}_i,\mathbf{P}_i)$ is unbiased and conservative, we have \begin{equation} \mathrm{E} [ (\mathbf{x}-\mathbf{\hat{x}}_i)(\mathbf{\hat{x}}_i-\mathbf{\hat{x}})^\mathrm{T} ] = \mathbf{0} \hspace{0.5mm}, \end{equation} \begin{equation} \mathbf{P}_i \geq \mathrm{E} [(\mathbf{x}-\mathbf{\hat{x}}_i)(\mathbf{x}-\mathbf{\hat{x}}_i)^\mathrm{T}] \hspace{0.5mm}. \end{equation} due to the unbiasedness and conservativeness, respectively. By decomposing $\mathbf{x}-\mathbf{\hat{x}}_i $ as $(\mathbf{x}-\mathbf{\hat{x}}_i)+(\mathbf{\hat{x}}_i-\mathbf{\hat{x}})$, we obtain $\mathrm{E} [(\mathbf{x}-\mathbf{\hat{x}})(\mathbf{x}-\mathbf{\hat{x}})^\mathrm{T}] \leq \mathbf{P}_i + (\mathbf{\hat{x}}-\mathbf{\hat{x}}_i)(\mathbf{\hat{x}}-\mathbf{\hat{x}}_i)^\mathrm{T}$ easily to finish the proof. \end{proof} \begin{lemma} \label{lemma2} Given a set of conservative estimate pairs $(\mathbf{\hat{x}},\tilde{\mathbf{P}}_i),i \in \mathcal{I}=\{1,2,\cdots\}$ of the same, unbiased estimate mean associated with possibly different error-covariance matrix, a sufficient condition for the fused estimate pair ($\mathbf{\hat{x}},\mathbf{P}$), to be conservative is given by \begin{equation} \label{} \mathbf{P} \geq \sum_{i \in \mathcal{I}} \omega_i \tilde{\mathbf{P}}_i \hspace{0.5mm}, \end{equation} where the non-negative scaling parameters $\omega_i \geq 0$, $\sum_{i \in \mathcal{I}} \omega_i = 1$ are called fusing weights hereafter. \end{lemma} \begin{proof} The conservativeness of fusing estimate pairs reads \begin{equation} \label{} \tilde{\mathbf{P}}_i \geq \mathrm{E} [(\mathbf{x}-\mathbf{\hat{x}}) (\mathbf{x}-\mathbf{\hat{x}})^\mathrm{T}], \hspace{1mm} \forall i \in \mathcal{I} \hspace{0.5mm}. \end{equation} The proof is simply done by multiplying both sides of \eqref{} by $\omega_i$ and summing up over all $i \in \mathcal{I}$, which leads to \begin{equation} \sum_{i \in \mathcal{I}} \omega_i \tilde{\mathbf{P}}_i \geq \mathrm{E} [(\mathbf{x}-\mathbf{\hat{x}}) (\mathbf{x}-\mathbf{\hat{x}})^\mathrm{T}] \hspace{0.5mm}. \end{equation} \end{proof} \textit{Definition 2 (Standard Mixture Reduction, SMR)}. Given a set of estimate pairs $(\mathbf{\hat{x}}_i,\mathbf{P}_i)$ weighted as $w_i, i \in \mathcal{I}$, respectively, the SMR scheme \cite{Salmond90} fuses them into a single estimate pair $(\mathbf{\hat{x}}_\mathrm{SMR}, \mathbf{P}_\mathrm{SMR})$ with weight $w_\mathrm{SMR}$, given by \begin{equation} \label{} w_\mathrm{SMR}=\sum_{i \in \mathcal{I}} w_i \hspace{0.5mm}, \end{equation} \begin{equation} \label{} \mathbf{\hat{x}}_\mathrm{SMR} = \frac{\sum_{i \in \mathcal{I}} w_i\mathbf{\hat{x}}_i}{\sum_{i \in \mathcal{I}} w_i} \hspace{0.5mm}, \end{equation} \begin{equation} \label{} \mathbf{P}_\mathrm{SMR} = \frac{\sum_{i \in \mathcal{I}} w_i\tilde{\mathbf{P}}_i}{\sum_{i \in \mathcal{I}} w_i} \hspace{0.5mm}, \end{equation} where the adjusted covariance matrix is given by (cf. \eqref{}) \begin{equation} \label{} \tilde{\mathbf{P}}_i=\mathbf{P}_i + (\mathbf{\hat{x}}_\mathrm{SMR}-\mathbf{\hat{x}}_i)(\mathbf{\hat{x}}_\mathrm{SMR}-\mathbf{\hat{x}}_i)^\mathrm{T} \hspace{0.5mm}. \end{equation} \begin{lemma} \label{lemma3} Given that all the fusing estimate pairs are unbiased and conservative, the resulting estimate pair given by the SMR scheme as in \eqref{} is unbiased and conservative. \end{lemma} \begin{proof} The proof is straightforwardly based on Lemmas \ref{lemma1} and \ref{lemma2}. First, unbiasedness is due to the convex combination. Second, given that each $(\mathbf{\hat{x}}_i,\mathbf{P}_i), \forall i \in \mathcal{I} = \{1,2,\cdots \}$ is unbiased and conservative, $(\mathbf{\hat{x}}_\mathrm{SMR},\mathbf{\tilde{P}}_i)$ is conservative according to Lemma \ref{lemma1}, and so their convex combination of $(\mathbf{\hat{x}}_\mathrm{SMR},\mathbf{\tilde{P}}_\mathrm{SMR})$ is conservative according to Lemma \ref{lemma2}. \end{proof} It is important to note that, considering ``conservativeness" only, the fusing weights used to get a conservative fused covariance matrix as in \eqref{} do not have to be the same as that to get the fused state as in \eqref{}. But instead, it is typically of interest to use different fusing weights to get an optimal fused covariance in some sense, while retaining conservativeness. For this, we have the following proposition, akin to the CI-based optimization \cite{Chen02,Niehsen02}. \textbf{Proposition 1}. The covariance-fusing weights for \eqref{} can be determined such that the trace of the resulting covariance matrix is minimized, i.e., $\big\{\omega_i\big\}_{i \in \mathcal{I}} = \underset{\omega_i,i \in \mathcal{I}}{\mathrm{argmin}} \hspace{0.5mm} \mathrm{tr}\Big(\sum_{i \in \mathcal{I}} \omega_i\tilde{\mathbf{P}}_i\Big)$. Thanks to the convex combination and positive trace of the matrices, the solution is simply given by $\omega_i=1, \omega_j=0, \forall j \neq i, j \in \mathcal{I}$ where $i=\underset{i \in \mathcal{I}}{\mathrm{argmin}} \hspace{0.5mm} \mathrm{tr} (\tilde{\mathbf{P}}_i)$. That is, the trace-minimal yet conservative fused covariance is given by \begin{equation} \label{} \mathbf{P}_\mathrm{OMR} = \underset{\tilde{\mathbf{P}}_i}{\mathrm{argmin}} \hspace{0.5mm} \mathrm{tr}\big(\tilde{\mathbf{P}}_i\big) \hspace{0.5mm}. \end{equation} Hereafter, we refer to the MR scheme based on \eqref{}, \eqref{} \eqref{} and \eqref{} as the optimal mixture reduction (OMR), which differs from the SMR only in the covariance-fusion part. It is a type of fusion seeking conservativeness, given that all fusing estimate pairs are unbiased and conservative. \subsection{GCI Fusion and KLA} Given a set of posteriors $f_i \in \Psi, i \in \mathcal{I}$ to be fused by the fusing weights $\omega_i \geq 0$, where $\Psi$ denotes the set of PDFs over the state space $\chi$, and $\mathcal{I}=\{1,2,\cdots\}$ denotes the fusing sensor set, the GCI/Chernoff fusion \cite{Mahler00} which resembles the logarithmic opinion pool \cite{Heskes98} and the belief consensus \cite{Olfati-Saber06} reads \begin{equation}\label{} f_\mathrm{GCI} \triangleq C^{-1}\prod_{i\in \mathcal{I}}{f_i^{\omega_i}} \hspace{0.5mm}, \end{equation} where $C$ is a normalization constant. The GCI fusing result is also known as GMD \cite{Bailey12} or EMD \cite{Julier06, Uney11, Uney13}, which actually minimizes the weighted sum of its KLD with respect to all posteriors $f_i, \forall i\in \mathcal{I}$, and is, therefore, also referred to as the weighted Kullback-Leibler average (KLA) \cite{Battistelli13, Battistelli14b}, i.e., \begin{equation}\label{} f_\mathrm{KLA} = \underset{f \in \Psi}{\mathrm{arginf}} \sum_i {\omega_i d_{KL}(f||f_i)} \hspace{0.5mm}, \end{equation} where $d_{KL} (f_a ||f_b) \triangleq \int f_a (X) \frac{f_a (X)}{f_b (X)}\delta X $ is the set-theoretical KLD of the intensities $f_a$ from $f_b$. Three challenges arise due to the exponentiation and product calculation when the posterior $f_i$ in \eqref{} is given by a mixture, such as the GM: \begin{enumerate} \item The fractional order exponential power of a GM does not provide a GM. Existing solutions are based on either analytical approximation that only appeals to special mixtures (e.g., components are well distant) \cite{Battistelli13, Battistelli14a, Battistelli14b, Battistelli15} or numerical approximation via important sampling \cite{Mariam07,Li17Nehorai} or sigma point method \cite{Gunay16}. \item The product rule is prone to mis-detection. Misdetection at one sensor will remarkably degrade the detection at the other sensors since any signal multiplied by a weak signal of almost zero energy will be greatly weakened. See also illustration given in \cite{Yu16,Yi17}. \item The GCI/KLA fusion will typically result in a multiplying number of fused GCs \cite{Battistelli13}, which is costly in both communication and computation. \end{enumerate} These problems can lead to disappointing results in certain case , which will be discussed within our simulation study in Section \ref{. To overcome these deficiencies, we propose alternatives without the intrinsic need for exponentiation and product calculation of mixtures while being ``conservative'' not only in fusion but also in communication. In the following distributed formulation, we will use subscripts $a$ and $b$ to distinguish between two neighboring sensors. Since all calculations regard the same filtering time $k$, we drop the subscript $k$ for notation simplicity. \section{Key Idea and Properties of Our Proposal} \label{} The section presents two ``conservative'' principles for distributed fusion algorithm design, which constitute the key idea of our approaches: \begin{itemize} \item \textit{P.1 Conservative communication}. Consensus should only be sought on the information of targets. To get this maximally respected, only the GC of significance (those that are highly likely to corresponding to the ``target'') should be disseminated, while the insignificant GC (those that are more like false alarms) should be the least involved for conservative consideration. We refer to this as the ``conservative communication'' principle. Here the ``conservativeness'' is not the same as the estimate ``conservativeness'' given in Definition 1. We assume that the reader will not be confused. \item \textit{P.2 Conservative fusion}. Only highly relevant information, namely that which corresponds to the same target as at least very likely, should be fused and the fused results should retain ``conservativeness'', to deal with the unknown correlation between sensors. We refer to this as the ``conservative fusion'' principle. \end{itemize} \subsection{Conservative Communication for Partial Consensus} First, the mixture reduction is carried out in local GM filters as usual at each filtering step, before network communication, to control the GM size. Second, only highly-weighted GCs that are more likely corresponding to real targets rather than false alarms are disseminated between neighbors. To this end, we propose two alternative rules to identify these target-likely GCs, referred to as \textit{rank rule} and \textit{threshold rule}, respectively. \begin{itemize} \item \textit{P.1.1 Rank rule}. Specify the number of GCs to be disseminated as equal to the intermediately estimated number of targets at each sensor using the closest integer to $W_k$ as in \eqref{eq:4} or more straightforwardly, specify a fixed number of GCs when a priori information (e.g., maximum) about the number of targets is known. Then, only the corresponding number of GCs with the greatest weights are transmitted to the neighbors. \item \textit{P.1.2 Threshold rule}. Specify a weight threshold $w_s$, and only the component that is weighted greater than that threshold will be transmitted. \end{itemize} It is also possible to use a hybrid, arguably more conservative, criterion such that only the GCs that fulfill both rules are chosen, or a hybrid, less conservative, criterion such that any GCs that fulfill either rules can be chosen. In any case, we refer to them as a separate GM, hereafter called Target-likely GM (T-GM) and denote the T-GM size at sensor $a$ as $n_a$, i.e., \begin{equation}\label{} D_{a,\mathrm{T}}(\mathbf{x})\triangleq \sum_{i=1}^{n_a} w_a^{(i)} \mathcal{N}(\mathbf{x};\mathbf{m}_a^{(i)},\mathbf{P}_a^{(i)}) \hspace{0.5mm}, \end{equation} of which the total weight ($\leq W_a $) is given as \begin{equation}\label{} W_{a,\mathrm{T}} = \sum_{i=1}^{n_a} w_a^{(i)} \hspace{0.5mm}. \end{equation} Correspondingly, the remaining GC at each sensor is called false alarm-suspicious GC (FA-GC), which will not be involved in the neighborhood communication. \textit{Definition 3 (partial consensus)} The consensus yielded by disseminating among sensors an incomplete part of the information they own, i.e., only target-likely GCs in our approaches, is called \textit{partial consensus}. \subsection{Conservative AMD Fusion} Different to the KLA optimality of the GMD as in \eqref{}, the AMD \cite{Bailey12} calculates the average of posterior multi-target density in the arithmetic sense \cite{Olfati-Saber07} rather than in the geometric average sense \cite{Olfati-Saber06}, or equivalently speaking, based on the linear opinion pool rather than the logarithmic opinion pool. \textit{Definition 4 (AMD)}. The AMD of multiple posteriors $f_i, i \in \mathcal{I}$, is given as follows: \begin{equation}\label{} f_\mathrm{AMD} \triangleq \sum_{i \in \mathcal{I}}{\omega_if_i} \hspace{0.5mm}, \end{equation} where the fusing weights $\omega_i \geq 0, \sum_{i\in \mathcal{I}} {\omega_i} =1$. As addressed, $f_i$ is only a partial PHD obtained at sensor $i$ in our work. As shown, the AMD is given by a convex union of multisensor posteriors, which does not double count information \cite{Bailey12} and is provably conservative (cf. Lemma 2 . It was further compared with the GMD in \cite{Bailey12} as that, ``the GMD is potentially inconsistent if a single component is inconsistent while the AMD is conservative if even a single component is consistent'' [cf. Lemma 1]. Indeed, the union-type AMD fusion is less prone to the problem of misdetection as it does not involve product calculation. More importantly, it does not fuse the information of different targets and of clutter unless they lie to each other too close. The AMD of GMs, can be easily realized through re-weighting (by using the fusing weights) and combining GMs in neighborhood. Similar idea has actually been applied \cite{Yu16} for pairwise gossip-based fusion and for averaging the ``generalized likelihood'' \cite{Streit08}. Basically, two key issues need to be solve . First, we need a proper mechanism to design the fusing weights. The most straightforward solution is given by uniform fusing weights, which may not guarantee efficient consensus convergence and appeals primarily to the case when only few P2P communication iterations are allowed. For faster convergence, the popular Metropolis weights \cite{Xiao04, Xiao05} approach is readily competent, given a large number of communication iterations. It determines the fusing weights for the information from sensor $b$ at the host sensor $a$ as follows \begin{equation}\label{} \omega_{b\rightarrow a} = \left\{ \begin{array}{ll} \frac{1}{1+\max{(|N_a|,|N_b|)}} & \mathrm{if} \hspace{1mm} b\in N_a, b\neq a \hspace{0.5mm},\\ 1-\sum_{l\in N_a}{\omega_{l\rightarrow a}} & \mathrm{if} \hspace{1mm} b=a\hspace{0.5mm},\\ 0 & \mathrm{if} \hspace{1mm} b\not\in N_a\hspace{0.5mm}, \end{array} \right. \end{equation} where $N_a$ denotes the set of neighbors of sensor $a$ (excluding $a$). Second, the AMD of $N$ GCs and $M$ GCs as in \eqref{} will have $N+M$ GCs, which is in general much smaller than $N\times M$ yielded by GCI. Still the local GM size grows linearly with the number of fusing sensors. In order to reduce the number of GCs to be transmitted and to maintain a stable overall GM size, we next present two conservative MR schemes for fusing the gathered T-GMs in a fully distributed fashion. \section{Conservative MR schemes} \label{} We use $t\in \mathbb{N}=\{0,1,2,...\}$ to denote the P2P communication iteration. $t=0$ means the original statue of the local sensor without any communication. This section presents two MR approaches in line with the conservative fusion principle, based on either OMR or pairwise GM averaging, which need to be executed at each P2P communication iteration. \subsection{Conservative Fusion P.2.1: GM Merging} The first MR protocol for T-GM fusion is given by combining the newly received and the local T-GMs into one set and merging the close T-GCs based on the proposed OMR. Before this, the GC weights should be scaled by using the fusing weights as addressed, according to their origination sensor. However, as shown in our simulation that this protocol typically bears high communication cost (which increases with $t$) and more iterations ($t>2$) do not yield significantly more benefit, we do not suggest a larger number of communication iterations. Therefore, uniform fusing weights are more preferable (especially for $t\leq2$). A key of MR/OMR is to determine the size of gate for fusing GCs to be merged. In our approach, the distance between two T-GCs, e.g., $\mathcal{N}(\mathbf{x};\mathbf{m}_a,\mathbf{P}_a)$ and $\mathcal{N}(\mathbf{x};\mathbf{m}_b,\mathbf{P}_b)$, is measured by the Mahalanobis-type distance as follows \begin{equation} \label{} C_{a,b}\triangleq\big(\mathbf{m}_a - \mathbf{m}_b\big) \mathbf{P}^{-1} \big(\mathbf{m}_a - \mathbf{m}_b\big)^\mathrm{T} \hspace{0.5mm}, \end{equation} where $\mathbf{P}$ is chosen as the covariance of the GC of higher weight. A gate threshold $\tau$ is needed to control the GC grouping such that only T-GCs that are of distance smaller than $\tau$ will be merged, for trade-off between the resultant GM size and merging error. The gate has a clear physical meaning as it indicates the distance no further than $\tau$ standard deviations from the state estimate that the real state lies in with a probability, or at least a lower bound on the probability. When the estimate is unbiased and inferred from Gaussian random variables, the probability that the real state $\mathbf{x}$ lies within $\tau$ standard deviations of the state estimate $\mathbf{\hat{x}}$ is given by \cite{Ye00} \begin{equation} \label{} \mathrm{Pr}\big[(\mathbf{x}-\mathbf{\hat{x}}) \mathbf{P}^{-1}(\mathbf{x}-\mathbf{\hat{x}}) \big] \leq \gamma \Big(\frac{d}{2},\frac{\tau^2}{2}\Big)\hspace{0.5mm}, \end{equation} where $\mathbf{P}$ is the error-covariance matrix of the estimate $\mathbf{\hat{x}}$, $\gamma$ is the lower incomplete Gamma distribution and $d$ is the cardinality of the state vector. Due to the uniform fusing weights, the T-GM combination and merging will certainly raise the weight sum at sensor $a$ to \begin{equation} \label{} \tilde{W}_a(t)=W_a(t-1)+\sum_{j\in {N_a}}W_{j,\mathrm{T}}(t-1)\hspace{0.5mm}, \end{equation} where $W_a(t-1)$ and $W_{j,\mathrm{T}}(t-1)$ are the whole GM and the T-GM at local sensor $a$ after $t-1$ iterations of P2P communication, respectively and $W_{j,\mathrm{T}}(0)$ is defined as in \eqref{}. As such, the weights of all GCs $w_a^{(i)}, i=1,2,\cdots, J_a(t)$ after merging at each iteration $t$ need to be re-scaled for correct cardinality estimation, where $J_a(t)$ is the GM size at iteration $t$ and we have $\tilde{W}_a(t) = \sum_{i=1}^{J_a(t)} w_a^{(i)}$. To this end, we may apply average consensus on the cardinality estimates, namely ``cardinality consensus'', which will be carried out simultaneously with the proposed T-GM consensus. This is feasible because the cardinality estimates yielded by the PHD filter \eqref{eq:4} are scalar-valued parameters, for which the standard average consensus based on Metropolis weights \cite{Xiao04, Xiao05} is straightforwardly applicable. To this end, the local GM weight sums will also be disseminated in neighborhood along with the T-GCs for consensus and we have the following proposition. \textbf{Proposition 2 (Cardinality Consensus)}. The Metropolis weights based average consensus is applied to update the local weight sum at each communication iteration as follows: \begin{equation}\label{eq:CC} {W}_a(t) = \sum_{l \in \{ a,N_a\}} \omega_{l\rightarrow a} W_l(t-1)\hspace{0.5mm}, \end{equation} which will be used for re-scaling the weights of all GCs at each communication iteration $t$, i.e., \begin{equation}\label{} w_a^{(i)} \leftarrow \beta_a w_a^{(i)}, \forall i=1,2,\cdots, J_a(t)\hspace{0.5mm}. \end{equation} where $\beta_a = \frac{W_a(t)}{\tilde{W}_a(t)}$. In order to analyze the change of the weight of FA-GCs due to \eqref{}, we make two approximate albeit reasonable assumptions: $W_a(t-1) \approx W_b(t-1)$ and $W_{b,\mathrm{T}}(t-1) \approx \alpha W_b(t-1)$, for all $b \in N_a$. Clearly, $\alpha < 1$. As such, \eqref{} and \eqref{eq:CC} reduce to $\tilde{W}_a(t) \approx (1+ \alpha|N_a|) W_a(t-1)$ and $W_a(t) \approx W_a(t-1)$, respectively. These read \begin{equation} \label{} \beta_a \approx \frac{1}{1+\alpha|N_a|} \end{equation} In most cases, the T-GCs take the majority of the weight sum, namely $\alpha > 0.5$. For example, when sensor $a$ has two neighbors namely $|N_a|=2$, $\beta_a \stackrel{\text{appr}}{<} 0.5$, which indicates that the weight of FA-GCs at local sensor $a$ will be approximately reduced to less than a half by \eqref{}. Comparably, the T-GCs merge with many others from neighbors which will counteract such reduction but instead their weight will likely be increased slightly. That is, the target-likely signal will be enhanced while the FA-suspicious signal will be weakened or even ultimately removed by pruning. This will give rise to the SNR at local sensors, reducing the possibility of causing false alarms and facilitating more accurate estimation. We refer to this fusion protocol as \textit{conservative GM merging} (CGMM). For illustrative purposes, CGMM operations including GC selection, transmission, merging and re-weighting are given, as shown in Fig.1, for a 1-dimensional state space model using two sensors. In the top row, the original PHDs at local sensors are given as GMs, each having two significant GCs that likely correspond to targets. The sensors share them with each other, and then GC merging (and pruning to remove very insignificant FA-GCs) and re-weighting are performed, as shown in the middle and bottom rows, respectively. The resulting GMs are reweighted such that they have the same weight sum for cardinality consensus. At the end, the T-GCs will become more significant due to merging while the FA-GCs are weakened, leading to an enhanced SNR. \begin{figure}\label{fig:1} \centering \includegraphics[width=8 cm]{1} \caption{Illustration of the proposed CGMM fusion between two sensors. The GCs originally formed by sensor $a$ are given in solid lines, while those formed by sensor $b$ are given in dash lines. Significant components of the GM are given in color while the insignificant ones in black} \end{figure} \subsection {Conservative Fusion P.2.2: Pairwise GM Averaging} CGMM can not guarantee all received GCs to be merged to the local T-GM unless a sufficiently high merging threshold $\tau$ is used which will in turn cause greater merging errors. As a result, the local T-GM size will likely grow against the P2P communication iteratio . As an alternative, we integrate the received T-GM to the local GM in a way such that each of the received T-GC is fused to the nearest host GC immediately if closely enough or otherwise abandoned. This will retain a promisingly constant local GM size during networking. To this end, we associate the received T-GMs from neighbors with the host T-GM based on Hungarian assignment (also called Munkres algorithm \cite{Munkres57}) and gating. Then, only associated GCs will be fused in the manner of ``averaging''. For clarity, denote the host sensor as $a$, one of its neighbors as $b \in N_a$, and the number of original T-GCs as $n_a$ and $n_b$, respectively. To carry out Hungarian assignment, a $n_a \times n_b$ cost matrix needs to be constructed as follows: if $n_a \leq n_b$ (otherwise transpose the matrix) \begin{equation} \label{eq:26} \left[ \begin{array}{ccc} C_{1,1} & \cdots & C_{1,n_b} \\ \cdots & \cdots & \cdots \\ C_{n_a,1} & \cdots & C_{n_a,n_b} \\ \end{array} \right] \hspace{0.5mm}, \end{equation} where $C_{i,j}$ is the Mahalanobis-type distance as in \eqref{} between GC $i$ from sensor $a$ and GC $j$ from sensor $b$ The optimal assignment is given by choosing one entry at each row of the cost matrix \eqref{eq:26}, all entries belonging to different columns, with a minimal sum. That is, the optimization cost function is given by \begin{equation} \label{eq:28} \underset{\pi \in \Pi_{n_b}}{\mathrm{argmin}} \sum_{i=1}^{n_a} {C_{i,\pi(i)}} \hspace{0.5mm}, \end{equation} where $\Pi_{n_b}$ is permutations of $n_b$ entries and $\pi(i)$ indicates the $i$th entry in the permutation $\pi$. The Hungarian algorithm has proven to be efficient in solving the above assignment problem in polynomial time \cite{Munkres57}. As a result, all the GCs in the smaller GM set will be assigned to one and only one GC from the larger GM set, while the GCs from the latter will be assigned to one or no GC from the former. We call this \textit{one to one-or-zero} assignment where the unassigned component will be unfused. Furthermore, a double-checking step is required so that only the assigned pair that are close enough are to be fused. Again, we use the Mahalanobis-type distance as given in \eqref{} to measure the distance between two GCs and the rule \eqref{} to design the gating threshold. Any assignment that does not fall in the valid gate will be canceled. The unassigned T-GCs will be abandoned and will not be involved in any fusion if it is received from the neighbor, otherwise it remains unchanged at local sensor. This will guarantee promisingly that the GM set of constant size at each local sensor. Finally, the pairwise assigned GCs from will be fused in the manner of Metropolis weights-based averaging at each iteration, as given below. First, Metropolis weights are used to re-scale all T-GC weights according to their origination sensor. Then, the associated GCs are labeled with the same, to say $\ell$, originating from whether sensor $a$ or its neighbors $N_a$. As addressed above, each sensor contributes a maximum of one GC to each group. We denote all the sensors that contribute one GC to group $\ell$ by a set $S_a^{[\ell]} \in \{a, N_a\}$ for which we have the following proposition for conservative fusion. \textbf{Proposition 3 (GM averaging)} All T-GCs associated in $S_a^{[\ell]}$ are averaged, resulting in a new single GC $\mathcal{N}(\mathbf{x};\mathbf{m} _a^{[\ell]}(t),\mathbf{P}_a^{[\ell]}(t))$ with weight $w_a^{[\ell]}(t)$ as follows: \begin{equation}\label{} w_a^{[\ell]}(t)= \frac{\sum_{l \in S_a^{[\ell]}} \omega_{l\rightarrow a} w_l^{[\ell]}(t-1)}{\sum_{l \in S_a^{[\ell]}} \omega_{l\rightarrow a}}\hspace{0.5mm}, \end{equation} \begin{equation}\label{} \mathbf{m}_a^{[\ell]}(t) = \frac {\sum_{l \in S_a^{[\ell]}} \omega_{l\rightarrow a} w_l^{[\ell]}(t-1)\mathbf{m}_l^{[\ell]}(t-1)} {\sum_{l \in S_a^{[\ell]}} \omega_{l\rightarrow a} w_l^{[\ell]}(t-1)}\hspace{0.5mm}, \end{equation} \begin{equation}\label{} \mathbf{P}_a^{[\ell]}(t) = \mathbf{P}_\mathrm{OMR} \hspace{0.5mm}, \end{equation} where $\mathbf{P}_\mathrm{OMR}$ is given as in \eqref{} by substituting $\mathcal{I} = S_a^{[\ell]}$ and $\tilde{\mathbf{P}}_i= \mathbf{P}_i^{[\ell]}(t-1) + \big(\mathbf{m}_i^{[\ell]}(t-1)-\mathbf{m}_a^{[\ell]}(t)\big) \big(\mathbf{m}_i^{[\ell]}(t-1)-\mathbf{m}_a^{[\ell]}(t)\big)^\mathrm{T}$ for all $i \in S_a^{[\ell]}$. As shown, the calculation of the fused state and covariance is akin to that of CGMM but they are different at the fused weight, which is an average in \eqref{} rather than a sum in \eqref{} in the CGMM. Therefore, \eqref{} does not hold here but instead roughly $\tilde{W}_a(t) \approx W_a(t-1)$ and $\beta_a \approx 1$ instead of \eqref{}. Comparably speaking, the FA-GCs will not be so significantly weakened in CGMA as in CGMM. However, the cardinality average consensus scheme as given in Proposition 2 can still be applied at each communication iteration to re-weight all GCs including the averaged one given in \eqref{} and the insignificant GC that is not involved in fusio . Overall, we refer to this consensus protocol as \textit{conservative GM averaging} (CGMA). \subsection{Potential Extensions} The proposed union-type, conservative fusion and partial-consensus-based distributed GM-PHD fusion can be extended in terms of both the communication protocol and the local filter. In the former, other consensus protocols other than averaging schemes (e.g., diffusion \cite{sayed14,Dedecius17}, flooding \cite{Li17flooding}) can be applied, while in the latter, multi-Bernoulli filters \cite{Guldogan14, Wang17, Suqi17} and even particle filter-based RFS filters can be employed based on novel mixture reduction or particle-resampling schemes. \section{Simulations} \label{} In this section, the proposed CGMM and CGMA approaches for distributed GM-PHD fusion are evaluated for tracking either a single target or simultaneous multiple targets, with comparison to the benchmark GCI/KLA fusion \cite{Battistelli13} and the pure cardinality consensus based on either flooding (CCF) \cite{Li17flooding} or Metropolis weights-based averaging (CCA). These different distributed filters will be evaluated on the same ground truths, sensor data series and sensor network setting up. For MR in all filters: GCs with a weight lower than $10^{-4}$ will be truncated, any two GCs closer than Mahalanobis-distance $\tau =5$ will be merged, and the maximum number of GCs is 50 in the case for tracking a single target and 100 in the case of multiple targets. The proposed partial consensus is carried out based on the rank rule P.1.1 for selecting the T-GCs. To save communication in GCI, we suggest a threshold $w_c = 0.005$ such that only the GC with a weight larger than $w_c$ will be disseminated to neighbors and then be considered in the subsequent fusion. The optimal sub-pattern assignment (OSPA) metric \cite{Schuhmacher08} is used to evaluate the estimation accuracy of the filter, with cut-off parameter $c=1000$ and order parameter $p=2$; for the meaning of these two parameters, please refer to \cite{Schuhmacher08}. We refer to the average of OSPAs obtained by all sensors in the network at each sampling step as \textit{Network OSPA}. The average of the Network OSPAs over all filtering steps is called \textit{Time-average Network OSPA}. To evaluate the communication cost, we record a GC that consists of a weight parameter (1 tuple), a 4-dimensional vector mean (4 tuples), and a 4$\times$4 -dimension matrix covariance (16 tuples) as data size 21 tuples and the scale-valued cardinality parameter as 1 tuple. Given that the covariance matrix is symmetric, only 10 tuples are needed here and then a GC only needs 15 tuples for data storing. Furthermore, to measure the efficiency of different consensus protocols, we define a consensus efficiency (CE) measure regarding the average OSPA reduction gained by sharing each tuple of network data as follows: \begin{equation} \label{eq:CE} \mathrm{CE}\triangleq\frac{\text{OSPA reduction due to communication}}{\text{Network communication cost (no. tuples)}}\hspace{0.5mm}. \end{equation} \begin{figure}\label{fig:2} \centering \includegraphics[width=7 cm]{2} \caption{Tracking scenario, target trajectories and sensor network} \end{figure} The simulations are set up in a scenario over the planar region $[-1000,1000]\mathrm{m}\times [-1000,1000]\mathrm{m}$ which is monitored by a randomly generated sensor network (with total 12 sensors and diameter $6$) as shown in Fig.2. We assume two different ground truths for the target trajectories, to be presented in the following two subsections respectively. To capture the average performance, we perform each simulation 100 MC runs with independently generated observation series for each run. Different numbers $t$ of P2P communication iterations from 0 (without applying any information disseminating) to 12 (twice the network diameter) are applied to all consensus schemes. To set up the local filter, the ground truth is simulated as follows: The target birth process follows a Poisson RFS with intensity function $ \gamma_k(\mathbf{x})= \sum_{i=1}^4 \lambda_i\mathcal{N}(.;\mathbf{m}_i,\mathbf{Q}_r)$, where the Poisson rate parameters $\lambda_1=\lambda_2=\lambda_3=\lambda_4$, the Gaussian parameters $\mathbf{m}_1=[0,0,0,0]^\mathrm{T}$, $\mathbf{m}_2=[-500, 0, -500, 0]^\mathrm{T}$, $\mathbf{m}_3=[0,0,500,0]^\mathrm{T}$, $\mathbf{m}_4=[500, 0, -500, 0]^\mathrm{T}$, $\mathbf{Q}_r=$ diag$([400,100,400,100]^\mathrm{T})$, and diag$(\mathbf{a})$ represents a diagonal matrix with diagonal $\mathbf{a}$. In addition, the target intensity function spawn from target $\mathbf{u}$ is given as $ b_k(\mathbf{x} | \mathbf{u})= 0.05\mathcal{N}(.;\mathbf{u},\mathbf{Q}_b)$, where $\mathbf{Q}_b=$ diag$([100,400,100,400]^\mathrm{T})$. Each target has a time-constant survival probability $p_S(\mathbf{x}_k)=0.99$ and the survival target follows a nearly constant velocity motion as given \begin{equation}\label{} \mathbf{x}_k= \left[ \begin{array}{cccc} 1 & \Delta & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & \Delta \\ 0 & 0 & 0 & 1 \\ \end{array} \right] \mathbf{x}_{k-1}+ \left[ \begin{array}{cc} \Delta^2/2 & 0 \\ \Delta & 0 \\ 0 & \Delta^2/2 \\ 0 & \Delta \\ \end{array} \right] \mathbf{u}_k\hspace{0.5mm}, \end{equation} where $\mathbf{x}_k=[p_{x,k},\dot{p}_{x,k},p_{y,k},\dot{p}_{y,k}]^\mathrm{T}$ with the position $[p_{x,k},p_{y,k}]^\mathrm{T}$ and the velocity $[\dot{p}_{x,k},\dot{p}_{y,k}]^\mathrm{T}$, the sampling interval $\Delta =1$s, and the process noise $\mathbf{u}_k \sim \mathcal{N}(\mathbf{0}_2,25\mathbf{I}_2)$. Without loss of generality, we employ a hybrid sensor network that consists of both linear and nonlinear observation sensors which run linear GM-PHD filter and unscented transform based nonlinear GM-PHD filter \cite{Vo06}, respectively. The sensors are ordered from 1 to 12, where the sensors no.1-6 generate linear observation (which are referred to as linear sensors, marked by square in Fig.2) while the rest (no. 7-12) generate nonlinear observation (referred to as nonlinear sensors, marked by circles in Fig.2). The linear sensors have the same time-constant target detect probability $p_D(\mathbf{x}_k)=0.95$ and the linear position observation model given as follows \begin{equation}\label{} \mathbf{z}_k= \left[ \begin{array}{cccc} 1 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ \end{array} \right] \mathbf{x}_k+ \left[ \begin{array}{c} v_{k,1} \\ v_{k,2} \\ \end{array} \right]\hspace{0.5mm}, \end{equation} with $v_{k,1}$ and $v_{k,2}$ as mutually independent zero-mean Gaussian noise with the same standard deviation of 10. The FOV of each nonlinear sensor is a disc of radius 3000m centralized with the sensor's position $[s_{n,x},s_{n,y}]^\mathrm{T},n=16,..,30$, which is able to fully cover the scenario. The target detection probability depends on the target position $[p_{x,k},p_{y,k}]^\mathrm{T}$, as given by $p_D(\mathbf{x}_k)=0.95 \mathcal{N}([|p_{x,k}-s_{n,x} |,|p_{y,k}-s_{n,y}|]^\mathrm{T}; \mathbf{0},6000^2 I_2)/\mathcal{N}(0;0,6000^2I_2)$ and the nonlinear range and bearing observation is given by \begin{equation} \label{} \mathbf{z}_k= \left[ \begin{array}{c} \sqrt{(p_{x,k}-s_{n,x} )^2+(p_{y,k}-s_{n,y})^2}\\ \arctan\big((p_{y,k}-s_{n,y})/(p_{x,k}-s_{n,x})\big) \\ \end{array} \right] + \mathbf{v}_k\hspace{0.5mm}, \end{equation} where $\mathbf{v}_k \sim \mathcal{N}( ;\mathbf{0},\mathbf{R}_k )$, with $\mathbf{R}_k=$ diag$\left([\sigma_r^2,\sigma_\theta^2 ]^\mathrm{T}\right), \sigma_r=10$m, $\sigma_\theta=\pi/90$ rad/s. Clutter is uniformly distributed over each sensor's FOV with an average rate of $r$ points per scan. For the nonlinear sensors, we set $r=5$ in both scenarios indicating a clutter intensity $\kappa_k = 5/3000/2\pi$ while for th linear sensors we set $r=5$ for the first single target scenario and $r=10$ for the second multiple target scenario, indicating clutter intensities $\kappa_k = 5/2000^2$ and $\kappa_k = 10/2000^2$, respectively. \begin{figure*}\label{fig:3} \centering \includegraphics[width=16 cm]{3} \caption{Network OSPA, online estimated number of targets and computing time of different consensus protocols for each filtering step when six iterations of P2P communication are applied} \end{figure*} \begin{figure*}\label{fig:4} \centering \includegraphics[width=16 cm]{4} \caption{Time-averaged network OSPA, network communication cost and CE against P2P communication iterations} \end{figure*} \subsection{Single Target Scenario} First, we limit the maximal number of targets that simultaneously exist in the scenario to one for generating the ground truth as that new target which can only appear after the existing target disappears. Also, there is no target spawning. That is to say, the tracking at any time actually involves maximally one target, which is favorable for CI/GCI. The network and the ground truth of the target trajectories are given in Fig.2. When a total of $t=6$ P2P communication iterations are applied, the Network OSPA, the online estimated number of targets, and the computing time of different consensus protocols for each filtering step are given in Fig.3, separately. For different numbers of communication iterations, the time-averaged network OSPA, time-averaged network communication cost, and CE of different consensus protocols are given in Fig.4, separately. We have the following key finding : \begin{enumerate} \item All consensus schemes converge with the increase of the number of P2P communication iterations; meanwhile, the more iterations, the higher the communication and computing cost and the lower OSPA. In particular, when $t=1$ (each sensors only share information with their immediate neighbors), GCI yields the best performance, providing the lowest Network OSPA and time-average OSPA over all. When $t\geq 2$ , CGMM yields the lowest OSPA over all which is even better than the GCI. \item When $t=6$, the computing time required by GCI is the most and is much higher than that of the others, while CGMA and CGMM come as the second and the third, respectively. \item On communication cost, CGMM costs slightly more than CGMA, both smaller than GCI, especially when few P2P communication iterations ($t<6$) are applied. \item Cardinality consensus has improved the cardinality estimation in all consensus schemes. However, the benefit of pure cardinality consensus is limited, whether CCF by flooding or CCA by averaging, which converges to a level that is significantly inferior to that of the others including GCI, CGMM, and CGMA. However, this is achieved at the price of significantly less computation and communication. \item The CE decreases with the increase of $t$ in all consensus schemes. Overall, CCA yields the highest CE and CCF comes second; comparably, CCF converges faster at the expense of more communication cost than CCA. When $t\geq 6$, the CCF achieves complete consensus/convergence \cite{Li17flooding} and so it will no further reduce the OSPA, leading to a zero CE. In regard to CE, CGMA slightly outperforms GCI and CGMM while the latter two perform similar. \item Local GM size remains constant favorably during network communication at each filtering step using CCF, CCA, and CGMA but varies (mainly increases with $t$) due to CGMM and GCI. This is one advantage that CGMA has over CGMM and GCI. \end{enumerate} To summarize the results in the single target case, CGMM is a fair alternative to GCI in favor of smaller OSPA and fewer fusion computation and communication, while CGMA is a better choice than GCI in favor of less computation and communication, and higher CE. More discussion will be given in Section \ref{. \subsection{Multiple Target Scenario} In this case, we extend the maximal number of targets that simultaneously exist in the scenario to three for generating a new ground truth. The trajectories of totally four targets are given in Fig.5 with the starting and ending times of each trajectory noted. To show the simulation result, similar contents given in Figs. 6 and 7 correspond to those in Figs. 3 and 4, respectively. While some of them give similar indication, e.g., the relative communication and computation cost of different protocols, key new findings are summarized as follows. \begin{figure}\label{fig:5} \centering \includegraphics[width=7.5 cm]{5} \caption{Trajectories of simultaneously appearing multiple targets} \end{figure} \begin{figure*}\label{fig:6} \centering \includegraphics[width=16 cm]{6} \caption{Network OSPA, online estimated number of targets and computing time of different consensus protocols for each filtering step when six iterations of P2P communication are applied} \end{figure*} \begin{figure*}\label{fig:7} \centering \includegraphics[width=16 cm]{7} \caption{Time-averaged network OSPA, network communication cost and CE for different numbers of consensus iterations} \end{figure*} \begin{enumerate} \item On filtering accuracy, CGMM gets the minimum network OSPA which significantly outperforms the others and CGMA comes second. In particular when $t=1$, CGMM that consumes the same communication as CGMA and smaller than GCI, yields the largest OSPA reduction, even more significant than that of the others by performing multiple iterations of communication. This simply indicates that, only immediate neighborhood T-GM information sharing for partial consensus outperforms sophisticated averaging fusion like GCI based on multiple iterations for complete consensus. We refer to this as ``\textit{many could be better than all}''. \item Somehow surprisingly, GCI does not benefit the filter accuracy when $t=1$. We leave here an attempt to explain the reason. \item On communication, CGMM costs less than GCI if $t\leq 3$ but more if otherwise. CGMA communicates always less than CGMM and GCI. \item On computation, GCI costs the most again while CGMM and CGMA perform similar except when the true number of targets is one for which CGMM may be more computing costly than GCI. \item GCI shows delay at detecting new born targets as it will even increase OSPA compared with the centralized filter with no sensor communication at time $k\in [7,11]$ and $k\in [58,60]$ as shown in the middle sub-figure of Fig.6. This obtuse capability in new target detection is indeed unfavorable, significantly reducing the filtering accuracy (as shown in the left sub-figure of Fig.6). \item CGMA and CGMM perform worse on CE than GCI except the first communication iteration, all inferior to CCF and CCA again. But, their achievements in reducing OSPA is to a large degree more significant than that of GCI and others. \end{enumerate} To summarize this multiple-target case, both CGMA and CGMM (in particular) afford better alternative to GCI in favor of smaller OSPA, less fusion computation and even less communication for the same OSPA reduction gain. \subsection{Further Discussion} Experimental findings reported in the literature are notable. The performance of GCI is greatest for few sensors and distant targets \cite{Uney13, Battistelli14a} or only a single target \cite{Battistelli13, Battistelli14b, Guldogan14}. Closely-distributed targets in dense clutter environment have not been particularly considered except few works such as \cite{Battistelli15, Wang17}, which just showed that GCI made worse result when targets are close. For example, the cardinality estimation is worse at around time $k=800$s when more iterations of GCI fusion are applied, as shown in Fig.5-7 of \cite{Battistelli15}. Delay has also been observed in estimating the number of targets when new targets appear in the scenario in \cite{Gunay16}. More specifically, the simulation given in Section V.A of \cite{Wang17} has explicitly demonstrated that GCI will degrade the local PHD filter in the case of close targets whose distance is under a specific threshold and/or in the case of low SNR. Deficiency of GCI for handling misdetection has been particularly noticed in \cite{Yu16,Yi17}. Relatively, the findings given in \cite{Hwang04} suggested that the arithmetic average method is most robust to incorrect information than the geometric average. It has also been demonstrated that the CI provides estimation error covariance that is not honest but pessimistic for track fusion with feedback, inferior to the minimum variance rule \cite{Mori12}. In summary of our findings and those given in the literature, the problems that a distributed multi-target filter may potentially suffer from due to GCI include: \begin{enumerate} \item Weakness to deal with closely distributed targets and/or low SNR background; \item Prone to mis-detection or local sensor failure; \item Delay in detecting new appearing targets; \item High communication and computation cost for complete consensus. \end{enumerate} It seems still unclear how to fix these problems on the basis of GCI, even some of the causes have been noted, nor was that our intension in this paper. We leave here direct simulation demonstration about the failures of GCI in complicated multi-target scenarios (e.g., new targets appear frequently, targets move closely or there is a high rate of mis-detection or clutter) in which our proposed approaches demonstrate more significant advantage. In particular, the straightforward arithmetic average based CGMM that can be easily implemented on different filter beds yields significant accuracy benefit with only one or two iterations of P2P diffusion. The merit of the presented partial consensus and conservative arithmetic average fusion is not only on reliable and significant consensus benefit, but also on inexpensive communication and computation for complying with the need of real time filtering. It is very crucial to note that, a key challenge in many large-scale WSN scenarios comes exactly from limitations imposed on the communication bandwidth/power allowance and the sensor computing capability because the nodes are low powered wireless devices. \section{Conclusion} \label{} For distributed GM-PHD fusion, this paper has proposed a notion of``partial consensus'' which abandons the ultimate goal that the estimate of each sensor converges to the estimation conditioned on all the information over the entire network but instead neighboring sensors shares only highly-weighted GCs with each other and at the end, the network achieves partially consensus. In addition to saving communication and computation, the local SNR at each sensor can be increased because of partial consensus, reducing the possibility to generate false alarms and facilitating more accurate estimation. To further reduce the communication cost, the disseminated significant GCs can be either pairwise averaged or locally merged in a fully distributed and conservative manner. In parallel, the arithmetic average consensus is sought on the GM weight sum at each communication iteration. Simulations based on both single target scenario and multiple target scenario have been provided to demonstrate the effectiveness and reliability of our approach with comparison to the GCI, which is the state of the art approach for distributed RFS filter fusion. Although the GCI works well in the single target scenario in the presence of low misdetection and clutter rates, it exhibits severe problems in complicated multi-target scenarios, such as delay in detecting new appearing targets, and incompetent to handle closely distributed targets, intensive clutter and mis-detection, in addition to its high communication and computation cost. For multi-target density fusion in the presence of significant clutter and misdetection, our final remarks are: \begin{itemize} \item \textit{Many could be better than all}: the concept of ``partial consensus'' is important as can not only save communication and computation but also benefit the accuracy more than the complete consensus. \item \textit{Union outperforms intersection}: Union-format arithmetic average fusion, as the original average consensus is, is computationally easier and provably more reliable than the Intersection-format geometric average fusion, while the former is also more conservative in general. \end{itemize} \ifCLASSOPTIONcaptionsoff \newpage \fi \bibliographystyle{IEEEtran}

\section{Introduction\label{sec_intro}} There exists a plethora of motion planning and control techniques for self-driving vehicles \cite{paden2016survey}. The diversity is caused by a core difficulty: the trade-off between model complexity and permitted online computation at short sampling times. Three popular control classes and recent vision-based end-to-end solutions are briefly summarized below. \subsection{Model-based control methods\label{subsec_Intro_mdlBased}} In \cite{karaman2011anytime} a \emph{sampling-based} anytime algorithm RRT$^*$ is discussed. Key notion is to refine an initial suboptimal path while it is followed. As demonstrated, this is feasible when driving towards a static goal in a static environment. However, it may be problematic in dynamic environments requiring to constantly replan paths, and where an online sampled suitable trajectory may not be returned in time. Other problems of online sampling-based methods are a limited model complexity and their tendency to produce jagged controls that require a smoothing step, e.g., via conjugate gradient \cite{dolgov2010path}. In \cite{mcnaughton2011motion}, a \emph{lattice-based} method is discussed. Such methods, and similarly also based on \emph{motion primitives} \cite{frazzoli1999hybrid,schouwenaars2003robust, gray2012predictive,liniger2015optimization}, are always limited by the size of the look-up table that can be searched in real-time. In \cite{mcnaughton2011motion}, a GPU is used for search. In \cite{falcone2007predictive}, \emph{linear time-varying model predictive control} (LTV-MPC) is discussed for autonomous vehicles. While appealing for its ability to incorporate constraints, MPC must trade-off model-complexity vs. computational burden from solving optimization problems online. Furthermore, MPC is dependent on state and input reference trajectories, typically for linearization of dynamics, but almost always also for providing a tracking reference. Therefore, a two-layered approach is often applied, with motion planning and tracking as the 2 layers \cite{paden2016survey}. See \cite{plessen2017spatial} for a method using geometric corridor planning in the first layer for reference generation and for the combinatorial decision taking on which side to overtake obstacles. As indicated in \cite[Sect. V-A]{falcone2007predictive} and further emphasized in \cite{plessen2017trajectory2}, the selection of reference velocities can become problematic for time-based MPC and motivated to use \emph{spatial-based} system modeling. Vehicle dynamics can be incorporated by inflating obstacles \cite{gray2012predictive}. For tight maneuvering, a linearization approach \cite{plessen2017trajectory} is more accurate, however, computationally more expensive. To summarize, 2 core observations are made. First, all methods (from sampling-based to MPC) are derived from vehicle \emph{models}. Second, all of above methods suffer from the real-time requirement of short sampling times. As a consequence, all methods make simplifications on the employed model. These include, e.g., omitting of dynamical effects, tire dynamics, vehicle dimensions, using inflated obstacles, pruning search graphs, solving optimization problems iteratively, or offline precomputing trajectories. \begin{figure \vspace{0.3cm} \centering \begin{tikzpicture} \node[color=black] (a) at (-1, 0.88) { \small{human route} }; \node[color=black] (a) at (-1.23, 0.57) { \small{selection} }; \draw[->,>=latex'] (-0.05, 0.72) -- (0.75,0.72); \draw[fill=white,solid] (0.75,0.45) rectangle (1.65,0.95); \node[color=black] (a) at (1.2, 0.72) { Navi }; \draw[->,>=latex'] (1.2, 0.45) -- (1.2,0.15); \draw[fill=white,solid] (0.75,-0.45) rectangle (1.65,0.15); \node[color=black] (a) at (1.2, -0.15) { Filter }; \draw[->,>=latex'] (1.65, -0.15) -- (2.65,-0.15); \node[color=black] (a) at (2.12, 0.03) { \small{\textbf{$s_t$}}}; \draw[fill=white,thick] (2.65,-0.45) rectangle (3.55,0.15); \node[color=black] (a) at (3.1, -0.15) { \textbf{C} }; \node[color=black] (a) at (4, 0) { \small{\textbf{$a_t$}}}; \draw[->,>=latex'] (3.55, -0.15) -- (4.4,-0.15); \draw[fill=white,solid] (4.4,-0.5) rectangle (5.4,0.2); \draw[fill=white,solid] (4.6,-0.3) rectangle (5.2,-0.15); \draw (4.74,-0.35) circle [radius=0.05]; \draw (5.04,-0.35) circle [radius=0.05]; \draw[-] (4.65, -0.15) -- (4.68,-0.02) -- (4.9,-0.02) -- (5.066,-0.15); \draw[-] (4.85, -0.15) -- (4.85,-0.02); \draw[->,>=latex'] (5.4, -0.15) -- (5.9,-0.15) -- (5.9,-0.75) -- (-0.05,-0.75) -- (-0.05,0.02)--(0.75,0.02); \draw[->,>=latex'] (-0.05, -0.32) -- (0.75,-0.32); \node[color=black] (a) at (2.95, -1.0) { \small{extero- \& proprioceptive measurements}}; \end{tikzpicture} \caption{Closed-loop control system architecture. ``Navi'' and ``Filter'' (not focus of this paper) map human route selections as well as extero- and proprioceptive measurements to feature vector $s_t$. This paper proposes a simple gradient-free algorithm (TSHC) for learning of controller C, which maps feature vector $s_t$ to control action $a_t$ to be applied to the vehicle.} \label{} \vspace{-0.6cm} \end{figure} \subsection{Vision-based methods\label{subsec_IntroVision}} In \cite{pomerleau1989alvinn} a pioneering end-to-end trained neural network labeled ALVINN was used for steering control of an autonomous vehicle. Video and range measurements are fed to a fully connected (FC)-network with a single hidden layer with 29 hidden units, and an output layer with 45 direction output units, i.e., discretized steering angles, plus one road intensity feedback unit. ALVINN does not control velocity and is trained using supervised learning based on road ``snapshots''. Similarly, recent DAVE-2 \cite{bojarski2016end} also only controls steering and is trained supervisedly. However, it outputs continuous steering action and is composed of a network including convolutional neural networks (CNN) as well as FC-layers with a total of 250000 parameters. During testing (i.e., after training), steering commands are generated from only a front-facing camera. Another end-to-end system based on only camera vision is presented in \cite{chen2017brain}. First, a driving intention (change to left lane, change to right lane, stay in lane and break) is determined, before steering angle is output from a recurrent neural network (RNN). Instead of mapping images to steering control, in \cite{chen2015deepdriving} and \cite{xu2016end}, affordance indicators (such as distance to cars in current and adjacent lanes etc.) and feasible driving actions (such as straight, stop, left-turn, right-turn) are output from neural networks, respectively. See also \cite{paxton2017combining} and their treatment of ``option policies''. To summarize, it is distinguished between (i) vision-based end-to-end control, and (ii) perception-driven approaches that attempt to extract useful features from images. Note that such features (e.g., obstacle positions) are implicitly required for all methods from Sect. \ref{subsec_Intro_mdlBased}. \subsection{Motivation and Contribution} This work is motivated by the following additional considerations. As noted in \cite{urmson2008autonomous}, localization relative to lane boundaries is more important than with respect to GPS-coordinates, which underlines the importance of lasers, lidars and cameras for automated driving. Second, vehicles are man- and woman-made products for which there exist decade-long experience in vehicle dynamics modeling \cite{gillespie1997vehicle},\cite{rajamani2011vehicle}. There is no reason to a priori entirely discard this knowledge (for manufacturers it is present even in form of construction plans). This motivates to leverage available vehicle models for control design. Consider also the position paper \cite{glasmachers2017limits} for general limitations of end-to-end learning. Third, a general purpose control setup is sought avoiding to switch between different vehicle models and algorithms for, e.g., highway driving and parking. There also exists only one real-world vehicle. In that perspective, a complex vehicle model encompassing all driving scenarios is in general preferable for control design. Also, a model mismatch on the planning and tracking layer can incur paths infeasible to track \cite{gray2012predictive}. Fourth, the most accident causes involving other mobile vehicles are rear-end collisions \cite{national2014traffic}, which most frequently are caused by inattentiveness or too close following distances. Control methods that enable minimal sampling times, such as feedforward control, can deterministically increase safety through minimal reaction times. In contrast, environment motion prediction (which can also increase safety) always remains stochastic. Fifth, small sampling times may contradict using complex vehicle models for control when applied for expensive online optimization or search problems. These considerations motivate a 2-step procedure: first learning of a controller during offline training based on an arbitrarily complicated mathematical system model, before online fast evaluation of the trained controller. In an automated vehicles settings, it implies that once trained, low-cost embedded hardware can be used online for evaluation of only few matrix vector multiplications. The contribution of this paper is a simple gradient-free algorithm for model-based deep reinforcement learning using task separation with hill climbing (TSHC). Therefore, it is specifically proposed to (i) simultaneously train on separate deterministic tasks with the purpose of encoding motion primitives in a neural network, and (ii) during training to employ maximally sparse rewards in combinations with virtual velocity constraints (VVCs) in setpoint proximity. This paper is organized as follows. Problem formulation, the proposed training algorithm and numerical simulation experiments are discussed in Sections \ref{sec_problFormulation}-\ref{sec_expts}, before concluding. \section{Problem Formulation and Preliminaries\label{sec_problFormulation}} \subsection{General setup} The problem formulation is visualized in Fig. \ref{}. Exteroceptive measurements are assumed to include inter-vehicular communication (car-2-car) sensings as well as the communication with a centralized or decentralized coordination service such that, in general, multi-automated vehicle coordination is also enabled~\cite{plessen2016multi}. For learning of controller C it is distinguished between 5 core aspects: the system model used for training, the neural network architecture used for function approximation, the training algorithm, the training tasks selection and the hardware/software implementation. Fundamental objective is to encode many desired motion primitives (training tasks) in a neural network. The main focus of this paper is on the training algorithm aspect, motivated within the context of motion planning for autonomous vehicles characterized by nonholonomic system models. \subsection{Illustrative system model for simulation experiments} For simplicity a simple Euler-discretized nonlinear kinematic bicycle model~\cite{rajamani2011vehicle} is assumed for simulation experiments of Sect. \ref{sec_expts}. Equations of motion are $x_{t+1} = x_t + T_s v_t \cos(\psi_t)$, $y_{t+1}=y_t+T_sv_t \sin(\psi_t)$, $\psi_{t+1}=\psi_t + T_s \frac{v_t}{l_f}\tan(\delta_t)$, with 3 states (position-coordinates and heading), 2 controls (steering angle $\delta$ and velocity $v$), 1 system parameter (wheelbase $l_f=3.5$m), and $t$ indexing sampling time $T_s$. Coordinates $x_t$ and $y_t$ describe the center of gravity (CoG) in the inertial frame and $\psi_t$ denotes the yaw angle relative to the inertial frame. Physical actuator absolute and rate constraints are treated as part of the vehicle model on which the network training is based on. Thus, the continuous control vector is defined as $a_t = [v_t,~\delta_t]$, with $\max( v_{t-1} + \dot{v}_{\min,t}T_s,~ v_{\min,t})\leq v_t \leq \min( v_{t-1} + \dot{v}_{\max,t}T_s,~v_{\max,t})$, and $\max( \delta_{t-1} + \dot{\delta}_{\min}T_s, \delta_{\min}) \leq \delta_t \leq \min( \delta_{t-1} + \dot{\delta}_{\max}T_s, \delta_{\max})$. The minimum velocity is negative to permit reverse driving. \subsection{Comments on feature vector selection $s_t$} While the mathematical system model used for training prescribes $a_t$, this is not the case for feature vector $s_t$. The dimension of $s_t$ may in general be much smaller than the system's state space. In general, $s_t$ may be an arbitrary function of filtered extero- and proprioceptive measurements according to Fig. \ref{}. Thus, a plethora of many different sensors may be compressed through the filtering to a low-dimensional $s_t$. Due to curse of dimensionality low-dimensional $s_t$ are favorable, since the easiest way to generate training tasks is to grid over the elements of $s_t$. Note further that for our purpose of encoding specific motion primitives, feature vector $s_t$ must always relate the current vehicle state with reference to a goal state (e.g., via a difference operator). Certificates about learnt control performance can be provided by statement of (i) the system model used for training, and (ii) the encoded motion primitives (training tasks) and their associated feature vectors. Ultimately, instead of only a single time-instant, $s_t$ may, in general, also represent a collection of multiple past time measurements (time-series) leading up to time $t$. \subsection{Comments on computation\label{subsec_implementability}} For perspective, deep learning using neural networks as function approximators is in general computationally very demanding. To underline remarkable dimensions and computational efforts in practice, note that, for example, in \cite{salimans2017evolution} training is distributed on 80 machines and 1440 CPU cores. In \cite{akiba2017extremely}, even more profoundly, 1024 Tesla P100 GPUs are used in parallel. For perspective, one Tesla P100 permits a double-precision performance of 4.7 TeraFLOPs \cite{teslaP100}. \section{Training Algorithm\label{sec_trainalg}} This section motivates a simple gradient-free algorithm for learning of neural network controllers according to Fig. \ref{}. \subsection{Neural network controller parametrization\label{subsec_netwCtrller}} The controller in Fig. \ref{} may be parameterized by any of, e.g., FCs, LSTM cells including peephole connections \cite{gers2002learning}, GRUs \cite{cho2014learning} and variants. All neural network parameter weights to be learnt are initialized by Gaussian-distributed variables with zero mean and a small standard deviation (e.g., $0.001$). Exceptions are adding a 1 to the LSTM's forget gate biases for LSTM cells, as recommended in \cite{jozefowicz2015empirical}, which are thus initialized with mean $1$. In proposed setting, the affine part of all FC-layers is followed by nonlinear tanh$(\cdot)$ activation functions acting elementwise. Because of their bounded outputs, saturating nonlinearities are preferred over ReLUs, which are used for the hidden layers in other RL settings \cite{lillicrap2015continuous}, but can result in large unbounded layer output changes. Before entering the neural network $s_t$ is normalized elementwise (accounting for the typical range of feature vector elements). The final FC-layer comprises a tanh$(\cdot)$ activation. It accordingly outputs bounded continous values, which are then affinely scaled to $a_t$ via physical actuator absolute and rate constraints valid at time $t$. So far, continuous $a_t$ was assumed. A remark with respect to gear selection is made. Electric vehicles, which appear suitable to curb urban pollution, do not require gearboxes. Nevertheless, in general $a_t$ can be extended to include \emph{discrete} gear as an additional decision variable. Suppose $N_\text{gears}$ gears are available. Then, the output layer can be extended by $N_\text{gears}$ channels, with each channel output representing a normalized probability of gear selection as a function of $s_t$, that can be trained by means of a softmax classifier. \subsection{Reward shaping\label{subsec_rt}} \begin{figure \vspace{0.3cm} \centering \begin{tikzpicture} \draw[->,>=latex',thick] (0, 0) -- (1, 0); \draw[->,>=latex',color=red,dashed, ultra thick] (0, 0) -- (0, 1); \node[color=black] (a) at (0.5, 1.35) {(a)}; \draw[->,>=latex',thick] (2, 0) -- (3, 0); \draw[->,>=latex',color=red,dashed, ultra thick] (2, 0.75) -- (3, 0.75); \node[color=black] (a) at (2.5, 1.35) {(b)}; \draw[->,>=latex',thick] (4.25, 0) -- (5, 0); \draw[->,>=latex',color=red,dashed, ultra thick] (5.75, 0) -- (6.5, 0); \draw[pattern=north west lines, pattern color=black!70] (5.25,-0.5) rectangle (5.5,1.25); \node[color=black] (a) at (4.625, 1.35) {(c)}; \end{tikzpicture} \caption{The problematic of rich rewards. Three scenarios (a), (b) and (c) indicating different start (black) and goal (red dashed) states (position and heading). For (c), an obstacle is added. See Sect. \ref{subsec_rt} for discussion.} \label{} \vspace{-0.5cm} \end{figure} Reward shaping is crucial for the success of learning by reinforcement signals \cite{sutton1998reinforcement}. However, reward shaping was found to be a far from trivial matter in practical problems. Therefore, our preferred choice is motivated in detail. In most practical control problems, a state $z_0$ is given at current time $t=0$, and a desired goal state $z^\text{goal}$ is known. Not known, however, is the shape of the best trajectory (w.r.t. a given criterion) and the control signals that realize that trajectory. Thus, by nature these problems offer a \emph{sparse} reward signal, $r_{\tilde{T}}(z^\text{goal})$, received only upon reaching the desired goal state at some time $\tilde{T}>0$. In the following, alternative \emph{rich reward} signals and \emph{curriculum learning} \cite{bengio2009curriculum} are discussed. \subsubsection{The problematic of designing rich reward signals} A reward signal $r_t(z_t,a_t,s_t)$, abbreviated by $r_t$, is labeled as \emph{rich} when it is time-varying as a function of states, controls and feature vector. Note that the design of any such signal is heuristic and motivated by the hope for accelerated learning through maximally frequent feedback. In the following, the problematic of rich rewards is exposed. First, let $e_{d,t} = \sqrt{ (x_t-x^\text{goal})^2 + (y_t-y^\text{goal})^2 }$, $e_{\psi,t}=|\psi_t - \psi^\text{goal} |$, and $e_{v,t} = |v_t - v^\text{goal} |$ relate states with desired goals, and let a binary flag indicate whether the desired goal pose is reached, \begin{equation} f_t^\text{goal} = \begin{cases} 1, & \text{if}~(e_{d,t}<\epsilon_d) \land (e_{\psi,t}<\epsilon_\psi) \land (e_{v,t}<\epsilon_v),\\ 0, & \text{otherwise}, \end{cases} \label{eq_ftgoal} \end{equation} where $(\epsilon_d,\epsilon_\psi,\epsilon_v)$ are small tolerance hyperparameters. Then, suppose a rich reward signal of the form $r_t = -(\alpha e_{d,t} + e_{\psi,t})$ is designed, which characterizes a weighted linear combination of different measures. This class of reward signals, trading-off various terms and providing feedback at \emph{every} $t$, occurs frequently in the literature \cite{randlov1998learning, koutnik2014online, lillicrap2015continuous, heess2017emergence}. However, as will be shown, for trajectory planning in an automotive setting (especially due to nonholonomic vehicle models), it may easily lead to undesirable behavior. Suppose case (a) in Fig. \ref{} and a maximum simulation time $T^\text{max}$. Then, omitting a discount factor for brevity, $-T^\text{max}\frac{\pi}{2} > - \sum_{t=0}^{\tilde{T}}\alpha e_{d,t}^\star + e_{\psi,t}^\star$, may be obtained for accumulated rewards. Thus, the no-movement solution may incur more accumulated reward, namely $-T^\text{max}\frac{\pi}{2}$, in comparison to the true solution, which is indicated on the right-hand side of the inequality sign. Similarly, for specific $(\alpha,T^\text{max})$, the second scenario (b) in Fig. \ref{} can return a no-movement solution since the initial angle is already coinciding with the target angle. Hence, for a specific $(\alpha,T^\text{max})$-combination, the accumulated reward when not moving may exceed the value of the actual solution. The third scenario (c) in Fig. \ref{} shows that even if reducing rich rewards to a single measure, e.g., $r_t=-e_{d,t}$, an undesired standstill may result. This occurs especially in the presence of obstacles (and maze-like situations in general). To summarize, for finite $T^\text{max}$, the design of rich reward signals is not straightforward and can easily result in solution trajectories that may even be globally optimal w.r.t. accumulated reward, however, prohibit to solve the original problem of determining a trajectory from initial to target state. \subsubsection{The problematic of curriculum learning} In \cite{bengio2009curriculum}, curriculum learning (CL) is discussed as a method to speed up learning by providing the learning agent first with simpler examples before gradually increasing complexity. Analogies to humans and animals are drawn. The same paper also acknowledges the difficulty of determining ``interesting'' examples \cite[Sect. 7]{bengio2009curriculum} that optimize learning progress. \begin{figure \vspace{0.3cm} \centering \begin{tikzpicture} \draw[->,>=latex',thick] (0, 0) -- (1, 0); \draw[->,>=latex',color=red,dashed, ultra thick] (2, 0) -- (2, 1); \node[color=black] (a) at (1.3, 1.25) {(a)}; \draw[->,>=latex',color=black!30] (4, 0) -- (7, 0); \draw[->,>=latex',thick] (4, 0) -- (5, 0); \draw[->,>=latex',color=blue,dashed, ultra thick] (6, 0) -- (7, 0); \node[color=black] (a) at (5.5, 1.2) {(b)}; \end{tikzpicture} \caption{The problematic of curriculum learning. The difficulty of selecting ``simple'' examples is illustrated, see Sect. \ref{subsec_rt} for discussion. The original problem with start (black) and goal (red dashed) state is denoted in (a). A ``simpler'' problem is given in (b).} \label{} \vspace{-0.5cm} \end{figure} Indeed, CL entails the following issues. First, ``simpler'' tasks need to be identified. This is not straightforward as discussed shortly. Second, these tasks must first be solved before their result can serve as initialization to more complex tasks. In contrast, without CL, the entire solution time can be devoted to the complex tasks rather than being partitioned into easier and difficult tasks. In experiments, this was found to be relevant. Third, the solution of an easier task does not necessarily represent a better initialization to a harder problem in comparison to an alternative random initialization. For example, consider the scenario in Fig. \ref{}. The solution of the simpler task does not serve as a better initialization than a purely random initialization of weights. This is since the final solution requires outreaching steering and possibly reversing of the vehicle. The simpler task just requires forward driving and stopping. This simple example illustrates the need for careful manual selection of suitable easier tasks for CL. \subsubsection{The benefits of maximal sparse rewards in combination with virtual velocity constraints} In the course of this work, many reward shaping methods were tested. These include, first, solving ``simpler'' tasks by first dismissing target angles limited to $30^\circ$-deviation from the initial heading. Second, $\epsilon$-tolerances were initially relaxed before gradually decreasing them. Third, it was tested to first solve a task for only the $\epsilon_d$-criterion, then both $(\epsilon_d,\epsilon_\psi)$, and only finally all of $(\epsilon_d,\epsilon_\psi, \epsilon_v)$. Here, also varying sequences (e.g., first $\epsilon_\psi$ instead of $\epsilon_d$) were tested. No consistent improvement could be observed for neither of these methods. On the contrary, solving allegedly simpler task reduced available solver time for the original ``hard'' problems. Without CL the entire solution time can be devoted to the complex tasks. Based on these findings, our preferred reward design method is maximally sparse and defined by \begin{align} r_t &= \begin{cases} -\infty, & \text{if}~f_t^\text{crash}==1,\\ -1, & \text{otherwise}, \end{cases}\label{eq_def_rt}\\ F_i &= \begin{cases} 1, & \text{if}~\sum_{\tau=t-T^\text{goal}+1}^t f_{\tau}^\text{goal}==T^\text{goal},\\ 0, & \text{otherwise},\label{eq_def_Fi} \end{cases} \end{align} where $f_t^\text{goal}$ from \eqref{eq_ftgoal}, and $f_t^\text{crash}$ being an indicator flag for a vehicle crash. Thus, upon $F_i=1$ the RL problem is considered as solved. In addition, the pathlength incurred for a transition from sampling time $t$ to $t+1$ is defined as \begin{equation} \Delta p_t = -\sqrt{ (x_{t+1}-x_t)^2 + (y_{t+1}- y_t)^2 }. \end{equation} As elaborated below, accumulated total pathlength is used to rank solution candidates solving all desired training tasks. The integral for $F_i$ is defined for generality, in particular for problems such as the inverted pendulum \cite{anderson1989learning} in mind, which are considered to be solved only after stabilization is demonstrated for sufficiently many consecutive $T^\text{goal}$ time steps. Note, however, that this is not required for an automotive setting. Here, it must be $T^\text{goal}=1$. Only then learning with $v^\text{goal}\neq 0$ is possible. Other criteria and trade-offs for $\Delta p_t$ are possible (e.g., accumulated curvature of resulting paths and a minmax objective therefore). The negation is introduced for maximization (``hill climbing''-convention). Note that the preferred reward signal is maximal sparse, returning $-1$, for all times up until reaching the target. It represents a \emph{tabula rasa} solution critizised in \cite{randlov1998learning} for its maximal sparsity. Indeed, standalone it was not sufficient to facilitate learning when also accounting for a velocity target $v^\text{goal}$. Therefore, \emph{virtual velocity constraints} (VVCs) in target proximity are introduced. Two variants are discussed. First, VVCs spatially dependent on $e_{d,t}$ can be defined as \begin{equation} v_t^m = \begin{cases} v^m, & \text{if}~e_{d,t}\geq R_{v}^\text{thresh},\\ v^\text{goal} + \frac{(v^m - v^\text{goal})}{R_{v}^\text{thresh}} e_{d,t}, & \text{otherwise}, \end{cases} \label{eq_virtualCstrt_v} \end{equation} where $m\in\{\text{max},\text{min}\}$, and $R_{v}^\text{thresh}$ is a hyperparameter (e.g., range-view length or a heuristic constant). Second and alternatively, VVCs may be defined as spatially invariant with a constant margin (e.g., 5km/h) around the target velocity. For both variants, the neural network output that regulates velocity is scaled with updated $v_t^\text{min}$ and $v_t^\text{max}$ constraints (i.e., using \eqref{eq_virtualCstrt_v} for spatially dependent VVCs). Let us further legitimize VVCs. Since speed is a decision variable it can always be constrained artificially. This justifies the introduction of VVCs. In \eqref{eq_virtualCstrt_v}, bounds are set to affinely converge towards $v^\text{goal}$ in the proximity of the goal location. This is a heuristic choice. Note that the affine choice do \emph{not} necessarily imply constant accelerations. This is since \eqref{eq_virtualCstrt_v} is spatially parameterized. Note further that physical actuator rate constraints still hold when $a_t$ is applied to the vehicle. It was also tested to constrain $\delta_t$. The final heading pose implies circles prohited from trespassing because of the nonholonomic vehicle dynamics. It was tested to add these as virtual obstacles. However, this did not accelerate learning. Finally, note that VVCs artificially introduce hard constraints and thus shape the learning result w.r.t velocity, at least towards the end of the trajectory. Two comments are made. First, in receding online operation, with additional frequent resetting of targets, this shaping effect is reduced since only the first control of a planned trajectory is applied. Second, in case of spatially dependent VVCs the influence of hyperparameter $R_v^\text{thres}$ only becomes apparent during parking when following the trajectory up until standstill. Here, however, no significant velocity changes are desired, such that the $R_v^\text{thres}$-choice is not decisive. Ultimately, note that sparse rewards naturally avoid the need to introduce trade-off hyperparameters for the weighting of states in different units. This permits solution trajectories between start and goal poses to naturally evolve without biasing them by provision of rich references to track. To summarize this section. It was illustrated that the design of \emph{rich reward} signals as well as \emph{curriculum learning} can be problematic. Therefore, \emph{maximal sparse rewards} in combination with \emph{virtual velocity constraints} are proposed. \subsection{The role of tolerances $\epsilon$\label{subsec_role_tolerances_eps}} Tolerances $\epsilon$ in \eqref{eq_ftgoal} hold an important role for 2 reasons. On one hand, nonzero $\epsilon$ result in deviations between actually learnt $\hat{z}^\text{goal}$ and originally desired goal pose $z^\text{goal}$. On the other hand, very small $\epsilon$ (e.g., $\epsilon_d=0.1$m, $\epsilon_\psi=1^\circ$ and $\epsilon_v=1$km/h) prolong learning time. Two scenarios apply. First, for a network trained on a large-scale and dense grid of training tasks and for small $\epsilon$, during online operation, suitable control commands are naturally interpolated even for setpoints not seen during training. The concept of natural interpolation through motion primitives encoded in neural networks is the core advantage over methods relying on look-up tables with stored trajectories, which require to solve time-critical search problems. For example, in \cite{mcnaughton2011motion} exhaustive search of the entire lattice-graph is conducted online on a GPU. In \cite{liniger2015optimization}, a total of about 100 motion primitives is considered. Then, online an integer program is solved by enumeration using maximal progress along the centerline as criterion for selection of the best motion primitive. In contrast, for control using neural networks as function approximators this search is not required. Second, the scenario was considered in which existing training hardware does (i) not permit large-scale encoding, and (ii) only permits to use larger $\epsilon$-tolerances to limit training time. Therefore, the following method is devised. First, tuples $(\hat{z}^\text{goal},z^\text{goal})$ are stored for each training task. Then, during online operation, for any setpoint, $z^\text{setpoint}$, the closest (according to a criterion) $\hat{z}^\text{goal}$ from the set of training tasks is searched, before the corresponding $z^\text{goal}$ is applied to the network controller. Two comments are made. First, in order to reach $\hat{z}^\text{goal}$ (with zero deviation), $z^\text{goal}$ must be applied to the network. Therefore, \emph{tuples} need to be stored. Second, eventhough this method now also includes a search, it still holds an important advantage over lattice-based methods. This is the compression of the look-up table in the network weights. Hence, only tuples need to be stored---not entire trajectories. This is especially relevant in view of limited hardware memory. Thus, through encoding, potetially many more motion primitives can be stored. In practice, the first scenario is preferable. It is also implementable for 2 reasons. First, see Sect. \ref{subsec_implementability} for computational opportunities. Second, neural networks have in principle unlimited function approximation capability~\cite{siegelmann1991turing}. Hence, the implementation of the first approach is purely a question of intelligent task setup, and computational power. \subsection{Main Algorithm -- TSHC\label{subsec_mainAlg_TSHC}} Algorithm \ref{alg_TSHC} is proposed for simple gradient-free model-based reinforcement learning. The name is derived from the fact of (i) learning from separate training tasks, and (ii) a hill climbing update of parameters (greedy local search). Let us elaborate on definitions. Analysis is provided in Sect. \ref{subsec_analysis_TSHC}. First, all network parameters are lumped into variable $\theta$. Second, the perturbation step 8 in Algorithm \ref{alg_TSHC} has to be intepreted accordingly. It implies parameter-wise affine perturbations with zero-mean Gaussian noise and spherical variance $\sigma_\text{pert}^2$. Third, $\mathcal{X}(\cdot)$, $\mathcal{R}(\cdot)$ and $\mathcal{Z}(\cdot)$ in Steps 14-16 denote functional mappings between properties defined in the preceding sections. Fourth, hyperparameters are stated in Step 1. While $N_\text{restarts}$, $N_\text{iter}^\text{max}$, $n$, $N_\text{tasks}$ and $T^\text{max}$ denote lengths of different iterations, $\beta>1$ is used for updating of $\sigma_\text{pert}$ in Step 35 and 37. Fifth, for every restart iteration, $i_\text{restart}$, multiple parameter iterations are conducted, at most $N_\text{iter}^\text{max}$ many. Sixth, in Steps 25 and 29 hill climbing is conducted, when (i) all tasks have been solved for current $i_\text{iter}$, or (ii) not all tasks have yet been solved, respectively. Seventh, there are 2 steps in which an early termination of iterations may occur: Step 21 and 41. The former is a must. Only then learning with $v^\text{goal}\neq 0$ is possible. The latter termination criterion in Step 41 is optional. If dismissed, a refinement step is implied. Thus, eventhough all $N_\text{tasks}$ tasks have been solved, parameter iterations (up until $N_\text{iter}^\text{max}$) are continued. Eighth, note that a discount hyperparameter $\gamma$, common to gradient-based RL methods \cite{schulman2017proximal}, is not required. This is since it is irrelevant in the maximally sparse reward setting. Ninth, nested parallelization is in principle possible with an inner and outer parallelization of Steps 10-22 and 7-22, respectively. The former refers to $N_\text{tasks}$ solutions for a given parameter vector $\theta_i$, whereas the latter parallelizes $n$ parameter perturbations. For final experiments, Steps 7-22 were implemented asynchronously. Finally, there are 3 options considered for $\sigma_\text{pert}$-selection. First, holding an initial $\sigma_\text{pert}$-selection constant throughout TSHC. Second, updating $\sigma_\text{pert}$ randomly (e.g., uniformly distributed between 10 and 1000), whereby this can be implemented either in Step 4 at every $i_ \text{restart}$, or in Step 6 at every $(i_\text{restart},i_\text{iter})$-combination. Third, $\sigma_\text{pert}$ may be adapted according to progress in $N^{\text{tasks},\star}$, as outlined in Algorithm \ref{alg_TSHC}. For the first 2 options of selecting $\sigma_\text{pert}$, Steps 34-37 are dismissed and at least $\beta$ can be dismissed from the list of hyperparameters in Step 1. \vspace{-0.2cm} \begin{algorithm} \begin{small} \DontPrintSemicolon \textbf{Input:} system model, network structure, $N_\text{tasks}$ training tasks; $N_\text{restarts}$, $N_\text{iter}^\text{max}$, $n$, $T^\text{max}$, $\epsilon$; and a method to update $\sigma_\text{pert}$: constant, random or adaptive based on ($\beta$, $\sigma_\text{pert}^\text{min}$, $\sigma_\text{pert}^\text{max}$).\; Initialize $\theta^\star \leftarrow \emptyset$, $N^\star\leftarrow 0$, $P^\star \leftarrow -\infty$, $J^\star \leftarrow 0$.\; \For{$i_\text{restart}=1,\dots,N_\text{restarts}$} { Initialize $\theta$ randomly, and $\sigma_\text{pert}\leftarrow \sigma_\text{pert}^\text{max},~N_\text{old}^{\text{tasks},\star}\leftarrow 0$.\; \For{$i_\text{iter}=1,\dots,N_\text{iter}^{\text{max}}$} { \% RUN ASYNCHRONOUSLY: \; \For {$i=1,\dots,n$} { Perturb $\theta_i \leftarrow \theta + \sigma_\text{pert}\zeta$, with $\zeta\sim\mathcal{N}(0,I)$.\; Initialize $N_i^{\text{tasks},\star}\leftarrow 0$, $P_i\leftarrow 0$, $J_i \leftarrow 0$.\; \For {$i_\text{task}=1,\dots,N_\text{tasks}$} { Initialize $z_0$ (and LSTM and GRU cells).\; \For {$t=0,\dots,T^\text{max}-1$} { Read $s_t$ from $i_\text{task}$-environment.\; $a_t \leftarrow \mathcal{X}(s_t,\theta_i)$.\; $(r_t,\Delta p_t,f_t^\text{goal}) \leftarrow \mathcal{R}(s_t,a_t)$.\; $z_{t+1} \leftarrow \mathcal{Z}(z_t,a_t)$.\; $P_i \leftarrow P_i + \Delta p_t$.\; $J_i \leftarrow J_i + r_t$.\; $F_i$ according to \eqref{eq_def_Fi}.\; \If {$(F_i==1) \lor(r_t==-\infty)$} {Break $t$-loop.\;} } $N_i^{\text{tasks},\star} \leftarrow N_i^{\text{tasks},\star} + F_i$.\; } } \% DETERMINE $i^\star$: \; \If {$\max\limits_i\{N_i^{\text{tasks},\star}\}_{i=1}^n==N_\text{tasks}$} { $i^\star= \text{arg}\max\limits_i \left\{P_i ~|~ N_i^{\text{tasks},\star}==N_\text{tasks}\right\}_{i=1}^n$.\; \If {$P_{i^\star}>P^\star$}{ $ (\theta^\star,N^\star,P^\star,J^\star ) \leftarrow ( \theta_{i^\star}, N_\text{tasks}, P_{i^\star},J_{i^\star})$.\; } } \Else { $i^\star= \text{arg}\max\limits_i \left\{ J_i \right\}_{i=1}^n$.\; \If {$(J_{i^\star}>J^\star) \land (P^\star==-\infty)$}{ $ (\theta^\star,N^\star,P^\star,J^\star ) \leftarrow ( \theta_{i^\star}, N_{i^\star}^\text{tasks}, P_{i^\star},J_{i^\star})$.\; } } $N^{\text{tasks},\star} \leftarrow N_{i^\star}^{\text{tasks},\star}$.\; \% UPDATE PARAMETERS: \; \If {$N^{\text{tasks},\star} > N_\text{old}^{\text{tasks},\star}$} { $\sigma_\text{pert} \leftarrow \max(\frac{1}{\beta}\sigma_\text{pert},~\sigma_\text{pert}^\text{min})$.\; } \ElseIf {$N^{\text{tasks},\star} < N_\text{old}^{\text{tasks},\star}$} { $\sigma_\text{pert} \leftarrow \min(\beta\sigma_\text{pert},~\sigma_\text{pert}^\text{max})$.\; } $\theta \leftarrow \theta_{i^\star}$ and $N_\text{old}^{\text{tasks},\star}\leftarrow N^{\text{tasks},\star}$.\; \% OPTIONAL: \; \If {$N_i^{\text{tasks},\star}==N_\text{tasks}$} { Break $i_\text{iter}$-loop. {\color{black} \% no further refinement step.}\;} } } \textbf{Output}: $(\theta^\star,N^\star,P^\star,J^\star )$.\; \caption{TSHC}\label{alg_TSHC} \end{small} \end{algorithm} \vspace{-0.5cm} \subsection{Analysis \label{subsec_analysis_TSHC}} According to classifications in \cite{fu2005simulation}, TSHC is a gradient-free instance-based simulation optimization method, generating new candiate solutions based on only the current solution and random search in its neighborhood. Because of its hill climbing (greedy) characteristic, it differs from (i) \emph{evolutionary} (population-based) methods that construct solution by combining others typically using weighted averaging \cite{wierstra2014natural, salimans2017evolution}, and (ii) from model-based methods that use \emph{probability distributions} on the space of solution candidates, see \cite{fu2005simulation} for a survey. In its high-level structure, Algorithm \ref{alg_TSHC} can be related to the COMPASS algorithm \cite{xu2010industrial}. Within a global stage, they identify several possible regions with locally optimal solutions. Then, they find local optimal solutions for each of the identified regions, before they select the best solution among all identified locally optimal solutions. In our setting, these regions are enforced as the separate training tasks and the best solution for all of these is selected. In combination with sufficiently large $n$, $\sigma_\text{pert}$ must be large enough to permit sufficient exploration such that a network parametrization solving all tasks can be found. In contrast, the effect of decreasing $\sigma_\text{pert}$ with an increasing number of solved tasks is that, ideally, a speedup in learning progress results from the assignment of more of solution candidates $\theta_i$ closer in variance to a promising $\theta$ (see Step 8 of TSHC). Steps 29-31 are discussed. For the case that for a specific $i_\text{iter}$-iteration not all tasks have yet been solved, $i^\star= \text{arg}\max\limits_i \{ N_i^{\text{tasks},\star} \}_{i=1}^n$ has been considered as an alternative criterion for Step 29. Several remarks can be made. First, Step 29 and the alternative are not equal. This is because, in general, different tasks are solved in a different number of time steps. However, the criteria are approximately equivalent for sparse rewards (since $J_i$ accumulates constants according to \eqref{eq_def_rt}), and especially for large $T^\text{max}$. The core advantage of employing Step 29 in TSHC is that it can, if desired, also be used in combination with rich rewards to accelerate learnig progress (\emph{if} a suitable rich reward signal can be generated). In such a scenario, $i^\star$ according to Step 29 is updated towards most promisining $J_{i^\star}$, then representing the accumulated rich reward. Thus, in contrast to \eqref{eq_def_rt}, a rich reward could be represented by a weighted sum of squared errors between state $z_t\in\mathbb{R}^{n_z}$ and a reference $z_t^\text{ref}\in\mathbb{R}^{n_z}$, \begin{equation} r_t = \begin{cases} -\infty, & \text{if}~f_t^\text{crash}==1,\\ - \sum_{l=0}^{n_z-1} \alpha_l (z_t(l) - z_t^\text{ref}(l))^2, & \text{otherwise}, \end{cases} \label{eq_rt_rich} \end{equation} where $\alpha_l$, are trade-off hyperparameters and scalar elements of vectors are indexed by $l$ in brackets. Another advantage of the design in Algorithm \ref{alg_TSHC} according to Step 29-31 is its \emph{anytime solution} character. Even if not all $N_\text{tasks}$ are solved, the solution returned for the tasks that are solved, typically is of good quality and optimized according to Steps 29-31. If for all $N_\text{tasks}$ tasks there exists a feasible solution for a given system model and a sufficiently expressive network structure parameterized by $\theta$, then Algorithm \ref{alg_TSHC} can find such parametrization for sufficiently large hyperparameters $N_\text{restarts}$, $N_\text{iter}^\text{max}$, $n$, $T^\text{max}$ and $\sigma_\text{pert}^\text{max}$. The solution parametrization $\theta^\star$ is the result from the initialization Step 4 and parameter perturbations according to Step 8, both nested within multiple iterations. As noted in \cite{hong2009brief}, for optimization via simulation, a global convergence guarantee provides little practical meaning other than reassuring a solution will be found ``eventually'' when simulation effort goes to infinity. However, the same reference also states that a convergence property is most meaningful if it can help in designing suitable stopping criteria. In our case, there are 2 such conceptual levels of stopping criteria: first, the solution of all training tasks, and second, the refinement of solutions. Control design is implemented hierarchically in 2 steps. First, suitable training tasks (desired motion primitives) are defined. Then, these are encoded in the network by the application of TSHC. This has practical implications. First, it encourages to train on deterministic tasks. Furthermore, at every $i_\text{iter}$, it is simultaneously trained on all of these separate tasks. This is beneficial in that the best parametrization, $\theta^\star$, is clearly defined via Step 25, maximizing the accumulated $P$-measure over \emph{all} tasks. Second, it enables to provide certificates on the learnt performance, which can be provided by stating (i) the employed vehicle model, and (ii) the list of encoded tasks (motion primitives). Note that such certificates cannot be given for the class of \emph{stochastic} continuous action RL algorithms that are derived from the Stochastic Policy Gradient Theorem \cite{sutton2000policy}. This class includes all stochastic actor-critic algorithms, including A3C \cite{mnih2016asynchronous} and PPO \cite{schulman2017proximal}. \subsection{Discussion and comparison with related RL work\label{subsec_relatedWork_TSHC}} Related continuous control methods that use neural network for function approximation are discussed, focusing on one stochastic \cite{schulman2017proximal}, one deterministic policy gradient method \cite{lillicrap2015continuous}, and one evolution strategy \cite{salimans2017evolution}. The methods are discussed in detail to underline aspects of TSHC. First, the \emph{stochastic policy gradient} method PPO \cite{schulman2017proximal} is discussed. Suppose that a stochastic continus control vector is sampled from a Gaussian distribution parameterized\footnote{In this setting, mean and variance of the Gaussian distribution are the output of a neural network whose parameters are summarized by lumped $\theta$.} by $\theta$ such that $a_t\sim \pi(a_t | s_t,\theta)$. Then, \begin{equation} J(\theta) = \underset{s_t,a_t}{\text{E}} \left[ g(s_t,a_t\sim \pi(a_t | s_t,\theta)) \right],\label{eq_Jtheta_stochastic} \end{equation} is defined as the expected accumulated and time-discounted reward when at $s_t$ drawing $a_t$, and following the stochastic policy for all subsequet times when acting in the simulation environment. Since function $g(s_t,a_t)$ is a priori not know, it is parameterized by $\theta_{V,\text{old}}$ and estimated. Using RL-terminology, in the PPO-setting, $g(s_t,a_t)$ represents the \emph{advantage function}. Then, using the ``log-likelihood trick'', and subsequently a first-order Taylor approximation of $\log( \pi(a_t | s_t,\theta) )$ around some reference $\pi(a_t | s_t,\theta_\text{old})$, the following parameterized cost function is obtained as an \emph{approximation} of \eqref{eq_Jtheta_stochastic}, \begin{equation} \tilde{J}(\theta) = \underset{s_t,a_t}{\text{E}} \left[ \hat{g}_t(\theta_{V,\text{old}}) \frac{\pi(a_t | s_t,\theta)}{\pi(a_t | s_t,\theta_\text{old})} \right]. \label{eq_Jtheta_stochasticPPO_approx} \end{equation} Finally, \eqref{eq_Jtheta_stochasticPPO_approx} is modified to the final PPO-cost function~\cite{schulman2017proximal} \begin{align} \hat{J}(\theta) =& \underset{s_t,a_t}{\text{E}} \big[\min\bigg( \hat{g}_t(\theta_{V,\text{old}}) \frac{\pi(a_t | s_t,\theta)}{\pi(a_t | s_t,\theta_\text{old})},\notag\\ & \text{clip}( \frac{\pi(a_t | s_t,\theta)}{\pi(a_t | s_t, \theta_\text{old})}, 1-\epsilon,1+\epsilon)\hat{g}_t(\theta_{V,\text{old}}) \bigg) \big], \label{eq_Jtheta_stochasticPPO_PPO} \end{align} whereby the advantage function is estimated by the policy parameterized by $(\theta_{V,\text{old}},\theta_{\text{old}})$, which is run for $T$ consecutive time steps such that for all $t$ the tuples $(s_t,a_t,r_t,s_{t+1},\hat{g}_t(\theta_{V,\text{old}}))$ can be added to a \emph{replay buffer}, from which later minibatches are drawn. According to \cite{schulman2017proximal}, the estimate is $\hat{g}_{T-1}(\theta_{V,\text{old}}) = \kappa_{T-1}$ with $\kappa_{T-1} = r_{T-1} + \gamma V(s_T,\theta_{V,\text{old}}) - V(s_{T-1},\theta_{V,\text{old}})$, $\hat{g}_{T-2}(\theta_{V,\text{old}}) = \kappa_{T-2} + \gamma \lambda \hat{g}_{T-1}(\theta_{V,\text{old}})$ and so forth until $\hat{g}_{0}(\theta_{V,\text{old}})$, and where $V(s,\theta_{V,\text{old}})$ represents a \emph{second}, the so-called critic neural network. Then, using uniform randomly drawn minibatches of size $M$, parameters $(\theta_V,\theta)$ of both networks are updated according to $\text{arg}\min\limits_{\theta_V}\frac{1}{M} \sum_{i=0}^{M-1} \left( V(s_i,\theta_V) - \left(\hat{g}_{i}(\theta_{V,\text{old}}) - V(s_i,\theta_{V,\text{old}})\right) \right)^2$ and $\text{arg}\max\limits_{\theta}\frac{1}{M} \sum_{i=0}^{M-1} \mathcal{G}_i(\theta)$, with $\mathcal{G}_i(\theta)$ denoting the argument of the expectation in \eqref{eq_Jtheta_stochasticPPO_PPO} evaluated at time-index $i$. This relatively detailed discussion is given to underline following observations. With first the introduction of a parameterized estimator, then a first-order Taylor approximation, and then clipping, \eqref{eq_Jtheta_stochasticPPO_PPO} is an arguably crude approximation of the original problem \eqref{eq_Jtheta_stochastic}. Second, the complexity with \emph{two} actor and critic networks is noted. Typically, both are of the same dimensions apart from the output layers. Hence, when not sharing weights between the networks, approximately \emph{twice} as many parameters are required. However, \emph{when} sharing any weights between actor and critic network, then optimization function \eqref{eq_Jtheta_stochasticPPO_PPO} must be extended accordingly, which introduces another approximation step. Third, note that gradients of both networks must be computed for backpropagation. Fourth, the dependence on \emph{rich} reward signals is stressed. As long as the current policy does not find a solution candidate, in a \emph{sparse} reward setting, all $r_i$ are uniform. Hence, there is no information permitting to find a suitable parameter update direction and all of the computational expensive gradient computations are essentially not usable\footnote{It is mentioned that typically the first, for example, 50000 samples are collected \emph{without} parameter update. However, even then that threshold must be selected, and the fundamental problem still perseveres.}. Thus, the network parameters are still updated entirely at random. Moreover, \emph{even if} a solution candidate trajectory was found, it is easily averaged out through the random minibatch update. This underlines the problematic of \emph{sparse} rewards for PPO. Fifth, A3C \cite{mnih2016asynchronous} and PPO \cite{schulman2017proximal} are by nature \emph{stochastic} policies, which \emph{draw} their controls from a Gaussian distribution (for which mean and variance are the output of a trained network with current state as its input). Hence, exact repetition of any task (e.g., the navigation between 2 locations) cannot be guaranteed. It can only be guaranteed if dismissing the variance component, and consequently using solely the mean for deterministic control. This can be done in practice, however, introduces another approximation step. \emph{Deterministic policy gradient} method DDPG \cite{lillicrap2015continuous} is discussed. Suppose a deterministic continuous control vector parameterized such that $a_t = \mu(s_t,\theta)$. Then, the following cost function is defined, \begin{equation*} J(\theta) = \underset{s_t,a_t}{\text{E}} \left[ g(s_t,a_t= \mu(s_t,\theta)) \right] = \underset{s_t}{\text{E}} \left[ g(s_t,\mu(s_t,\theta)) \right]. \end{equation*} Its gradient can now be computed by applying the chain-rule for derivatives \cite{silver2014deterministic}. Introducing a parameterized estimate of $g(s_t,a_t)$, which here represents the \emph{Q-function} or \emph{action value function} (in contrast to the advantage function in above stochastic setting), the final DDPG-cost function \cite{lillicrap2015continuous} is \begin{equation*} J(\theta) = \underset{s_t}{\text{E}} \left[ \hat{g}(s_t,\mu(s_t,\theta),\theta_Q) \right]. \end{equation*} Then, using minibatches, critic and actor network parameters $(\theta_Q,\theta)$ are updated as $\text{arg}\min\limits_{\theta_Q}\frac{1}{M} \sum_{i=0}^{M-1} ( \hat{g}(s_i,a_i,\theta_Q) -\left(r_i + \gamma Q(s_{i+1},\mu(s_{i+1},\theta_\text{old}),\theta_{Q,\text{old}})\right) )^2$ and $\text{arg}\min\limits_{\theta}\frac{1}{M} \sum_{i=0}^{M-1} Q(s_i,\mu(s_i,\theta),\theta_Q)$, with slowly tracking target network parameters $(\theta_{Q,\text{old}},\theta_\text{old})$. Several remarks can be made. First, the Q-function is updated towards only its \emph{one-step} ahead target. It is obvious that rewards are therefore propagated \emph{very} slowly. For sparse rewards this is even more problematic than for rich rewards, especially because of the additional danger of averaging out important update directions though random minibatch sampling. Furthermore, and analogous to the stochastic setting, for the sparse reward setting, as long as no solution trajectory was found, all of the gradient computations are not usable and all network parameters are still updated entirely at random. DDPG is an \emph{off-policy} algorithm. In \cite{lillicrap2015continuous}, exploration of the simulation environment is achieved according to the current policy plus additive noise following an \emph{Ornstein-Uhlenbeck} process. This is a mean-reverting linear stochastic differential equation \cite{geering2011stochastic}. A first-order Euler approximation thereof can be expressed as the action exploration rule $a_t = \mu(s_t,\theta)(1-P_\theta^\text{OU}) + P_\sigma^\text{OU} \epsilon,~\epsilon\sim\mathcal{N}(0,I)$, with hyperparameters $(P_\theta^\text{OU},P_\sigma^\text{OU})=(0.15,0.2)$ in \cite{lillicrap2015continuous}. This detail is provided to stress a key difference between policy gradient methods (both stochastic and deterministic), and methods such as \cite{salimans2017evolution} and TSHC. Namely, while the former methods sample controls from the stochastic policy or according to \emph{heuristic} exploration noise before updating parameters using minibatches of \emph{incremental} tuples $(s_i,a_i,r_i,s_{i+1})$ plus $\hat{g}_i(\theta_{V,\text{old}})$ for PPO, the latter directly work in the \emph{parameter space} via local perturbations, see Step 8 of Algorithm \ref{alg_TSHC}. This approach appears particularly suitable when dealing with sparse rewards. As outlined above, in such setting, parameter updates according to policy gradient methods are also entirely at random, however, with the computationally significant difference of first an approximately four times as large parameter space and, second, the unnecessary costly solution of non-convex optimization problems as long as no solution trajectory has been found. A well-known issue in training neural networks is the problem of vanishing or exploding gradients. It is particularly relevant for networks with saturating nonlinearities and can be addressed by batch \cite{ioffe2015batch} and layer normalization \cite{ba2016layer}. In both normalization approaches, \emph{additional} parameters are introduced to the network which must be learnt (bias and gains). These issues are not relevant for the proposed gradient-free approach. \begin{figure*} \centering% \includegraphics[scale=0.06]{ex1_ddpg.png} \includegraphics[scale=0.06]{ex1_ppo.png} \includegraphics[scale=0.06]{ex1_tshc.png} \caption{Experiment 1. 1000 training trajectories resulting from the application DDPG (Left), PPO (Middle) and TSHC (Right), respectively. The effect of virtual constraints on velocity is particularly visible for DDPG. For the given hyperparameter setting \cite[Tab. 3]{schulman2017proximal}, the trajectories for PPO have little spread and are favoring reverse driving. TSHC has a much better exploration strategy resulting from noise perturbations in the parameter space. The task is solved by TSHC in only 2.1s of learning time, when terminating upon the first solution found (no refinement step, no additional restart).} \label{fig_ex1_ddpgppotshc} \vspace{-0.5cm} \end{figure*} This paper is originally inspired by and most closely related to \cite{salimans2017evolution}. The main differences are discussed. The latter evolutionary (population-based) strategy updates parameters using a \emph{stochastic gradient estimate}. Thus, it updates $\theta \leftarrow \theta + \alpha \frac{1}{n \sigma}\sum_{i=1}^n R_i \zeta_i$, where hyperparameters $\alpha$ and $\sigma$ denote the learning rate and noise standard deviation, and where $R_i$ here indicates the stochastic scalar return provided by the simulation environment. This weighted averaging approach for the stochastic gradient estimate is not suitable for our control design method when using separate deterministic training tasks in combination with maximally sparse rewards. Here, hill climbing is more appropriate. This is since most of the $n$ trajectory candidates do not end up at $z^\text{goal}$ and are therefore not useful. Note also that only the introduction of virtual velocity constraints permitted us to quickly train with maximally sparse rewards. It is well known that for gradient-based training, especially of RNNs, the learning rate ($\alpha$ in \cite{salimans2017evolution}) is a critical hyperparameter choice. In the hill climbing setting this issue does not occur. Likewise, \emph{fitness shaping} \cite{wierstra2014natural}, also used in \cite{salimans2017evolution}, is not required. Note that above $\sigma$ has the same role as $\sigma_\text{pert}$. Except, in our setting, it additionally is adaptive according to Steps 29 and 31 in Algorithm \ref{alg_TSHC}. As implemented, this is only possible when training on multiple separate tasks. Other differences include the parallelization method in \cite{salimans2017evolution}, where random seeds shared among workers permit each worker to only need to send and receive the scalar return of an episode to and from each other worker. All perturbations and parameters are then reconstructed locally by each worker. Thus, for $n$ workers there are $n$ reconstructions at each parameter-iteration step. This requires precise control of each worker and can in rare cases lead to differing CPU utilizations among workers due to differing episode lengths. Therefore, they use a capping strategy on maximal episode length. In contrast, our proposed method is less sophisticated with one synchronized parameter update, which is then sent to all workers. \section{Numerical Simulations\label{sec_expts}} This section highlights different aspects of Algorithm \ref{alg_TSHC}. Numerical simulations of Sect. \ref{subsec_expt1} and \ref{subsec_expt2} were conducted on a laptop with an Intel Core i7 CPU @2.80GHz$\times$8, 15.6GB of memory, and with the only libraries employed Python's \texttt{numpy} and \texttt{multiprocessing}. Furthermore, in Sect. \ref{subsec_expt1} for the implementation of 2 comparative policy gradient methods, Tensorflow (without GPU-support) was used. Using these (for deep learning) very limited ressources enabled to evaluate the method's potential when significant computational power is not available. For more complex problems the latter is a necessity. Therefore, in Sect. \ref{subsec_expt3} TSHC is implemented in \texttt{Cuda C++} and 1 GPU is used. \subsection{Experiment 1: Comparison with policy gradient methods\label{subsec_expt1}} \begin{table \begin{center} \vspace{0.2cm} \caption{Experiment 1. Number of scalar parameters (weights) that need to be identified for DDPG, PPO and TSHC, respectively. TSHC requires to identify the least by a large margin, roughly by a factor 4. The fact that PPO here requires exactly four times the number of parameters of TSHC is a special case for $n_a=2$ controls (not generalizable for arbitrary $n_a$).} \label{tab_ex1_Nvar} \bgroup \def\arraystretch{1 \begin{tabular}{|c|c|c|c|} \hline \rowcolor[gray]{0.8} & DDPG & PPO & TSHC \\ \hline % $N_\text{var}$ & 19078 & 18440 & 4610 \\ \hline \end{tabular} \egroup \end{center} \vspace{-0.7cm} \end{table} To underline conceptual differences between TSHC and 2 policy gradient methods DDPG \cite{lillicrap2015continuous} and PPO \cite{schulman2017proximal}, a freeform navigation task with $z_0=[0,0,0,0]$ and $z^\text{goal}=[20,0,\pi/4,0]$ was considered, where vector $z=[x,~y,~\psi,~v]$ summarizes four of the vehicle's states. The same network architecture from \cite{schulman2017proximal} is used: a fully-connected MLP with 2 hidden layers of 64 units before the output layer. Eventhough this is the basic setup, considerable differences between DDPG, PPO and TSHC are implied. Both DDPG and PPO are each composed of a total of four networks: one actor, one critic, one actor target and one critic target network. For DDPG, further parameters result from batch normalization \cite{ioffe2015batch}. The number of parameters $N_\text{var}$ that need to be identified are indicated in Table \ref{tab_ex1_Nvar}. To enable a fair comparison, all of DPPG, PPO and TSHC are permitted to train on 1000 full rollouts according to their methods, whereby each rollout lasts at most $100$ timesteps. Thus, for TSHC, $N_\text{restarts}=1$ and $n=1000$ are set. For both PPO and DDPG, this implies 1000 iterations. Results are summarized in Fig. \ref{fig_ex1_ddpgppotshc}. The following observations can be made. First, in comparison to TSHC, for both DDPG and PPO significantly more parameters need to be identified, see Table \ref{tab_ex1_Nvar}. Second, DDPG and PPO do \emph{not} solve the task based on 1000 training simulations. In contrast, as Fig. \ref{fig_ex1_ddpgppotshc} demonstrates, TSHC has a much better exploration strategy resulting from noise perturbations in the \emph{parameter space}. It solves the task in just 2.1s. Finally, note that no $\sigma_\text{pert}$-iteration is conducted. It is not applicable since a single task is solved with an initial $\sigma_\text{pert}^\text{max}=10$. Because of these findings (other target poses were tested with qualitatively equivalent results) and the discussion in Sect. \ref{subsec_relatedWork_TSHC} about the handling of \emph{sparse rewards} and the fact that DDPG and PPO have no useful gradient direction for their parameter update or may average these out through random minibatch sampling, the focus in the subsequent sections is on TSHC and its analysis. \subsection{Experiment 2: Inverted Pendulum\label{subsec_expt2}} The discussion of tolerance levels in Sect. \ref{subsec_role_tolerances_eps} motivated to consider an alternative approach for tasks requiring stabilization. An analogy to optimal control is drawn. In linear finite horizon MPC, closed-loop stability can be guaranteed through a terminal state constraint set which is invariant for a \emph{terminal controller}, often a linear quadratic regulator (LQR), see \cite{mayne2000constrained}. In a RL setting, the following procedure was considered. First, design a LQR for stabilization. Second, compute the region of attraction of the LQR controller \cite[Sect. 3.1.1]{tedrake2010lqr}. Third, use this region of attraction as stopping criterion, \emph{replacing} the heuristic $\epsilon$-tolerance selection. For evaluation, the inverted pendulum system equations and parameters from \cite{anderson1989learning} were adopted (four states, one input). However, in contrast to \cite{anderson1989learning}, which assumes just 2 discrete actions (maximum and minimum actuation force), here a continuous control variable is assumed which is limited by the 2 bounds, respectively. There are 2 basic problems: stabilization in the upright position with initial state in the same position, as well as a swing-up from the hanging position plus consequent stabilization in the upright position. For the application of TSHC, $(N_\text{restarts},N_\text{iter}^\text{max},n,T^\text{max},\beta,\sigma_\text{pert}^\text{max})=(3,100,100,500,2,10)$ are set, and the same MLP-architecture from Sect. \ref{subsec_expt1} is used. The following remarks can be made. First, the swing-up plus stabilization task was solved in $43.5$s runtime of TSHC (without refinement step) and using sparse rewards (obtained in the upright position $\pm 12^\circ$). For all three restarts a valid solution was generated. Note that $T^\text{max}=500$ in combination with a sampling time~\cite{anderson1989learning} of 0.02s corresponds to 10s simulation time. Stabilization in the upright position was achieved from 2.9s on. Rich reward signals were also tested, exploiting the deviation from current to goal angle as measure. However, rich rewards did not accelerate learning. In a second experiment, the objective was to \emph{simultaneously} encode the following 2 tasks in the network: stabilization in the upright position with initial state in the same position \emph{and} a swing-up from the hanging position plus consequent stabilization in the upright position. The runtime of TSHC (without refinement step) was $264.4$s, with 2 of 3 restarts returning a valid solution and using sparse rewards. Instead of learning both tasks simultaneously according to TSHC, it was also attempted to learn them by selecting one of the 2 tasks at random at every $i_\text{iter}$, and consequently conducting Step 6-41. Since the 2 tasks are quite different, this procedure could not encode a solution for \emph{both} tasks. This is mentioned to exemplify the importance of training \emph{simultaneously} on separate tasks, rather than training on a single tasks with $(z_0,z^\text{goal})$-combinations varying over $i_\text{iter}$. Finally, for the system parameters from \cite{anderson1989learning}, it was observed that the continuous control signal was operating mostly at saturated actuation bounds (switching in-between). This is mentioned for 2 reasons. First, aforementioned LQR-strategy could therefore never be applied since LQR assumes absence of state and input constraints. Second, it exemplifies the ease of RL-workflow with TSHC for quick nonlinear control design, even without significant system insights. \begin{figure} \vspace{0.3cm} \centering% \includegraphics[scale=1.0]{figex3.pdf} \caption{Experiment 3. (Left) Display of \emph{all} 181 trajectories learnt and encoded in one MLP-[5,8,2]. Trajectories for each task are visualized in separate colors. (Right) Display of learnt result for only the most complex of the 181 training tasks, i.e., for $\psi^\text{goal}=180^\circ$. Recall that $\epsilon_\psi=1^\circ$ and $\epsilon_d=0.25$m. The vehicle's start and end CoG-position is indicated by red and black balls, respectively. As indicated, the transition involves frequent forward and backward driving but is constrained locally around the $(x,y)$-origin. See Sect. \ref{subsec_expt3} for further discussion.} \label{} \vspace{-0.6cm} \end{figure} \subsection{Experiment 3 and 4: GPU-based training\label{subsec_expt3}} Experiment 3 is characterized by transitioning from $z_0=[0,0,0,0]$ to $z^\text{goal}=[0,0,\psi^\text{goal},0]$ with $\psi^\text{goal}\in\{0,1,\dots,180\}$ measured in $[^\circ]$. This implies $N_\text{tasks}=181$. The feature vector is selected as $s_t=[\frac{x^\text{goal}-x_t}{\Delta x_n},\frac{y^\text{goal}-y_t}{\Delta y_n},\frac{\psi^\text{goal}-\psi_t}{\Delta \psi_n},\frac{v^\text{goal}-v_t}{\Delta v_n},a_t[0]]$ with normalization constants in the denominators and $a_t[0]\in[-1,1]$ indicating the steering angle-related network output (before scaling to $\delta_t$). A high-resolution tolerance of $\epsilon_\psi=1^\circ$ is set. In addition, $\epsilon_d=0.25$m and $\epsilon_v=5$km/h. Sampling time is 0.01s. As neural network, a MLP-[5,8,2] is used, which implies 1 hidden layer with 8 units. For selections $N_\text{restarts}=10$ and $N_\text{iter}^\text{max}=20$, MLP-[5,8,2] was the smallest possible network found to simultaneously encode all $181$ training tasks. The second variant of VVCs discussed in Sect. \ref{subsec_rt} is employed. Furthermore, $\sigma_\text{pert}\sim\mathcal{U}[10,1000]$, i.e., uniformly distributed at every $(i_\text{restart},i_\text{iter})$-combination. \begin{figure} \vspace{0.3cm} \centering% \includegraphics[scale=1.0]{figex4.pdf} \caption{Experiment 4. Display of learnt trajectory when encoding just \emph{one} task in a MLP-[5,8,2] for the transition from $\psi_0=0^\circ$ to $\psi^\text{goal}=180^\circ$. Brackets (1) and (3) imply forward, (2) reverse driving and their sequence. A black indicator visualizes the vehicle's final heading. Total learning time (including 10 restarts) was 10.6s. See Sect. \ref{subsec_expt3} for further discussion.} \label{} \vspace{-0.6cm} \end{figure} Several comments can be made. First, by application of \emph{control mirroring} w.r.t. steering the trained network enables to reach also all of $\psi^\text{goal}\in\{181,\dots,360\}$. Second, the total learning time (runtime of TSHC) to encode all 181 training tasks was 31.1min. MLP-[5,8,2] implies a total of 66 parameters to learn. It is remarkable that such a small network has enough function approximation capability to encode all 181 tasks within limited training time and $\epsilon_\psi=1^\circ$. Third, note that the TSHC-trained network controller permits repeatable precision. As mentioned in of Sect. \ref{subsec_relatedWork_TSHC}, this is not attainable for stochastic policy gradient-based algorithms, which \emph{draw} their control signals, typically from a Gaussian distribution. Fourth, the learning results are visualized in Fig. \ref{}. These motivated to conduct an additional experiment with identical basic training setup (TSHC-settings, MLP-[5,8,2], etc.), however, now encoding only \emph{one} task for the transition from $\psi_0=0^\circ$ to $\psi^\text{goal}=180^\circ$. The result is visualized in Fig. \ref{}. Notice the much reduced number of switches between forward and backward driving, and the different $(x,y)$-range. The comparison of Fig. \ref{} and \ref{} emphasizes the interesting observation that the more motion primitives are encoded in a single network the less performant the single learnt motion trajectories are. This is believed to illustrate potential of partitioning the total number of designated tasks into \emph{subsets} of training tasks for which separate networks are then learnt using TSHC. The promised advantages include faster overall trainig times, higher performance of learnt trajectories, and ability to employ tiny networks with few parameters for each subset. Work in this perspective is subject of ongoing work. \section{Conclusion\label{sec_conclusion}} Within the context of automated vehicles, for the design of model-based controllers parameterized by neural networks a simple gradient-free reinforcement learning algorithm labeled TSHC was proposed. The concept of (i) training on separate tasks with the purpose of encoding motion primitives, and (ii) employing sparse rewards in combinations with virtual velocity constraints in setpoint proximity were specifically advocated. Aspects of TSHC were illustrated in 4 numerical experiments. The presented method is not limited to automated driving. Most real-world learning applications for control systems, especially in robotics, are characterized by sparse rewards and the availability of high-fidelity system models that can be leveraged for offline training. Subject of future work is focus on system models of various complexity (e.g., kinematic vs. dynamic vehicle models), the partitioninig of tasks into separate subsets of tasks for which separate network parametrizations are learnt, analysis of different feature vectors and closed-loop evaluation. \nocite{*} \bibliographystyle{ieeetr}

\section{Introduction} Standard regression theory assumes that explanatory variables are deterministic or error-free, but this assumption is quite unrealistic for many biological processes and replicated observations of covariates are often obtained to quantify the variability induced by the presence of measurement error (ME). Indeed covariates measured with error are considered a serious danger for inferences drawn from regression models. The most well known effect of measurement error is the bias towards zero induced by additive i.i.d. measurement error, but under more general measurement error specifications (as considered in this paper), different types of misspecification errors are to be expected \citep{, Loken584}. This is particularly true for semiparametric additive models, where the functional shape of the relation between responses and covariates is specified adaptively and therefore is also more prone to disturbances induced by ME. Recent papers advocate the hierarchical Bayesian modeling approach as a natural route for accommodating ME uncertainty in regression models. In particular, in the context of semiparametric additive models, \cite{sarkar2014} provide a Bayesian model based on B-spline mixtures and Dirichlet process mixtures relaxing some assumptions about the ME model, such as normality and homoscedasticity of the underlying true covariate values. Indeed many authors rely on almost flat normal priors for the true covariate, as \cite{muff2015} who frame ME adjustments into Bayesian inference for latent Gaussian models with the integrated nested Laplace approximation (INLA). The INLA framework allows to incorporate various types of random effects into generalized linear mixed models, such as independent or conditional auto-regressive models to account for continuous or discrete spatial structures. Relatively few articles have explicitly addressed covariate ME in the context of spatial modeling. A notable exception is given by the work of \cite{arima2017} who proposed a parametric spatial model for multivariate Bayesian small area estimation, accounting for heteroscedastic ME in some of the covariates. Continuous spatial data are approached by \cite{huque2016} to assess the relationship between a covariate with ME and a spatially correlated outcome in a non-Bayesian semiparametric regression context, assuming that the true unobserved covariate can be modeled by a smooth function of the spatial coordinates. In this paper we introduce a functional ME modeling approach allowing for replicated covariates with ME within a flexible class of regression models recently introduced by \cite{}, namely structured additive distributional regression models. In this modeling framework, each parameter of a class of potentially complex response distributions is modeled by an additive composition of different types of covariate effects, e.g. non-linear effects of continuous covariates, random effects, spatial effects or interaction effects. Structured additive distributional regression models are therefore in between the simplistic framework of exponential family mean regression and distribution-free approaches as quantile regression. We allow for quite general measurement error specifications including multiple replicates with heterogeneous dependence structure. From a computational point of view, based on the seminal work by \cite{berry2002} for Gaussian scatterplot smoothing and \cite{Kneib2010} for general semiparametric exponential family and hazard regression models, we develop a flexible fully Bayesian ME correction procedure based on Markov chain Monte Carlo (MCMC) techniques to generate observations from the joint posterior distribution of structured additive distributional regression models. ME correction is obtained by the imputation of unobserved error-free covariate values in an additional sampling step. Our implementation is based on an efficient binning strategy that avoids recomputing the complete design matrix after imputing true covariate values and combines this with efficient storage and computation schemes for sparse matrices. The main motivation of our investigation comes from a case study on the use of proximal soil-crop sensor technologies to analyze the within-field spatio-temporal variation of soil-plant relationships in view of the implementation of efficient agricultural management practices. More precisely, we analyze the relationship between multi-depth soil information indirectly assessed through the use of high resolution geophysical soil proximal sensing technology and data of forage ground-cover variation measured by a multispectral radiometer within a seven hectares Alfalfa stand in South Italy. Observations of both quantities were made using sensors with very refined spatial resolution: ground-cover data were obtained at four sampling occasions with point locations changing over time, while soil data were sampled only once for three different depth layers. Estimating a functional relation between ground-cover and soil with the data at hand involves addressing several issues, also linked to the spatial and temporal misalignment and the large data size. The nonlinear relation between crop productivity and soil is estimated by additive distributional regression models with structured additive predictor and measurement error correction. While distributional regression allows to deal with the heterogeneity of the response scale at the four sampling occasions, the ME correction is motivated by observations of covariates being replicated along a depth gradient and extends the model proposed by \cite{Kneib2010}, accounting for heterogeneous variances and possibly dependent replicates of the soil covariate. The paper is structured as follows: in Section~\ref{}, we introduce the Bayesian additive distributional regression model and the ME specification. Section~\ref{} reports details of the MCMC estimation algorithm and different tools for model choice and comparison. Simulation-based investigations of the model performance are the contents of Section~\ref{}. Section~\ref{} is dedicated to review some issues of the sampling design, describe the spatial interpolation method, report summary features of the data at hand, interpret estimation results and comment on possible Alfalfa management strategies. The final Section~\ref{} summarizes our main findings and includes comments on directions for future research. \section{Measurement Error Correction in Distributional Regression}\label{} The main motivation for our methodological developments comes from the need to estimate the nonlinear dependence of ground-cover on soil information by a smooth function, accounting for the heterogeneity in the position and scale of the response due to the sampling time, for the repeated measurements of the soil covariate and for the residual variation of unobserved spatial features. A detailed discussion of the data collection and pre-processing is given in Section~\ref{}. There we address ground-cover differences in both location and scale by structured additive distributional regression models, in which both parameters of the response distribution are related to additive regression predictors \citep[][and references therein]{}. The explicit consideration of replicated covariate observations avoids erroneous conclusions about the precise functional form of the relationship between ground-cover and soil variables or even about the presence of any association. In this section, the distributional regression model is set up, structured additive predictors and the measurement error correction are introduced with the relevant prior distributions. \subsection{Distributional Regression}\label{} Our treatment of Bayesian measurement error correction is embedded into the general framework of structured additive distributional regression. Assume that independent observations $(y_i,\text{\boldmath$\nu$}_i)$ $i=1,\ldots,n$ are available on the response $y_i$ and covariates $\text{\boldmath$\nu$}_i$ and that the conditional distribution of the responses belongs to a $K$-parametric family of distributions such that $y_i|\text{\boldmath$\nu$}_i\sim\mathcal D(\text{\boldmath$\vartheta$}(\text{\boldmath$\nu$}_i))$ and the $K$-dimensional parameter vector $\text{\boldmath$\vartheta$}(\text{\boldmath$\nu$}_i) = (\vartheta_1(\text{\boldmath$\nu$}_i),\allowbreak\ldots,\allowbreak\vartheta_K(\text{\boldmath$\nu$}_i))'$ is determined based on the covariate vector $\text{\boldmath$\nu$}_i$. More specifically, we assume that each parameter is supplemented with a regression specification \[ \vartheta_k(\text{\boldmath$\nu$}_i) = h_k(\eta^{\vartheta_k}(\text{\boldmath$\nu$}_i)) \] where $h_k$ is a response function that ensures restrictions on the parameter space and $\eta^{\vartheta_k}(\text{\boldmath$\nu$}_i)$ is a regression predictor. In our analyses, we will consider two specific special cases where $y_i\sim\ND(\mu(\text{\boldmath$\nu$}_i), \allowbreak\sigma(\text{\boldmath$\nu$}_i)^2)$, i.e. the responses are conditionally normal with covariate-dependent mean and variance and $y_i\sim\BetaD(\mu(\text{\boldmath$\nu$}_i), \sigma^2(\text{\boldmath$\nu$}_i))$, i.e. conditionally beta distributed responses with regression effects on location and scale. To ensure $\sigma(\text{\boldmath$\nu$}_i)^2>0$ for the variance of the normal distribution, we use the exponential link function, i.e. $\sigma(\text{\boldmath$\nu$}_i)^2 = \exp(\eta^{\sigma^2}(\text{\boldmath$\nu$}_i))$ while for both parameters $\mu(\text{\boldmath$\nu$}_i)$ and $\sigma(\text{\boldmath$\nu$}_i)^2$ of the beta distribution we employ a logit link since they are restricted to the unit interval \citep{FerCri2004}. \subsection{Structured Additive Predictor}\label{} For each of the predictors, we assume an additive decomposition as \[ \eta^{\vartheta_k}(\text{\boldmath$\nu$}_i) = \beta_0^{\vartheta_k} + f_1^{\vartheta_k}(\text{\boldmath$\nu$}_i) + \ldots + f_{J_k}^{\vartheta_k}(\text{\boldmath$\nu$}_i) \] i.e. each predictor consists of a total of $J_k$ potentially nonlinear effects $f_j^{\vartheta_k}(\text{\boldmath$\nu$}_i)$, $j=1,\ldots,J_k$, and an additional overall intercept $\beta_0^{\vartheta_k}$. The nonlinear effects $f_j^{\vartheta_k}(\text{\boldmath$\nu$}_i)$ are a generic representation for a variety of different effect types (including nonlinear effects of continuous covariates, interaction surfaces, spatial effects, etc., see below for some more details on examples that are relevant in our application). Any of these effects can be approximated in terms of a linear combination of basis functions as \[ f(\text{\boldmath$\nu$}_i) = \sum_{l=1}^L\beta_lB_l(\text{\boldmath$\nu$}_i) = \text{\boldmath$b$}} \def \mB {\text{\boldmath$B$}_i'\text{\boldmath$\beta$} \] where we dropped both the function index $j$ and the parameter index $\vartheta_k$ for simplicity, $B_l(\text{\boldmath$\nu$}_i)$ denotes the different basis functions with basis coefficients $\beta_l$ and $\text{\boldmath$b$}} \def \mB {\text{\boldmath$B$}_i=(B_1(\text{\boldmath$\nu$}_i),\ldots,B_L(\text{\boldmath$\nu$}_i))'$ and $\text{\boldmath$\beta$}=(\beta_1,\ldots,\beta_l)'$ denote the corresponding vectors of basis function evaluations and basis coefficients, respectively. Since in many cases the number of basis functions will be large, we assign informative multivariate Gaussian priors \[ p(\text{\boldmath$\beta$}|\text{\boldmath$\theta$}) \propto \exp\left(-\frac{1}{2}\text{\boldmath$\beta$}'\mK(\text{\boldmath$\theta$})\text{\boldmath$\beta$}\right) \] to the basis coefficients to enforce certain properties such as smoothness or shrinkage. The specific properties are determined based on the prior precision matrix $\mK(\text{\boldmath$\theta$})$ which itself depends on further hyperparameters $\text{\boldmath$\theta$}$. To make things more concrete, we discuss some special cases that we will also use later in our application: \begin{itemize} \item Linear effects: for parametric, linear effects, the basis functions are only extracting certain covariate elements from the vector $\text{\boldmath$\nu$}_i$ such that $B_l(\text{\boldmath$\nu$}_i) = \nu_{il}$. In case of no prior information, flat priors are obtained by $\mK(\text{\boldmath$\theta$})=\mathbf{0}$ which reduces the multivariate normal prior to a multivariate flat prior. In case of a rather large number of parametric effects, a Bayesian ridge prior can be obtained by $\mK(\text{\boldmath$\theta$}) = \tau^2\mI_L$ where $\tau^2$ is an additional prior variance parameter that can, for example, be supplemented an inverse gamma prior. \item Penalised splines for nonlinear effects $f(x)$ of continuous covariates $x$: for nonlinear effects of continuous covariates, we follow the idea of Bayesian penalized splines \citep{BreLan2006} where B-spline basis functions are combined with a random walk prior on the regression coefficients. In this case, the prior precision matrix is of the form $\mK(\text{\boldmath$\theta$}) = \tau^2\mD'\mD$ where again $\tau^2$ is a prior variance parameter and $\mD$ is a difference matrix. To obtain a data-driven amount of smoothness, we will assign an inverse gamma prior $\tau^2\sim\IGD(a,b)$ to $\tau^2$ with $a=b=0.001$ as a default choice. \item Tensor product splines for coordinate-based spatial effects $f(s_x, s_y)$: for modelling spatial surfaces based on coordinate information, we utilize tensor product penalized splines where each of the basis functions is constructed as $B_l(s_x,s_y) = B_{l,x}(s_x)B_{l,y}(s_y)$ with univariate B-spline bases $B_x(\cdot)$ and $B_y(\cdot)$. The penalty matrix is then given by \[ \mK(\text{\boldmath$\theta$}) = \frac{1}{\tau^2}\left[\omega\mK_x \otimes \mI_{y} + (1-\omega)\mI_{x}\otimes\mK_y\right] \] where $\mK_x$ is a random walk penalty matrix for a univariate spline in $s_x$, $\mK_y$ is a random walk penalty matrix for a univariate spline in $s_y$ and $\mI_x$ and $\mI_y$ are identity matrices of appropriate dimension. In this case, the prior precision matrix comprises two hyperparameters: $\tau^2$ is again a prior variance parameter (and can therefore still be assumed to follow an inverse gamma distribution) while $\omega\in(0,1)$ is an anisotropy parameter that allows for varying amounts of smoothness along the $x$ and $y$ coordinates of the spatial effect. For the latter, we assume a discrete prior following the approach discussed in \citet{KneKleLanUml2017}. \end{itemize} For some alternative model components comprise spatial effects based on regional data, random intercepts and random slopes or varying coefficient terms, see \citet[][Ch.~8]{} \subsection{Measurement Error}\label{subsec:me} In our application, we are facing the situation that a continuous covariate with potentially nonlinear effect modelled as a penalized spline has been observed with measurement error. Hence, we are interested in estimating the nonlinear effect $f(x)$ in one of the predictors of a distributional regression model where instead of the continuous covariate $x$ we observe $M$ replicates \[ \tilde{x}_i^{(m)} = x_i + u_i^{(m)}, \quad m=1,\ldots,M, \] contaminated with measurement error $u_i^{(m)}$. For the measurement error, we consider a multivariate Gaussian model such that \[ \text{\boldmath$u$}} \def \mU {\text{\boldmath$U$}_i\sim\ND_M(\mathbf{0}, \mathbf{\Sigma}_{u,i}) \] where $\text{\boldmath$u$}} \def \mU {\text{\boldmath$U$}_i = (u_i^{(1)},\ldots,u_i^{(M)})'$ and $\mathbf{\Sigma}_{u,i}$ is a known, pre-specified covariance matrix. Independent replicates with constant variance were considered in \citet{Kneib2010} for a mean regression model specification, with $\boldsymbol{\Sigma}_{u;i}=\sigma^2_{u}\mathbf{I}_M\label{. This assumption is here relaxed in the context of distributional regression, considering correlated replicates with potentially heterogeneous variances and covariances, leaving $\boldsymbol{\Sigma}_{u;i}$ unstructured. Of course, it would be conceptually straightforward to include priors on unknown parameters in $\mathbf{\Sigma}_{u,i}$, but this would require large amounts of data to obtain reliable estimates. We therefore restrict ourselves to the case of a known covariance matrix. The basic idea in Bayesian measurement error correction is now to include the unknown, true covariate values $x_i$ as additional unknowns to be imputed by MCMC simulations along with estimating the other parameters in the model. This requires that we assign a prior distribution to $x_i$ as well and rely on the simplest version \[ x_i\sim\ND(\mu_x, \tau_x^2) \] where we achieve flexibility by adding a further level in the prior hierarchy via \[ \mu_x\sim\ND(0, \tau_\mu^2), \qquad \tau_x^2\sim\IGD(a_x, b_x). \] To obtain diffuse priors on these hyperparameters, we use $\tau_\mu^2=1000^2$ and $a_x=b_x=0.001$ as default settings. \section{Bayesian Inference}\label{} All inferences rely on MCMC simulations implemented within the free, open source software BayesX \citep{}. As described in \citet{}, the software makes use of efficient storing even for large data sets and sparse matrix algorithms for sampling from multivariate Gaussian proposal distributions. In the following, we mostly focus on the required sampling steps for the measurement error part, since the remaining parts are mostly unchanged compared to standard distributional regression models \citep[see][for details]{KleSohKneLan2015}. \subsection{Measurement Error Correction}\label{} For the general case of distributional regression models, no closed form full conditional is obtained for the true covariate values $x_i$ such that we rely on a Metropolis Hastings update. Proposals are generated based on a random walk proposal \begin{equation}\label{} x_i^p = x_i^c + e_i, \qquad e_i\sim\ND(0, f \cdot \trace(\mathbf{\Sigma}_{u,i})/M^2) \end{equation} where $x_i^p$ denotes the proposal, $x_i^c$ is the current value and the variance of the random walk is determined by the measurement error variability (quantified by the trace of the covariance matrix $\mathbf{\Sigma}_{u,i}$), the number of replicates $M$ and a user-defined scaling factor $f$ that can be used to determine appropriate acceptance rates. The proposed value is then accepted with probability \[ \alpha(x_i^p|x_i^c) = \frac{p(y_i|x_i^p)}{p(y_i|x_i^c)} \frac{\exp\left(-\frac{1}{2\tau_x^2}(x_i^p-\mu_x)^2\right)}{\exp\left(-\frac{1}{2\tau_x^2}(x_i^c-\mu_x)^2\right)} \frac{\exp\left(-\frac{1}{2}(\text{\boldmath$\tilde x$}} \def \mtildeX {\text{\boldmath$\tilde X$}_i-x_i^p\mathbf{1}_M)'\mathbf{\Sigma}_{u,i}^{-1}(\text{\boldmath$\tilde x$}} \def \mtildeX {\text{\boldmath$\tilde X$}_i-x_i^p\mathbf{1}_M)\right)}{\exp\left(-\frac{1}{2}(\text{\boldmath$\tilde x$}} \def \mtildeX {\text{\boldmath$\tilde X$}_i-x_i^c\mathbf{1}_M)'\mathbf{\Sigma}_{u,i}^{-1}(\text{\boldmath$\tilde x$}} \def \mtildeX {\text{\boldmath$\tilde X$}_i-x_i^c\mathbf{1}_M)\right)} \] where the first term is the ratio of likelihood contributions given the proposed and current values of the covariate, the second term is the ratio of the measurement error priors and the third term is the ratio of the measurement error likelihoods. The ratio of the proposal densities cancels due to the symmetry of the random walk proposal. In contrast, the updates of the hyperparameters are standard due to the conjugacy between the Gaussian prior for the true covariate values and the hyperprior specifications. We therefore obtain \[ \mu_x|\cdot\sim\ND\left(\frac{n\bar{x}\tau_\mu^2}{n\tau_{\mu}^2 + \tau_x^2},\frac{\tau_x^2\tau_\mu^2}{n\tau_{\mu}^2 + \tau_x^2}\right) \] and \[ \tau_x^2|\cdot\sim\IGD\left(a_x + \frac{n}{2}, b_x + \frac{1}{2}\sum_{i=1}^n(x_i-\mu_x)^2\right) \] As mentioned before, priors could be assigned to incorporate uncertainty about elements of the covariance matrices $\mathbf{\Sigma}_{u;i}$ of the measurement error model. However, reliable estimation of $\mathbf{\Sigma}_{u;i}$ would typically require a large number of repeated measurements on the covariates, therefore we will restrict our attention to the case of known measurement error covariance matrices $\mathbf{\Sigma}_{u;i}$. \subsection{Updating the Structured Addditive Predictor}\label{} Updating the components of a structured additive predictor basically follows the same steps as in any distributional regression model \citep[see][for details]{KleSohKneLan2015,KneKleLanUml2017}. One additional difficulty arises from the fact that the imputation of true covariate values in each iteration implies that the associated spline design matrix would have to be recomputed in each iteration which would considerably slow down the MCMC algorithm. To avoid this, we utilize a binning approach \citep[see][]{LanUmlWecHarKne2012} where we assign each exact covariate value to a small interval. Using a large number of intervals allows us to control for potential rounding errors. On the positive side, the binning approach allows us to precompute the design matrix for all potential intervals and to only re-assign the observations based on an index vector in each iteration instead of recomputing the exact design matrix. \subsection{Model Evaluation}\label{} We combine different tools to determine a final model considering the quality of estimation and the predictive ability in terms of probabilistic forecasts. We rely on the deviance information criterion \citep[DIC,][]{}, the Watanabe-Akaike information criterion \citep[WAIC,][]{}, proper scoring rules \citep{} as well as normalized quantile residuals \citep{}. Measures of predictive accuracy are generally referred to as \emph{information criteria} and are defined based on the deviance (the log predictive density of the data given a point estimate of the fitted model, multiplied by -2). Both DIC and WAIC adjust the log predictive density of the observed data by subtracting an approximate bias correction. However, while DIC conditions the log predictive density of the data on the posterior mean $\mbox{E}(\text{\boldmath$\theta$}|\text{\boldmath$y$}} \def \mY {\text{\boldmath$Y$})$ of all model parameters $\text{\boldmath$\theta$}$ given the data $\text{\boldmath$y$}} \def \mY {\text{\boldmath$Y$}$, WAIC averages it over the posterior distribution $p(\text{\boldmath$\theta$}|\text{\boldmath$y$}} \def \mY {\text{\boldmath$Y$})$. Then, compared to DIC, WAIC has the desirable property of averaging over the whole posterior distribution rather than conditioning on a point estimate. This is especially relevant in a predictive Bayesian context, as WAIC evaluates the predictions that are actually being used for new data while DIC estimates the performance of the plug-in predictive density \citep{Gelman2014}. As measures of the out-of-sample predictive accuracy we consider three \emph{proper scoring rules} based on the logarithmic score $S(f_r,y_r)=\log(f_r(y_r))$, the spherical score $S(f_r,y_r)=f_r(y_r)/\|f_r(y_r)\|_2$ and the quadratic score $S(f_r,y_r)=2f_r(y_r)-\|f_r(y_r)\|_2^2$, with $\|f_r(y_r)\|_2^2=\int f_r(\omega)^2 d\omega$. Here $y_r$ is an element of the hold out sample $y_1,\ldots,y_R$ and $f_r$ is the predictive distribution of $y_r$ obtained in the current cross-validation fold. In practice the data set is approximately divided into $R$ non-overlapping subsets of equal size and predictions for one of the subsets are obtained from estimates based on all the remaining subsets. The predictive ability of the models is then compared by the aggregated average score $S_R=(1/R) \sum_{r=1}^R S(f_r,y_r)$. Higher logarithmic, spherical and quadratic scores deliver better forecasts when comparing two competing models. With respect to the others, the logarithmic scoring rule is usually more susceptible to extreme observations that introduce large contributions in the log-likelihood \citep{}. If $F$ and $y_i$ are respectively the assumed cumulative distribution and a realization of a continuous random variable and $\text{\boldmath$\hat \vartheta$}$ is an estimate of the distribution parameters, \emph{quantile residuals} are given by $\hat r_i=\Phi^{-1}(u_i)$, where $\Phi^{-1}$ is the inverse cumulative distribution function of a standard normal distribution and $u_i=F(y_i|\text{\boldmath$\hat \vartheta$}_i)$. If the estimated model is close to the true model, then quantile residuals approximately follow a standard normal distribution. Quantile residuals can be assessed graphically in terms of quantile-quantile-plots and can be an effective tool for deciding between different distributional options \citep{}. \section{Simulation experiment}\label{} In this section, we present a simulation trial designed to show some advantages in the performance of the proposed approach with respect to two alternative specifications of structured additive distributional regression. As a baseline, we consider Gaussian regression to describe the conditional distribution of the response, assuming that $y_{i}$'s follow a Gaussian law. Like in our case study, Beta regression is a useful tool to describe the conditional distribution of responses that take values in a pre-specified interval such as (0, 1). Thus, alternatively, we assume that $y_{i}$'s follow a Beta law. With both distributional assumptions, data are simulated from two scenarios corresponding to uncorrelated and correlated covariate repeated measurements. For the two distributional assumptions and the two scenarios, we compare three different structured additive distributional regression model settings: (1) as a \emph{benchmark}, we consider a model based on the ``true'' covariate values (i.e. those without measurement error), (2) a \emph{naive} model that averages upon repeated measurements and ignores ME and (3) a model implementing the proposed \emph{ME} adjustment. We expect that results from the latter are closer to the true model while the naive approach makes distinct errors (larger MSE, too narrow credible intervals). \subsection{Simulation settings}\label{} For 100 simulations, $n=500$ samples of the ``true'' covariate $x_i$ ($i=1,\ldots,n$) are generated from $N(0,5)$. Then 3 replicates with measurement error are obtained for each $x_i$ as $\tilde{\mathbf{x}}_i\thicksim N_3\left(x_i \mathbf{1}_3, \Sigma_{u;i}\right)$ with \begin{equation}\label{} \Sigma_{u;i}=\sigma^2_{u;i}\left(\begin{array}{ccc} 1 & c_u & c_u \\ c_u & 1 & c_u \\ c_u & c_u & 1\end{array}\right) \end{equation} We allow for ME heteroscedasticity setting $\sigma^2_{u;i=1,\ldots,n/2}=1$ and $\sigma^2_{u;i=n/2+1,\ldots,n}=2$ and we consider two alternative ME scenarios: Scenario 1 with uncorrelated replicates, i.e. $c_u=0$, and Scenario 2 where $c_u=0.8$. For the Gaussian observation model, we simulate 500 values of the response with $y_{i}\thicksim\mbox{N}(\mu_{i},\sigma_{i}^2)$, where $\mu_{i}=\sin(x_i)$ sets the nonlinear dependence with the covariate and we introduce response heteroscedasticity by \[ \mbox{Var}(y_i)=\sigma^2_{i}=\left\{\begin{array}{cc}0.5 & v_i = 1 \\0.3 & v_i = 0\end{array}\right.,\mbox{ with }v_i \sim \mbox{Bernoulli}(0.5). \] We proceed in like vein for the Beta observation model and generate $n$ samples from $\mbox{Beta}(p_{i},q_{i})$ with $\mu_{i}=p_i/(p_i+q_i)$ and $\sigma^2_i=1/(p_i+q_i+1)=\mbox{Var}(y_i)/(\mu_i(1-\mu_i))$, where $\mu_{i}=\exp(\sin(x_i))/(1+\exp(\sin(x_i))$ and $\sigma^2_i$ is specified as in the Gaussian case. \subsection{Model settings} As for the estimation model, we consider Gaussian and Beta regression with the following settings. Assuming $y_i$'s are Gaussian with $\mbox{N}(\mu_i,\sigma_i^2)$, model parameters are linked to regression predictors by the identity and the log link respectively: $\mu_i=\eta^\mu(x_{i})$ and $\sigma^2_i=\exp(\eta^{\sigma^2}(x_{i}))$. Alternatively, let $y_i$'s have Beta law $\mbox{Beta}(p_{i},q_{i})$, then both model parameters $\mu_i$ and $\sigma^2_i$ given in Section \ref{} are linked to respective regression predictors $\eta^\mu(x_{i})$ and $\eta^{\sigma^2}(x_{i})$ by the logit link, such that $\mu_i=\exp(\eta^\mu(x_{i}))/(1+\exp(\eta^\mu(x_{i})))$ and $\sigma^2_i=\exp(\eta^{\sigma^2}(x_{i}))/(1+\exp(\eta^{\sigma^2}(x_{i})))$. Under both distributional assumptions, we consider the two predictors as simply given by $\eta_i^\mu(x_{i})=\beta_0^\mu+f(x_{i})$ and $\eta_i^{\sigma^2}(x_{i})=\beta_0^{\sigma^2}+\beta_1^{\sigma^2}v_i$, with $f(\cdot)$ a penalized spline term and $v_i$ defined as in Section~\ref{}. As for the measurement error, we consider the \emph{benchmark}, \emph{naive} and \emph{ME} models defined at the beginning of Section~\ref{}. For the latter we specify $\Sigma_{u;i}$ as in Equation \ref{}, for the two ME scenarios. Finally, we set the scaling factor $f$ in Equation \ref{} to $1$ in both distributional settings. \begin{figure} \centering \includegraphics[scale=0.45]{SmoothGaussSim.pdf} \includegraphics[scale=0.45]{ICwGaussSim.pdf} \caption{Average estimates of the smooth effect (left) and width of 95\% credibility intervals (right) obtained with 100 simulations of the three model settings and the two scenarios by the Gaussian distribution model.}\label{} \end{figure} \begin{figure} \centering \includegraphics[scale=0.45]{boxDICGaussSim.pdf} \includegraphics[scale=0.45]{boxRMSEGaussSim.pdf} \caption{Boxplots of DIC's (left) and RMSE's (right) of the Gaussian distribution model for 100 simulations of the three model settings and the two scenarios.}\label{} \end{figure} \begin{figure} \centering \includegraphics[scale=0.45]{SmoothBetaSim.pdf} \includegraphics[scale=0.45]{ICwBetaSim.pdf} \caption{Average estimates of the smooth effect (left) and width of 95\% credibility intervals (right) obtained with 100 simulations of the three model settings and the two scenarios by the Beta distribution model.}\label{} \end{figure} \begin{figure} \centering \includegraphics[scale=0.45]{boxDICBetaSim.pdf} \includegraphics[scale=0.45]{boxRMSEBetaSim.pdf} \caption{Boxplots of DIC's (left) and RMSE's (right) of the Beta distribution model for 100 simulations of the three model settings and the two scenarios.}\label{} \end{figure} \subsection{Simulation results} In Fig. \ref{} we show the averages and quantile ranges of smooth effect estimates obtained with 100 simulated samples for the three Gaussian model settings and the two scenarios. While the \emph{benchmark} model setting obviously outperforms the other two, the \emph{ME} correction provides a sharper fit with respect to the \emph{naive} consideration of replicate averages. In general, severe smoothing is indeed only found with the \emph{naive} model in the case of correlated measurement errors (Scenario 2). Notice that the \emph{naive} model setting provides credible intervals as narrow as those obtained when the ME is not present (Fig. \ref{}, right panel), thus underestimating the smooth effect variability in the presence of ME. Preference for the \emph{ME} corrected model setting is also accorded by DIC's and RMSE's, as shown in Fig. \ref{}. The Beta distribution model obtains very similar results, with even stronger evidence for underestimation of the smooth effect variability by the \emph{naive} model in Fig. \ref{} (right). With the \emph{ME} proposal, the considerable gain in terms of fit and reliability of smooth effects variability does not fully overcome the increased complexity register by DIC's in the Beta distribution case. In this case the improvement attained in terms of RMSE's justifies the use of this type of model compared to the \emph{naive} approach. \section{Analysis of Sensor Data on Soil-Plant Variability}\label{} Alfalfa (Medicago sativa L.) plays a key role in forage production systems all over the world and increasing its competitiveness is one of the European Union's agricultural priorities \citep{visser2014}. Optimizing the use of crop inputs by applying site-specific management strategies can be a cost-effective way to increase crop profitability and stabilize yield. To achieve this goal, precise management activities such as precision irrigation, variable-rate N targeting, etc. are required. Precision Agriculture (PA) strategies in deep root perennials like alfalfa require information on deep soil variability. This is due to the higher dependence of these crops, compared to annuals, on deep water reserves often influenced by deep soil resources \citep{}. Soil texture and structure are among the most important soil properties affecting plant growth, since they are both strongly related to soil-nutrients and soil-water relationships \citep{}. On-the-go geophysical soil sensing has been successfully used to map the soil spatial variability at high resolution, covering hectares in a day of work and simultaneously investigating multiple soil depths \citep{, }. The non-destructive measurement of soil electrical resistivity (ER), or its inverse soil electrical conductivity, \citep{} is correlated to many crop-relevant soil characteristics \citep{}, such as the texture \citep{} and structure \citep{}. We focus on the relation between plant growth and soil features as measured by two sensors within a seven hectares Alfalfa stand in Palomonte, South Italy, with average elevation of 210 \emph{m} a.s.l.. The optical sensor GreenSeeker$^{\tiny \mbox{TM}}$ (NTech Industries Inc., Ukiah, California, USA) measures the normalized density vegetation index (NDVI), while Automatic Resistivity Profiling (ARP\copyright, Geocarta, Paris, F) provides multi-depth readings of the soil electrical resistivity. NDVI field measurements were taken at four time points in different seasons, while ER measurements were taken only once at three depth layers: 0.5\emph{m}, 1\emph{m} and 2\emph{m}. Ground-truth calibration \citep{} showed that the resistitivity distribution was linked to permanent soil features such as texture, gravel content and the presence of hardpans. As shown in Table~\ref{} and Figure~\ref{}, NDVI point locations change with time and are not aligned with ER samples. Hence NDVI and ER samples are misaligned in both space and time. Other data issues include response space-time dependence with spatially dense data \citep[big $n$ problem, ][]{JonaLasinio2013} and repeated covariate measurements. \begin{table \begin{center} \begin{tabular}{|l||c|c|c|c||c|} \hline & \multicolumn{4}{|c||}{NDVI} & ER \\\hline month/year & 9/2013& 11/2013 & 10/2014 & 6/2015 & 6/2013\\ \hline \# points &186667& 202172&91438&222278&120261\\ \hline \end{tabular} \end{center} \caption{Sampling time and number of sampled locations for NDVI and ER.}\label{} \end{table} \begin{figure} \centering \includegraphics[scale=0.27]{mappe_tutte2} \caption{Sampled NDVI, ER and grid point locations.}\label{} \end{figure} \subsection{Data pre-processing} Big $n$ and spatial misalignment suggest to change the support of both NDVI and ER samples. The change of support problem \citep[COSP, see][and references therein]{gelfand2010} for continuous spatial phenomena, commonly observed at the point and/or area level, involves a change of the spatial scale that can be required for any of several reasons, such as predicting the process of interest at a new resolution or to fuse data coming from several sources and characterized by different spatial resolutions. Bayesian inference with COSP may be a computationally demanding task, as it usually involves stochastic integration of the continuous spatial process over the new support. For this reason, in case of highly complex models or huge data sets, some adjustments and model simplifications have been proposed to make MCMC sampling feasible \citep[see][]{gelfand2010, cameletti2013}. Although relatively efficient, these proposals don't seem to fully adapt to our setting, mostly because of the need to overcome the linear paradigm. The most commonly used block average approach \citep[see for example][]{banerjee2014, cressie2015} would become computationally infeasible with a semi-parametric definition of the relation between NDVI and ER. Given the aim of this work and the data size, COSP is here addressed by a non-standard approach: the spatial resolution was downscaled by interpolating samples to a 2574 cells square lattice overlaying the study area. Given the different number of sampled points corresponding to each sampling occasion (NDVI) and survey (ER), we used a proportional nearest neighbors neighborhood structure to compute the downscaled values. More precisely, 27 neighbors were just enough to obtain non-empty cells at all grid points with the least numerous NDVI series (at the 3$^{\mbox{\textit{rd}}}$ time point). We then modified this number proportionally to the samples sizes, obtaining 55, 59, and 65 neighbors respectively for NDVI at the 1$^{\mbox{\textit{st}}}$, 2$^{\mbox{\textit{nd}}}$ and 4$^{\mbox{\textit{th}}}$ time points and 35 neighbors for ER. At each grid point we calculated the neighbors' means for both NDVI and ER, while neighbors' variances and covariances between depth layers were obtained for ER. Summary measures of the scale and correlation of ER repeated measures at each of the 2574 grid points provide valuable information that enables us to increase the model complexity with no additional costs in terms of parameters, i.e. degrees of freedom (see the prior specification later in this section). Such a by-product of the downscaling of the original data is plugged into the model likelihood. Exploratory analysis of the interpolated data shows some interesting features that we assume to drive the specification of the distributional regression model: NDVI has much higher position and smaller variability at the second time point, when it approached saturation (Figure~\ref{}, left); ER does not show a strong systematic variation along depth (Figure~\ref{}, right); the functional shape of the nonlinear relation between NDVI and ER is common to all NDVI sampling occasions and ER depth layers (Figure~\ref{}). \begin{figure} \centering \includegraphics[scale=0.35]{box_NDVI.pdf} \includegraphics[scale=0.35]{box_ER.pdf} \caption{NDVI distributions along time (left) and ER distributions along depth (right).}\label{} \end{figure} \begin{figure} \centering \includegraphics[scale=0.60]{NDVI_ER_smooth} \caption{Dependence of NDVI on ER at different time points (rows) and depth layers (columns). Lowess curves in green.}\label{} \end{figure} \subsection{Additive distributional regression models for location and scale} For available NDVI recordings, we consider Gaussian and Beta distributional regression models as those defined in Section~\ref{} and specify the two predictors as follows. For $s=1,\ldots,2574$ grid points and $t=1,\ldots,4$ time points, the structured additive predictor of the location parameter $\eta^\mu(\text{\boldmath$\nu$}_{st})$ is determined as an additive combination of three linear and functional effects, such as a linear seasonal effect, a spatial effect and a nonlinear effect of the continuous covariate ER: \begin{equation}\label{} \eta^\mu_{st}(\text{\boldmath$\nu$}_{st})=\beta_0^{\mu}+\boldsymbol{i}'_t\boldsymbol{\beta}_1^{\mu}+f_1^\mu(x_s)+f_2^\mu(lon_s,lat_s) \end{equation} where $\boldsymbol{i}_t$ is a vector of seasonal indicator variables with cornerpoint parametrization corresponding to the first sampling time, $f(x_s)$ is a nonlinear smooth function of latent replicate-free ER recordings $x_s$ (as in Section~\ref{subsec:me}) and $g(lon_s, lat_s)$ is a bivariate nonlinear smooth function of geographical coordinates $lon_s$, $lat_s$. The linear predictor of the scale parameter $\eta^{\sigma^2}(\text{\boldmath$\nu$}_{st})$ is assumed to depend only on the effect of time, thus allowing heteroscedasticity of seasonal NDVI recordings: \begin{equation}\label{} \eta^{\sigma^2}(\text{\boldmath$\nu$}_{st})=\beta_0^{\sigma^2}+\boldsymbol{i}'_t\boldsymbol{\beta}_1^{\sigma^2} \end{equation} for $t=1,\ldots,4$ time points where $\beta_0^{\sigma^2}$ and $\boldsymbol{\beta}_1^{\sigma^2}$ represent the overall level of the predictor and the vector of seasonal effects on the transformed scale parameter. Fixed effects $\boldsymbol{\beta}_1^{\mu}$ and $\boldsymbol{\beta}_1^{\sigma^2}$ in (\ref{}) and (\ref{}) respectively account for mean effects and heteroscedasticity of NDVI seasonal recordings. While conjugacy allowed to use Gibbs sampling to simulate from the full conditionals of the Gaussian models for the location parameter, the Metropolis-Hastings algorithm was required to sample the posteriors of the Gaussian models for the scale parameter and those of the Beta models. A different simulation setup was then adopted in the two cases: for Gaussian models we obtained 10000 simulations with 5000 burnin and thinning by 5, while Beta models required longer runs of 50000 iterations with 35000 burnin and thinning by 15. In all cases, convergence was reached and checked by visual inspection of the trace plots and standard diagnostic tools. Fine tuning of hyperparameters lead us to 10 and 8 equidistant knots for each of the two components of the tensor product spatial smooth in the Gaussian and Beta case, respectively. The additive distributional regression model was compared to standard additive mean regression with the same mean predictor, applying the proposed measurement error correction to both models, under the Gaussian and Beta assumptions. By this comparison we show that adding a structured predictor for the scale parameter improves both the in-sample and out-of-sample predictive accuracy. In the following, M1 is an additive regression model with mean predictor as in (\ref{}), while M2 is an additive distributional regression model with the same mean predictor and scale predictor given by (\ref{}). \subsection{Results} Model comparison shows some interesting features of the proposed alternative model specifications. Concerning the distributional assumption, DIC, WAIC and the three proper scoring rules clearly favor the Beta models (Tab. \ref{tabFit}), showing a better compliance with in-sample and out-of-sample predictive accuracy. As far as additive distributional regression is concerned, information criteria and scoring rules agree in assessing the proposed model (M2) as performing better than a simple additive mean regression (M1). \begin{table}[ht] \begin{center} \begin{tabular}{|l|ccccc|} \hline Model & DIC & WAIC & LS & SS & QS\\\hline GM1 & -21333.3 & -21331.2 & 1.0340 & 1.8427 &3.3947\\ GM2 & -27088.4 & -27076.2& 1.1234 & 1.8986 &3.6935\\ \hline BM1 & -22841.0 & -22835.9 & 1.1085 & 1.9298 & 3.7893\\ BM2 & -27131.4& -27126.4& 1.2140& 2.0182 & 4.2902\\\hline \end{tabular} \caption{Model fit statistics for Gaussian (G) and Beta (B) distribution additive mean (M1) and distributional (M2) regression models: deviance information criterion, Watanabe-Akaike information criterion, logarithmic score, spherical score and quadratic score. }\label{tabFit} \end{center} \end{table} Quantile residuals of model M2 under the two distributional assumptions show a generally good behavior, with a substantial reduction in scale in the Beta case and only a slightly better compliance with the latter distributional shape (Figure~\ref{}). When comparing the two distributional assumptions, it should be recalled that Gaussian models are by far much more convenient from the computational point of view. \begin{figure} \centering \includegraphics[scale=0.38]{quanRes1.pdf} \includegraphics[scale=0.38]{quanRes2.pdf} \caption{Quantile residuals for models M1 (left) and M2 (right) under the Gaussian (top panel) and Beta (bottom panel) assumptions: boxplots at different time points (left panel) and Normal q-q plots (right panel).}\label{} \end{figure} Values of the fixed time effects estimates of the mean (\ref{}) and variance (\ref{}) predictors and their 95\% credibility intervals for the two models (Tab. \ref{tabFixed}) are expressed in the scale of the linear predictor and cannot be compared, due to different link functions being implied. However, in both the Gaussian and Beta case their relative variation clearly reproduces the NDVI behavior in Figure~\ref{}, left. \begin{table}[ht] \begin{center} {\footnotesize \begin{tabular}{|c|c|cccc|} \hline Model & Parameter & 9/13 & 11/13 & 10/14 & 6/15 \\\hline \multirow{4}{*}{GM2} & \multirow{2}{*}{$\mu$} & 0.6037 & 0.8224 & 0.5215 & 0.5410 \\ && {\tiny 0.5964, 0.6113} & {\tiny 0.8157, 0.8295 }& {\tiny 0.5127 0.5294} & {\tiny 0.5318 0.5503}\\ & \multirow{2}{*}{$\log(\sigma^2)$} &-5.0995 & -8.3773 & -4.5524 & -3.8718 \\ && {\tiny -5.1588,-5.0401} & {\tiny -8.4321, -8.3190} & {\tiny -4.6071, -4.4992} & {\tiny -3.9287 -3.8171}\\\hline \multirow{4}{*}{BM2} & \multirow{2}{*}{$\mbox{logit}(\mu)$} &0.3943 & 1.5231 & 0.0528 & 0.1365 \\ && {\tiny 0.3528, 0.4329} & {\tiny 1.4825, 1.5636} & {\tiny0.0104 , 0.0940} & {\tiny 0.0919, 0.1799}\\ & \multirow{2}{*}{$\mbox{logit}(\sigma^2)$} &-3.7627 & -6.2897 & -3.2067 & -2.4696 \\ && {\tiny 3.8204 ,-3.7062} & {\tiny -6.3542 ,-6.2249} & {\tiny -3.2591, -3.1520} & {\tiny -2.5245, -2.4175}\\\hline \end{tabular}} \caption{Fixed time effects estimates of mean and variance predictors with 95\% credibility intervals. To facilitate interpretation, estimates were transformed avoiding the cornerpoint parametrization.}\label{tabFixed} \end{center} \end{table} Estimates of smooth effects of ER and of the nonlinear trend surface (Figure~\ref{}) have again different scale and common shapes for the two distributional assumptions. The inclusion of the nonlinear trend surface (Figure~\ref{}, right), besides accounting for the spatial pattern and lack of independence between nearby observations, allows to separate the effect of ER from any other source of NDVI spatial variability including erratic and deterministic components (such as slope and/or elevation). The estimated nonlinear effect of ER on NDVI shows a monotonically increasing relation up to approximately 10 \emph{Ohm m} with a subsequent steep decline up to approximately 20 \emph{Ohm m}. After dropping to lower values, the smooth function declines more slowly. Based on the resulting estimated smooth functions (Figure~\ref{}, left), two ER cut-offs (at 10 and 20 \emph{Ohm m}) are proposed that can be used to split the field in three areas characterized by a different monotonic soil-plant relationship: \begin{itemize} \item \textbf{Zone i: ER $<$ 10 \emph{Ohm m}}, where NDVI grows with ER and very low ER readings correspond to intermediate to high NDVI values (the former correspond to the presence of poorly drained soils and the consequent risk of waterlogging; crop management needs to take into account in-season rain patterns to minimize the risks of waterlogging damages in wet years); \item \textbf{Zone ii: 10 \emph{Ohm m} $<$ ER $<$ 20 \emph{Ohm m}}, where ER is negatively related to NDVI and soil factors affecting ER act almost linearly and consistently on plant performance (precision management can be applied as a function of ER, i.e. the resistivity map itself can be used as a prescription map in the corresponding areas); \item \textbf{Zone iii: ER $>$ 20 \emph{Ohm m}}, where despite the large variation in ER there is a limited NDVI-soil responsiveness and NDVI is constantly low (corresponds to the presence of the hardpans and management criteria should differ accordingly). \end{itemize} Each zone conveys information on the shape and strength of the association between soil and crop variability, thus the proposed field zonation helps discerning areas where even a little change in soil properties can affect plant productivity (zone ii) from areas where soil environment is not practically alterable (zone iii) or in-season evaluations are possibly needed (zone i). \begin{figure} \centering \includegraphics[scale=0.55]{estSmooth2.pdf} \caption{Smooth estimates of ER effects (left) and residual spatial effects (right) for models GM2 (top), and BM2 (bottom). Notice that while GM2 effects are on the scale of the response (NDVI), BM2 effects are estimated on the logit scale. Dotted red vertical lines locate ER cut-offs corresponding to different monotonic soil-plant relationships.}\label{} \end{figure} \section{Concluding remarks and directions of future work}\label{} The work described in this paper was motivated by the analysis of a database characterized by some complex features: response space-time dependence with spatially dense data, data misalignment in both space and time and repeated covariate measurements. These data features were addressed by first changing the spatial support of the data, then proposing an extension of structured additive distributional regression models by the introduction of a replicated covariate measured with error. Within a fully Bayesian implementation, measurement error is dealt with in the context of the functional modeling approach, accounting for possibly heteroscedastic and correlated covariate replicates. In the paper we only allow for Gaussian and Beta distributed responses, but the proposed correction is implemented to accommodate for potentially any K-parametric family of response distributions. With a simulation experiment we show some advantages in the performance of the proposed ME correction with respect to two alternative less ambitious ME specifications. The proposed extension of the of the ME correction in \cite{Kneib2010} to structured additive regression models proves to be essential for the case study on soil-plant sensor data, where both mean and variability effects have to be modeled. Indeed in this case the proposed approach outperforms the simpler use of the ME correction with a mean regression model with the same (mean) predictor. In the Bayesian framework, a straightforward extension would be the consideration of other types of (potentially non-normal) measurement error structures, as in \citet{sarkar2014} under a structural ME approach. Given that a fully specified measurement error is given, this will only lead to a minor adaptation of the acceptance probability in our MCMC algorithm. A more demanding extension would be to include inference on the unknown parameters in the measurement error model, such as the covariance structure in our approach based on multivariate normal measurement error. Such parameters will typically be hard to identify empirically unless the number of replicates and/or the sample size is large. In the case of big spatial data, the computational burden induced by spatial correlation could be reduced integrating the ME correction with a low rank approach \citep[see for instance][and references therein]{banerjee2014, datta2016}. While in this paper we have considered measurement error in a covariate that enters the predictor of interest via a univariate penalized spline, other situations are also easily conceivable. One option would be to develop a Bayesian alternative to the simulation and extrapolation algorithm developed in \citet{KueMwaLes2006} to correct for misclassification in discrete covariates. Another route could be the consideration of measurement error in one or both covariates entering an interaction surface modeled as a bivariate tensor product spline.\\\\ \textbf{Acknowledgments}: Alessio Pollice and Giovanna Jona Lasinio were partially supported by the PRIN2015 project "Environmental processes and human activities: capturing their interactions via statistical methods (EPHASTAT)" funded by MIUR - Italian Ministry of University and Research. \bibliographystyle{natbib}

\section{The early years} In a celebrated paper \cite{Serrin} Serrin initiates the study of elliptic equations under overdetermined boundary conditions. He establishes in particular the radial symmetry of the solution to the following overdetermined Poisson problem. Let $\Omega$ be a { bounded, } smooth, open, connected set of { $\mathbb R^n$} , and { let} $\nu_x$ the outward normal { at $x \in \partial \Omega$}, if $u$ is a smooth solution to \begin{equation}\label{delta} \left\{ \begin{array}{ll} \Delta u=-1 & \mathrm{in}\,\,\Omega \\ u=0 & \mathrm{on}\,\,\partial \Omega \\ \displaystyle\frac{\partial u}{\partial \nu_x}=const(=c) & \mathrm{on}\,\,\partial \Omega, \end{array} \right. \end{equation} then $u=\frac{R^2-|x|^2}{2n}$ up to a translation and therefore $\Omega$ is a ball with radius $R$. The main tool of his proof is a technique introduced by Alexandrov \cite{A1,A2} known as \emph {moving plane} (in a completely different context to established that { the only compact, embedded $(n-1)$-dimensional smooth hypersurfaces in $\mathbb R^n$ with constant mean curvature are the spheres}) combined with a clever refinement of the maximum principle (see Lemma \ref{maxpri}). Right after Serrin's paper, Weinberger \cite{Wein} came out with a very short proof of the same result using the maximum principle applied to an auxiliary function. However in spite of its simplicity Weinberger's proof on one hand seems to rely on the linearity of the Laplace operator and was not elementarily generalizable to nonlinear ones, on the other hand is restricted to constant righthand side in the Poisson problem. Serrin's proof has in fact the great advantage of being easily stretchable to a wide range of fully nonlinear elliptic operators with fairly general data.\\ Shortly after, these early papers resulted in a wide research field which nowadays it is still very prominent. It is also very important to mention that Serrin's approach inspired other { fundamental }results concerning symmetry in PDE's. Among the others a seminal paper by Gidas, Ni and Nirenberg \cite{GNN} which unfortunately is beyond the scope of the present survey. Our goal indeed is to summarize in a concise but self contained way both Serrin's and Weinberger's proofs along with some of the alternative results which came out more recently. Nonlinear problems, stability issues, possibility to extend the symmetry result in case of lack of regularity, overdetermined problems in exterior domains, different overdetermined boundary conditions, are only few of the interests which became popular during the last few decades. It is impossible to give an exhaustive list of all the results hence we will restrict our attention just on the original Poisson problem \eqref{delta}. We shall stress pros and cons of every approach and mention major applications to different settings. As Serrin explains \cite{Serrin}, his work originated form physical motivations: \begin{quote} \it Consider a viscous incompressible fluid moving in straight parallel streamlines through a straight pipe of given cross sectional form $\Omega$. If we fix rectangular coordinates in space with the $z$ axis directed along the pipe, it is well known that the flow velocity $u$ is then a function of $x$, $y$ alone satisfying the Poisson differential equation (for $n = 2$) $$\Delta u=-A \quad\mbox{in $\Omega$}$$ where $A$ is a constant related to the viscosity and density of the fluid and to the rate of change of pressure per unit length along the pipe. Supplementary to the differential equation one has the adherence condition $$u=0 \quad\mbox{on $\partial\Omega$}.$$ Finally, the tangential stress per unit area on the pipe wall is given by the quantity $\mu\frac{\partial u}{\partial \nu_x}$ where $\mu$ is the viscosity. Our result states that the tangential stress on the pipe wall is the same at all points of the wall if and only if the pipe has a circular cross section. Exactly the same differential equation and boundary condition arise in the linear theory of torsion of a solid straight bar of cross section $\Omega$, (...) when a solid straight bar is subject to torsion, the magnitude of the resulting traction which occurs at the surface of the bar is independent of position if and only if the bar has a circular cross section. \end{quote} In order to understand why the boundary overdetermination is so interesting in physical context one has to notice that { it may arise in optimal control theory}. Following for instance the analogy with the torsion problem, we can ask what is the shape of a prismatic bar that maximizes the torsional rigidity when the cross sectional area is assigned. This is the famous Saint-Venant problem and the answer is the provided by the bar of circular cross section \cite{PSZ}. A necessary condition that a smooth cross section $\Omega$ has to satisfy, for being the bar a maximizer of the torsion, is stationarity among smooth domain variations. The Torsion becomes a so called shape functional and the problem is recast in the framework of the shape optimization via domain derivative \cite{HenPi}. { As mentioned by Serrin the torsion problem consists in finding a function $u$ (called torsion function) which solves $$\Delta u=-A \quad\mbox{in $\Omega$}$$ and $$u=0 \quad\mbox{on $\partial\Omega$}.$$} The well known Hadamard formula for the torsional rigidity enforces the gradient of $u$ to be constant on the boundary of $\Omega$ (i.e. the bar has constant shear stress) and here comes the overdetermination. In view of Serrin's result we can state that, when optimizing the torsion of a prismatic bar with respect to area preserving smooth variations of the cross section, the circular shaft is the unique stationary point. \subsection{Serrin's { result} \cite{Serrin}} { Before stating the result we observe that every solution to \eqref{delta} is positive in $\Omega$ and that divergence theorem together with the fact that $Du=c\nu$ on $\partial \Omega$ give \[ c= -\dfrac{|\Omega|}{|\partial \Omega|}. \] } The main Theorem reads as follows. \begin{theorem}\label{S1} Whenever $\Omega$ is a $C^2$ bounded domain (bounded open and connected) of $\mathbb R^n$ and $u\in C^2(\overline\Omega)$ is a solution to problem \eqref{delta} then, up to a translation, $u=\frac{R^2-|x|^2}{2n}$ and $\Omega$ is a ball with radius $R$. \end{theorem} We start by recalling Serrin's proof of Theorem \ref{S1}, which relies on the moving planes method together with the strong maximum principle. \begin{proof} We denote by $H_\nu$ an open halfspace with unit outer normal $\nu$ and we move this halfspace along the direction $\nu$ until it intersects $\Omega$. We still denote by $H_\nu$ the halfspace after its motion and by $\Omega_\nu= \Omega \cap H_\nu$. For every cap $\Omega_\nu$ let us denote by $\Sigma(\Omega_\nu)$ its reflection with respect to $\partial H_\nu$ and let us move $H_\nu$ until $\Sigma(\Omega_\nu)\subset \Omega$. When the motion stops then one of these two { cases} occur: \medskip \begin{enumerate} \item $\Sigma(\Omega_\nu)$ becomes internally tangent to $\partial \Omega$ at a point $\bar x$ not belonging to $\partial H_\nu$; \\ \item $H_\nu$ reaches a position such that $\nu$ is tangent to $\partial \Omega$ at some point $\bar y$ . \end{enumerate} We denote by $H'_\nu$ the halfspace when it reaches one of these positions and by $\Omega'_\nu$ the respective cap. The goal is to prove that $\Omega$ is symmetric with respect to the hyperplane $\partial H_\nu$. Once this fact is proved then the theorem follows, since for every direction $\nu$, $\Omega$ would be symmetric with respect to the hyperplane normal to $\nu$. Moreover, by construction, $\Omega$ would also be simply connected, then it has to be a ball and the unique solution to \eqref{delta} is the paraboloid. Let now $\Sigma(\cdot)$ denote the reflection across $\partial H'_\nu$. { We set $v$ the function defined in $\Sigma(\Omega'_\nu)$ by \[ v(x)= u(\Sigma(x)) \quad x \in \Sigma(\Omega'_\nu). \]} Obviously $v$ satisfies: \begin{equation}\label{reflection} \left\{ \begin{array}{ll} \Delta v=-1 & \mathrm{in}\,\,\Sigma(\Omega'_\nu) \\ \\ v=u & \mathrm{on}\,\,\partial \Sigma(\Omega'_\nu) \cap \partial H'_\nu \\ \\ v=0\quad \mathrm{ and}\quad \displaystyle\frac{\partial v}{\partial \nu_x}=c & \mathrm{on}\,\,\partial \Sigma(\Omega'_\nu) \setminus \partial H'_\nu. \end{array} \right. \end{equation} Since $\Sigma(\Omega'_\nu)$ is contained in $\Omega$ one can consider the function $u-v$ and (recalling that $u>0$ in $\Omega$) observe that it satisfies \begin{equation}\label{difference} \left\{ \begin{array}{ll} \Delta (u-v)=0 & \mathrm{in}\,\,\Sigma(\Omega'_\nu) \\ \\ u-v =0 & \mathrm{on}\,\,\partial \Sigma(\Omega'_\nu) \cap \partial H'_\nu \\ \\ u-v \ge 0, & \mathrm{on}\,\,\partial \Sigma(\Omega'_\nu) \setminus \partial H'_\nu. \end{array} \right. \notag \end{equation} At this point the strong maximum principle { gives} either \begin{equation} \label{positive} u-v>0 \quad \mathrm{in}\,\, \Sigma(\Omega'_\nu) \end{equation} or $u \equiv v$ in $ \Sigma(\Omega'_\nu)$. The latter case would imply that $\Omega$ is symmetric about $\partial H'_\nu$. Assume that { case} (1) occurs, that is $\Sigma(\Omega'_\nu)$ is internally tangent to $\partial \Omega$ at a point $\bar x$ not belonging to $\partial H_\nu$ and assume by contradiction that \eqref{positive} holds true. Then Hopf Lemma ensures that \[ \displaystyle\frac{\partial }{\partial \nu_{\bar x}}(u-v) >0, \] but this contradicts the fact that \eqref{delta} and in \eqref{reflection} yield \[ \displaystyle\frac{\partial u }{\partial \nu_{\bar x}}=\displaystyle\frac{\partial v}{\partial \nu_{\bar x}} =c. \] We conclude that \eqref{positive} cannot occur in case (1). Case (2) is much more { complicated }since Hopf Lemma cannot apply. The proof makes use of a refinement of the maximum principle, see Lemma \ref{maxpri} below (for its proof see \cite{Serrin}). The goal is to prove that $u-v$ has in $\bar y$ a second order zero. To do this we fix a coordinate system with the origin at $\bar y$, the $x_n$ axis in the direction of the inward normal to $\partial \Omega$ at $\bar y$ (that is $-\nu_{\bar y}$), and the $x_1$ axis in the direction of $\nu$, that is normal to $\partial H'_\nu$. In this coordinates system the boundary of $\Omega$ is locally given by \[ x_n= \phi(x_1,x_2, \cdots,x_{n-1}) \quad \phi \in C^2. \] Since $u \in C^2$ the boundary conditions, { $u=0$ on $\partial \Omega$ and $\displaystyle\frac{\partial u }{\partial \nu_{x}}=c$ on $\partial \Omega$,} can be written as \begin{equation} \label{bc1} u(x_1,x_2, \cdots,x_{n-1},\phi)=0, \end{equation} and \begin{equation} \label{bc2} \displaystyle\frac{\partial u }{\partial x_n} - \displaystyle\sum_{k=1}^{n-1} \displaystyle\frac{\partial u }{\partial x_k}\displaystyle\frac{\partial \phi }{\partial x_k}= c \left \{ 1+ \sum_{k=1}^{n-1} \left(\displaystyle\frac{\partial \phi }{\partial x_k}\right)^2\right \}^{1/2}, \end{equation} respectively. Differentiating \eqref{bc1} with respect to $x_i$, for { $ i=1,\dots, n-1$,} we have \begin{equation} \label{diff1} \displaystyle\frac{\partial u }{\partial x_i}+\displaystyle\frac{\partial u }{\partial x_n}\displaystyle\frac{\partial \phi }{\partial x_i}=0. \end{equation} Evaluating \eqref{diff1} and \eqref{bc2} at $\bar y$ and recalling that $\displaystyle\frac{\partial \phi }{\partial x_i}(\bar y) = 0$ for { $i=1, \dots, n-1$} we have \begin{equation} \label{gradzero} \displaystyle\frac{\partial u }{\partial x_i}(\bar y)=0 \quad \displaystyle\frac{\partial u }{\partial x_n}(\bar y)=c. \end{equation} Differentiating \eqref{diff1} with respect to $x_j$, we get for { $i,j=1, \dots, n-1$} \begin{equation} \label{diff2} \displaystyle\frac{\partial^2 u }{\partial x_i\partial x_j}(\bar y) +c \displaystyle\frac{\partial^2 \phi }{\partial x_i\partial x_j}(\bar y)=0 \end{equation} while differentiating \eqref{bc2} with respect to $x_i$, for $i=1, \cdots n-1$ and using \eqref{gradzero} we obtain \begin{equation} \label{diffn} \displaystyle\frac{\partial^2 u }{\partial x_n\partial x_i}(\bar y) =0. \end{equation} From \eqref{diff2} \begin{equation} \label{diffnn} \displaystyle\frac{\partial^2 u }{\partial x_n^2}(\bar y) = -\displaystyle\sum_{i=1}^{n-1} \displaystyle\frac{\partial^2 u }{\partial x_i^2}(\bar y) -1= c \Delta \phi(\bar y) -1. \end{equation} { By construction} $\Sigma(\Omega'_\nu) \subseteq \Omega$ { and} all the second derivatives $\displaystyle\frac{\partial^2 \phi }{\partial x_1\partial x_j}(\bar y)=0$ for $j=2, \cdots, n-1$, { because} $\displaystyle\frac{\partial \phi }{\partial x_1}$ has an extremum point at $\bar y$ with respect to all but the first coordinates directions. Setting \[ v(x_1,x_2,\cdots,x_n) = u(-x_1,x_2,\cdots,x_n), \] by \eqref{diff2}, \eqref{diffn} and the last remark we have that all the first and second derivatives of $u$ and $v$ coincide at $\bar y$. The function $w= u-v$ satisfies { \[ \Delta w= 0 \quad {\rm in } \>\> \Sigma(\Omega'_\nu), \]} { \[ w>0 \quad {\rm in } \>\> \Sigma(\Omega'_\nu), \]} and $w(\bar y)=0$. If $\theta $ is any direction not parallel to $\nu$ Lemma \ref{maxpri} ensures that { either} \[ \displaystyle\frac{\partial (u-v) }{\partial \theta}(\bar y)>0 \quad {\rm or } \>\> \displaystyle\frac{\partial^2 (u-v) }{\partial \theta^2}(\bar y)>0, \] which is a contradiction since all the first and second derivatives of $u$ and $v$ coincide at $\bar y$. \end{proof} \noindent The following Lemma is a refinement of Hopf Lemma. We omit its proof which is contained in \cite{Serrin} \begin{lemma} \label{maxpri} Let $\Omega$ be a $C^2$ bounded domain (bounded open and connected) of $\mathbb R^n$ and let $\nu$ a direction such that $<\nu,\nu_y>=0$, $y \in \partial \Omega$. Let $H_\nu$ be an open halfspace with unit outer normal $\nu$, $\Omega_\nu= \Omega \cap H_\nu$ and let $w \in C^2(\bar \Omega_\nu)$ satisfy $$ \Delta w \le 0 \quad {\rm in} \>\> \Omega_\nu, $$ $w\ge 0$ in $\Omega_\nu$ and $w(y)=0$. If $\theta$ is a direction in $y$ entering $\Omega_\nu$ such that $<\theta,\nu_y>\neq 0$, then either \[ \frac{\partial w}{\partial \theta}(y) >0 \quad {\rm or} \> \> \frac{\partial^2 w}{\partial \theta^2}(y) >0 \] unless $w \equiv 0$. \end{lemma} \subsubsection{Remark on the proof and generalization.} The great advantage of Serrin's proof with respect to all other techniques that we are going to analyze is that it works out of the box on a massive number of other problems. The main ingredients used are: \begin{itemize} \item The problem is invariant under reflection \item In any boundary point of $\Omega$, in a framework where one of the axis points into the normal direction, the second derivative of $u$ can be determined in terms of the other second order derivative. \item Maximum principle and boundary point maximum principle hold. \end{itemize} If for instance we consider \begin{equation*} \left\{ \begin{array}{ll} \Delta u=f(u,|Du|) & \mathrm{in}\,\,\Omega \\ u=0 & \mathrm{on}\,\,\partial \Omega \\ \displaystyle\frac{\partial u}{\partial \nu_x}=const(=c) & \mathrm{on}\,\,\partial \Omega. \end{array} \right. \end{equation*} then $\Omega$ is a ball and $u$ is radially symmetric, provided $f$ is differentiable and $u>0$. The condition $u>0$ { is unavoidable in order to apply the moving plane.} The eigenvalue problem serves as a counterexample. No symmetry of solutions can be established via moving planes for \begin{equation}\label{Schiffer} \left\{ \begin{array}{ll} \Delta u=-\lambda u & \mathrm{in}\,\,\Omega \\ u=0 & \mathrm{on}\,\,\partial \Omega \\ \displaystyle\frac{\partial u}{\partial \nu_x}=const(=c) & \mathrm{on}\,\,\partial \Omega. \end{array} \right. \end{equation} unless we know that we are dealing with the first eigenvalue $\lambda$ { where $u$ has constant sign}. The radial symmetry of solutions to the overdetermined eigenvalue problem \eqref{Schiffer} is known as Schiffer conjecture. Serrin moving plane technique can be generalized also to many nonlinear elliptic operators (such as $p$-Laplacian), but the effectiveness of the proof depends upon the fine structure of the equation, and it is not possible to give an exhaustive list of the nonlinearity covered. Finally we notice that it is possible to consider also different boundary conditions. For instance replace the constant $c$ in \eqref{delta} with a smooth monotone non decreasing function of the mean curvature of $\partial\Omega$. \subsection{Weinberger's proof \cite{Wein}}\label{subswein} The proof makes use of an integral identity (Poho\v{z}aev identity), and of the strong maximum principle applied to an auxiliary function called $P$-function. We recall the Poho\v{z}aev identity \begin{proposition} Let $g\in C^1(\mathbb R)$ be a nonnegative function and let $G(u)=\int_u^0 g(s)ds$. If $u \in C^2(\Omega)\cap C^1(\overline{\Omega})$ is a solution to the problem $$\left\{\begin{array}{ll} \Delta u=g(u)&\mathrm{in } \>\Omega\\ u=0&\mathrm{on }\> \partial\Omega \end{array} \right.$$ in a smooth domain $\Omega$ of $\mathbb R^n$, then \begin{equation} \label{pohozaev} \frac{n-2}{2}\int_\Omega |Du|^2 \>dx +\frac{1}{2}\oint_{\partial\Omega}<x,\nu_x>|Du|^2= n \int_\Omega G(u)\>dx. \end{equation} \end{proposition} \begin{proof} For the proof see for instance \cite{Struwe}, \cite{Poh65}. \end{proof} \begin{proof}[Proof of Theorem \ref{S1}] Let $u$ be a solution to \eqref{delta}, by \eqref{pohozaev} we have \[ \frac{n-2}{2}\int_\Omega |Du|^2 \>dx + \frac{c^2}{2} \oint_{\partial \Omega} <x, \nu_x> = n \int_\Omega u \>dx. \] Equation in \eqref{delta} and the divergence theorem give: \begin{equation} \label{uguale} \int_\Omega |Du|^2 \> dx=\int_\Omega u \>dx \quad { \rm and}\quad \oint_{\partial \Omega} <x, \nu_x>= n |\Omega|. \end{equation} { Therefore we get} \begin{equation} \label{mean} (n+2) \int_\Omega u \> dx = n c^2 |\Omega|. \end{equation} The classical Schwarz's inequality and equation in \eqref{delta} give \begin{equation} \label{schwarz} 1= (\Delta u)^2 \leq n \sum_{i=1}^{n} \left(\frac{\partial^2 u}{\partial x_i^2}\right)^2 \leq n \sum_{i,j=1}^{n} \left(\frac{\partial^2 u}{\partial x_i \partial x_j}\right)^2, \end{equation} so the function $P= |Du|^2 + \frac{2}{n}u$ satisfies \begin{equation} \label{pfunction} \Delta \left(|Du|^2 + \frac{2}{n}u\right) = 2 \sum_{i,j=1}^{n} \left(\frac{\partial^2 u}{\partial x_i \partial x_j}\right)^2 - \frac{2}{n} \ge 0. \end{equation} From the strong maximum principle, since $|Du|^2 + \frac{2}{n}u = c^2$ on $\partial \Omega$, we conclude that either \[ |Du|^2 + \frac{2}{n}u < c^2 \quad {\rm in} \quad \Omega \] or \[ |Du|^2 + \frac{2}{n}u \equiv c^2 \quad {\rm in} \quad \Omega. \] { In the first case by \eqref{uguale} we have} \[ \frac{n+2}{n} \int_\Omega u \> dx < c^2 |\Omega|, \] which contradicts \eqref{mean}. { Therefore $P$ is constant in $\Omega$. This implies equality in both \eqref{pfunction} and \eqref{schwarz}, and we deduce that \[ \frac{\partial^2 u}{\partial x_i \partial x_j}= \frac{\delta_{ij}}{n}. \] Consequently \[ u=\frac{R^2-|x|^2}{2n} \] up to translations and $\Omega$ is a ball of radius $R$.} \end{proof} \subsubsection{Remark on the proof and generalization{ s} .} Weinberger's proof is particularly attractive for its elementary arguments. With respect to Serrin's proof it requires less regularity. Indeed already only interior maximum principle for the auxiliary P-function and the Poho\v{z}aev identity are needed. For this reason $u\in C^2(\Omega)\cap C^1(\bar\Omega)$ is enough. This also means that Weinberger broadens the class of domain among which the symmetry result can be established. Moreover Garofalo and Lewis showed in \cite{GL} that is is possible to assemble via $P$-function a Weinberger argument also for $p$-Laplacian type operators and recast the problem in the Sobolev $W^{1,p}$ settings. This paper opened new perspectives on a technique which for many years have been prescribed to the linear case. Operator in divergence form of $p$-Laplacian type have been later considered for instance in \cite{FaKa,FGK}, and even the special case of the $\infty$-Laplacian has been handled in \cite{BuKaw,CraFra}. \section{More recent alternative proofs } \subsection{The duality Theorem \cite{PaSc}} The duality Theorem shows a deep connection between Serrin's overdetermined problem and the mean value theorem for harmonic functions. It is well known that the average of an harmonic function in a ball always equals the average on its boundary. Serrin's result established that the mean value theorem can be true only on balls in the sense that, if the average on a smooth, { bounded} domain $\Omega$ equals the one on $\partial \Omega$ regardless the harmonic function we consider, then $\Omega$ must be a ball. In what follows $\Omega$ is a smooth domain. \begin{theorem} \label{D1} { Let $u \in C^2(\Omega) \cap C^1(\bar \Omega)$ be the solution to $-\Delta u=1$ in $\Omega$, and $u=0$ on $\partial \Omega$. The following statements are equivalent:} \begin{enumerate}[(i)] \item { $u$ is a solution to \eqref{delta}.} \item ${ \dfrac1{|\Omega|}\displaystyle\int_\Omega h \> dx = \dfrac1{|\partial\Omega|} \displaystyle\oint_{\partial \Omega} h } \quad{ \rm for \>all \> functions} \>h \in { C^{0}({\bar \Omega})} \>{ \rm harmonic \>in\>} \Omega.$ \end{enumerate} \end{theorem} \begin{proof} Assume (i). Divergence theorem immediately implies that for every $h$ harmonic in $\Omega$ \[ \int_\Omega h \>dx = \int_\Omega (-\Delta u) h \> dx = \displaystyle\oint_{\partial \Omega} - \frac{\partial u}{\partial \nu} \,h = { \dfrac{|\Omega|}{|\partial \Omega|} }\displaystyle\oint_{\partial \Omega} \,h \] and then (ii). Conversely assume (ii) and let { $u \in C^2(\Omega) \cap C^1(\bar \Omega)$ }be such that $-\Delta u=1$ in $\Omega$, $u=0$ on $\partial \Omega$. Then \[ 0=\int_\Omega h \>dx -{ \dfrac{|\Omega|}{|\partial \Omega|} } \displaystyle\oint_{\partial \Omega} h \> dx= { \displaystyle\oint_{\partial \Omega} - (\frac{\partial u}{\partial \nu} + \dfrac{|\Omega|}{|\partial \Omega|} )h} \] which implies $\displaystyle\frac{\partial u}{\partial \nu}= -\dfrac{|\Omega|}{|\partial \Omega|} $ choosing $ { \displaystyle h=\frac{\partial u}{\partial \nu} + \dfrac{|\Omega|}{|\partial \Omega|}}$ on $\partial\Omega$. \end{proof} \begin{theorem} \label{D2} If (ii) holds true then $\Omega$ is a ball. \end{theorem} \begin{proof} { By Theorem \ref{D1} there exists $u \in C^2(\Omega) \cap C^1(\bar \Omega)$ solution to \eqref{delta}. Let $h= <x,Du> -2u$. The function $h$ is harmonic in $\Omega$ (observe that it is harmonic in the distributional sense thanks to the differential identity $\Delta(<x,Dv>)= <D(\Delta v),x> + 2\Delta v$ and by classical regularity results $h$ is smooth in $\Omega$).} { (ii) together with $Du = c \nu$ on $\partial \Omega$} give \[ \int_\Omega <x,Du>\> dx -2 \int_\Omega u \> dx = -c^2\oint_{\partial\Omega} <x, \nu_x> =-c^2n |\Omega|. \] { In view of the divergence theorem} \begin{equation} \label{equality33} c^2n |\Omega| -(n+2) \int_\Omega u \> dx =0. \end{equation} Let $P= |Du|^2 +\frac{2}{n}u$, since \[ \int_\Omega u (\Delta P) \> dx = -\int_\Omega P\> dx -c\oint_{\partial\Omega }P, \] { the fact that $u$ is a solution to \eqref{delta} together with \eqref{equality33} yields} \[ \int_\Omega u (\Delta P) \> dx = -\left(\frac{n+2}{n}\right) \int_\Omega u \> dx -c^3 |\partial \Omega|= - \left(\frac{n+2}{n}\right) \int_\Omega u \> dx + c^2|\Omega|=0. \] Here we have used that $c = -\dfrac {|\Omega|}{|\partial\Omega|}$. \noindent The strong maximum principle { leads} $u>0$ in $\Omega$, then \eqref{pfunction} implies that $\Delta P=0$ in $\Omega$ and the proof concludes as in Subsection \ref{subswein}. \end{proof} \subsubsection{Remark on the proof and generalization.} To our knowledge the proof by duality theorem is the first one which does not make explicit use of maximum principle. We face a flavor of the maximum principle when $u$ is assumed { to have constant sign}. Nevertheless the proof is reminiscent of Weinberger's one and indeed both share the same regularity of $u$. Finally there is also an interesting generalization due to Bennett \cite{Be} where fourth order overdetermined problem for biharmonic operator are taken into account. \subsection{The domain derivative \cite{ChoHen}} As we have seen in the first section, there is a deep connection between overdetermination and shape optimization. Throughout this section if $\Omega$ is an open subset of $\mathbb R^n$ satisfying the assumptions of Theorem \ref{S1} such that there exists a solution to problem \eqref{delta}, we say that $\Omega$ is a solution to Serrin's problem. Following \cite{ChoHen} we are going to show how to construct a shape functional which is minimized by solutions to Serrin's Problem and then infer the uniqueness of the minimizer. For every $\omega \subset \mathbb R^n$ with $C^2$ boundary we denote by $u_\omega$ the solution to \begin{equation}\label{deltino} \left\{ \begin{array}{ll} \Delta u_\omega=-1 & \mathrm{in}\,\,\omega \\ u_\omega \in H_0^1(\omega) \\ \end{array} \right. \end{equation} and we consider the functional \[ J(\omega)= n \oint_{\partial \omega} |Du_\omega|^3 -(n+2) \int_\omega |Du_\omega|^2 \> dx. \] The strategy of the proof consists in proving that every solution to Serrin's problem minimizes the functional $J$. Indeed the following lemma holds true \begin{lemma} \label{funct} $J(\omega) \ge 0$ for every $\omega \subset \mathbb R^n$ with $C^2$ boundary. If $\omega$ is a solution to Serrin's problem then $J(\omega)=0$. \end{lemma} \begin{proof} Multiplying \eqref{schwarz} by $u_\omega$ ($u_\omega>0$ in $\omega$) and recalling that $$\Delta(|Du_\omega|^2) = 2 \sum_{i,j=1}^{i=n} \left(\frac{\partial^2 u_\omega}{\partial x_i \partial x_j}\right)^2 $$ we have \begin{equation} \label{integral1} \int_\omega u_\omega \>dx \leq \frac{n}{2} \int_\omega u_{\omega}\Delta(|Du_\omega|^2) \> dx. \end{equation} { The divergence} theorem, the fact that $u_\omega=0$ and the fact that $\frac{\partial u_\omega}{\partial \nu_x}= -|Du_\omega|$ on $\partial \omega$ { bring} \begin{equation} \label{integral2} \int_\omega u_\omega \>dx \leq \frac{n}{2} \left[\oint_{\partial\omega} |Du_{\omega}|^3 + \int_\omega |Du_{\omega}|^2 (\Delta u_\omega)\> dx \right]. \end{equation} { Equation} in \eqref{deltino}, { together with} \begin{equation} \label{aster} \int_\omega u_\omega \>dx= \int_\omega |Du_{\omega}|^2 \> dx \end{equation} { carry} \begin{equation} \label{integrale3} 0 \leq \frac{n}{2} \oint_{\partial\omega} |Du_{\omega}|^3 - \left(\frac{n}{2}+1\right)\int_\omega |Du_{\omega}|^2 \> dx, \end{equation} { which proves the inequality}. Now assume that $\omega$ is a solution to Serrin's problem, then $c= \frac{\partial u_\omega}{\partial \nu_x}= -|Du_\omega|$ on $\partial \omega$. From \eqref{mean} and \eqref{aster} we have that \[ (n+2) \int_\omega |Du_{\omega}|^2 \> dx = nc^2 |\omega|, \] { therefore} \[ J(\omega) = -nc^3 |\partial \omega| -nc^2|\omega|. \] { Since} $c =- \frac{|\omega|}{|\partial \omega|},$ the thesis follows. \end{proof} We briefly recall the definition of shape derivative and Hadamard formula (we refer for instance to \cite{had, HenPi}){ .} Let $\omega$ be a smooth open set in $\mathbb R^n$, and let $\theta \in C^2(\mathbb R^n;\mathbb R^n)$ and denote by $\omega_t= \{ x+t \theta(x), x \in \omega \}, \>\> t>0$. The derivative of $J$ at $\omega$ in the direction $\theta$ is \begin{equation} \label{shapederiv} dJ(\omega, \theta) = \lim_{t \rightarrow 0^+} \frac{J(\omega_t)-J(\omega)}{t}. \end{equation} The computation of \eqref{shapederiv} leads to calculate also the derivative of $u_\omega$ with respect to the domain. Such a derivative denoted by $u_\omega'$ satisfies \begin{equation}\label{deltino2} \left\{ \begin{array}{ll} \Delta u'_\omega=0 & \mathrm{in}\,\,\omega \\ u'_\omega = -\displaystyle\frac{\partial u}{\partial \nu_x} <\theta,\nu_x> & \mathrm{on}\,\,\partial \omega. \end{array} \right. \end{equation} \begin{lemma} \label{sd} The derivative of the functional $J$ at $\omega$ in the direction $\theta$ is given by \begin{equation} \label{derivJ} dJ(\omega,\theta) = \oint_{\partial\omega} \left (\left[(2n-2)|Du_\omega|^2 -2n(n-1)H|Du_\omega|^3 \right] <\theta,\nu_x> -3n |Du_\omega|^2 \displaystyle\frac{\partial u'_\omega}{\partial \nu_x}\right) \end{equation} where $H$ is the mean curvature of $\partial \omega$ and $u'_\omega$ is defined in \eqref{deltino2}. \end{lemma} \begin{proof} The proof follows { from} Hadamard formula (see \cite{had,HenPi}). { For the function $j_1(\omega)= \displaystyle\int_\omega f(\omega) \>dx$ such a formula reads \[ dj_1(\omega,\theta)= \int_\omega f'(\omega)\> dx + \oint_{\partial \omega} f(\omega) <\theta,\nu_x> , \] while for $j_2(\omega) = \displaystyle\oint_{\partial \omega} g(\omega) $ we have \[ dj_2(\omega,\theta)= \oint_{\partial \omega} g'(\omega)\> + \oint_{\partial \omega} (n-1)H \,g(\omega) <\theta,\nu_x> + \oint_{\partial \omega} \frac{\partial g(\omega)}{\partial \nu_x} \, <\theta,\nu_x>. \] Here $f'(\omega)$ and $g'(\omega)$ denote the derivatives with respect to the domain of $f$ and $g$, respectively.} From these two formulae applied to $J$ we get \begin{equation}\label{der1} \begin{array}{ll} dJ(\omega,\theta)= & 3n \displaystyle\oint_{\partial \omega}|Du_\omega| <Du_\omega,Du'_\omega> + n(n-1) \oint_{\partial \omega} |Du_\omega|^3 H <\theta,\nu_x> \\ \\ &+ n \displaystyle\oint_{\partial \omega} \frac{\partial |Du_\omega|^3}{\partial \nu_x} \, <\theta,\nu_x> -2(n+2) \int_\omega <Du_\omega,Du'_\omega> \>dx \\\\ & -(n+2) \displaystyle\oint_{\partial \omega} |Du_\omega|^2 \, <\theta,\nu_x> . \end{array} \end{equation} By { the divergence theorem, Problem \eqref{deltino}, and Problem \eqref{deltino2}, we get} \begin{equation} \label{der2} \int_\omega <Du_\omega,Du'_\omega> \>dx = \displaystyle\oint_{\partial \omega} u_\omega \frac{\partial u'_\omega}{\partial \nu_x}-\int_\omega u_\omega \Delta u'_\omega \>dx =0. \end{equation} On the other hand \begin{equation} \label{der3} \displaystyle\oint_{\partial \omega} |Du_\omega| <Du_\omega,Du'_\omega>= -\displaystyle\oint_{\partial \omega} |Du_\omega|^2 \frac{\partial u'_\omega}{\partial \nu_x}. \end{equation} Finally \[ \frac{\partial |Du_\omega|^3}{\partial \nu_x}= 3 |Du_\omega|^2 <D(|D u_\omega|), \nu_x>= -3 |Du_\omega| <D(|D u_\omega|), D u_\omega>. \] { Bearing in mind} that for $\partial \omega =\{ x: u_\omega(x)=0\}$ it holds \[ (n-1) H = -{\rm div}\left(\displaystyle\frac{Du_\omega}{|Du_\omega|}\right)= \frac{1}{|Du_\omega|}+ \frac{1}{|Du_\omega|^2}<D(|D u_\omega|), D u_\omega>, \] \noindent we have \begin{equation} \label{der4} \frac{\partial |Du_\omega|^3}{\partial \nu_x}= -3 |Du_\omega| \left( (n-1) H - \frac{1}{|Du_\omega|} \right). \end{equation} Plugging \eqref{der2}, \eqref{der3}, \eqref{der4} { into} \eqref{der1} we obtain \eqref{derivJ}. \end{proof} \begin{proof}[Proof of Theorem \ref{S1}] Let $\Omega$ be a solution to Serrin problem. By lemma \ref{funct}, $\Omega$ is a minimizer of $J$, then for every vector field $\theta \in C^2(\mathbb R^n;\mathbb R^n)$ we must have $$ dJ(\Omega,\theta)=0. $$ Using \eqref{derivJ} together with $c= \dfrac{\partial u_\Omega}{\partial \nu_x}= -|Du_\Omega|$ on $\partial \Omega$ yields \[ dJ(\Omega,\theta)=2c^2(n-1) \oint_{\partial\Omega} [ 1+nHc] <\theta,\nu_x> - 3nc^2 \oint_{\partial\Omega} \displaystyle\frac{\partial u'_\omega}{\partial \nu_x}. \] { Equation \eqref{deltino2} gives} $$\oint_{\partial\Omega} \displaystyle\frac{\partial u'_\omega}{\partial \nu_x}=0,$$ and then \[ dJ(\Omega,\theta)=2c^2(n-1) \oint_{\partial\Omega} [ 1+nHc] <\theta,\nu_x> =0, \quad \forall \>\> \theta \in C^2(\mathbb R^n;\mathbb R^n). \] Hence the mean curvature of $\partial \Omega$ is constant and Alexandrov theorem (see \cite{A1,A2}) implies that $\Omega$ is a ball. This concludes the proof. \end{proof} \subsubsection{Remark on the proof and { generalizations}.} The proof via shape derivative is another nice example of proof which does not uses the maximum principle explicitly. Again however the constant sign of the solution $u$ is used. It is also interesting to notice that it uses Alexandrov theorem \cite{A1,A2} which in turn, at least in the original version, relies on the moving planes. Recently even a deeper connection between Alexandrov Theorem and Serrin problem has been exploited in \cite{CiMa, MP} on the wake of \cite{Ros}. The shape derivative technique requires (following \cite{ChoHen}) somewhat more regularity than Weinberger's ones. It has been however successfully applied in other contexts for instance to obtain partial result toward the solution to the Schiffer conjecture (see \cite{ChaHen}). \subsection{An integral approach via arithmetic-geometric mean inequality \cite{BNST}} This idea stems from the need to extend Serrin overdetermined result to non uniformly elliptic operators of Hessian type. It is a fairly simple proof once we get acquainted with the notation used. We denote by $A=(a_{ij})$ a matrix in the space $\mathcal{S}_n$ of the real symmetric $n\times n$ matrices, and by $\lambda_1,...,\lambda_n$ its eigenvalues, we define the first and the second elementary symmetric function of its eigenvalues as $$S_1(A) = Tr(A), \quad S_2(A)=S_2(\lambda_1,...,\lambda_n)=\sum_{1\leq i_1 <i_2\leq n } \lambda_{i_1} \lambda_{i_2}.$$ Note that $S_2(A)$ is just the sum of all $2 \times 2$ principal minors of $A$, and in dimension $2$ is nothing but ${\rm Det} A$. Denoting by $$ S_2^{ij}(A) = \frac{\partial }{\partial a_{ij}}S_2(A),$$ Euler identity for homogeneous functions gives $$ S_2(A) = \frac{1}{2} S_2^{ij}(A) a_{ij}, $$ here we are adopting the Einstein summation convention for repeated indices. Then the following inequality, known as Newton inequality, holds true in the class of matrices whose trace is nonnegative \begin{equation}\label{rel3} (S_1(A))^2 \ge \frac{2n}{n-1} S_2 (A); \end{equation} equality in \eqref{rel3} implies $\lambda_1= \lambda_2=...=\lambda_n $ (see \cite{HLP}). Given a $C^2$ function $u$, the $k$-Hessian operators $S_k\left(D^2u\right)$ ($k=1,2$) are defined as the $k$-th elementary symmetric function of $D^2u$. Observe that with this notation $$S_1(D^2u)=\Delta u.$$ \noindent A direct computation yields that $(S_2^{1j}(D^2u),\dots,S_2^{nj}(D^2u))$ is divergence free, i.e. \begin{equation} \label{div0} \frac{\partial}{\partial x_i}S_2^{ij}=0; \end{equation} hence $S_2(D^2 u)$ can be written in the following divergence form \begin{equation}\label{divk} S_2(D^2 u) = \frac{1}{2} S_2^{ij}(D^2 u) u_{ij} = \frac{1}{2} (S_2^{ij}(D^2 u) u_j)_i, \end{equation} (from now on subscripts stand for partial differentiations). Let $t$ be a regular value of $u$ and let $L=\{u \ge t\}$. If, with an abuse of notation, we denote by $H= -{\rm div}\left(\displaystyle\frac{Du}{|Du|}\right)$, $(n-1)$ times the curvature of the level set $\partial L$ at the point $x$, then $$-\Delta u=H |Du|-\frac{u_{ij}u_iu_j}{|Du|^2}.$$ { This} means that the value of $\Delta u$ at any regular point (i.e. { a point with} non vanishing gradient) only involves derivates of $u$ along the direction of steepest descent and the mean curvature $H/(n-1)$ of the level surface through that point. Finally the following pointwise identity holds (see \cite{Reilly}) \begin{equation}\label{H-S} -H=\frac{S_2^{ij}(D^2 u)u_iu_j}{|Du|^{3}}. \end{equation} \begin{proof}[Proof of Theorem \ref{S1}] { First we observe that} \begin{eqnarray} &\displaystyle\int_\Omega|Du|^2 &=\int_\Omega |Du|^2 (-\Delta u) \notag \\ &&=2\int_\Omega u_{ij}u_iu_j -\oint_{\partial \Omega}|Du|^2<Du,\nu_x> \notag \\ &&=2\int_\Omega\left[\Delta u|Du|^2+H|Du|^3\right] + c^3|\partial \Omega|. \notag \end{eqnarray} Then, using the equation in \eqref{delta} and the fact that $-c = \frac{|\Omega|}{|\partial \Omega|}$ we have \begin{equation}\label{H} \int_\Omega H|Du|^3=\frac{3}{2}\int_\Omega |Du|^2-\frac{c^2 }{2}|\Omega|. \end{equation} { Plugging \eqref{uguale} and \eqref{mean} into \eqref{H} we obtain \begin{equation}\label{H_Omega} \int_\Omega H|Du|^3= \frac{(n-1)}{ (n+2)}c^2|\Omega|. \end{equation} Using that $u >0$ in $\Omega$, equations \eqref{divk}, \eqref{H-S}, \eqref{H_Omega}, and inequality \eqref{rel3} yield} $$ \frac{(n-1)}{ (n+2)}c^2|\Omega|=\int_\Omega H|Du|^3=2 \int_\Omega u S_2(D^2 u) \le \frac{n-1}{n}\int_\Omega u=\frac{(n-1)}{ (n+2)}c^2|\Omega|. $$ This implies that equality holds true in \eqref{rel3} so \begin{equation} S_2(D^2 u)=\frac{n-1}{2n} \quad {\rm in } \>\> \Omega, \end{equation} and the Hessian matrix $D^2 u$ has all equal eigenvalues at every point of $\Omega$. This fact implies that $D^2 u$ is { a constant times} the identity matrix and the thesis follows. \end{proof} \subsubsection{Remark on the proof and generalization.} Here is another example where besides the constant sign of the solution $u$ there is no shade of maximum principle. Basically the only ingredient of the proof is the geometric mean inequality. Once again the proof only needs the regularity required by the Poho\v{z}aev inequality i.e.: $u\in C^2(\Omega)\cap C^1(\bar\Omega)$. There is a deep connection between this proof and the Weinberger's proof, since the first one consists somehow in evaluating the integral over $\Omega$ of $\Delta u$ times the $P$-function. However no maximum principle on $P$ is established and everything is kept in integral form. Even if the proof was successfully applied to nonlinear operator of Hessian type, it turned out that the main advantage of this approach is that it does not use any pointwise argument. By means of this technique, stability theorem for Serrin problem like those in \cite{ABR} were improved in \cite{BNST2}. Moreover the technique is well designed when dealing with anisotropic overdetermined problem \cite{CiaSal}, where intrinsic asymmetry and lack of regularity advise against Serrin's and Weinberger's proofs.

\section{Introduction} \IEEEPARstart{I}{t} Cloud Radio Access Network (Cloud-RAN) is a network architecture proposed to meet the explosive growth of mobile data traffic. An important problem of Cloud-RAN is the energy efficiency consideration, due to the increasing power consumption of a large number of Remote Radio Heads (RRHs) as well as the fronthaul links. We focus on the power consumption problem of green Cloud-RAN by jointly involving the power consumption of the transport network and RRHs. Several methods have been proposed to solve the Cloud-RAN power minimization problem. The network power minimization problem can be formulated into a Mixed Integer Programming (MIP) problem, which is mainly solved by three strategies. First of all, a global optimal solution can achieve by the branch-and-bound method~\cite{cheng2013joint}, but it may suffer from an exponential worst-case complexity and work slowly in practice. In order to alleviate the computational burden, Yang et. al.~\cite{cheng2013joint} derives an approximation of the MIP problem by relaxing the binary constraint to a $[0, 1]$ box constraint. The most related method is a three-stage approach named Group Sparse Beamforming (GSBF) algorithm~\cite{Hong2013Joint, Dai2017Sparse}, which balances between the computational complexity and the accuracy of solution. This algorithm exploits the group sparsity structure of beamformers with the priori knowledge. Specifically, in the first stage, it solves a convex weighted $\ell_1/\ell_2$ norm relaxation of the MIP problem to induce the sparsity of the beamformers. The second stage generates an ordering rule to decide which RRH has a higher priority to be switched off. In the third stage, a selection procedure is performed to determine the best combination of the active and the sleep set of RRHs. However, the GSBF algorithm generally can not guarantee to provide a high accuracy solution. In this paper, we propose a new formulation of the Cloud-RAN power consumption problem along with a dual ascent method combined with an inexact Majorization-Minimization (MM) algorithm. The major idea of this recast is the $\ell_2$-box technique, introduced recently in~\cite{wu2016ell_p} as a continuous equivalent formulation of the binary constraints. By using this technique to replace the binary constraint with the intersection of a box and $\ell_2$ sphere, we obtain a new formulation of the Cloud-RAN power consumption problem. As a result, a local optimal solution can be quickly found by continuous algorithms, while for the original mixed-binary problem, excessive computational effort may be needed to find a comparable solution. Therefore, the solution of the proposed reformulated problem can be employed to initialize the bi-section GSBF algorithm, which can be a more powerful sparsity-promoting tool than the weighted $\ell_1/\ell_2$ norm relaxation. It should be emphasized that our exact and continuous formulation of the network power consumption problem, in contrast to a relaxation model, can enable a better solution be reached. The reformulated problem can be addressed by our proposed MM dual ascent algorithm and test by the numerical experiments. The numerical results manifest that our proposed framework obviously improves the network energy efficiency, especially in the case of more RRHs but fewer users. \section{PROBLEM DESCRIPTION} \subsection{System model} We consider a Cloud-RAN with $L$ RRHs and $K$ single-antenna Mobile Users (MUs), where the $l$-th RRH is equipped with $N_l$ antennas. In this architecture, all the Baseband Units (BBU) are moved in to a single BBU pool, creating a set of shared processing resources, and enabling efficient interference management and mobility management. All the RRHs are connected to the BBU pool through fronthaul links. In a beamforming framework, let $\bm{v}_{lm} \in \mathbb{C}^{N_l}$ be the transmit beamforming vector from the $l$-th RRH to the $k$-th user, and $s_k$ be the data symbol for user $k$ with $\emph{E}[\left | s_k\right |^2 =1]$. The transmit signal at RRH $l$ is given by \begin{equation} \bm{x}_{l}=\sum_{k=1}^{K}\bm{v}_{lm}s_k,\ \forall l \in \mathcal{L} \label{eq1}. \end{equation} The channel propagation between user $k$ and RRH $l$ is denoted as $\bm{h}_{lm}\in \mathbb{C}^{N_l}$, and $n_k \in \mathcal{CN}(0, \sigma_k^2)$ is the additive Gaussian noise at user $k$. Therefore, the received signal at user $k$ is then \begin{equation} y_k = \sum _{l \in \mathcal{L}}\bm{h}_{kl}^{\sf{H}}\bm{v}_{lk}s_k+\sum_{i\neq k}\sum_{l \in \mathcal{L}}\bm{h}_{kl}^{\sf{H}}\bm{v}_{li}s_{i}+n_k \label{eq2}. \end{equation} We assume that all the users treat the interference as noise~\cite{cadambe2008interference}. The corresponding signal-to-interference-plus-noise ratio (SINR) for user $k$ is \begin{equation} \Gamma_k = \frac{\left | \sum _{l \in \mathcal{L}}\bm{h}_{kl}^{\sf{H}}\bm{v}_{lk} \right |^2}{\sum_{i\neq k}\left | \sum_{l \in \mathcal{L}}\bm{h}_{kl}^{\sf{H}}\bm{v}_{li} \right |^2+\sigma _k^2}. \label{eq3} \end{equation} Each RRH has its own transmit power constraint \begin{equation} \sum_{k=1}^{K} \| \bm{v}_{lk} \|_{2}^2 \leq \sqrt{P_l},\ \forall l \in \mathcal{L}, \label{eq4} \end{equation} where $P_l$ is the maximum transmit power of the $l$-th RRH. \subsection{Network power consumption minimization} Due to the high density of RRHs and their joint transmission, the energy used for signal transmission can be reduced significantly. However, the power consumption of the transport network becomes numerous and cannot be ignored. In order to reduce the network power consumption, it is essential to put some RRHs into sleep whenever possible. We introduce a binary vector $\bm{z} = (z_1,..., z_L)^T$ to represent the active RRH, i.e., $z_l = 1$ denotes the $l$-th RRH is active, and $z_l = 0$ means the $l$-th RRH is sleeping. Denote the relative fronthaul link power consumption by $P_l^c$, and the inefficient of drain efficiency of the radio frequency power amplifier by $\eta_l$. Then the network power consumption $p (\bm{z}, \bm{v})$ is the sum of total transmit power consumption and the total relative fronthaul links power consumption: \begin{equation} p (\bm{z}, \bm{v}) = \sum_{l \in \mathcal{L}}P_l^c z_l + \sum_{l \in \mathcal{L}}\frac{1}{\eta_l} \| \bm{\tilde{v}}_{l} \|_2^2. \label{eq5} \end{equation} where, for convenience, let $\bm{\tilde{v}}_{l}=[\bm{v}_{l1}^T,..., \bm{v}_{lK}^T]^T\in \mathbb{C}^{KN_l\times 1}$. With target SINRs $\bm{\gamma} = (\gamma_1,..., \gamma_K)^T$, the SINR constraint for user $k$ as a second-order cone (SOC) constraint~\cite{shi2014group} must be satisfied. Therefore, the power minimization problem can be formulated as a MIP problem~\cite{lee2011mixed} \begin{equation} \begin{aligned} & \underset{(\bm{z},\bm{v})}{\operatorname{min}} & & p (\bm{z}, \bm{v}) \\ & \text{s.t.} & & \sqrt{\sum_{i\neq k} \| \bm{h}_k^{\sf{H}} \bm{v}_i \|_2^2 + \sigma _k^2}\leq \frac{1}{\gamma _k}\Re (\bm{h}_k^{\sf{H}} \bm{v}_k), k \in \mathcal{S},\\ &&& \| \bm{\tilde{v}}_{l} \| \leq z_l \sqrt{P_l}, z_l = \{0, 1\} , l \in \mathcal{L} , \end{aligned} \label{eq6} \end{equation} where $\Re(\cdot)$ denotes the real part. \section{$\ell_2$-box optimization reformulation} In this section, we propose a new formulation of the Cloud-RAN power consumption problem. An $\ell_2$-box technique is proposed to replace the binary by the intersection between a box and an $\ell_2$ sphere as described in~\eqref{eq7}. A geometric illustration of the $\ell_2$-box technique is depicted in Fig. 1. \begin{equation} \bm{x} \in \{0, 1\}^n \Leftrightarrow \bm{x} \in [0, 1]^n \cap \left\{\bm{x}: \| \bm{x}- \frac{1}{2} \bm{\emph{1}}_n \|_2^2 = \frac{n}{4} \right \}, \label{eq7} \end{equation} \begin{figure}[t!] \centering \centerline{\includegraphics[width=0.4\linewidth]{lpbox.png}} \caption{Geometric illustration of the $\ell_2$-box technique. The hollow circle is the intersection of the box and $\ell_2$-norm.} \label{} \end{figure} where $\bm{\emph{1}}_n$ is an $n$-dimension all-one vector. Therefore, the MIP~\eqref{eq6} can be recast into \begin{equation} \begin{aligned} & \underset{(\bm{z},\bm{v})}{\operatorname{min}} & & \sum_{l \in \mathcal{L}}P_l^c z_l + \sum_{l \in \mathcal{L}}\frac{1}{\eta_l} \| \bm{\tilde{v}}_{l} \|_2^2 \\ & \text{s.t.} & & \sqrt{\sum_{i\neq k} \| \bm{h}_k^{\sf{H}} \bm{v}_i \|_2^2 + \sigma _k^2}\leq \frac{1}{\gamma _k}\Re (\bm{h}_k^{\sf{H}} \bm{v}_k), k \in \mathcal{S} ,\\ &&& \| \bm{\tilde{v}}_{l} \| \leq z_l \sqrt{P_l}, 0 \leq z_l \leq 1 , l \in \mathcal{L},\\ &&& \| \bm{z} - \frac{1}{2}\bm{\emph{1}}_L \|_2^2 = \frac{L}{4}. \end{aligned} \label{eq8} \end{equation} The main difficulty of this problem comes from the nonconvexity of the sphere constraint, and we are well aware that it may be effort-consuming to find the global optimal solution. However, rather than solve~\eqref{eq8} directly for the global optimal solution, rather than use the global nonlinear method to solve~\eqref{eq8} until global optimal, we only use local nonlinear algorithm to address~\eqref{eq8} and use the (local) solution as the initial point in the first stage of the bi-section GSBF algorithm, which can further induce the group sparsity of the beamformers. \section{MM Dual ascent Algorithm} \label{section4} In this section, we design a dual ascent algorithm~\cite{boyd2011distributed} incorporated with an inexact MM algorithm to solve our proposed $\ell_2$-box Cloud-RAN power minimization problem. Notice that~\eqref{eq8} is a convex problem except for the nonconvex $\ell_2$ sphere constraint. Therefore, we focus on dealing with the sphere constraint to construct our algorithm. For simplicity, let \begin{equation} \phi(\bm{v},\bm{z}) = \sum_{l \in \mathcal{L}}P_l^c z_l + \sum_{l \in \mathcal{L}}\frac{1}{\eta_l} \| \bm{\tilde{v}}_{l} \|_2^2 , \label{eq9} \end{equation} and $\Omega = \{(\bm{z},\bm{v})|\sqrt{\sum_{i\neq k} \| h_k^{\sf{H}} \bm{v}_i \|_2^2 + \sigma _k^2}\leq \frac{1}{\gamma _k}\Re (h_k^{\sf{H}} \bm{v}_k), k \in \mathcal{K} ;\\ \| \bm{\tilde{v}}_{l} \| _2\leq z_l \sqrt{P_l},0 \leq z_l \leq 1,l \in \mathcal{L} \}.$ Notice that $\Omega$ is a convex set. Now~\eqref{eq8} can be stated as \begin{equation} \begin{aligned} & \underset{{(\bm{z},\bm{v})\in \Omega}}{\operatorname{min}} & & \phi(\bm{z},\bm{v}) \\ & \text{s.t.} && \| \bm{z} - \frac{1}{2}\bm{\emph{1}}_L \|_2^2 = \frac{L}{4}. \\ \end{aligned} \label{eq10} \end{equation} A natural way to solve such a problem is to dualize the sphere constraint. Letting $\lambda$ be the multiplier associated with the sphere constraint, the Lagrangian of~\eqref{eq10} is defined as \begin{equation} L(\bm{z},\bm{v},\lambda) = \phi(\bm{z},\bm{v})+\lambda(\frac{L}{4}- \| \bm{z} - \frac{1}{2}\bm{\emph{1}}_L \|_2^2),\\ \label{eq11} \end{equation} for $(\bm{z},\bm{v})\in \Omega$. An alternative option is to use the augmented Lagrangian, but we do not suggest such approach since it will severely increase the nonlinearity of the resulted subproblems by introducing a fourth-order polynomial in the objective. The dual of objective is then given by~\eqref{eq12} \begin{equation} g(\lambda) = \inf_{(\bm{z},\bm{v})\in \Omega} L(\bm{z},\bm{v},\lambda), \label{eq12} \end{equation} and we have the dual problem~\eqref{eq13} \begin{equation} \max_{\lambda} g(\lambda) = \max_{\lambda}\inf_{(\bm{z},\bm{v})\in \Omega}L(\bm{z},\bm{v},\lambda). \label{eq13} \end{equation} Now we are ready to provide our dual ascent framework. The dual ascent method consists of two stages: the first stage is to update the primal variables by minimizing the Lagrangian for a fixed dual variable $\lambda$, \begin{equation} (\bm{z}^{t+1},\bm{v}^{t+1}) = \text{arg}\min_{(\bm{z},\bm{v})\in \Omega} L(\bm{z},\bm{v},\lambda^t), \label{eq14} \end{equation} and then update the dual variable based on the constraint residual \begin{equation} \lambda^{t+1} = \lambda^t + \alpha^t(\frac{L}{4}- \| \bm{z}^{t+1} - \frac{1}{2}\bm{\emph{1}}_L \|_2^2), \label{eq15} \end{equation} where $\alpha^t>0$ is the step size of the dual update. Since $\bm{z}$ is restricted in $\Omega$, it holds true $\bm{z}^{t+1}\in [0, 1]^L$. It follows that $(\frac{L}{4}- \| \bm{z}^{t+1} - \frac{1}{2}\bm{\emph{1}}_L \|_2^2) \geq 0$, where the equality holds true if and only if $\bm{z} \in \{0, 1\}^L$. As a result, the dual variable $\lambda^0 $ should be initialized to be positive to penalize the violation of the sphere constraint. Since the step size $\alpha^t > 0$, $\lambda$ is maintaining positive and increasing incrementally during the solution. Consequently, $\lambda^t \| \bm{z} - \frac{1}{2}\bm{\emph{1}}_L \|_2^2$ is kept convex with respect to $\bm{z}$. In other words, the subproblem~\eqref{eq14} is a DC problem~\cite{hartman1959functions}. At the $s$-th iteration of MM algorithm~\cite{sun2017majorization}, a convex surrogate objective $\hat{L}(\bm{z},\bm{v},\lambda^t,\bm{z}^{(s)})$ is generated by linearizing the second convex function while keeping the first function unchanged \begin{equation} \begin{aligned} \hat{L}(\bm{z},\bm{v},\lambda^t,\bm{z}^{(s)}) = \phi(\bm{z},\bm{v})+ \lambda^t(\frac{L}{4}- \| \bm{z}^{(s) }- \frac{1}{2}\bm{\emph{1}}_L \|_2^2\\ -2(\bm{z}-\frac{1}{2}\bm{\emph{1}}_L)^T(\bm{z}-\bm{z}^{(s)})), \label{eq16} \end{aligned} \end{equation} where $(\bm{z},\bm{v})\in \Omega$ and $\{( \bm {z^{(s)}}, \bm{v^{(s)}})\}$ represents a sequence of primal iterates for the subproblems. To obtain $( \bm {z^{(s+1)}}, \bm{v^{(s+1)}})$ in each iteration of MM algorithm, we need to solve \begin{equation} \min_{(\bm{z},\bm{v}) \in \Omega}\hat{L}(\bm{z},\bm{v},\lambda^t,\bm{z}^{(s)}), \label{eq17} \end{equation} which can easily be solved by CVX solver~\cite{grant2008cvx}. It should be noticed that generally the subproblem does not need to be solved exactly if sufficient improvement on the primal variable can be achieved. Therefore, we also solve~\eqref{eq14} inexactly, meaning we only solve a few subproblems~\eqref{eq17}. This inexact strategy has proven to be able to reduce the computational cost substantially The description of the entire MM dual ascent algorithm is stated in Algorithm~\ref{alg.MMdual}. The convergence analysis of dual ascent method is provided by~\cite{Bertsekas1999Nonlinear}. Since MM algorithm is proposed~\cite{Kiers2016Discussion} as a generalization of the EM algorithm, MM algorithm inherits the convergence properties of the EM algorithm~\cite{vaida2005parameter}. The convergence results of the EM algorithm includes: the likelihood sequence of the EM algorithm is nondecreasing and convergent~\cite{Dempster1977Maximum}, and that the limit points of the EM algorithm are stationary points of the likelihood~\cite{Wu1983On}. After obtaining the sparse beamformer $\bm{{v^*}}$, we use its group sparsity to generate the ordering criterion, and then adopt the same binary search procedure as the bi-Section GSBF algorithm in~\cite{shi2014group} to obtain the final results. \begin{algorithm}[htbp] \caption{MM dual ascent algorithm} \label{alg.MMdual} \begin{algorithmic}[1] \STATE Given the tolerances $\epsilon_1>0,\ \epsilon_2>0$ and $\epsilon_3>0$. \STATE Initialize $\bm{z}, \bm{v}$ and $\lambda>0$. \WHILE{$| \lambda^{t+1}-\lambda^t |\geq \epsilon_2$ or $\| \bm{z}^{t+1}-\bm{z}^t \|_2 + \| \bm{v}^{t+1}-\bm{v}^t \|_F \geq \epsilon_3$} \WHILE{$ \| \bm{z}^{(s+1)}-\bm{z}^{(s)} \|_2 + \| \bm{v}^{(s+1)}-\bm{v}^{(s)} \|_F\geq \epsilon_1$} \STATE \textbf{Update $(\bm{z},\bm{v})$}:\\ $(\bm{z}^{(s+1)},\bm{v}^{(s+1)}) = \text{arg}\min_{\bm{z},\bm{v}} \hat{L}(\bm{z},\bm{v},\lambda^t,\bm{z}^{(s)}) $ \STATE $s = s+1$ \ENDWHILE \STATE $\bm{z}^t = \bm{z}^{(s)}, \bm{v}^t = \bm{v}^{(s)}$ \STATE \textbf{Update $\lambda$}:\ $\lambda^{t+1} = \lambda^t + \alpha^t(\frac{L}{4}- \| \bm{z}^t - \frac{1}{2}\bm{\emph{1}}_L \|_2^2)$ \STATE Set $t = t+1$ \ENDWHILE \STATE \textbf{return} $\bm{\tilde{v}}_{l}$, for $l = 1,...,L$. \end{algorithmic} \end{algorithm} \section{Simulation Results} In this section, we describe the experimental setting including the initial point and the algorithm parameters. In our experiment we check the convergence of the proposed method, and exhibits the effectiveness of our proposed method compared with contemporary methods. The initial point plays a significantly important role while solving the nonconvex problems. Instead of randomly choosing initial point, we remove the sphere constraint and solve the approximation problem \begin{equation} (\bm{z}^0, \bm{v}^0) = \text{arg}\min_{(\bm{z},\bm{v})\in \Omega} \phi(\bm{z},\bm{v}) \label{eq18} \end{equation} to derive the initial point. This generally renders better estimate than random initial point of the solution. The DC subproblem is solved inexactly under the stopping criterion $\| \bm{z}^{t+1}-\bm{z}^t \|_2 \leq \epsilon_1$ with $\epsilon_1 = 10^ {-5}$. The main algorithm is terminated whenever the primal iterates or the dual iterates converge, i.e., we use termination criterion $| \lambda^{t+1}-\lambda^t |\leq \epsilon_2$ or $\| \bm{z}^{s+1}-\bm{z}^s \|_2 + \| \bm{v}^{s+1}-\bm{v}^s \|_F \leq \epsilon_3$ with $\epsilon_2 =10^{-2}, \epsilon_3 =10^{-3}$. In our experiment, we consider a network with $L = 10$, $K = 6$, $2$-antenna RRHs and single-antenna MUs uniformly and independently distributed in the square region $[-1000,\ 1000]\times[-1000,\ 1000]$ meters. We set all the relative transport link power consumption to be $P_l^c = 13W,\ l = 1, . . . , L$, and the inefficient of power amplifier~\cite{auer2011much} at each RRH is $\eta_l = \frac{1}{4}$. In our first experiment, we show the efficiency of Algorithm~\ref{alg.MMdual}. The evolution of $ tol_1^t = \log{\| \bm{z}^{t+1}-\bm{z}^t \|_2} $ and $ tol_2^t = \log{\| \bm{v}^{t+1}-\bm{v}^t \|_F}$, where $\|\cdot\|_F$ is the Frobenius norm, is depicted in Fig.~\ref{ and (b) to represent the differences between the current and the previous iterates. As shown in Fig.~\ref{}, both of the primal variables $\bm{z}$ and $\bm{v}$ are converging efficiency. In particular, $ {tol}_1 $ and $ {tol}_2$ decrease dramatically about $10^{-5}$ in the first five iterations. \begin{figure}[hbt] \begin{minipage}[b]{0.495\linewidth} \centering \centerline{\includegraphics[width=4cm]{z.png}} \centerline{(a) Evolution of $ {tol}_1 $}\medskip \label{} \end{minipage} \begin{minipage}[b]{0.495\linewidth} \centering \centerline{\includegraphics[width=4cm]{v.png}} \centerline{(b) Evolution of $ {tol}_2 $}\medskip \label{} \end{minipage} \caption{Convergence of the primal and dual variables.} \label{} \end{figure} We also compare our proposed method with the existing methods including: MIP which is the branch-and-bound algorithm for solving the MIP problem~\eqref{eq6} for global optimal solution, RMIP which is the algorithm in~\cite{cheng2013joint} for solving the relaxed MIP problem, and GSBF which is the bi-section GSBF algorithm~\cite{Hong2013Joint, Dai2017Sparse}. The average network power consumption with different target SINR is shown in Fig.~\ref{}. The simulation results indicate that the $\ell_2$-box algorithm outperforms the GSBF and RMIP algorithm for different target SINR. This advantage becomes obvious in situations with smaller SINR. \begin{figure}[t!] \begin{minipage}[b]{1.0\linewidth} \centering \centerline{\includegraphics[width=5.5cm]{result.png}} \end{minipage} \caption{Average network power consumption versus target SINR.} \label{} % \end{figure} \section{Conclusion and Future Work} In this paper, we have proposed a new formulation of the Cloud-RAN power consumption problem by using the $\ell_2$-box technique, which replaces the binary constraint to two continuous constraints: a box constraint and a sphere constraint. We design a dual ascent algorithm to solve the new $\ell_2$-box optimization problem leading a sequence of DC subproblems. We apply MM algorithm to inexactly solve the subproblem. The effectiveness of our proposed reformulation and algorithm is demonstrated in numerical experiment. Our method exhibited lower network consumptions of different target SINR than the GSBF algorithm. Our investigation leads to a variety of open questions. The final solution found by a nonlinear solver is often sensitive to the initial point. Therefore, it would be useful to explore better estimate of the global optimal solution to initialize our algorithm. Furthermore, the binary constraint is also equivalent to the intersection of a box and an $\ell_p$ sphere with $p \in (0,+\infty)$. It would be interesting to investigate the performance of other $\ell_p$-box techniques, e.g., $\ell_1$-box or $\ell_{\frac{1}{2}}$-box. Moreover, besides dual ascent method, there are many other options for solving the proposed nonlinear problem, we leave it a subject for future work to investigate the performance other existing nonlinear solvers for our proposed problem. \bibliographystyle{IEEEtran}

\section{Introduction} Reinforcement learning (RL) is the dominant class of algorithms to learn sequential decision-making from data. In RL we start with zero prior knowledge and need to actively collect our own data. Therefore, we should not settle on a policy too early, instead of trying out actions we have not properly explored yet. However, we neither want to continue exploring sub-optimal actions, when we already know what is best. This challenge is known as the exploration/exploitation trade-off. Most state-of-the-art reinforcement learning implementations use {\it undirected} forms of exploration, such as $\epsilon$-greedy or Boltzmann exploration. These methods act on {\it point estimates} of the mean action-value, usually applying some random perturbation to avoid only selecting the currently optimal action. However, these undirected methods are known to be highly inefficient \citep{osband2014generalization}. By only tracking point estimates of the mean state-action value, these algorithms lack the information to, for example, discriminate between an action that has never been tried before (and requires exploration) and an action that has been tried extensively and deemed sub-optimal (and can be avoided). A natural solution to this problem originates from tracking uncertainties/distributions. The intuition is that with limited data and large uncertainty there is reason to explore, while narrow distributions naturally transfer to exploitation (see Appendix \ref{illustration} for a detailed illustration). For this work we identify two types of uncertainties/distributions that are interesting for exploration: \begin{itemize} \item {\it Parametric uncertainty}: This is the classical statistical uncertainty which is a function of the number of available data points. The cardinal example is the posterior distribution of the mean (action-value). \item {\it Return uncertainty}: This is the distribution over returns from a state-action pair given the policy. For this work we focus on deterministic domains, which makes the return distribution entirely induced by the (exploratory) stochastic policy. \end{itemize} We argue that - for deterministic environments - we can explore by acting probabilistically optimal with respect to {\it both} distributions (see Section \ref{valuefunctions}). We identify neural network methods to estimate each of them separately, and subsequently show that both can be combined in one network, which we call the Double Uncertain Value Network (DUVN). To the best of our knowledge, we are the first to 1) distinguish between uncertainty due to limited data (parametric) and uncertainty over the return distribution, 2) propagate both through the Bellman equation, 3) track both with neural networks (i.e., high-capacity function approximators), and 4) use both to improve exploration.\footnote{As a side contribution, we introduce the Initial Return Entropy (IRE) as a measure of task exploration difficulty. See Appendix \ref{ire}.} The remainder of this paper is organized as follows. In Section 2 we provide a general introduction to Bayesian deep learning and distributional reinforcement learning. In Section 3, we discuss parametric and return uncertainty, and identify their potential for exploration. Section 4 discusses their implementations for policy evaluation with neural networks, and also discusses how to derive a policy from the learned distributions based on Thompson sampling. Sections 4, 5 and 6 show experimental results, discuss future work, and draw conclusions, respectively. \section{Pre-liminaries} \subsection{Bayesian deep learning} \label{bayesian} Bayesian neural networks \citep{mackay2003information} represent the uncertainty in the model through posterior distributions over the model parameters. Assume we observe some random variables $X$ and $Y$ and are interested in the conditional distribution $p(Y|X)$. We introduced a neural network $p_\phi(Y|X)$ with parameters $\phi \in \Phi$ to estimate this conditional distribution. In the Bayesian setting, we treat the model parameters $\phi$ as random variables themselves. Given an observed dataset $\mathcal{H}$, we may use the posterior distribution over model parameters $p(\phi|\mathcal{H})$ to obtain the posterior predictive distribution \begin{equation} p(y^\star|x^\star,\mathcal{H}) = \int p(y^\star|x^\star,\phi) p(\phi|\mathcal{H}) \mathrm{d}\phi \label{eq_postpred} \end{equation} for a new observed datapoint $x^\star$. In the non-linear neural networks of practical interest, the posterior distribution $p(\phi|\mathcal{H})$ is analytically intractable. \citet{gal2016dropout} showed that the well-known empirical procedure drop-out actually produces a Monte-Carlo approximation to Eq. \ref{eq_postpred}, providing samples from the posterior predictive distribution by simply retaining drop-out during test time (prediction). We use this technique in this paper, and discuss alternative methods for Bayesian inference in neural networks in the Future work section. \subsection{Distributional reinforcement learning} \label{distr_rl} In reinforcement learning (RL) \citep{sutton1998reinforcement} agents are studied that interact with an unknown environment with the goal to optimize some long-term performance measure. The framework adopts a Markov Decision Process (MDP) given by the tuple $\{\mathcal{S},\mathcal{A},\mathcal{T},\mathcal{R},\gamma\}$. At every time-step $t$ we observe a state $s_t \in \mathcal{S}$ and pick an action $a_t \in \mathcal{A} = \{ 1...N_\mathcal{A} \} $, for $N_\mathcal{A}$ available discrete actions. The MDP follows the transition dynamics $s_{t+1} = \mathcal{T}(\cdot|s_t,a_t) \in \mathcal{S}$ and returns rewards $r_t = \mathcal{R}(s_t,a_t) \in \mathbb{R}$. For this work, we assume a discrete action space and deterministic transition and reward functions. We act in the MDP according to a stochastic policy $\pi$, i.e. $a \sim \pi(\cdot|s) \in \mathcal{P}(\mathcal{A})$. The discounted return $Z^\pi(s,a)$ from a state-action pair $(s,a)$ is a {\it random process} given by \begin{equation} Z^\pi(s,a) = \sum_{t=0}^{\infty} \gamma^t r_t, \quad \quad s_{t+1} = \mathcal{T}(\cdot|s_t,a_t),a_{t+1} \sim \pi(\cdot|s_{t+1}),s_0=s,a_0=a \label{eq_return} \end{equation} for discount factor $\gamma \in [0,1]$. We emphasize that the return $Z^\pi$ is a random variable, where the distribution of $Z^\pi$ is induced by the stochastic policy (as we assume a deterministic environment). We may rewrite equation \ref{eq_return} into a recursive form, known as the {\it distributional Bellman equation} (omitting the $\pi$ superscript from now on): \begin{equation} Z(s,a) = r_t + \gamma Z(s',a'), \quad \quad s' = \mathcal{T}(\cdot|s,a),a' \sim \pi(\cdot|s'). \label{eq_distr_bellman} \end{equation} Note that the equality sign represents {\it distributional equality} here \citep{engel2005reinforcement}. We are now ready to define the action-value function. Denote by $\mathbb{E}_\pi$ the expectation over all traces induced by the policy $\pi$. Applying this operator to $Z(s,a)$ defines the state-action value $Q(s,a) = \mathbb{E}_\pi [Z(s,a)]$. Applying this operator to Eq. \ref{eq_return} gives \begin{equation} Q(s,a) = r_t + \gamma \mathbb{E}_{s' = \mathcal{T}(\cdot|s,a),a' \sim \pi(\cdot|s')} [Q(s',a') ] \label{eq_bellman} \end{equation} which is known as the Bellman equation \citep{sutton1998reinforcement}. Most RL papers actually start-off from Eq. \ref{eq_bellman}. We present the current introduction to emphasize that the mean action value $Q(s,a)$ is a quantity that we estimate by sampling from an underlying return distribution $p(Z|s,a)$.\footnote{We empirically observe that the shape of this return distribution strongly differs between domains. This matters because the shape of the return distribution also influences how easily we can estimate its expectation, or some other quantity like an upper confidence bound, from samples. For example, a long, thin right tail in the return distribution - as frequently the case in RL with only a few `good' traces - may give our mean estimate high variance (it would actually need importance sampling). In Appendix \ref{ire} we visualize return distributions for some well-known domains, and also introduce the {\it Initial Return Entropy} as a measure of task exploration difficulty.} \iffalse \begin{figure}[t] \centering \includegraphics[width = 1.0\textwidth]{ire_paper} \caption{\small Return distributions from the initial state in different environments for a random policy. Histogram produced over 50.000 traces of maximum 500 steps. The first 3 domains are directly taken from the OpenAI Gym. The Chain domain is introduced in Appendix \ref{chain}. Orange line is a kernel density estimate, with the vertical dashed line its empirical mean (a Monte Carlo estimate of $Q(s,a)$ under a uniform random policy). The top-right of each plot displays the initial return entropy (IRE) estimate of the domain, which we propose as a measure of exploration difficulty (see Appendix \ref{ire}).} \label{fig_ire_paper} \end{figure} \fi We approximate the action-value (distribution) with a deep neural network. We write $Q_\phi(s,a)$ for a network predicting a (point estimate) action-value, and $p_\phi(Z|s,a)$ for a network approximating the entire return distribution. To learn the state-action value RL algorithms follow variants of a scheme known as {\it generalized policy iteration} (GPI) \citep{sutton1998reinforcement}. GPI iterates between policy evaluation, in which we calculate a new estimates $\Psi(s,a)$ of the state-action value based on (new) sample data (e.g., for one-step SARSA $\Psi(s,a) = r(s,a) + \gamma Q(s',a')$), and policy improvement, in which we use the estimate $\Psi(s,a)$ to improve the policy (whether with a value-based, actor-critic or policy gradient algorithm). \section{Distributional perspective on exploration} \label{valuefunctions} We will now argue for a probabilistic perspective on value functions and exploration. There are two distributions that might be useful from an exploration point of view: 1) the statistical parametric uncertainty of the mean action value, and 2) the distribution of the return. \paragraph{Parametric uncertainty of the mean} Given a policy the state-action value $Q(s,a)$ is a scalar number by definition, as it is an expectation over all possible future traces. However, from a statistical point of view it makes sense to treat our {\it estimate} of $Q(s,a)$ as a random variable, as we need to approximate it from a finite number of samples. We call this the parametric uncertainty. \newline Parametric exploration, i.e. acting optimistic with respect to the uncertainty of the mean action-value, has been very successful in the bandit setting. However, it has been sparsely applied to RL (see Appendix \ref{relatedwork} for related work). We believe this is due to a fundamental complication regarding uncertainties in RL, which has only been identified by \citet{dearden1998bayesian} before. Bandits are one-step decision problems with pay-offs originating from a stationary distribution, which makes the value approximation an ordinary supervised learning problem. However, in RL the target distribution is highly non-stationary. A standard target like $r + \gamma Q(s',a')$ falsely assumes that $Q(s',a')$ is known, while it is actually uncertain itself. Therefore, repeatedly visiting a state-action pair should not makes us certain about its value if we are still uncertain about what to do next. In other words: {\it the state-action value certainty depends on the future policy certainty}. Standard parametric uncertainty cannot account for this problem (the `local' parametric uncertainty will converge as if it is supervised learning), and we somehow need to propagate the uncertainty of future state-action pairs' value (the `global' uncertainty) back through the Bellman equation. An illustration is seen in Fig. \ref{fig_duvn}b right, where the uncertainty in $\phi$ influences both the current and future value estimates (ignoring the $Z$ distributions in that graph for now, as the need to propagate the parametric uncertainty $p(\phi)$ over timesteps already applies when we learn mean action-values). \paragraph{Return distribution} Standard RL, and also the parametric uncertainty introduced above, usually deal with the {\it mean} action-value $Q(s,a)$. However, from an exploration point of view, it makes more sense to learn the full {\it return distribution} $p(Z|s,a)$. Note that we still focus on deterministic environments. Therefore, the distribution over returns is solely induced by our own policy. As we may modify our own policy, it makes sense to act optimistically with respect to the return distribution we observe. As an illustration, consider a state-action pair with particular mean action value estimate $Q(s,a)$. It really matters whether this average originates from a highly varying return, or from consistently the same return. It matters because our policy may {\it influence} the shape of this distribution, i.e. for the highly varying returns we may actively transform the distribution towards the good returns. In other words, what we really care about in deterministic domains is the best return, or the upper end of the return distribution.\footnote{For stochastic domains the return distribution has additional noise for which we do want to act on the expectation.} It turns out that both challenges focus around propagating either parametric uncertainty and/or return distributions through the Bellman equation (Fig \ref{fig_duvn}b). The overall idea is to memorize the propagating {\it global} MDP uncertainties in a neural network, which makes them {\it locally} available at action selection time. We thereby avoid the need for any forward planning (to get global information), and our approach is entirely model-free. \begin{figure}[t] \centering \includegraphics[width = 0.9\textwidth]{duvn.png} \caption{\small {\bf a)} Three types of neural networks with different uncertanties/probabilitiy distributions. Circles are probabilistic nodes. Left: parametric uncertainty over the mean action-value. Middle: propagating (return) distributions for point estimate parameters. Right: parametric uncertainty over propagating distribution (= Double Uncertain Value Network). {\bf b)} Illustration of propagating distributions. Subscripts identify unique state-action pairs. We initialize all state-action pairs with a prior parametric uncertainty $p(\phi)$ and prior output distribution $p_\phi(Z)$. Then, for a new observed transition, we want to update our estimates of $p_\phi(Z)$ at the current state action pair by propagating the distribution of the next node $p(Z')$ through the Bellman operator $T$ (instead of just propagating the mean). For this work, we consider two quantities to propagate: i) the return distribution at the next node $p_\phi(Z')$ (for point estimate $\phi$), or ii) the parametric uncertain return distribution at the next node $p(Z') = \int p_\phi(Z') p(\phi|\mathcal{H}) \mathrm{d} \phi$. Arrows point backwards because we focus on the direction of uncertainty propagation/back-up (which runs in the different direction than our exploration). } \label{fig_duvn} \end{figure} \section{Double Uncertain Value Networks} \label{seq_duvn} \subsection{Policy evaluation} We now discuss three probabilistic policy evaluation approaches that incorporate the uncertainties introduced in the previous section: 1) (local) parametric uncertainty only, 2) return distribution only, and 3) both combined. The respective network structures are illustrated in Fig. \ref{fig_duvn}a. Implementation details are provided in Appendix \ref{appendix_implementation}. \paragraph{Parametric uncertainty only} To estimate our parametric uncertainty we may use any type of Bayesian inference method suitable for neural networks. For this paper we consider the Bayesian dropout \citep{gal2016improving}, as it has a very simple practical implementation (see Sec. \ref{bayesian}). This gives us a sample from the posterior predictive distribution of the mean action value: $p(Q|s,a,\mathcal{H}) = \int Q_\phi(s,a) p(\phi|\mathcal{H}) \mathrm{d}\phi$. The associated network structure is visualized in Figure \ref{fig_duvn}a, left. \paragraph{Return distribution only} We next consider the problem of learning return distributions instead of mean action-values. For this work we will assume that the return distribution $p(Z|s,a)$ can be approximated by a Gaussian. Therefore, we modify our neural network to output the distribution parameters $\mu^Z(s,a)$ and $\sigma^Z(s,a)$, where clearly $\mu^Z(s,a) = Q(s,a)$. Note that the network parameters $\phi$ are point estimates now. The associated network structure is visualized in Figure \ref{fig_duvn}a, middle. \newline During policy evaluation we need to estimate distributional targets instead of point estimate targets. We will construct bootstrap estimators based on the distributional Bellman equation (Eq. \ref{eq_distr_bellman}). The derivation for the mean $\mu^Z(s,a) = Q(s,a)$ is well-known from standard RL, so we focus on propagating the return standard deviation through the distributional Bellman equation: \begin{align} \mathrm{Sd} \Big[ p(Z|s,a) \Big] &= \mathrm{Sd}\Big[ r(s,a) + \gamma \sum_{a' \in \mathcal{A}} \pi(a'|s') p(Z|s',a')\Big] = \gamma \sum_{a' \in \mathcal{A}} \pi(a'|s') \mathrm{Sd}\Big[p(Z|s',a')\Big] \end{align} because $\gamma \geq 0$, $\pi(a|s) \geq 0$, and we assume the next state distributions are independent so we may ignore the covariances.\footnote{For random variables $X,Y$ and scalar constants $a,b,c$ we have: $\mathrm{Var}[a + b X + c Y] = b^2 \hspace{0.1cm} \mathrm{Var}[X] + c^2 \hspace{0.1cm} \mathrm{Var}[Y] + 2bc \hspace{0.1cm} \mathrm{Cov}[X,Y]$.} We see that the standard deviation of $p(Z|s,a)$ is a linear combination of the standard deviations $p(Z|s',a')$ (one timestep ahead), reweighted by the policy probabilities and shrunken by $\gamma$. We approximate the sum over the policy probabilities $\pi(a'|s')$ by sampling from our policy (as is the usual solution in RL, which will be right in expectation over multiple traces). The network may then be trained to move the current predictions closer to these targets, for example with a squared loss \begin{equation} L(\phi) = \Big(r(s,a) + \gamma \mu_\phi^Z(s',a') - \mu_\phi^Z(s,a)\Big)^2 + \Big(\gamma \sigma_\phi^Z(s',a') - \sigma_\phi^Z(s,a)\Big)^2 \label{eq_loss} \end{equation} where we as usual fix the bootstrap predictions at $(s',a')$, i.e. the training gradients w.r.t. $\phi$ are blocked there. This approach can be seen as a form of analytic approximate return propagation with a (heuristic) distributional loss (see Appendix \ref{related_return} for other distributional losses). Similar ideas with approximate return propagation were recently explored with discrete network output distributions \citep{bellemare2017distributional}, which may also accommodate for propagating multimodality. A second, more simple propagation method which we also experimented with is sampling-based propagation. In that setting we sample $M$ values $z_m(s',a') \sim p_\phi(Z|s',a')$, push these through the Bellman operator to construct $\Psi_m(s,a) = r(s,a) + \gamma z_m(s',a')$, and train our network on this collection of samples with, e.g., a maximum likelihood loss. This may require more samples and be less accurate, but it will also work for complicated network output distributions (like deep generative models) for which analytic propagation and projection is infeasible. Results of this approach are not shown, but were comparable to the results with approximate return propagation shown in Section \ref{results}. \paragraph{Parametric uncertainty over return distributions} We finish with the observation that both ideas may actually naturally be combined in one function approximator (Fig \ref{fig_duvn}a, right). Note that we can now propagate both the return distribution and its parametric uncertainty at the next timestep, i.e. we are effectively propagating {\it uncertain return distributions} (the parametric uncertainty over the network output distribution). Starting from a sampled transition now, we want to propagate the return distribution weighted over the parametric uncertainty at the next timestep: \begin{equation} Z(s,a) = \int \Big[r + \gamma Z_\phi(s',a') \Big] p(\phi|\mathcal{H}) \mathrm{d} \phi = r + \gamma \int Z_\phi(s',a') p(\phi|\mathcal{H}) \mathrm{d} \phi \end{equation} Besides that, the same distributional Bellman propagating machinery applies as above.\footnote{Sample from $p(\phi|\mathcal{H})$ at the next time-step, make network predictions $\mu_\phi^Z(s',a')$ and $\sigma_\phi^Z(s',a')$, and do the Bellman propagation. Repeated sampling of $\phi$ does Monte Carlo integration over $p(\phi|\mathcal{H})$, as a numerical integration like in \citet{dearden1998bayesian} is infeasible for the neural network setting.} We refer to the general mechanism of uncertainty propagation (parametric, return or both) as {\it Bellman uncertainty}. \newline The appearance of the network, with both uncertainty over the network parameters $\phi$ and over the output distribution to track the propagating (uncertain) return distributions, makes us refer to it as the Double Uncertain Value Network (DUVN) (Fig \ref{fig_duvn}a, right). The intuition is that during early learning we will mostly be propagating uncertainty, while with converging distributions we will eventually start propagating true return distributions. In summary, we identified three types of probabilistic policy evaluation algorithms (with the three associated network structures visualized in Fig. \ref{fig_duvn}a):\footnote{We could think of another algorithm that does propagate (i.e., has a probabilistic network output), but only propagates the parametric uncertainty of the mean at the next time-step $p(Q'|\mathcal{H}) = \int Q'_\phi p(\phi|\mathcal{H}) \mathrm{d}\phi$ (and not the entire return distribution). We did not come up with such an algorithm, but concurrently with our work, \citet{o2017uncertainty} did focus on this problem. See Related Work.} \begin{enumerate} \item The (local) parametric uncertainty of the mean value: $p(Q|s,a,\mathcal{H}) = \int Q_\phi(s,a) p(\phi|\mathcal{H}) \mathrm{d}\phi$. \item The (propagating) distribution of the return: $p_\phi(Z|s,a,\mathcal{H})$ (with point parameters $\phi$). \item Both, (propagating) uncertain return distr.: $p(Z|s,a,\mathcal{H}) = \int p_\phi(Z|s,a,\mathcal{H}) p(\phi|\mathcal{H}) \mathrm{d}\phi$. \end{enumerate} \subsection{Policy improvement} We now describe how to use any of these distributions to naturally balance exploration versus exploitation, based on Thompson sampling \citep{thompson1933likelihood} (see Appendix \ref{illustration} as well). To generalize notation, we introduce a new random variable $\Theta$ with distribution $p(\Theta|s,a)$ to capture any of the three policy evaluation distributions introduced in the previous section. We write $p(\boldsymbol{\Theta}|s) = \prod_{a^\star \in \mathcal{A}} p(\Theta|s,a^\star)$ for the joint action-value distribution in a state $s$, where we assume the posterior distributions per action are independent. Thompson sampling selects action $a$ with probability equal to: \begin{equation} \pi(a|s) = \int p(\Theta_{a}>\Theta_{a^\star \ne a}) p(\boldsymbol{\Theta}|s) \mathrm{d} \boldsymbol{\Theta} \end{equation} where $\Theta_a = \Theta(s,a)$ and $\Theta_{a^\star \ne a}$ notational convention for $\Theta(s,a^\star) \forall a^\star \in \mathcal{A},a^\star \ne a$. In words, we choose action $a$ with probability equal to the probability that the specific action is the optimal one when averaging over all uncertainty in the joint distribution $p(\boldsymbol{\Theta}|s)$. The practical implementation of Thompson sampling is very simple, as we may just sample from $p(\Theta|s,a)$ for every $a$ and argmax over these values: \begin{enumerate} \item Sample $\phi \sim p(\phi)$ (or equivalently a dropout mask). \item Sample $Z(s,a^\star) \sim p_{\phi}(Z|s,a^\star) \quad \forall \quad a^\star \in \mathcal{A}$. \item Select $a = \argmax_{a^\star \in \mathcal{A}} Z(s,a^\star)$. \end{enumerate} If we do not consider parametric uncertainty, then we ignore the first sampling step and just use the current parameter point estimates. If we do not consider the Bellman uncertainty, then we replace the second sampling step with a deterministic prediction $Q_\phi(s,a)$. Thompson sampling is not the only possible choice to make decisions under uncertainty, but it has shown good empirical performance in the bandit literature \citep{chapelle2011empirical}. It naturally performs policy improvement, as it gradually starts to prefer the better actions when the distributions start narrowing/converging. We thereby hope to improve on the instability of greedy policy improvement (see also \citet{bellemare2017distributional}) or undirected exploration. Ideally, the uncertain return distribution would gradually narrow and for a deterministic environment eventually converge to a Dirac distribution on the optimal value function. \begin{figure}[t] \centering \includegraphics[width = 0.82\textwidth]{chain_results} \caption{\small Learning curves on Chain domain for Thompson sampling on parametric uncertainty, return distribution and uncertain return distribution versus $\epsilon$-greedy exploration ($\epsilon = 0.05$). Plots progress row-wise for increased depth of the Chain, i.e. increased exploration difficulty. Note that the correct action at each state in the chain is initialized at random (i.e. not always action 2, as in the visualization in Fig. \ref{chainfigure}). Results averaged over 5 repetitions.} \label{fig_results_chain} \end{figure} \section{Experiments} \label{results} We now evaluate the different types of probabilistic policy evaluation in combination with Thompson sampling exploration. We refer to Thompson sampling on the three types of discussed policy evaluation as {\it parametric exploration}, {\it return exploration}, and {\it uncertain return exploration}. Experimental details are provided in Appendix \ref{appendix_implementation}. We first consider the Chain domain (Appendix \ref{chain}, Figure \ref{chainfigure}). The domain consists of a chain of states of length $N$, with two available actions at each state. The only trace giving a positive, non-zero reward is to select the `correct' action at every step. The correct action per step is determined at domain initialization by sampling from a uniform Bernoulli. The domain has a strong exploration challenge, which grows exponentially with the length of the chain (see Appendix \ref{chain}). Learning curves for the Chain domain are shown in Fig. \ref{fig_results_chain}, for different lengths of the chain. First of all, we note that the $\epsilon$-greedy strategy does not learn in this domain at all (not even for the short length). The three probabilistic approaches do explore, with best performance for the uncertain return exploration. In the longest chain, of length 100, all probabilistic exploration methods also get trouble solving the domain. However, they do see some rewards, which makes us hypothesize this could be an issue of stabilization (more than that the exploration does not work). See Appendix \ref{extra_results} for results when the correct action is always the same, as in the original variants of this problem \citep{osband2016deep}. We next test our method on a set of tasks from the OpenAI Gym repository (Fig. \ref{fig_results_other}). We see that our exploration methods manage to learn on all domains. The achieved end policies all reflect good policies on each problem. $\epsilon$-greedy exploration is a bit unstable on some domains (CartPole, LunarLander), but generally performs reasonable as well. We note that the uncertainty exploration methods, which have a completely different exploration mechanism compared to $\epsilon$-greedy exploration, never really perform worse on these domains. We hypothesize these domains have too much structure and are not challenging enough to show the same exploration difference as seen for the Chain domain. Future work should address more challenging (high-dimensional) exploration problems. We also want to stress that probabilistic exploration will not always outperform undirected methods, especially not on domains with relatively simple exploration. Uncertainty methods will generally create a cautious agent, that first wants to properly verify all parts of the domain. In contrast, undirected exploration agents may exploit sooner, which can be beneficial in domains with non-deep exploration (i.e., with quick rewards). \begin{figure}[t] \centering \includegraphics[width = 0.82\textwidth]{combined_results} \caption{\small Learning curves for parametric exploration, return exploration and uncertain return exploration on different OpenAI Gym environments. Results averaged over 5 repetitions.} \label{fig_results_other} \end{figure} \section{Future work} \label{future} We identify several directions for future work: \begin{enumerate} \item {\bf Other types of Bayesian inference in neural networks} (for parametric uncertainty): we hypothesize that the Bayesian drop-out may be too unstable and tedious to tune, as we sometimes observed in our experiments as well. Potentially, different methods to approximate the posterior over the network parameters (e.g., \citet{welling2011bayesian}) may improve estimation of parametric uncertainty. \item {\bf More expressive output distributions} (for Bellman uncertainty propagation): for this work we only experimented with Gaussian distributions for propagation. Recently, \citet{bellemare2017distributional} studied return distribution propagation with categorical distributions, which more naturally accommodate for multi-modality (see Related Work as well). Another extension could involve more expressive continuous network distributions, e.g. based on conditional variational inference \citep{moerland2017learning}. \item {\bf Continuous action-spaces}: the current implementations only focussed on discrete action spaces, where Thompson sampling can easily be applied by maintaining a distribution per action and enumerating all actions for action selection. Extension to the continuous setting would require either directly propagating policy uncertainty, or learning a parametric policy whose distribution mimics the uncertainty in the value function. \item {\bf Stochastic environments}: this paper entirely focussed on domains with deterministic reward and transition functions, which makes the return distribution only induced by the stochastic policy. In stochastic domains the return distribution will have additional noise for which we do want to act on the expectation, to prevent continuing to act optimistically with respect to something we can't influence. \end{enumerate} Finally, we want to stress that the RL algorithms in this paper are entirely model-free. The uncertainty theme also appears in model-based RL, where it is useful/necessary to track the uncertainty on an estimated transition and/or reward function \citep{deisenroth2011pilco,depeweg2017uncertainty}. This parametric {\it model} uncertainty is different from the parametric {\it value/policy} uncertainty studied in this work, but our ideas may be extended to the model-based setting as well (which would add another source of uncertainty). \section{Conclusion} This paper introduced Double Uncertain Value Networks (DUVN), which, to the best of our knowledge, is the first algorithm that 1) distinguishes between uncertainty due to limited data (parametric) and uncertainty over the return distribution, 2) propagates both through the Bellman equation, 3) tracks both with neural networks (i.e., high-capacity function approximators), and 4) uses both to improve exploration. We implemented the DUVN algorithm with Bayesian dropout for the parametric uncertainty and a Gaussian distribution for the Bellman uncertainty propagation. The main appeal of this implementation is its simplicity: any deep Q-network implementation can be easily extended as in this work by adding drop-out to the neural network layers and specifying a Gaussian output distribution instead of a mean-only prediction. This should take no more then a few lines of code in most automatic differentiation software packages. We showed that, even for the vanilla implementation, we at least match or improve undirected exploration performance on a variety of problems, and drastically improve performance on an exploration heavy domain (Chain). We believe further improvements in the distributional approach to RL, e.g. with more expressive network output distributions that capture multi-modality, is a promising direction for RL exploration research. \clearpage {\footnotesize \bibliographystyle{apalike}

\section{Discussion and Future Work}\label{} In this ongoing work, we proposed a method based on black-box optimization and RL to deal with the problem of adapting speech-enhancement algorithms to varying input signal quality. Our work is related to hyperparameter optimization in deep learning and machine learning\footnote{Here, hyperparameters refer to training parameters such as learning rate, decay function or a network structure such as the number of layers, number of hidden units, type of layers \textit{etc.}}, which has been extensively studied in the literature. Methods like random search~\citep{Bergstra2012}, Bayesian optimization with probabilistic surrogates (\textit{e.g.}, Gaussian processes~\citep{Snoek2012, HenrandezLobato2014}) or deterministic surrogates (\textit{e.g.}, radial basis functions~\citep{ilievski2017efficient}) have been used to find the best setup for the model hyperparameters. However, once these methods find a set of parameters for a given model offline, the set typically remains fixed throughout the inference process. In contrast, our RL approach adaptively changes parameters of the underlying (data-driven or analytical) algorithm at inference time, achieving the best performance under all input signal conditions. Furthermore, in this particular paper, we demonstrated how to apply our dynamic parameter-adaptation technique to the problem of speech enhancement~\citep{Tashev2009}. To the best of our knowledge, black-box optimization using reinforcement learning for real-time application such as speech enhancement has not been conducted before, the previous work~\citep{chen17eLTLWGD} only studies a simple synthetic task. Based on experiments with real user data, we showed that our RL agent is very effective in changing algorithmic parameters at a frame level, enabling existing speech-enhancement algorithms to adapt to changing input signal quality and denoising performance. However, there are still hurdles that need to be overcome in the design of a reliable reward function that helps us achieve the best algorithmic performance across a diverse range of metrics including WER, SER and PESQ. We intend to address this challenge in future work, in addition to reducing the overhead of RL computation during training. \section{Experimental Results}\label{} We evaluated the performance of our methodology with single-channel recordings based on real user queries to the Microsoft Cortana Voice Assistant. We split studio-level clean recordings into training, validation and test sets comprising $7500$, $1500$ and $1500$ queries, respectively. Further, we mixed these clean recordings with noise data (collected from 25 different real-world environments), while accounting for distortions due to room characteristics and distances from the microphone. Thus, we convolved the resulting noisy recordings with specific room-impulse responses, and scaled them to achieve a wide input SNR range of 0-30 dB. Each (clean and noisy) query has on average more than $4500$ audio frames of spectral amplitudes, each lasting $16$ ms. We applied a Hann weighting window to the frames allowing accurate reconstruction with a $50\%$ overlap. These audio frames in the spectral domain formed the features for our algorithm. Since we utilized a $512$-point short-time Fourier Transform (STFT), each feature vector was a positive real number of dimensionality $256$. To train our network model, we employed a first-order stochastic gradient-based optimization method, Adam ~\citep{KingmaB14}, with a learning rate that was adjusted on the validation set. Furthermore, we used a single layer LSTM with $196$ hidden units and number of steps equal to the number of frame. We trained it with a batch size of $1$ given each of input file has different number of frames. To compute the reward function, we employed the negated mean-squared error (MSE) between the ground-truth clean signals $g_t$ and denoised input signals $\hat{g}_t$ as follows: \begin{equation} \label{} r_t = - \Vert g_t - \hat{g}_t \Vert^2_2. \end{equation} Further, to avoid instability during training, we normalized this reward function to lie between $[-1,1]$. Our underlying speech-enhancement algorithm~\citep{Tashev2009}, i.e black-box, was based on a generalization of the decision-directed approach, first defined in \citep{Ephraim:84}. To quantify the performance of speech enhancement, we employed the following metrics: \begin{itemize} \setlength{\itemsep}{-1pt} \item Signal-to-noise ratio (SNR) dB \item Log spectral distance (LSD) \item Mean squared error in time domain (MSE) \item Word error rate (WER) \item Sentence error rate (SER) \item Perceptual evaluation of speech quality (PESQ) \end{itemize} \begin{table}[t] \centering \begin{tabular}{cllllll} \toprule Method & {\sc SNR}(dB) & {\sc LSD} &{\sc MSE}& {\sc WER} & {\sc SER} & {\sc PESQ}\\ \hline \\ [-3pt] \textbf{Noisy Data} & $15.18$ & $23.07$ & $0.04399$ & $15.4$ & $25.07$ & $2.26$ \\ [3pt] \textbf{Baseline~\citep{Tashev2009}} & $18.82$ & $22.24$ & $0.03985$ & $14.77$ & $25.93$ & $2.40$ \\ [2pt] \textbf{Proposed (Unbiased Estimation)} & $\mathbf{26.16}$ & $\mathbf{21.48}$ & $\mathbf{0.03749}$ & $17.38$ & $31.87$ & $2.40$ \\ [2pt] \textbf{Proposed (Baselined Estimation)} & \multirow{1}{*}{$\mathbf{26.68}$} & \multirow{1}{*}{$\mathbf{21.12}$} & \multirow{1}{*}{$\mathbf{0.03756}$} & \multirow{1}{*}{$18.97$} & \multirow{1}{*}{$32.73$} & \multirow{1}{*}{$2.38$} \\ [3pt] \textbf{Clean Data} & $57.31$ & $1.01$ & $0.0$ & $2.19$ & $7.4$ & $4.48$ \\ \bottomrule \end{tabular} \vspace*{10pt} \caption{The proposed RL approach improves MSE, LSD and SNR with no algorithmic changes to the baseline speech-enhancement process, except frame-level adjustment of the control parameters.} \label{} \vspace*{-5pt} \end{table} A larger value is desirable for the first and last metrics, while a lower value is better for the rest. To compute the WER and SER, we employed a production-level automatic speech recognition (ASR) algorithm, whose acoustical model was trained separately on a different dataset that had similar statistics as our training examples. Thus, the ASR algorithm was not re-trained during our speech-enhancement experiments. Results of evaluating our model on the test data are shown in Table~\ref{}. In the baseline approach~\citep{Tashev2009}, we utilized a non-linear solver to find the set of algorithmic parameters that achieved the best score for a multi-variable function that equally weighted all of the above metrics. This unconstrained optimization was performed \textit{offline} once across the training data, which resulted in a parameter set that achieved the best trade-off for all metrics and feature vectors across the input SNR range. This parameter set was held constant when the speech-enhancement algorithm~\citep{Tashev2009} was applied to the test audio frames. However, in the proposed approaches (third and fourth rows in the table), the parameter set was adjusted depending on the action proposed by our RL meta-network. The third row corresponds to utilizing LSTM-based RL alone on top of the baseline speech-enhancement algorithm [also known as the unbiased estimator that utilizes the reward function of Eq.~\eqref{}]. While in the fourth row, we add an additional step of reducing the variance of gradient estimation by baselining Eq.~\eqref{}. From Table~\ref{}, we see that the proposed models show better performance on MSE, LSD and SNR, improving them by up to $16\%$, $4\%$, and $42\%$, respectively. However, they do not show improvement on the other metrics (WER, SER and PESQ). In fact, these results are expected because our RL network only employs a measure for signal distortion as the reward function [see Eq.~\eqref{}]. Thus, it is able to only optimize metrics that are related to this measure (\textit{i.e.}, MSE, SNR and LSD). The difficulty of including WER, SER and PESQ in the RL optimization process lies in the fact that these metrics do not provide a direct way of quantifying representation error. There is also no good signal-level proxy for them that can be computed with a low processing cost, which is necessary to train the RL algorithm in practical amounts of time. Thus, although our network already deals with a hard and complex optimization problem due to the black-box optimization and policy gradients, as part of future work, we are continuing to investigate different methods of incorporating WER, SER and PESQ functions into the RL reward signal. \section{Introduction}\label{} Noise-suppression algorithms for a single-channel of audio data employ machine-learning or statistical methods based on the amplitude of the short-term Fourier Transform of the input signal \citep{Ephraim:84, ephraim1995signal, boll1979suppression, xu2014experimental}. Although the approach we propose can be applied to the entire gamut of noise-suppression techniques, in this paper, we only illustrate its benefits with the classical algorithms for speech enhancement that are based on spectral restoration~\citep{Tashev2009}. Such algorithms typically comprise four components~\citep{Tashev2009}: (1) voice-activity detection, (2) noise-variance estimation, (3) suppression rule, and (4) signal amplification. The first two components help gather statistics on the target speech signal in the input audio, while the third and fourth components allow us to utilize these statistics to distill out the estimated speech signal. Despite these components being based on sound mathematical principles~\citep{Tashev2009}, their performance is directly influenced by a sizeable set of parameters such as those that control the gain, geometry weighting, estimator bias, voice- and noise-energy thresholds, and \textit{etc}. Consequently, the combined set of these parameters plays a critical role in achieving the best performance of the end-to-end speech-enhancement process. Needless to say, the numerical values of these parameters are heavily influenced by the input signal and noise characteristics. Furthermore, thanks to the complex interdependency between statistical models~\citep{Tashev2009}, there is no known best value for these parameters that works well across all levels of input signal quality. Therefore any offline optimization process, such as the simplex method~\citep{Nash20020}, is only a \textit{sub-optimal solution} that tends to achieve good performance across only a small range of instantaneous input signal-to-noise ratios (SNRs). This is the status quo that we intend to break. \section{Proposed Approach}\label{} We develop data-driven techniques that allow us to adjust control parameters of a classical speech-enhancement algorithm~\citep{Tashev2009} dynamically at a frame level, depending just on simple feedback from the underlying algorithm. Specifically, we rely on reinforcement-learning (RL) based on a network of long-short term memory (LSTM) cells \citep{Hochreiter}. We show that our method can achieve the best performance of speech enhancement across a broad range of input SNRs. \begin{figure}[tb] \centering \centerline{\includegraphics[width=0.7\textwidth]{figures/ourModel2}} \caption{Proposed architecture. $g_t$ and $f_t$ are clean and noisy input frames, while $s_t$ and $r_t$ are action and reward values at time step $t$, respectively. At each time step, our model utilizes $s_t$ and $f_t$ to find the best set of control parameters that maximize $r_t$ of the speech-enhancement black-box algorithm.} \label{} \end{figure} The neural-network model that we propose is shown in Fig.~\ref{}. As mentioned before, it treats the classical speech-enhancement algorithm~\citep{Tashev2009} as a black box. Suppose $s_t$ and $r_t$ represent the action proposed by our network model and the reward signal that is available to it from the algorithm at an instance of time $t$. We set up an objective function as follows: \begin{equation} \label{} J^\pi(\theta) = \mathbb{E}_{\pi_\theta} \bigg[\sum_{t=0}^T r_t(s_t)\bigg] \end{equation} where $\pi_{\theta}$ is the policy that our model tries to learn. The goal of this function is to maximize the expected reward over time. Thus, in order to solve Eq.~\eqref{}, we employ the REINFORCE algorithm \citep{Williams, ZarembaS15, RanzatoCAZ15, NIPS2014_5542,baVald2015,s2sIlya}. At each time step, our network picks an action at a given state of the model (i.e. given it policy $\pi_{\theta}$) that causes some change to the set of control parameters that are applied to the black box. In other words, each parameter of the speech-enhancement algorithm~\citep{Tashev2009} is mapped to an action. Thus, an action can result in an increase or decrease of the parameter value by a specific step size. It could also lead to no change in the parameter values. Note that within the black-box algorithm, we also utilize the clean signal $g_t$ in addition to the noisy frame at every time step. The reason we do this is because once the underlying speech-enhancement algorithm~\citep{Tashev2009} is done with the denoising process, it relies on the ground-truth clean signal $g_t$ to compute a score that represents the reward function for the RL agent. $g_t$ is not used within the black box in any other way.

\section{Introduction} Starbursts dominate the total luminosity of massive, compact galaxies known as HIIGx. The closely related giant extragalactic HII~regions (GEHR) also undergo massive bursts of star formation, but tend to be located predominantly at the periphery of late-type galaxies. In both environments, the ionized hydrogen is characterized by physically similar conditions (Melnick et al. 1987), producing optical spectra with strong Balmer $\mathrm{H}\alpha$ and $\mathrm{H}\beta$ emission lines that are indistinguishable between these two groups of sources (Searle \& Sargent 1972; Bergeron 1977; Terlevich \& Melnick 1981; Kunth \& \"{O}stlin 2000). Since both the number of ionizing photons and the turbulent velocity of the gas in these objects increase as the starburst becomes more massive, HIIGx and GEHR have been recognized as possible standard candles, a rather exciting prospect given that the very high starburst luminosity facilitates their detection up to a redshift $z\sim 3$ or higher (e.g., Melnick et al. 2000; Siegel et al. 2005). The exact cause of the correlation between the luminosity $L(\mathrm{H}\beta)$ in $\mathrm{H}\beta$ and the ionized gas velocity dispersion $\sigma$ is not yet fully understood, though an explanation may be found in the fact that the gas dynamics is almost certainly dominated by the gravitational potential of the ionizing star and its surrounding environment (Terlevich \& Melnick 1981). These sources may therefore function as standard candles because the scatter in the $L(\mathrm{H}\beta)$ versus $\sigma$ relation appears to be small enough for HIIGx and GEHRs to probe the cosmic distance scale independently of $z$ (Melnick et al. 1987; Melnick et al. 1988; Fuentes-Masip et al. 2000; Melnick et al. 2000; Bosch et al. 2002; Telles 2003; Siegel et al. 2005; Bordalo \& Telles 2011; Plionis et al. 2011; Mania \& Ratra 2012; Ch{\'a}vez et al. 2012, 2014; Terlevich et al. 2015). Over the past several decades, HIIGx and GEHRs have been used to measure the local Hubble constant $H_0$ (Melnick et al. 1988; Ch{\'a}vez et al. 2012), and to sample the expansion rate at intermediate redshifts (Melnick et al. 2000; Siegel et al. 2005). More recently, Plionis et al. (2011) and Terlevich et al. (2015) demonstrated that the $L(\mathrm{H}\beta)-\sigma$ correlation is a viable high-$z$ tracer, and used a compilation of 156 combined sources, including 24 GEHRs, 107 local HIIGx, and 25 high-$z$ HIIGx, to constrain the parameters in $\Lambda$CDM, producing results consistent with Type Ia SNe. Most recently, we (Wei et al. 2017) extended this very promising work even further by demonstrating that GEHRs and HIIGx may be utilized, not only to refine and confirm the parameters in the standard model but---perhaps more importantly---to compare and test the predictions of competing cosmologies, such as $\Lambda$CDM and the $R_{\rm h}=ct$ universe (Melia 2003, 2007, 2013a, 2016, 2017a; Melia \& Abdelqader 2009; Melia \& Shevchuk 2012). These two models have been examined critically using diverse sets of data, including high-$z$ quasars (e.g., Kauffmann \& Haehnelt 2000; Wyithe \& Loeb 2003; Melia 2013b, 2014; Melia \& McClintock 2015b), cosmic chronometers (e.g., Jimenez \& Loeb 2002; Simon et al. 2005; Melia \& Maier 2013; Melia \& McClintock 2015a), gamma-ray bursts (e.g., Dai et al. 2004; Ghirlanda et al. 2004; Wei et al. 2013), Type Ia supernovae (e.g., Perlmutter et al. 1998; Riess et al. 1998; Schmidt et al. 1998; Melia 2012; Wei et al. 2015b), and Type Ic superluminous supernovae (e.g., Inserra \& Smart 2014; Wei et al. 2015a). Their predictions have also been compared using the age measurements of passively evolving galaxies (e.g., Alcaniz \& Lima 1999; Lima \& Alcaniz 2000; Wei et al. 2015c). A more complete summary of these comparisons, now based on over 20 different types of observation, may be found in Table~1 of Melia (2017b). The application of HIIGx and GEHRs as standard candles has provided one of the more compelling outcomes of this comparative study involving $\Lambda$CDM and $R_{\rm h}=ct$ (Wei et al. 2016). Using the combined sample of Ch\'avez et al. (2014) and Terlevich et al. (2015), we constructed the Hubble diagram extending to redshifts $z\sim 3$, beyond the current reach of Type Ia SNe, and confirmed that the proposed correlation between $L(\mathrm{H}\beta)$ and $\sigma$ is a viable luminosity indicator in both models. This sample is already large enough to demonstrate that $R_{\rm h}=ct$ is favored over $\Lambda$CDM with a likelihood $\gtrsim 99\%$ versus only $\lesssim 1\%$, corresponding to a confidence level approaching $3\sigma$. These results, however, come with two important caveats, which partially motivate the complementary approach we are taking in this paper. Not surprisingly, the cosmological parameters are most sensitive to the high-$z$ data, so the constraints resulting from this work are heavily weighted by the high-$z$ sample of only 25 HIIGx. Given how sensitive the results are to the sub-sample of high-$z$ HIIGx data, one would want to increase the significance of this analysis by increasing the number of HIIGx-related measurements. Indeed, with the K-band Multi Object Spectrograph at the Very Large Telescope, a larger sample of high-$z$ HIIGx high-quality measurements may be available soon (Terlevich et al. 2015). The second caveat attached to the analysis of Wei et al. (2017) is that we do not yet have a full grasp of the systematic uncertainties in the $L(\mathrm{H}\beta)-\sigma$ correlation; these no doubt impact the use of HIIGx as cosmological probes. They include the burst size, its age, the oxygen abundance of HIIGx, and the internal extinction correction (Ch{\'a}vez et al. 2016). An example of a non-ignorable systematic uncertainty arises from the fact that the $L(H\beta)-\sigma$ relation correlates the ionizing flux from massive stars with random velocities in the potential well created by all the stars and the surrounding gas. Thus, any systematic variation in the initial mass function would alter the mass-luminosity ratio, and therefore also the zero point and slope of the relation (Ch\'avez et al. 2014). In spite of the fact that the high-$z$ sample of HIIGx is still relatively small, we can nonetheless further test the previous results by probing this compilation more deeply (than has been attempted before) using a two-point diagnostic, $\Delta\mu(z_i,z_j)$, defined in Equation~(9) below. Quite generally, two-point diagnostics such as this differ from parametric fitting approaches in several distinct ways. They facilitate the comparative analysis of measurements in a pairwise fashion. One may use them with $n$ measurements of a particular variable to generate $n(n-1)/2$ comparisons for each pair of data. The benefits are twofold: (1) one can test how well each pair of data fits the models, and (2) assess how closely the published error bars fit a normal distribution, thereby providing some indication of possible contamination by correlated systematic uncertainties. Zheng et al. (2016) recently used such an approach to conclude that the stated errors in cosmic chronometer data are strongly non-Gaussian, suggesting that the quoted measurement uncertainties are almost certainly not based exclusively on statistical randomness (see also Leaf \& Melia 2017). As we shall see, the diagnostic $\Delta$$\mu$$(z_i,z_j)$ is expected to be zero if the model being tested is the correct cosmology. To allow for possible non-Gaussianity in the published errors, we shall use both weighted-mean and median statistics to determine the degree to which each model's distribution of $\Delta$$\mu$$(z_i,z_j)$ values is consistent with this null result. So while Wei et al. (2016) optimized the overall $\Lambda$CDM and $R_{\rm h}=ct$ parametric fits to the HII~galaxy Hubble diagram, here we will test the consistency of each fit with individual pairs of data. We will begin with a brief description of the data in \S~2, and then define and apply the diagnostic $\Delta$$\mu$$(z_i,z_j)$ in \S~3. The outcome of our analysis will be discussed in \S~4, followed by our conclusions in \S~5. \section{Observational Data and Methodology} We base our analysis on the methodology described in Ch\'avez et al. (2012, 2014) and Terlevich et al. (2015), using their total sample of 156 sources, including 107 local HII~galaxies, 24 giant extragalactic HII~regions, and 25 high-$z$ HII~galaxies. The correlation between the emission-line luminosity and the ionized gas velocity dispersion may be written as (Ch\'avez et al. 2012; Ch\'avez et al. 2014; Terlevich et al. 2015) \begin{equation} \log L(\mathrm{H}\beta)=\alpha \log \sigma(\mathrm{H}\beta)+\kappa\;, \end{equation} where $\alpha$ is the slope and the constant $\kappa$ represents the logarithmic luminosity at $\log \sigma(\mathrm{H}\beta)=0$. As noted, previous applications of this relation have produced a very small scatter in the correlation for $L(\mathrm{H}\beta)$, making it a viable luminosity indicator for cosmology. But one cannot completely avoid its cosmology dependence because the $\mathrm{H}\beta$ luminosity is calculated using the expression \begin{equation} L(\mathrm{H}\beta) = 4 \pi D_L^2(z) F(\mathrm{H}\beta)\;, \end{equation} where $D_L$ is the model-dependent luminosity distance at redshift $z$ and $F(\mathrm{H}\beta)$ is the reddening corrected $\mathrm{H}\beta$ flux. From Equation~(1), we may then obtain the distance modulus of an HII~galaxy according to \begin{equation} \mu_{\rm obs}=2.5\left[\kappa +\alpha \log \sigma(\mathrm{H}\beta) - \log F(\mathrm{H}\beta)\right]-100.2 \;, \end{equation} with an associated error \begin{equation}\label{sigma} \sigma_{\mu_{\rm obs}}=2.5\left[\left(\alpha \sigma_{\log \sigma}\right)^{2}+ \left(\sigma_{\log F}\right)^{2}\right]^{1/2} \;, \end{equation} in terms of $\sigma_{\log \sigma}$ and $\sigma_{\log F}$, these being the $1\sigma$ uncertainties in $\log \sigma(\mathrm{H}\beta)$ and $\log F(\mathrm{H}\beta)$, respectively. This is to be compared with the theoretical distance modulus \begin{equation} \mu_{\rm th}\equiv5 \log\left[\frac{D_{L}(z)}{\rm Mpc}\right]+25\;, \end{equation} as a function of the cosmology-dependent luminosity distance $D_L$. \begin{table*} \begin{center} \caption{Parameters optimized via maximization of the likelihood function} \begin{tabular}{lccccc} &&&& \\ \hline\hline &&& \\ Model& $\alpha$ & $\delta$ & $\Omega_{\rm m}$ & $\Omega_{\rm de}$ & $w_{\rm de}$ \\ &&&& \\ \hline &&&& \\ $R_{\rm h}=ct$ &$ 4.78_{-0.09}^{+0.07}$ & $32.01_{-0.30}^{+0.32}$ & -- & -- & -- \\ &&&& \\ {\it Planck} $\Lambda$CDM & $4.86_{-0.08}^{+0.08}$ & $32.27_{-0.31}^{+0.22}$ & $0.3089$ & $1.0-\Omega_{\rm m}$ & $-1$ \\ &&&& \\ $\Lambda$CDM & $4.86_{-0.10}^{+0.09}$ & $32.27_{-0.36}^{+0.34}$ & $0.32_{-0.06}^{+0.09}$ & $1.0-\Omega_{\rm m}$ & $-1$ \\ &&&& \\ \hline\hline \end{tabular} \end{center} \end{table*} In $\Lambda$CDM, the luminosity distance may be written \begin{eqnarray}\label{DL_LCDM} \qquad D_{L}^{\Lambda {\rm CDM}}(z) &=& {c\over H_{0}}\,{(1+z)\over\sqrt{\mid\Omega_{k}\mid}}\times {\rm sinn}\Biggl\{\mid \Omega_{k}\mid^{1/2}\times\cr \null &\null&\hskip-0.6in \int_{0}^{z}{dz\over\sqrt{\Omega_{\rm m}(1+z)^{3}+\Omega_{k}(1+z)^{2}+ \Omega_{\rm de}(1+z)^{3(1+w_{\rm de})}}}\Biggr\}\;, \end{eqnarray} where $p_{\rm de}=w_{\rm de}\rho_{\rm de}$ is the dark-energy equation of state; radiation is ignored in the local Universe. Also, $\Omega_i\equiv \rho_i/\rho_c$, for matter (m), radiation (r) and dark energy (de), while $\Omega_{k}=1-\Omega_{\rm m}-\Omega_{\rm de}$ incorporates the spatial curvature of the Universe, and sinn is $\sinh$ when $\Omega_{k}>0$ and $\sin$ when $\Omega_{k}<0$. Today's critical density is $\rho_c\equiv 3c^2 H_0^2/8\pi G$. Since we are here assuming a flat Universe, (i.e., $\Omega_{k}=0$), the right side of this equation becomes $(1+z)c/H_{0}$ times the integral. For the $R_{\rm h}=ct$ cosmology (Melia 2003, 2007, 2013a, 2016a, 2016b; Melia \& Abdelqader 2009; Melia \& Shevchuk 2012), the luminosity distance is given by the much simpler expression \begin{equation}\label{DL_Rh} D_L^{R_{\rm h}=ct}(z)={c\over H_0}(1+z)\ln(1+z)\;. \end{equation} Here we follow Wei et al's. (2015b,2016) approach and circumvent circularity issues by optimizing the coefficients $\alpha$ and $\kappa$ individually for each model, via maximization of the likelihood function. With this approach, $H_{0}$ and $\kappa$ are not independent of each other; one may vary either $H_{0}$ or $\kappa$, but not both. For the purpose of maximizing the likelihood function, it is therefore useful to define a combined parameter, \begin{equation} \delta\equiv -2.5\kappa-5\log H_0+125.2\;, \end{equation} where $\delta$ is the ``$H_0$-free" logarithmic luminosity and the Hubble constant $H_0$ is in units of km $\rm s^{-1}$ $\rm Mpc^{-1}$. The constants $\alpha$ and $\delta$ are statistical ``nuisance" parameters, analogous to the adjustable coefficients characterizing the lightcurve in Type Ia SNe. The best-fit parameters obtained in this fashion are shown in Table~1, for three models we will compare: {\it Planck} $\Lambda$CDM (Planck Collaboration 2016), $\Lambda$CDM with a re-optimized matter density $\Omega_{\rm m}$, and the $R_{\rm h}=ct$ universe. A quick inspection of Equations~(3) and (5) shows that the two-point diagnostic \begin{eqnarray} \qquad{\Delta}{\mu}{(z_i,z_j)}&\equiv&\dfrac{-{\delta}+2.5{\alpha}\log\sigma_i-2.5\log F_i}{5\log \left[\dfrac{D_L(z_i)}{1\;{\rm Mpc}}\right]}-\cr \null&\null&\qquad\dfrac{-{\delta}+2.5{\alpha}\log\sigma_j- 2.5\log F_j}{5\log\left[\dfrac{D_L(z_j)}{1\;{\rm Mpc}}\right]} \end{eqnarray} is expected to be zero for any pair of HII~data at redshifts $z_i$ and $z_j$ if the cosmology used to calculate $D_L$ is correct. As one can see, the value of $H_0$ does not affect this constraint and is absorbed into the optimized coefficient $\delta$. For the sake of normalizing the various quantities, however, we simply use the {\it Planck} value $67.74$ km s$^{-1}$ Mpc$^{-1}$ throughout this analysis. Notice in passing that $\alpha$ and $\delta$ are similar between the different cosmologies, varying between them by $\lesssim 4\%$, i.e., well within $1\sigma$. Thus, since $H_0$ is also not a factor in $\Delta\mu(z_i,z_j)$, Equation~(9) represents a powerful diagnostic for comparing the viability of different models. The application of this 2-point diagnostic will be described in the next section. Finally, to improve the statistics even further, we have removed seventeen points (including one GEHR source at z=000001) from our complete sample whose measurement places them more than $3\sigma$ away from the best fit curves. We have also chosen to remove the other GEHR source at z=0.00001. While this point is only $~2\sigma$ from the best fit curve, it is the lowest-redshift measurement in the catalog, which, by the nature of 2-point diagnostics, causes it to drastically alter the statistical results. These anomalous points are identical for all three models, so their removal does not bias either of them. The final reduced sample therefore contains 138 measurements that are used to determine the best fits reported in Table 1. The eighteen eliminated sources are the two GEHRs at z=0.00001, and J162152+151855, J132347-013252, J211527-075951, J002339-094848, J094000+203122, J142342+225728, J094252+354725,\break J094254+340411, J001647-104742, J002425+140410, J103509+094516, J003218+150014, J105032+153806, WISP173-205, J084000+180531, and Q2343-BM133. \section{Application of the Two-point Diagnostic} As discussed in more detail in Leaf \& Melia (2017), the use of two-point diagnostics necessitates special care when analyzing the statistics they produce. First, the weighted mean of all $n(n-1)/2$ $\Delta$$\mu(z_i,z_j)$ values may be calculated using the expression \begin{equation} \mu=\dfrac{{\Sigma_{i=1}^{n-1}{\Sigma}_{j=i+1}^{n}{\Delta}\mu(z_i,z_j)/ {\sigma}_{{\Delta}\mu_{ij}}^2}}{{{\Sigma}_{i=1}^{n-1}\Sigma_{j=i+1}^{n}1/{\sigma}_{{\Delta\mu_{i,j}}}^2}}\;, \end{equation} in which ${\sigma}_{{\Delta\mu_{i,j}}}$ is the error for a single application of Equation~(9), found using standard error propagation. The error in the mean, however, must be calculated by carefully considering the correlation introduced from the repeated use of individual points in different pairs. For this purpose, we rewrite the weighted mean in the equivalent form \begin{equation} \mu=\dfrac{\Sigma_{i=1}^{n}{\beta_i}M(z_i)}{{\Sigma}_{i=1}^{n-1} \Sigma_{j=i+1}^{n}1/{\sigma}_{{\Delta\mu_{i,j}}}^2}\;, \end{equation} with each $\beta$ value given by the expression \begin{equation} \beta_i=\Sigma_{j=1}^{i-1}\dfrac{1}{\sigma_{\Delta\mu_{i,j}}^2}-\Sigma_{k=1+i}^{N}\dfrac{1} {\sigma_{\Delta\mu_{i,k}}^2}\;. \end{equation} In addition, we have defined the quantity \begin{equation} M(z_i)=\dfrac{-{\delta}+2.5{\alpha}\log\sigma_i-2.5\log F_i}{5\log \left[\dfrac{D_L(z_i)}{1\;{\rm Mpc}}\right]}\;. \end{equation} With the values of $\beta$ thus calculated, the variance then follows and is given as \begin{equation} {\Delta}{\sigma^2}_{\rm w.m.}=\dfrac{\Sigma_{i=1}^{n}{\beta^2_i} \sigma_{M}^2(z_i)}{({\Sigma}_{i=1}^{n-1}\Sigma_{j=i+1}^{n}1/{\sigma}_{\Delta\mu_{i,j}}^2)^2}\;. \end{equation} Knowing the standard deviation of the mean, we now have a measure of the consistency of the measurements with a given model. In the case of the $\Delta\mu$ diagnostic, we expect the weighted mean to be statistically consistent with zero if the applied model is the correct cosmology. Note that we do not introduce the errors in the fitted parameters in this analysis. This is due to the error affecting both halves of the two-point diagnostics in a very similar manner. That is, if the value of $\alpha$ is slightly too low, it would have the effect of reducing both `single-points', the net effect of which ends up being statistically insignificant. When non-Gaussian errors are suspected, however, such situations motivate the use of `median statistics,' pioneered by Gott et al. (2001), in which error propagation is neither required nor assumed. This approach takes advantage of the fact that for any measurement based on some distribution function, there is a $50\%$ chance of it being above the true median of the underlying distribution, without any need to know its form. Thus, for $N$ ranked measurements, the true median has a probability \begin{equation} P_i=\dfrac{2^{-N}N!}{i!(N-i)!} \end{equation} (i.e., the binomial distribution) of being found between measurements $i$ and $i+1$. One can use this to construct confidence regions about the median of the data, analogous to a standard deviation in Gaussian statistics, and assign to them a formal probability of finding the true median of the underlying distribution. However, it would be incorrect to apply this to, say all $n(n-1)/2$ diagnostics at once, for the same reasons noted in Leaf \& Melia (2017). The fact that each measurement contributes to N-1 diagnostics means that the data are correlated; as a result, a single measurement can move the median farther than in the case where the 2-point values are truly randomly distributed. We propose a remedy that takes advantage of the binomial properties of the median, but instead of considering all the diagnostics simultaneously, we construct a random sub-sample, in which each realization of the diagnostic is used exactly once, except for the one that was omitted. Therefore none of the diagnostic values is used more than once, completely avoiding any possible correlation. Following this, we record the median of the diagnostics of this uncorrelated sample, as well as the standard deviation of the realization. Next, we generate a large number (here, 1 million) of these realizations, and report the overall median of all the individual medians in Table 2. \begin{table*} \begin{center} \caption{Statistical analysis of the 2-point diagnostic $\Delta\mu(z_i,z_j)$} \begin{tabular}{lccccccc} &&&&&& \\ \hline\hline &&&&&& \\ Model& {\rm Weighted mean} & $1\sigma$ {\rm Error} & $|$Mean$|$ / $\sigma$ &$|N_\sigma|<1$ & {\rm Median} & Std. Dev. of the Median & $|$Median$|$ / Std. Dev.\\ &&&&&& \\ \hline &&&&&& \\ $R_{\rm h}=ct$ &$-0.00242$ & $0.00218$& $1.11$ & $51.3\%$ & $-0.00425$ & $0.00336 $ & $1.26$ \\ &&&&& \\ {\it Planck} $\Lambda$CDM & $-0.00340$ & $0.00221$& $1.54$ & $52.3\%$ & $-0.00483$ & $0.00363$ & $1.33$\\ &&&&& \\ $\Lambda$CDM & $-0.00330$ & $0.00220$& $1.50$ & $52.2\%$ & $-0.00476$ & $0.00342$ & $1.39$\\ &&&&&& \\ \hline\hline \end{tabular} \end{center} \end{table*} \begin{figure} \begin{center} \includegraphics[width=1.0\linewidth]{fig1.eps} \end{center} \caption{Unweighted histogram of all 9,453 $\Delta\mu$ diagnostic values for the $R_{\rm h}=ct$ universe (see Eq.~9). The $y$-axis gives the number of diagnostic values per bin.} \end{figure} \begin{figure} \begin{center} \includegraphics[width=1.0\linewidth]{fig2.eps} \end{center} \caption{Same as fig.~1, except now for {\it Planck} $\Lambda$CDM.} \end{figure} \begin{figure} \begin{center} \includegraphics[width=1.0\linewidth]{fig3.eps} \end{center} \caption{Same as fig.~1, except now for $\Lambda$CDM with a re-optimized value of $\Omega_{\rm m}$ (see Table~1).} \end{figure} In table 2, we also report the standard deviation of the median. This value is different from the overall standard deviation of the set of all 1 million medians. It is fundamentally related to the error in the mean of any set of data, in that it is some distinct factor smaller than the standard deviation of the data, dependent on the size of the data set. However, the exact relationship that exists between the standard deviation of the medians and the number of sources used to determine the median of all the realizations is not empirically known. In order to address this deficiency, we have have used the following approach, based on Monte-Carlo simulations with mock data to find this relationship to reasonable accuracy. We construct a mock data set by drawing at random from some probability distribution function, with the same number (i.e., 138) of points as in the real data set. Then, we construct a random set of 2-point diagnostics following the same method used with the real data. We record the median and standard eviation of the realization, repeating this process a sufficiently large number of times (say, $20,000$). Then, we repeat the process with a new random set of mock data drawn from the same distribution, and repeat this 5,000 times. Next, we determine the standard deviation of the set of 5,000 medians, as well as the mean of the 5,000 standard deviations. Finally, we compare the actual standard deviation of the median of all realizations with the mean of the standard deviations of each realization. We run this simulation with three different probability density functions---a normal distribution, a skew normal distribution with shape parameter $\alpha$=4, and a flat distribution over an interval. In all three cases the relationship between the standard deviation of the median and the mean standard deviation of each realization is found to be statistically consistent, and apparently dependent only on the number of sources chosen. For a sample of 138, the multiplicative factor is 1.822, always yielding a standard deviation of the medians smaller than the mean of the standard deviations by this factor. The values reported in Table 2 for the standard deviation of the median are therefore determined by taking the standard deviations of the million medians and dividing them by the corresponding factor. While this does technically include an implicit assumption that all data are sampled from a single underlying statistical distribution, we argue that by focusing on the median of these (instead of the mean), and the fact that there must certainly exist a single true cosmological model, this assumption is reasonable. The two-point diagnostic we have introduced in Equation~(9) is expected to be zero for the correct cosmology. The degree by which a given model's median is consistent with zero is therefore a measure of its consistency with the observations. We discuss the results of this analysis in the next section. \section{Discussion} In Table~2 and figs.~1-6 we report the results of both our weighted-mean and median statistical analyses, described in \S\S~2 and 3 above. One of the principal benefits of 2-point diagnostics constructed with regard to redshift ordering lies not only in determining how well a set of data fits a model, as revealed, e.g., with the use of information criteria but, also in providing insight into whether or not the low-$z$ sources are consistent with the same model as that preferred by the higher-$z$ sources. Our complete sample of 138 sources constitutes the original 156 minus the 18 outliers, as detailed in \S~2. As one can see from Table 1, the optimized value of $\alpha$ is about $4.8$ in every case, statistically consistent with the results of previous analyses by Ch\'avez et al. (2012, 2014), Terlevich et al. (2015), and Wei et al. (2016). For these 138 measurements, we constructed for each model the 9453 unique 2-point diagnostics and calculated the weighted mean and corresponding 1-$\sigma$ error based on the reported uncertainties (see figs.~1-3 for the complete unweighted histograms). For the $R_h=ct$ universe (fig.~4), the weighted mean is found to be consistent with zero at about $1\sigma$. There is mild tension for {\it Planck }$\Lambda$CDM (fig.~5) and the best-fit $\Lambda$CDM cosmology (fig.~6), however, in that the weighted mean is inconsistent with zero at about $1.5\sigma$ (compare the entries in columns 2, 3 and 4 of Table 2). Perhaps more importantly, fewer than the expected $68.3\%$ of the diagnostics lie within $1\sigma$ of the weighted mean (column 5) for all three models, implying that the reported errors are probably not purely Gaussian and that there may be an additional source of error not accounted for in this analysis. It is therefore helpful to circumvent this possible non-Gaussianity by also analyzing the 2-point diagnostics using median statistics, as described above. With this approach, the three models show a similar inconsistency with a zero median (columns 6, 7 and 8 of Table 2), with a negative value in every case, roughly $1.3\sigma$ different from zero. The fact that both the weighted mean and the median are negative for all the models suggests that the luminosity distance at low-$z$ is generally greater than that predicted by these cosmologies, or that it is smaller than expected at high-$z$. The implication is that either (i) none of the models are completely correct, or (ii) there may be some systematic problems with the data at high-$z$ or (more likely) at low-$z$. Thus, while a discrepancy smaller than $2\sigma$ may not be definitive, it nonetheless motivates further analysis involving a possible contamination by non-Gaussian systematic errors. Along these lines, we point out that some authors have speculated on the possibility that a local ``Hubble bubble" (Shi 1997; Keenan et al. 2013; Romano 2016) might be influencing the local dynamics within a distance $\sim 300$ Mpc (i.e., $z\lesssim 0.07$). If true, such a fluctuation might lead to anomalous velocities within this region, causing the nearby expansion to deviate somewhat from a pure Hubble flow. This effect could be the reason we are seeing a slight negative bias for the weighted mean and median of the 2-point diagnostic for every model, since nearby velocities would be slightly larger than Hubble, implying larger than expected luminosity distances at redshifts smaller than $\sim 0.07$. In addition, the existence of local peculiar velocities would imply that the errors associated with low-$z$ measurements should be bigger than quoted, increasing the number of 2-point diagnostics that fall within $1\sigma$ of the expected dispersion, possibly `filling' the distributions in figs.~1-3 sufficiently to produce entries in column 5 of Table~2 closer to the value ($\sim 68.3\%$) expected of a true Gaussian distribution. \begin{figure} \begin{center} \includegraphics[width=1.0\linewidth]{fig4.eps} \end{center} \caption{Histogram of the medians found in one million random realizations of the 2-point diagnostic for the $R_{\rm h}=ct$ universe. The $y$-axis denotes the number of times ($\times 1000$) that the median of a realization falls within the range given on the $x$-axis.} \end{figure} \begin{figure} \begin{center} \includegraphics[width=1.0\linewidth]{fig5.eps} \end{center} \caption{Same as figure~4, except for {\it Planck} $\Lambda$CDM.} \end{figure} \begin{figure} \begin{center} \includegraphics[width=1.0\linewidth]{fig6.eps} \end{center} \caption{Same as figure~4, except for $\Lambda$CDM with a re-optimized $\Omega_{\rm m}$, as indicated in Table~1.} \end{figure} \section{Conclusions} The totality of the results shown in Tables~1 and 2, and illustrated in figs.~1-6, supports the use of HIIGx and GEHR sources as standard candles for cosmological testing, though the analysis based on 2-point diagnostics has probed the measurement errors in greater detail than was possible solely via parametric fits to the data, the subject of our previous paper on this subject (Wei et al. 2016). In this paper, we have proposed a new 2-point diagnostic for analyzing HIIGx and GEHR data with the inclusion of median statistics, which circumvents the need for assuming Gaussian errors in the measurements. This approach may be used alongside, and compared, with the better understood weighted mean method. We have shown that these two types of analysis give generally consistent results, insofar as the HII data are concerned. Broadly speaking, one of the principal conclusions of this analysis is that employing the entire compilation of HIIGx and GEHR sources (with the exception of several outliers) produces slight tension between the cosmological parameters favoured by the data at low and high redshifts. We believe this is circumstantial evidence in support of the proposal by Shi (1997), Keenan et al. (2013) and Romano (2016) of a dynamical influence due to a local Hubble bubble extending out to $z\sim 0.07$, which produces local peculiar velocities comparable to those in the Hubble flow at low redshifts. Nonetheless, probing the HIIGx and GEHR data with 2-point diagnostics has not changed the essential conclusions drawn by Wei et al. (2016), whose cosmological tests based on these sources favoured the $R_{\rm h}=ct$ model over $\Lambda$CDM. Our comparison using the HII sample has shown that $R_{\rm h}=ct$ is favoured over both {\it Planck} $\Lambda$CDM and $\Lambda$CDM with a variable $\Omega_{\rm m}$, at least when viewed in terms of weigthed mean statistics. The caveat, however, is that an approach based on median statistics produces less differentiation between the three models. In addition, we have found in all cases that our 2-point diagnostic with the weighted mean approach yields fewer values within individual $1\sigma$ error regions than the $68.3\%$ required of a true Gaussian distribution. This may be an indication that the reported errors are not purely statistical, which may happen, e.g., when the uncertainties are contaminated by systematic effects, including at least a partially non-Gaussian component, or when there is an additional source of uncertainty, other than what we considered in this analysis. \section*{Acknowledgments} We are very grateful to the anonymous referee for providing valuable insights and suggested improvements to the manuscript. FM is supported by Chinese Academy of Sciences Visiting Professorships for Senior International Scientists under grant 2012T1J0011, and the Chinese State Administration of Foreign Experts Affairs under grant GDJ20120491013.

\section{Introduction} \label{Section 1} Quasinormal modes of rapidly rotating isolated and accreting compact stars act as sensitive probes for general relativistic effects such as gravitational waves and also of the properties of ultradense matter. Temporal changes in the rotational period of neutron stars (NSs) can reveal the internal changes of the stars with time. Gravitational waves from rotational instabilities of pulsars can provide an insight about the nature of the high density Equation of State (EoS). Detecting these waves by LIGO and VIRGO will provide a new pathway in the field of asteroseismology. Rotational instabilities in NSs come in different flavours, but they have one general feature in common: they can be directly associated with unstable modes of oscillation \cite{Andersson1998,Andersson2003,Freidman1998,Provost1981,Andersson2001,Bondarescu2007}. In the present work the r-mode instability has been discussed with reference to the EoS obtained using the density dependent M3Y (DDM3Y) effective nucleon-nucleon (NN) interaction. The discovery of r-mode oscillation in neutron star (NS) by Anderson \cite{Andersson1998} and confirmed by Friedman and Morsink \cite{Freidman1998} opened the window for study of the gravitational wave emitted by NSs by using advance detecting system. Also it provides the possible explanation for the spin down mechanism in the hot young NSs as well as in spin up cold old accreting NSs. The r-mode oscillation is analogous to Rossby wave in the ocean and results from perturbation in velocity field of the star with little disturbance in the star's density. In a non-rotating star the r-modes are neutral rotational motions. In a rotating star Coriolis effects provide a weak restoring force that gives them genuine dynamics. The r-mode frequency always has different signs in the inertial and rotating frames. That is, although the modes appear retrograde in the rotating system, an observer in the inertial frame shall view them as prograde. To the leading order, the pattern speed of the mode is \cite{Andersson1999,Papaloizou1978} \begin{equation} \sigma=\frac{\left(l-1\right)\left(l+2\right)}{l\left(l+1\right)} \Omega \label{eqn1} \end{equation} Since, $0<\sigma<\Omega$ for all $l\geq 2$, where $\Omega$ is the angular velocity of the star in the inertial frame, the r-modes are destabilized by the standard Chandrasekhar-Friedman-Schutz (CFS) mechanism and are unstable because of the emission of gravitational waves. The gravitational radiation that the r-modes emit comes from their time-dependent mass currents. This is the gravitational analogue of magnetic monopole radiation. The quadrupole $l=2$ r-mode is more strongly unstable to gravitational radiation than any other mode in neutron stars. Further, these modes exist with velocity perturbation if and only if $l=m$ mode \cite{Provost1981,Andersson1999}. This emission in gravitational waves causes a growth in the mode energy $E_{rot}$ in the rotating frame, despite decrease in the inertial-frame energy $E_{inertial}$. This puzzling effect can be understood from the relation between the two energies, \begin{equation} E_{rot}=E_{inertial}-\Omega J \label{eqn2} \end{equation} where the angular momentum of the star is J. From this it is clear that $E_{rot}$ may increase if both $E_{inertial}$ and J decrease. The frequencies of these r-modes, in the lowest order terms in an expansion in terms of angular velocity $\Omega$ is \cite{Papaloizou1978,Lindblom1998} \begin{equation} \omega=-\frac{\left(l-1\right)\left(l+2\right)}{l+1} \Omega. \label{eqn3} \end{equation} The instability in the mode grows because of gravitational wave emission which is opposed by the viscosity \cite{Lindblom1987}. For the instability to be relevant, it must grow fast than it is damped out by the viscosity. So the time scale for gravitationally driven instability needs to be sufficiently short to the viscous damping time scale. The amplitude of r-modes evolves with time dependence $ e^{i\omega t - t/\tau } $ as a consequence of ordinary hydrodynamics and the influence of the various dissipative processes. The imaginary part of the frequency $1/\tau$ is determined by the effects of gravitational radiation, viscosity, etc \cite{Lindblom2000,Lindblom1998,Owen1998}. The time-scale associated with the different process involve the actual physical parameters of the neutron star. In computing these physical parameters the role of nuclear physics comes into picture, where one gets a platform to constrain the uncertainties existing in the nuclear EoS. The present knowledge on nuclear EoS under highly isospin asymmetric dense situation is quite uncertain. So correlating the predictions of the EoSs obtained under systematic variation of the physical properties, to the r-mode observables can be of help in constraining the uncertainity associated with the EoS. \noindent \section{Dissipative time scales and stability of the r-modes} \label{Section 2} The concern here is to study the evolution of the r-modes due to the competition of gravitational radiation and dissipative influence of viscosity. For this purpose it is necessary to consider the effects of radiation on the evolution of mode energy. This is expressed as the integral of the fluid perturbation \cite{Lindblom1998,Lindblom1999}, \begin{equation} \widetilde{E}=\frac{1}{2}\int{\left[ \rho \delta \vec{v}.\delta \vec{v}^{*}+\left(\frac{\delta p}{\rho}-\delta \Phi \right)\delta \rho^{*}\right]}d^{3}r, \label{eq4} \end{equation} with $\rho$ being the mass density profile of the star, $\delta \vec{v}$, $\delta p$, $\delta \Phi$ and $\delta \rho $ are perturbations in the velocity, pressure, gravitational potential and density due to oscillation of the mode respectively. The dissipative time scale of an r-mode is \cite{Lindblom1998}, \begin{equation} \frac{1}{\tau_{i}}=-\frac{1}{2\widetilde{E}}\left(\frac{d\widetilde{E}}{dt}\right)_{i}, \label{eq5} \end{equation} where the index `$i$' refers to the various dissipative mechanisms, i.e., gravitational wave emissions and viscosity (bulk and shear). For the lowest order expressions for the r-mode $\delta \vec{v}$ and $\delta \rho$ the expression for energy of the mode in Eq.(4) can be reduced to a one-dimensional integral \cite{Lindblom1998,Vidana2012} \begin{equation} \widetilde{E}=\frac{1}{2}\alpha_r^{2} R^{-2l+2} \Omega^{2} \int^{R}_{0} \rho(r) r^{2l+2} dr, \label{eq6} \end{equation} where R is the radius of the NS, $\alpha_r$ is the dimensionless amplitude of the mode, $\Omega$ is the angular velocity of the NS and $\rho(r)$ is the radial dependance of the mass density of NS. Since the expression of $(\frac{d\widetilde{E}}{dt})$ due to gravitational radiation \cite{Thorne1980,Owen1998} and viscosity \cite {Lindblom1991,Owen1998,Lindblom2000} are known, Eq.(5) can be used to evaluate the imaginary part $\frac{1}{\tau}$. It is convenient to decompose $\frac{1}{\tau}$ as \begin{equation} \frac{1}{\tau(\Omega,T)}=\frac{1}{\tau_{GR}(\Omega,T)}+\frac{1}{\tau_{BV}(\Omega,T)}+\frac{1}{\tau_{SV}(\Omega,T)}, \label{eq7} \end{equation} where $1/\tau_{GR}$, $1/\tau_{BV}$ and $1/\tau_{SV}$ are the contributions from gravitational radiation, bulk viscosity and shear viscosity, respectively, and are given by \cite{Owen1998,Lindblom2000} \begin{eqnarray} \frac{1}{\tau_{GR}}=-\frac{32 \pi G \Omega^{2l+2}}{c^{2l+3}} \frac{(l-1)^{2l}}{[(2l+1)!!]^2}\left(\frac{l+2}{l+1}\right)^{(2l+2)}\nonumber \\ \times\int^{R_{c}}_{0}\rho(r)r^{2l+2} dr, \label{eq8} \end{eqnarray} \begin{equation} \frac{1}{\tau_{BV}}\approx \frac{4 R^{2l-2}}{(l+1)^2} \int {\xi |{\frac{\delta \rho}{\rho}}|^{2}} d^{3}r\left(\int^{R_{c}}_{0}\rho(r)r^{2l+2} dr\right)^{-1}, \label{eq9} \end{equation} \begin{eqnarray} \frac{1}{\tau_{SV}}=\left[\frac{1}{2\Omega} \frac{2^{l+3/2}(l+1)!}{l(2l+1)!!I_{l}}\sqrt{\frac{2\Omega R_{c}^{2} \rho_{c}}{\eta_c}}\right]^{-1}\nonumber \\ \times\left[\int^{R_{c}}_{0} \frac{\rho(r)}{\rho_{c}}\left(\frac{r}{R_{c}}\right)^{2l+2} \frac{dr}{R_c}\right]^{-1}, \label{eq10} \end{eqnarray} where G and c are the gravitational constant and velocity of light respectively; $\delta \rho$ in Eq.(9) is the density perturbation associated with r-modes and $\xi$ is the bulk viscosity of the fluid, $R_{c}$, $\rho_{c}$, $\eta_{c}$ in Eq.(10) are the radius, density and shear viscosity of the fluid at the outer edge of the core respectively. The expression for $\frac{1}{\tau_{BV}}$ in Eq.(9) is approximate and the exact expression should contain the Lagrangian density perturbation $\Delta \rho$ instead of Eulerian perturbation $\delta \rho$. The shear viscosity time scale in Eq.(10) is obtained by considering the shear dissipation in the viscous boundary layer between solid crust and the liquid core with the assumption that the crust is rigid and hence static in rotating frame \cite{Lindblom2000}. The motion of the crust due to mechanical coupling to the core effectively increases $\tau_{SV}$ by $(\frac{ \Delta v}{v})^{-2}$, where $\frac{\Delta v}{v}$ is the difference in the velocities in the inner edge of the crust and outer edge of the core divided by the velocity of the core \cite{Levin2001}. Bildsten and Ushomirsky \cite{Bildsten2000} have first estimated this effect of solid crust on r-mode instability and shown that the shear dissipation in this viscous boundary layer decreases the viscous damping time scale by more than $10^5$ in old acreting neutron stars and more than $10^7$ in hot, young neutron stars. $I_{l}$ in Eq.(10) has the value $I_{2}=0.80411$, for $l=2$ \cite{Lindblom2000}. Moreover, the bulk viscous dissipation is not significant for temperature of the star below $10^{10}$ K and in this range of temperature the shear viscosity is the dominant dissipative mechanism, We have restricted our study in this work to the range of the temperature $T<10^{10}$ K and included only shear dissipative mechanism. The studies is similar to the one done by Moustakidis \cite{Moustakidis2015}, where we have mainly examined the influence of neutron star EoS and the gravitational mass on the instability boundary and other relevant quantities, such as, critical frequency and temperature, etc. for a neutron star using the DDM3Y effective interaction \cite{BCS08}. As mentioned above, we have studied the instability within $T \leq 10^{10}$ K, the dominant dissipation mechanism is the shear viscosity in the boundary layer for which the time scale is given in Eq.(7), where $\eta_c$ is the viscosity of the fluid. In the temperature range $T \geq 10^9$ K, the dominant contribution to shear is from neutron-neutron (nn) scattering and below $T \leq 10^9$, it is the electron-electron (ee) scattering that contributes to shear primarily \cite{Lindblom2000}. Therefore, \begin{equation} \frac{1}{\tau_{SV}}=\frac{1}{\tau_{ee}}+\frac{1}{\tau_{nn}}, \label{eq11} \end{equation} where $\tau_{ee}$ and $\tau_{nn}$ can be obtained from Eq.(10) using the corresponding value of $\eta_{SV}^{ee}$ and $\eta_{SV}^{nn}$. These are given by \cite{Flowers1979,Cutler1987} \begin{equation} \eta_{SV}^{ee}=6 \times 10^{6} \rho^{2} T^{-2} ~~~~~({\rm g~cm^{-1}~s^{-1}}), \label{eq12} \end{equation} \begin{equation} \eta_{SV}^{nn}=347 \rho^{9/4} T^{-2} ~~~~~({\rm g~cm^{-1}~s^{-1}}), \label{eq13} \end{equation} where all the quantities are given in CGS units and T is measured in K. In order to have transparent visualisation of the role of angular velocity and temperature on various time scales, it is useful to factor them out by defining fiducial time scales. Thus, we define fiducial shear viscous time scale $\widetilde{\tau}_{SV}$ such that \cite{Lindblom2000,Lindblom1998}, \begin{equation} \tau_{SV}=\widetilde{\tau}_{SV} \left(\frac{\Omega_0}{\Omega}\right)^{1/2} \left(\frac{T}{10^8 K}\right), \label{eq14} \end{equation} and fiducial gravitational radiation time scale $\widetilde{\tau}_{GR}$ is defined through the relation \cite{Lindblom2000,Lindblom1998}, \begin{equation} \tau_{GR}=\widetilde{\tau}_{GR} \left(\frac{\Omega_0}{\Omega}\right)^{2l+2}, \label{eq15} \end{equation} where $\Omega_0=\sqrt{ \pi G \bar{\rho}}$ and $\bar{\rho}= 3M/4 \pi R^3$ is the mean density of NS having mass $M$ and radius $R$. Thus Eq.(7) (neglecting bulk viscosity contributions) becomes \begin{equation} \frac{1}{\tau(\Omega,T)}=\frac{1}{\widetilde{\tau}_{GR}}\left(\frac{\Omega}{\Omega_0}\right)^{2l+2}+\frac{1}{\widetilde{\tau}_{SV}}\left(\frac{\Omega}{\Omega_0}\right)^{1/2} \left(\frac{10^8 K}{T}\right). \label{eq16} \end{equation} Dissipative effects cause the mode to decay exponentially as $e^{-t/\tau}$ i.e. the mode is stable, as long as $\tau>0$. From Eq.(8) and Eq.(10) it can be seen that $\widetilde{\tau}_{SV}>0$, while $\widetilde{\tau}_{GR}<0$. Thus gravitational radiation drives these modes towards instability while viscosity tries to stabilize them. For small $\Omega$ the gravitational radiation contribution to $1/\tau$ is very small since it is proportional to $\Omega^{2l+2}$. Thus for sufficiently small angular velocity, viscosity dominates and the mode is stable. But for sufficiently large $\Omega$ gravitational radiation will dominate and drive the mode unstable. For a given temperature and mode $l$ the equation for critical angular velocity $\Omega_c$ is obtained from the condition $\frac{1}{\tau(\Omega_c,T)}=0$. At a given T and mode $l$, the equation for the critical velocity is a polynomial of order $l+1$ in $\Omega_c^{2}$ and thus each mode has its own characteristic $\Omega_c$. Since the smallest of these, i.e. $l=2$, is the dominant contributor, study is being done for this mode only. The critical angular velocity $\Omega_c$ for this mode is obtained to be \begin{equation} \left(\frac{\Omega_c}{\Omega_0}\right)=\left(-\frac{\widetilde{\tau}_{GR}}{\widetilde{\tau}_{SV}}\right)^{2/11}\left(\frac{10^8 K}{T}\right)^{2/11}. \label{eq17} \end{equation} The angular velocity of a neutron star can never exceed the Kepler velocity $\Omega_K\approx\frac{2}{3}\Omega_0$. Thus, there is a critical temperature below which the gravitational radiation is completely suppressed by viscosity. This critical temperature is given by \cite{Lindblom2000} \begin{equation} \frac{T_c}{10^8 K}=\left(\frac{\Omega_0}{\Omega_c}\right)^{11/2} \left(-\frac{\tilde{\tau}_{GR}}{\widetilde{\tau}_{SV}}\right)\approx(3/2)^{11/2}\left(-\frac{\widetilde{\tau}_{GR}}{\widetilde{\tau}_{SV}}\right). \label{eq18} \end{equation} The critical angular velocity is now expressed in terms of critical temperature from Eq.(14) and Eq.(15) as \begin{equation} \left(\frac{\Omega_c}{\Omega_0}\right)=\frac{\Omega_K}{\Omega_0}\left( \frac{T_c}{T}\right)^{2/11}\approx(2/3)\left( \frac{T_c}{T}\right)^{2/11}. \label{eq19} \end{equation} So, once the neutron star EoS is ascertained, then all physical quantities necessary for the calculation of r-mode instability can be performed. Further, following the work of Owen et al. \cite{Owen1998} the evolution of the angular velocity, as the angular momentum is radiated to infinity by the gravitational radiation is given by \begin{equation} \frac{d\Omega}{dt}=\frac{2\Omega}{\tau_{GR}}\frac{\alpha_r^2Q}{1-\alpha_r^2Q}, \label{eq20} \end{equation} where $\alpha_r$ is the dimensionless r-mode amplitude and $Q=3 \widetilde{J}/2 \widetilde{I}$ with, \begin{equation} \widetilde{J}=\frac{1}{MR^4}\int^{R}_{0}\rho(r)r^{6} dr \label{eq21} \end{equation} and \begin{equation} \widetilde{I}=\frac{8\pi}{3MR^2}\int^{R}_{0}\rho(r)r^{4} dr. \label{eq22} \end{equation} $\alpha_r$ is treated as free parameter whose value varies within a wide range $1-10^{-8}$. Under the ideal consideration that the heat generated by the shear viscosity is same as that taken out by the emission of neutrinos \cite{Bondarescu2009,Moustakidis2015}, Eq.(20) can be solved for the angular frequency $\Omega(t)$ as \begin{equation} \Omega(t)=\left(\Omega^{-6}_{in}-\textsl{C}t\right)^{-1/6}, \label{eq23} \end{equation} where \begin{equation} \textsl{C}=\frac{12\alpha_r^2Q}{\widetilde{\tau}_{GR}\left(1-\alpha_r^2Q\right)}\frac{1}{\Omega_0^6}, \label{eq24} \end{equation} and $\Omega_{in}$ is considered as a free parameter whose value corresponds to be the initial angular velocity. The spin down rate can be obtained from Eq.(20) to be, \begin{equation} \frac{d\Omega}{dt}=\frac{\textsl{C}}{6}\left(\Omega^{-6}_{in}-\textsl{C}t\right)^{-7/6}. \label{eq25} \end{equation} The neutron star spin shall decrease continually until it approaches its critical angular velocity $\Omega_c$. The time $t_c$ taken by neutron star to evolve from its initial value $\Omega_{in}$ to its minimum value $\Omega_{c}$ is given by \begin{equation} t_c=\frac{1}{\textsl{C}}\left(\Omega_{in}^{-6}-\Omega_{c}^{-6}\right). \label{eq26} \end{equation} \noindent \section{Nuclear Equation of State} \label{Section 3} The EoS for nuclear matter is obtained by using the isoscalar and the isovector \cite{Sa83} components of M3Y effective NN interaction along with its density dependence. The nuclear matter calculation is then performed which enables complete determination of this density dependence. The minimization of energy per nucleon determines the equilibrium density of the symmetric nuclear matter (SNM). The variation of the zero range potential with energy, over the entire range of the energy per nucleon $\epsilon$, is treated properly by allowing it to vary freely with the kinetic energy part $\epsilon^{kin}$ of $\epsilon$. This treatment is more plausible as well as provides excellent result for the SNM incompressibility $K_\infty$. Moreover, the EoS for SNM is not plagued with the superluminosity problem. The energy per nucleon $\epsilon$ for isospin asymmetric nuclear matter (IANM) can be derived within a Fermi gas model of interacting neutrons and protons as \cite{BCS08} \begin{equation} \epsilon(\rho,X) = [\frac{3\hbar^2k_F^2}{10m}] F(X) + (\frac{\rho J_v C}{2}) (1 - \beta\rho^n) \label{eq27} \end{equation} \noindent where isospin asymmetry $X= \frac{\rho_n - \rho_p} {\rho_n + \rho_p}$, $\rho = \rho_n+\rho_p$ with $\rho_n$, $\rho_p$ and $\rho$ being the neutron, proton and nucleonic densities respectively, $m$ is the nucleonic mass, $k_F$=$(1.5\pi^2\rho)^{\frac{1}{3}}$ which equals Fermi momentum in case of SNM, $\epsilon^{kin}=[\frac{3\hbar^2k_F^2}{10m}] F(X)$ with $F(X)$=$[\frac{(1+X)^{5/3} + (1-X)^{5/3}}{2}]$ and $J_v$=$J_{v00} + X^2 J_{v01}$, $J_{v00}$ and $J_{v01}$ represent the volume integrals of the isoscalar and the isovector parts of the M3Y interaction. The isoscalar and isovector components $t_{00}^{M3Y}$ and $t_{01}^{M3Y}$ of the M3Y effective NN interaction are given by $t_{00}^{M3Y}(s, \epsilon)=7999\frac{\exp( - 4s)}{4s}-2134\frac{\exp( - 2.5s)}{2.5s}+J_{00}(1-\alpha\epsilon)\delta(s)$ and $t_{01}^{M3Y}(s, \epsilon)=-4886\frac{\exp( - 4s)}{4s}+1176\frac{\exp( - 2.5s)}{2.5s}+J_{01}(1-\alpha\epsilon)\delta(s)$, respectively, with $J_{00}$=-276 MeVfm$^3$, $J_{01}$=228 MeVfm$^3$, $\alpha=0.005$MeV$^{-1}$. The DDM3Y effective NN interaction is given by $v_{0i}(s,\rho, \epsilon) = t_{0i}^{M3Y}(s, \epsilon) g(\rho)$ where $g(\rho) = C (1 - \beta \rho^n)$ is the density dependence with $C$ and $\beta$ being the constants of density dependence. Differentiating Eq.(27) with respect to $\rho$ one obtains equation for $X=0$: \begin{equation} \frac{\partial\epsilon}{\partial\rho} = [\frac{\hbar^2k_F^2}{5m\rho}] + \frac{J_{v00} C}{2} [1 - (n+1)\beta\rho^n] -\alpha J_{00} C [1 - \beta\rho^n] [\frac{\hbar^2k_F^2}{10m}]. \label{eq28} \end{equation} \noindent The saturation condition $\frac{\partial\epsilon}{\partial\rho} = 0$ at $\rho = \rho_{0}$, $\epsilon = \epsilon_{0}$, determines the equilibrium density of the cold SNM. Then for fixed values of the saturation energy per nucleon $\epsilon_0$ and the saturation density $\rho_{0}$ of the cold SNM, Eq.(27) and Eq.(28) with the saturation condition can be solved simultaneously to obtain the values of $\beta$ and $C$ which are given by \begin{equation} \beta = \frac{[(1-p)+(q-\frac{3q}{p})]\rho_{0}^{-n}}{[(3n+1)-(n+1)p+(q-\frac{3q}{p})]} \label{eq29} \end{equation} \noindent \begin{equation} {\rm where}~~~~p = \frac{[10m\epsilon_0]}{[\hbar^2k_{F_0}^2]},~q=\frac{2\alpha\epsilon_0J_{00}}{J^0_{v00}} \label{eq30} \end{equation} \noindent where $J^0_{v00} = J_{v00}(\epsilon^{kin}_0)$ which means $J_{v00}$ is evaluated at $\epsilon^{kin}=\epsilon^{kin}_0$, the kinetic energy part of the saturation energy per nucleon of SNM, $k_{F_0} = [1.5\pi^2\rho_0]^{1/3}$ and \begin{equation} C = -\frac{[2\hbar^2k_{F_0}^2] }{ 5mJ^0_{v00} \rho_0 [1 - (n+1)\beta\rho_0^n -\frac{q\hbar^2k_{F_0}^2 (1-\beta\rho_0^n)}{10m\epsilon_0}]}, \label{eq31} \end{equation} \noindent respectively. Obviously, the constants $C$ and $\beta$ determined by this methodology depend upon $\epsilon_0$, $\rho_{0}$, the index $n$ of the density dependent part and through the volume integral $J^0_{v00}$, on the strengths of the M3Y interaction . The calculations have been carried out by using the values of saturation density $\rho_0=0.1533$ fm$^{-3}$ \cite{Sa89} and saturation energy per nucleon $\epsilon_0=-15.26$ MeV \cite{CB06} for the SNM. $\epsilon_0$ is the co-efficient $a_v$ of the volume term of Bethe-Weizs\"acker mass formula, calculated by fitting the recent experimental and estimated Audi-Wapstra-Thibault atomic mass excesses \cite{Au03}. This term has been obtained by minimizing the mean square deviation incorporating correction for the electronic binding energy \cite{Lu03}. In a similar work, including surface symmetry energy term, Wigner term, shell correction and also the proton form factor correction to Coulomb energy, $a_v$ turns out to be 15.4496 MeV \cite{Ro06} ($a_v=14.8497$ MeV when $A^0$ and $A^{1/3}$ terms are also included). Using the standard values of $\alpha=0.005$ MeV$^{-1}$ for the parameter of energy dependence of the zero range potential and $n$=2/3, the values deduced for the constants $C$ and $\beta$ and the SNM incompressibility $K_\infty$ are, respectively, 2.2497, 1.5934 fm$^2$ and 274.7 MeV. The term $\epsilon_0$ is $a_v$ and its value of $-15.26\pm0.52$ MeV encompasses, more or less, the entire range of values. For this value of $a_v$ now the values of the constants of density dependence are $C=2.2497\pm0.0420$, $\beta=1.5934\pm0.0085$ fm$^2$ and the SNM incompressibility $K_\infty$ turns out to be $274.7\pm7.4$ MeV. \subsection{Symmetric and isospin asymmetric nuclear matter} The EoSs of the SNM and the IANM which describes energy per nucleon $\epsilon$ as a function of nucleonic density $\rho$ can be obtained by setting isospin asymmetry $X=0$ and $X \neq 0$, respectively, in Eq.(27). It is observed that the energy per nucleon $\epsilon$ for SNM is negative (bound) up to nucleonic density of $\sim2\rho_0$ while for pure neutron matter (PNM) $\epsilon>0$ and is always unbound by nuclear forces. The compression modulus or incompressibility of the SNM, which is a measure of the curvature of an EoS at saturation density and is defined as $k_F^2\frac{\partial^2\epsilon}{\partial{k_F^2}} \mid_{k_F=k_{F_0}}$. It measures the stiffness of an EoS and can be theoretically obtained by using Eq.(27) for X=0. The IANM incompressibilities are evaluated at saturation densities $\rho_s$ with the condition of vanishing pressure which is $\frac{\partial\epsilon}{\partial\rho}|_{\rho=\rho_s}=0$. The incompressibility $K_s$ for IANM is therefore expressed as \begin{eqnarray} &&K_s = -\frac{3\hbar^2k_{F_s}^2}{5m} F(X) - \frac{9 J^s_v C n(n+1) \beta\rho_s^{n+1}}{2} \nonumber \\ &&- 9\alpha J C [1-(n+1)\beta\rho_s^n] [\frac{\rho_s\hbar^2k_{F_s}^2}{5m}] F(X) \nonumber \\ &&+ [\frac{3\rho_s\alpha J C (1-\beta\rho_s^n)\hbar^2k_{F_s}^2}{10m}] F(X), \label{eq32} \end{eqnarray} \noindent where $k_{F_s}$ implies that the $k_F$ is calculated at saturation density $\rho_s$. The term $J^s_v=J^s_{v00} + X^2 J^s_{v01}$ is $J_v$ evaluated at $\epsilon^{kin}=\epsilon^{kin}_s$ which is the kinetic energy part of the saturation energy per nucleon $\epsilon_s$ and $J=J_{00} + X^2 J_{01}$. In Table-I, IANM incompressibility $K_s$ as a function of $X$, for the standard value of $n=2/3$ and energy dependence parameter $\alpha=0.005$ MeV$^{-1}$, is provided. The magnitude of the IANM incompressibility $K_s$ decreases with $X$ due to lowering of the saturation densities $\rho_s$ with the isospin asymmetry $X$ as well as decrease in the EoS curvature. At high values of $X$, the IANM does not have a minimum which signify that it can never be bound by itself due to interaction of nuclear force. However, the $\beta$ equilibrated nuclear matter which is a highly neutron rich IANM exists in the core of the neutron stars since its energy per nucleon is lower than that of SNM at high densities. Although it is unbound by the nuclear interaction but can be bound due to very high gravitational field that can be realized inside neutron stars. \noindent \begin{table} \centering \caption{IANM incompressibility at different isospin asymmetry $X$ using the usual values of $n=\frac{2}{3}$ and $\alpha=0.005$ MeV$^{-1}$.} \begin{tabular}{ccc} \hline \hline $X$&$\rho_s$& $K_s$ \\ \hline & fm$^{-3}$ &MeV \\ \hline 0.0&0.1533&274.69 \\ 0.1&0.1525&270.44 \\ 0.2&0.1500&257.68 \\ 0.3&0.1457&236.64 \\ 0.4&0.1392&207.62 \\ 0.5&0.1300&171.16 \\ 0.6&0.1170&127.84 \\ 0.7&0.0980&78.38 \\ \hline \hline \label{table1} \end{tabular} \end{table} It is worthwhile to mention that the RMF-NL3 incompressibility for SNM is 271.76 MeV \cite{La97,La99} which is very close to 274.7$\pm$7.4 MeV obtained from the present calculation. In spite of the fact that the parameters of the density dependence of DDM3Y interaction have been tuned to reproduce the saturation energy per nucleon $\epsilon_0$ and the saturation density $\rho_{0}$ of the cold SNM that are obtained from finite nuclei, the agreement of the present EoS with the experimental flow data, where the high density behavior looks phenomenologically confirmed, justifies its extrapolation to high density. The SNM incompressibility is experimentally determined from the compression modes isoscalar giant monopole resonance (ISGMR) and isoscalar giant dipole resonance (ISGDR) of nuclei. The violations of self consistency in HF-RPA calculations \cite{Sh06} of the strength functions of ISGMR and ISGDR cause shifts in the calculated values of the centroid energies. These shifts may be larger in magnitude than the current experimental uncertainties. In fact, due to the use of a not fully self-consistent calculations with Skyrme interactions \cite{Sh06}, the low values of $K_\infty$ in the range of 210-220 MeV were predicted. Skyrme parmetrizations of SLy4 type predict $K_\infty$ values lying in the range of 230-240 MeV \cite{Sh06} when this drawback is corrected. Besides that bona fide Skyrme forces can be built so that the $K_\infty$ for SNM is rather close to the relativistic value of $\sim$ 250-270 MeV. Conclusion may, therefore, be drawn from the ISGMR experimental data that the magnitude of $K_\infty \approx$ 240 $\pm$ 20 MeV. The lower values \cite{Lu04,Yo04} for $K_\infty$ are usually predicted by the ISGDR data. However, it is generally agreed upon that the extraction of $K_\infty$ in this case more problematic for various reasons. Particularly, for excitation energies \cite{Sh06} above 30 and 26 MeV for $^{116}$Sn and $^{208}$Pb, respectively, the maximum cross-section for ISGDR at high excitation energy decreases very strongly and can even fall below the range of current experimental sensitivity. The upper limit of the recent values \cite{Yo05} for the nuclear incompressibility $K_\infty$ for SNM extracted from experiments is rather close to the present non-relativistic mean field model estimate employing DDM3Y interaction which is also in agreement with the theoretical estimates of relativistic mean field (RMF) models. With Gogny effective interactions \cite{Bl80}, which include nuclei where pairing correlations are important, the results of microscopic calculations reproduce experimental data on heavy nuclei for $K_\infty$ in the range about 220 MeV \cite{Bl95}. It may, therefore, be concluded that the calculated value of 274.7$\pm$7.4 MeV is a good theoretical result and is only slightly too high compared to the recent acceptable value \cite{Vr03,Sh09} of $K_\infty$ for SNM in the range of 250-270 MeV. \begin{figure}[t] \vspace{-0.0cm} \eject\centerline{\epsfig{file=fig01.eps,height=8cm,width=8cm}} \caption{Plots of the nuclear symmetry energy NSE is as a function of $\rho/\rho_0$ for the present calculation using DDM3Y interaction and its comparison, with those for Akmal-Pandharipande-Ravenhall (APR) \cite{Ak98} and the MDI interactions for the variable x$=$0.0, 0.5 defined in Ref. \cite{Zh09}.} \label{fig1} \vspace{-0.0cm} \end{figure} \begin{figure}[t] \vspace{0.0cm} \eject\centerline{\epsfig{file=fig02.eps,height=8cm,width=8cm}} \caption{The plots of $K_\tau$ versus $K_\infty$ ($K_{inf}$). Present calculation (DDM3Y) is compared with other predictions as tabulated in Refs. \cite{Pi09,Sa07} and the dotted rectangular region encompasses the values of $K_\infty=250-270$ MeV \cite{Sh09} and $K_\tau=-370\pm120$ MeV \cite{Ch09}.} \label{fig2} \vspace{0.0cm} \end{figure} \subsection{Nuclear symmetry energy $\&$ its slope, incompressibility and isobaric incompressibility} The EoS of IANM, given by Eq.(27) can be expanded, in general, as \begin{equation} \epsilon(\rho,X) = \epsilon(\rho,0) + E_{sym}(\rho) X^2 + O ( X^4) \label{eq33} \end{equation} \noindent where $E_{sym}(\rho)= \frac{1}{2} \frac{\partial^2\epsilon(\rho,X)}{\partial{X^2}} \mid_{X=0}$ is termed as the nuclear symmetry energy (NSE). The exchange symmetry between protons and neutrons in nuclear matter when one neglects the Coulomb interaction and assumes the charge symmetry of nuclear forces results in the absence of odd-order terms in $X$ in Eq.(33). To a good approximation, the density-dependent NSE $E_{sym}(\rho)$ can be obtained using the following equation \cite{Kl06} \begin{equation} E_{sym}(\rho)=\epsilon(\rho,1) -\epsilon(\rho,0) \label{eq34} \end{equation} \noindent as the higher-order terms in $X$ are negligible. The above equation can be obtained using Eq.(27). It represents a penalty levied on the system as it departs from the symmetric limit of equal number of protons and neutrons. Thus, it can be defined as the energy required per nucleon to change the SNM to PNM. In Fig.-1 the plot of NSE as a function of $\rho/\rho_0$ is shown for the present calculation (DDM3Y) and compared with those for Akmal-Pandharipande-Ravenhall \cite{Ak98} and MDI interactions \cite{Zh09}. A constraint on the NSE at nuclear saturation density $E_{sym}(\rho_0)$ is provided by the volume symmetry energy coefficient $S_v$ which can be extracted from measured atomic mass excesses. The theoretical estimate for value of the NSE at saturation density $E_{sym}(\rho_0)$=30.71$\pm$0.26 MeV obtained from the present calculations (DDM3Y) is reasonably close to the value of $S_v$=30.048 $\pm$0.004 MeV extracted \cite{Mu07} from the measured atomic mass excesses of 2228 nuclei. The value of NSE at $\rho_0$ remains mostly the same which is 30.03$\pm$0.26 MeV if one uses the mathematical definition of $E_{sym}(\rho)= \frac{1}{2} \frac{\partial^2\epsilon(\rho,X)} {\partial{X^2}} \mid_{X=0}$ alternatively. The value of $E_{sym}(\rho_0)\approx$ 30 MeV \cite{St05,Da03,Po03} appears well established empirically. The different parameterizations of RMF models, which fit observables of isospin symmetric nuclei nicely, steers to a comparatively wide range of predictions of 24-40 MeV for $E_{sym}(\rho_0)$ theoretically. Our present result (DDM3Y) of 30.71$\pm$0.26 MeV is reasonably close to that obtained using Skyrme interaction SkMP (29.9 MeV) \cite{Be89}, Av18+$\delta v$+UIX$^*$ variational calculation (30.1 MeV) \cite{Ak98}. The NSE $E_{sym}(\rho)$ can be expanded around the nuclear matter saturation density $\rho_0$ as \begin{equation} E_{sym}(\rho)= E_{sym}(\rho_0) + \frac{L}{3} {\Big (}\frac{\rho - \rho_0}{\rho_0}{\Big )}+ \frac{K_{sym}}{18}{\Big (}\frac{\rho - \rho_0}{\rho_0}{\Big )}^2 \label{eq35} \end{equation} \noindent up to second order in density where $L$ and $K_{sym}$ represents the slope and curvature parameters of NSE at $\rho_0$ and hence $L= 3\rho_0 \frac{\partial E_{sym}(\rho)}{\partial\rho} \mid_{\rho=\rho_0}$ and $K_{sym}= 9\rho_0^2 \frac{\partial^2 E_{sym}(\rho)}{\partial {\rho^2}} \mid_{\rho=\rho_0}$. The $K_{sym}$ and $L$ highlights the density dependence of NSE around $\rho_0$ and carry important information at both high and low densities on the properties of NSE. Particularly, it is found that the slope parameter $L$ correlate linearly with neutron-skin thickness of heavy nuclei and it can be obtained from the measured thickness of neutron skin of heavy nuclei \cite{Pi05,Ch05,SL05}. Although there are large uncertainties in the experimental measurements, this has been possible \cite{Ce09} recently. Differentiation of Eq.(34) twice with respect to the nucleonic density $\rho$ using Eq.(27) provides \cite{BCM14} \begin{eqnarray} &&\frac{\partial E_{sym}}{\partial \rho} =\frac{2}{5}(2^{2/3}-1)\frac{E^0_F}{\rho}(\frac{\rho}{\rho_0})^{2/3}+\frac{C}{2}[1-(n+1)\beta\rho^n] \nonumber \\ &&\times J_{v01}(\epsilon^{kin}_{X=1})-\frac{\alpha J_{01}C}{5}E^0_F(\frac{\rho}{\rho_0})^{2/3} [1-\beta\rho^n]F(1) \nonumber \\ &&-(2^{2/3}-1)\frac{\alpha J_{00}C}{5}E^0_F(\frac{\rho}{\rho_0})^{2/3}[1-\beta\rho^n] \nonumber \\ &&-\frac{3}{10}(2^{2/3}-1)\alpha J_{00}CE^0_F(\frac{\rho}{\rho_0})^{2/3}[1-(n+1)\beta\rho^n] \label{eq36} \end{eqnarray} \noindent \begin{eqnarray} &&\frac{\partial^2 E_{sym}}{\partial \rho^2} =-\frac{2}{15}(2^{2/3}-1)\frac{E^0_F}{\rho^2}(\frac{\rho}{\rho_0})^{2/3}-\frac{C}{2}n(n+1)\beta\rho^{n-1}\nonumber \\ &&\times J_{v01} (\epsilon^{kin}_{X=1}) -\frac{2\alpha J_{01}C}{5}\frac{E^0_F}{\rho}(\frac{\rho}{\rho_0})^{2/3}[1-(n+1)\beta\rho^n]F(1) \nonumber \\ &&+\frac{\alpha J_{01}C}{15}\frac{E^0_F}{\rho}(\frac{\rho}{\rho_0})^{2/3}[1-\beta\rho^n]F(1) \nonumber \\ &&+(2^{2/3}-1)\frac{\alpha J_{00}C}{15}\frac{E^0_F}{\rho}(\frac{\rho}{\rho_0})^{2/3} [1-\beta\rho^n] \nonumber \\ &&-\frac{2}{5}(2^{2/3}-1)\alpha J_{00}C\frac{E^0_F}{\rho}(\frac{\rho}{\rho_0})^{2/3} [1-(n+1) \beta\rho^n] \nonumber \\ &&+\frac{3}{10}(2^{2/3}-1)\alpha J_{00}CE^0_F(\frac{\rho}{\rho_0})^{2/3}n(n+1)\beta\rho^{n-1}. \label{eq37} \end{eqnarray} \noindent Here the Fermi energy is $E^0_F=\frac{\hbar^2k_{F_0}^2}{2m}$ for the SNM at ground state and to evaluate the values of $L$ and $K_{sym}$, the definitions of which are provided after Eq.(35), along with Eqs.(36,37) at $\rho$=$\rho_0$ have been used. The isobaric incompressibility $K_\infty (X)$ for infinite IANM can be expanded in the power series of isospin asymmetry parameter $X$ as $K_\infty (X) = K_\infty + K_\tau X^2 + K_4 X^4 + O(X^6)$. Compared to $K_\tau$ \cite{Ch09} the magnitude of the higher-order $K_4$ parameter is quite small in general. The former essentially characterizes the isospin dependence of the incompressibility at $\rho_0$ and expressed as $K_\tau= K_{sym}-6L- \frac{Q_0}{K_\infty}L =K_{asy}-\frac{Q_0}{K_\infty}L$ where the third-order derivative parameter of SNM at $\rho_0$ is $Q_0$, given by \begin{equation} Q_0 = 27\rho_0^3 \frac{\partial^3 \epsilon(\rho,0)}{\partial {\rho^3}} \mid_{\rho=\rho_0}. \label{eq38} \end{equation} \noindent One obtains, using Eq.(27), the following \begin{eqnarray} &&\frac{\partial^3 \epsilon(\rho,X)}{\partial {\rho^3}} =-\frac{CJ_v(\epsilon^{kin})n(n+1)(n-1)\beta\rho^{n-2}}{2} \nonumber \\ &&+\frac{8}{45}\frac{E^0_F}{\rho^3}F(X)(\frac{\rho}{\rho_0})^\frac{2}{3} +\frac{3\alpha JC}{5}n(n+1)\beta\rho^{n-1}\frac{E^0_F}{\rho} \nonumber \\ &&\times F(X)(\frac{\rho}{\rho_0})^\frac{2}{3} +\frac{\alpha JC}{5}[1-(n+1)\beta\rho^n] \frac{E^0_F}{\rho^2}F(X)(\frac{\rho}{\rho_0})^\frac{2}{3} \nonumber \\ &&-\frac{4\alpha JC}{45}[1-\beta\rho^n]\frac{E^0_F}{\rho^2}F(X)(\frac{\rho}{\rho_0})^\frac{2}{3} \label{eq39} \end{eqnarray} \noindent where the Fermi energy $E^0_F$=$\frac{\hbar^2k_{F_0}^2}{2m}$ for the SNM at ground state, $k_{F_0}$=$(1.5\pi^2\rho_0)^{\frac{1}{3}}$ and $J$=$J_{00}$+$X^2J_{01}$. Thus \begin{eqnarray} &&\frac{\partial^3 \epsilon(\rho,0)}{\partial {\rho^3}} \mid_{\rho=\rho_0} =-\frac{CJ_{v00}(\epsilon_0^{kin})n(n+1)(n-1)\beta\rho_0^{n-2}}{2} \nonumber \\ &&+\frac{8}{45}\frac{E^0_F}{\rho_0^3}+\frac{3\alpha J_{00}C}{5}n(n+1)\beta\rho_0^{n-1} \frac{E^0_F}{\rho_0}+\frac{\alpha J_{00}C}{5} \nonumber \\ &&\times [1-(n+1)\beta\rho_0^n]\frac{E^0_F}{\rho_0^2}-\frac{4\alpha J_{00}C}{45}[1-\beta\rho_0^n]\frac{E^0_F}{\rho_0^2}. \label{eq40} \end{eqnarray} \noindent \begin{figure}[t] \vspace{0.0cm} \eject\centerline{\epsfig{file=fig03.eps,height=8cm,width=8cm}} \caption{Plots for pressure P of dense nuclear matter as functions of $\rho/\rho_0$. The continuous line represents the pure neutron matter and the dashed line represents the $\beta$-equilibrated neutron star matter. The dotted line represents the same for the A18 model using variational chain summation of Akmal et al. \cite{Ak98}. The areas enclosed by the continuous and the dashed lines correspond to the pressure regions for neutron matter consistent with the experimental flow data after inclusion of the pressures from asymmetry terms with weak (soft NM) and strong (stiff NM) density dependences, respectively \cite{Da02}.} \label{fig3} \vspace{-0.14cm} \end{figure} \noindent \begin{figure}[t] \vspace{0.0cm} \eject\centerline{\epsfig{file=fig04.eps,height=8cm,width=8cm}} \caption {Mass versus central baryonic density plot of slowly rotating neutron stars for the DDM3Y EoS. } \label{fig4} \vspace{-0.14cm} \end{figure} \noindent For the calculations of $K_\infty$, $E_{sym}(\rho_0)$, $L$, $K_{sym}$ and $K_{\tau}$, the values of $\rho_0$=0.1533 fm$^{-3}$, $\epsilon_0=-15.26\pm0.52$ MeV for the SNM and $n=\frac{2}{3}$ \cite{CBS09} have been used. Using the improved quantum molecular dynamics transport model, the collisions involving $^{112}$Sn and $^{124}$Sn nuclei can be simulated to reproduce isospin diffusion data from two different observables and the ratios of proton and neutron spectra. The constraints on the density dependence of the NSE at subnormal density can be obtained \cite{Ts09} by comparing these data to calculations performed over a range of NSEs at $\rho_0$ and different representations of the density dependence of the NSE. The results for $K_\infty$, $L$, $E_{sym}(\rho_0)$ and density dependence of $E_{sym}(\rho)$ \cite{CBS09} of the present calculations are consistent with these constraints \cite{Ts09}. In Table-II, the values of $K_\infty$, $E_{sym}(\rho_0)$, $L$, $K_{sym}$ and $K_\tau$ are tabulated and compared with the corresponding quantities obtained from the RMF models \cite{Pi09}. What is a reasonable value of incompressibility \cite{Sh06} remains controversial. In the following we present our results in the backdrop of others, without justifying any particular value for $K_\infty$, but for an objective view of the current situation which, we stress, is still progressing. In Fig.-2, the plot of $K_\tau$ versus $K_\infty$ for the present calculation (DDM3Y) has been compared with the predictions of SkI3, SkI4, SLy4, SkM, SkM*, FSUGold, NL3, Hybrid \cite{Pi09}, NLSH, TM1, TM2, DDME1 and DDME2 as given in Table-I of Ref. \cite{Sa07}. The recent values of $K_\infty=250-270$ MeV \cite{Sh09} and $K_\tau=-370\pm120$ MeV \cite{Ch09} are enclosed by the dotted rectangular region. Though both DDM3Y and SkI3 are within the above region, unlike DDM3Y the $L$ value for SkI3 is 100.49 MeV which is much above the acceptable limit of $58.9\pm16$ MeV \cite{Wa09,Ag13,Li13,Ag17}. Another recent review \cite{Kl17} also finds that $E_{sym}(\rho_0)=31.7\pm3.2$ MeV and $L=58.7\pm28.1$ MeV with an error for $L$ that is considerably larger than that for $E_{sym}(\rho_0)$. However, DDME2 is reasonably close to the rectangular region which has $L=51$ MeV. It is worthwhile to mention here that the DDM3Y interaction with the same ranges, strengths and density dependence that provides $L=45.11\pm0.02$, allows good descriptions of elastic and inelastic scattering, proton radioactivity \cite{BCS08} and $\alpha$ radioactivity of superheavy elements \cite{CSB06,scb07}. The present NSE increases initially with nucleonic density up to about 2$\rho_0$ and then decreases monotonically (hence `soft') and becomes negative at higher densities (about 4.7$\rho_0$) \cite{BCS08,CBS09} (hence `super-soft'). It is consistent with the recent evidence for a soft NSE at suprasaturation densities \cite{Zh09} and with the fact that the super-soft nuclear symmetry energy preferred by the FOPI/GSI experimental data on the $\pi^+/\pi^-$ ratio in relativistic heavy-ion reactions can readily keep neutron stars stable if the non-Newtonian gravity proposed in the grand unification theories is considered \cite{We09}. \begin{table*}[htbp] \centering \caption{Comparison of the present results obtained using DDM3Y effective interaction with those of RMF models \cite{Pi09} for SNM incompressibility $K_\infty$, NSE at saturation density $E_{sym}(\rho_0)$, slope $L$ and the curvature $K_{sym}$ parameters of NSE, $K_{asy}$ and isobaric incompressibility $K_\tau$ of IANM (all in MeV).} \begin{tabular}{||c|c|c|c|c|c||} \hline \hline Model&$K_\infty$&$E_{sym}(\rho_0)$&$L$&$K_{sym}$&$K_{asy}$\\ \hline \hline This work &$274.7\pm7.4$&$30.71\pm0.26$&$45.11\pm0.02$&$-183.7\pm3.6$&$-454.4\pm3.5$\\ &&&&$Q_0$=$-276.5\pm10.5$&$K_\tau$=$-408.97\pm3.01$ \\ \hline FSUGold&230.0&32.59&60.5&-51.3&-414.3\\ &&&&$Q_0$=$-523.4$&$K_\tau$=$-276.77$\\ \hline NL3&271.5&37.29&118.2&+100.9&-608.3\\ &&&&$Q_0$=$+204.2$&$K_\tau$=$-697.36$\\ \hline Hybrid&230.0&37.30&118.6&+110.9&-600.7\\ &&&&$Q_0$=$-71.5$&$K_\tau$=$-563.86$\\ \hline \hline \end{tabular} \label{table2} \end{table*} \noindent \begin{table*}[htbp] \centering \caption{Results of the present calculations (DDM3Y) of symmetric nuclear matter incompressibility $K_\infty$, nuclear symmetry energy at saturation density $E_{sym}(\rho_0)$, the slope $L$ and isospin dependent part $K_\tau$ of the isobaric incompressibility (all in MeV) \cite{CBS09} are tabulated along with the saturation density and the density, pressure and proton fraction at the core-crust transition for $\beta$-equilibrated neutron star matter.} \begin{tabular}{||c|c|c|c||} \hline \hline $K_\infty$&$E_{sym}(\rho_0)$&$L$&$K_\tau$ \\ \hline $274.7\pm7.4$&$30.71\pm0.26$&$45.11\pm0.02$&$-408.97\pm3.01$ \\ \hline $\rho_0$&$\rho_t$& P$_t$ & x$_{p(t)}$ \\ \hline 0.1533 fm$^{-3}$&0.0938 fm$^{-3}$&0.5006 MeV fm$^{-3}$& 0.0308 \\ \hline \hline \end{tabular} \label{table3} \end{table*} \noindent \noindent \section{$\beta$-equilibrated neutron star matter} \label{Section 4} The $\beta$-equilibrated nuclear matter EoS is obtained by evaluating the asymmetric nuclear matter EoS at the isospin asymmetry $X$ determined from the $\beta$-equilibrium proton fraction $x_p =\frac{\rho_p}{\rho}$, obtained approximately by solving \begin{equation} \hbar c (3 \pi^2\rho x_p)^{1/3}= 4E_{sym}(\rho) (1 - 2 x_p). \label{eq41} \end{equation} \noindent The exact way of obtaining $\beta$-equilibrium proton fraction is by solving \begin{equation} \hbar c (3 \pi^2\rho x_p)^{1/3} = -\frac{\partial \epsilon(\rho,x_p)}{\partial x_p} = +2\frac{\partial \epsilon}{\partial X}, \label{eq42} \end{equation} \noindent where isospin asymmetry $X=1-2x_p$. The pressure $P$ of PNM and $\beta$-equilibrated neutron star matter are plotted in Fig.-3 as functions of $\rho/\rho_0$. The continuous line represents the PNM and the dashed line (almost merges with the continuous line) represents the $\beta$-equilibrated neutron star matter (present calculations) whereas the dotted line represents the same using the A18 model using variational chain summation (VCS) of Akmal et al. \cite{Ak98} for the PNM. The areas enclosed by the continuous and the dashed lines in Fig.-3 correspond to the pressure regions for neutron matter consistent with the experimental flow data after inclusion of the pressures from asymmetry terms with weak (soft NM) and strong (stiff NM) density dependences, respectively \cite{Da02}. Although, the parameters of the density dependence of DDM3Y interaction have been tuned to reproduce $\rho_0$ and $\epsilon_0$ which are obtained from finite nuclei, the agreement of the present EoS with the experimental flow data, where the high density behaviour looks phenomenologically confirmed, justifies its extrapolation to high density. It is interesting to note that the RMF-NL3 incompressibility for SNM is 271.76 MeV \cite{La97,La99} which is about the same as 274.7$\pm$7.4 MeV obtained from the present calculation but the plot of $P$ versus $\rho/\rho_0$ for PNM of RMF using NL3 parameter set \cite{La97} does not pass through the pressure regions for neutron matter consistent with the experimental flow data \cite{Da02}. The stability of the $\beta$-equilibrated dense matter in neutron stars is investigated and the location of the inner edge of their crusts and core-crust transition density and pressure are determined using the DDM3Y effective nucleon-nucleon interaction \cite{At14}. The stability of any single phase, also called intrinsic stability, is ensured by the convexity of $\epsilon(\rho,x_p)$. The thermodynamical inequalities allow us to express the requirement in terms of $V_{thermal}=\rho^2\Big[2\rho\frac{\partial \epsilon}{\partial \rho}+\rho^2\frac{\partial^2 \epsilon}{\partial \rho^2}-\rho^2\frac{(\frac{\partial^2 \epsilon}{{\partial\rho \partial x_p}})^2}{\frac{\partial^2 \epsilon}{\partial x_p^2}}\Big]$. The condition for core-crust transition is obtained by making $V_{thermal}=0$. The results for the transition density, pressure and proton fraction at the inner edge separating the liquid core from the solid crust of neutron stars are calculated and presented in Table-III for $n$=2/3. The symmetric nuclear matter incompressibility $K_\infty$, nuclear symmetry energy at saturation density $E_{sym}(\rho_0)$, the slope $L$ and isospin dependent part $K_\tau$ of the isobaric incompressibility are also tabulated since these are all in excellent agreement with the constraints recently extracted from measured isotopic dependence of the giant monopole resonances in even-A Sn isotopes, from the neutron skin thickness of nuclei, and from analyses of experimental data on isospin diffusion and isotopic scaling in intermediate energy heavy-ion collisions. \begin{figure}[t] \vspace{0.0cm} \eject\centerline{\epsfig{file=fig05.eps,height=8cm,width=8cm}} \caption{Mass-equatorial radius plot of slowly rotating neutron stars for the DDM3Y EoS. } \label{fig5} \vspace{0.0cm} \end{figure} The r-mode instability occurs when the gravitational-radiation driving time scale of the r-mode is shorter than the time scales of the various dissipation mechanisms that may occur in the interior of a neutron star. The nuclear EoS affects the time scales associated with the r-mode in two different ways viz. EoS defines the radial dependence of the mass density distribution which is the basic ingredient of the relevant integrals and defines the core-crust transition density and core radius which is the upper limit of these integrals. In Fig.-4, plot for mass versus central baryonic density of slowly rotating neutron stars is shown using the DDM3Y EoS for solving the Tolman-Oppenheimer-Volkoff Equation while in Fig.-5, the plot for mass-radius relation is shown. The mass-radius relation can be obtained for fixed values of $\frac{\Delta I}{I}$, the core-crust transition density $\rho_t$ and transition pressure P$_t$. This is then plotted in the same figure for $\frac{\Delta I}{I}$ equal to 0.014. For Vela pulsar, the constraint $\frac{\Delta I}{I}>1.4\%$ implies that allowed mass-radius lie to the right of the line defined by $\frac{\Delta I}{I} = 0.014$ (for $\rho_t$ = 0.0938 fm$^{-3}$ and P$_t$ = 0.5006 MeV fm$^{-3}$) \cite{At17}. The newer observational data \cite{Ho12} on Vela pulsar claims slightly higher estimate for $\frac{\Delta I}{I}>1.6\%$ based on glitch activity. This minute change neither affects the conclusions nor warrants any new idea of the neutron superfluidity extending partially into the core. However, if the phenomenon of crustal entrainment due to the Bragg reflection of unbound neutrons by the lattice ions is taken into account then \cite{An12,Ch13} a much higher fraction of the moment of inertia (7$\%$ instead of 1.4-1.6$\%$) has to be associated to the crust. The only reasonable constraint on mass of Vela pulsar is that it should exceed about one solar mass according to core-collapse supernova simulations. Therefore, the present calculations suggest that without entrainment, the crust is enough to explain the Vela glitch data and with entrainment, the crust is not enough since the mass of Vela pulsar would be below one solar mass (Fig.-5), in accordance with other studies \cite{An12,Ch13,Ho15,De16,Do16}. It is recently conjectured that there may be a good correlation between the core-crust transition density and the symmetry energy slope $L$ and it is predicted that this behaviour should not depend on the relation between $L$ and $K_\tau$ \cite{Du10}. On the contrary, no correlation of the transition pressure with $L$ was obtained \cite{Du10}. In Table-IV, variations of different quantities with parameter $n$ which controls the nuclear matter incompressibility are listed. It is worthwhile to mention here that the incompressibility increases with $n$. The standard value of $n$=2/3 used here has a unique importance because then the constant of density dependence $\beta$ has the dimension of cross section and can be interpreted as the isospin averaged effective nucleon-nucleon interaction cross section in ground state symmetric nuclear medium. For a nucleon in ground state nuclear matter $k_F\approx$ 1.3 fm$^{-1}$ and $q_0 \sim \hbar k_F c \approx$ 260 MeV and the present result for the `in medium' effective cross section is reasonably close to the value obtained from a rigorous Dirac-Brueckner-Hartree-Fock \cite{Sa06} calculations corresponding to such $k_F$ and $q_0$ values which is $\approx$ 12 mb. Using the value of $\beta$=1.5934 fm$^2$ along with the nucleonic density 0.1533 fm$^{-3}$, the value obtained for the nuclear mean free path $\lambda$ is about 4 fm which is in excellent agreement with that obtained using another method \cite{Si83}. Moreover, comparison of the theoretical values of symmetric nuclear matter incompressibility and isobaric incompressibility with the recent experimental values for $K_\infty=250-270$ MeV \cite{Sh09,St14} and $K_\tau=-370\pm120$ MeV \cite{Chen09} further justifies importance for our choice of $n$=2/3. It is interesting to mention here that the present EoS for $n$=2/3, provides the maximum mass for the static case is 1.9227 M$_\odot$ with radius $\sim$9.75 km and for the star rotating with Kepler's frequency it is 2.27 M$_\odot$ with equatorial radius $\sim$13.1 km \cite{Ch10}. However, for stars rotating with maximum frequency limited by the r-mode instability, the maximum mass turns out to be 1.95 (1.94) M$_\odot$ corresponding to rotational period of 1.5 (2.0) ms with radius about 9.9 (9.8) kilometers \cite{APP12} which reconcile with the recent observations of the massive compact stars $\sim$2 M$_\odot$ \cite{De10,An13}. \begin{table*}[htbp] \centering \caption{Variations of the core-crust transition density, pressure and proton fraction for $\beta$-equilibrated neutron star matter, symmetric nuclear matter incompressibility $K_\infty$, isospin dependent part $K_\tau$ of isobaric incompressibility, maximum neutron star mass, corresponding radius and crustal thickness with parameter $n$.} \begin{tabular}{||c|c|c|c|c|c|c|c|c||} \hline \hline $n$&$\rho_t$& P$_t$ & x$_{p(t)}$&$K_\infty$&$K_\tau$&Maximum NS mass & Radius& Crustal thickness \\ \hline &&&&&&&& \\ 1/6&0.0797 fm$^{-3}$&0.4134 MeV fm$^{-3}$& 0.0288&182.13 MeV&-293.42 MeV &1.4336 M$_\odot$&8.5671 km&0.4009 km \\ \hline &&&&&&&&\\ 1/3&0.0855 fm$^{-3}$&0.4520 MeV fm$^{-3}$& 0.0296&212.98 MeV&-332.16 MeV &1.6002 M$_\odot$&8.9572 km&0.3743 km \\ \hline &&&&&&&&\\ 1/2&0.0901 fm$^{-3}$&0.4801 MeV fm$^{-3}$& 0.0303&243.84 MeV&-370.65MeV &1.7634 M$_\odot$&9.3561 km&0.3515 km \\ \hline &&&&&&&&\\ 2/3&0.0938 fm$^{-3}$&0.5006 MeV fm$^{-3}$& 0.0308&274.69 MeV&-408.97 MeV &1.9227 M$_\odot$&9.7559 km&0.3318km \\ \hline &&&&&&&&\\ 1 &0.0995 fm$^{-3}$&0.5264 MeV fm$^{-3}$& 0.0316&336.40 MeV&-485.28 MeV &2.2335 M$_\odot$&10.6408 km&0.3088 km \\ \hline \hline \end{tabular} \label{table4} \vspace{3.5cm} \end{table*} \noindent \noindent \section{Theoretical Calculations} \label{Section 5} The quantity which is of crucial importance in the evaluation of various times scales, as can be seen from Eq.(8) and Eq.(10), is the integral $\int_{0}^{R_c} \rho(r) r^6 dr$. This integral can be re-written in terms of energy density $\epsilon(r)=\rho(r)c^2$ and expressed in dimensionless form as \begin{equation} I({R_{c}})=\int_{0}^{R_c} \left[\frac{\epsilon(r)}{\rm MeV fm^{-3}}\right]\left(\frac{r}{\rm km}\right)^6 d\left(\frac{r}{\rm km}\right) \label{eq43} \end{equation} \begin{figure}[htbp] \vspace{0.0cm} \eject\centerline{\epsfig{file=fig06.eps,height=8cm,width=8cm}} \caption{Plots of fiducial timescales with gravitational mass of neutron stars with DDM3Y EoS.} \label{fig6} \vspace{0.0cm} \end{figure} \begin{figure}[htbp] \vspace{0.0cm} \eject\centerline{\epsfig{file=fig07.eps,height=8cm,width=8cm}} \caption{Plots of critical angular frequency with temperature for different masses of neutron stars.} \label{fig7} \vspace{0.0cm} \end{figure} The fiducial gravitational radiation timescale $\widetilde{\tau}_{GR}$ from Eq.(8) and Eq.(15), is given by \begin{equation} \widetilde{\tau}_{GR}=-0.7429\left[\frac{R}{\rm km}\right]^{9} \left[\frac{1M_{\odot}}{M}\right]^{3}\left[I(R_c)\right]^{-1} ({\rm s}) \label{eq44} \end{equation} where $R$ and $r$ are in km and $M$ in $M_{\odot}$. The fiducial shear viscous timescale $\widetilde{\tau}_{SV}$ for electron-electron scattering and neutron-neutron scattering can be obtained from Eq.(10), Eqs.(12-14) as \begin{eqnarray} \widetilde{\tau}_{ee}=0.1446\times10^8\left[\frac{R}{\rm km}\right]^{3/4}\left[\frac{1M_{\odot}}{M}\right]^{1/4} \left[\frac{\rm km}{R_c}\right]^{6}\nonumber \\ \times\left[\frac{\rm g~cm^{-3}}{\rho_t}\right]^{1/2}\left[\frac{\rm MeV fm^{-3}}{\epsilon_t}\right] \left[I(R_c)\right] ({\rm s}) \label{eq45} \end{eqnarray} \begin{eqnarray} \widetilde{\tau}_{nn}=19\times10^8\left[\frac{R}{\rm km}\right]^{3/4}\left[\frac{1M_{\odot}}{M}\right]^{1/4}\left[\frac{\rm km}{R_c}\right]^{6}\nonumber \\ \times\left[\frac{\rm g~cm^{-3}}{\rho_t}\right]^{5/8}\left[\frac{\rm MeV fm^{-3}}{\epsilon_t}\right] \left[I(R_c)\right] ({\rm s}) \label{eq46} \end{eqnarray} where the transition density $\rho_t$ is expressed in g cm$^{-3}$ and $\epsilon_t$ is the energy density expressed in MeV fm$^{-3}$ at transition density. \begin{figure}[t] \vspace{0.0cm} \eject\centerline{\epsfig{file=fig08.eps,height=8cm,width=8cm}} \caption{Plots of critical temperature versus mass.} \label{fig8} \vspace{0.2cm} \end{figure} \noindent \section{Results and discussion} \label{Section 6} In Fig.-6 plots of the fiducial timescales with the gravitational masses of neutron stars are shown for the DDM3Y EoS. It is seen that the gravitational radiation timescale falls rapidly with increasing mass while the viscous damping timescales increase approximately linearly. By knowing the fiducial gravitational radiation and shear viscous timescales, the temperature T dependence of the critical angular velocity $\Omega_c$ of the r-mode $(l=2)$ can be studied from Eq.(17). In Fig.-7, $\frac{\Omega_c}{\Omega_0}$ is shown as a function of temperature T for several masses of neutron stars for the DDM3Y EoS. The plots act as boundaries of the r-mode instability windows. Neutron stars lying above the plots (whose angular frequency is greater than the critical frequency) possess unstable r-modes and hence emit gravitational waves, thus reducing their angular frequencies. Once their angular velocities reach the critical frequency they enter the region below the plots, where the r-modes become stable and hence stop emitting gravitational radiation. In computing the instability windows in Fig.-7, the fiducial shear viscous timescale $\widetilde \tau_{ee}$ given in Eq.(45) is substituted for $\widetilde \tau_{SV}$ in Eq.(17) for temperatures T $\leq$ $10^9$ K and $\tau_{nn}$ from Eq.(46) is used for T $>$ $10^9$ K. Fig.-8 depicts the plot of the critical temperature as a function of mass. The electron-electron scattering shear viscosity timescale is used for the calculation of $T_c$. We see that the critical temperature rapidly decreases with mass. The explanation is straightforward. From Fig.-7 we see that for fixed $T$, $\frac{\Omega_c}{\Omega_0}$ rapidly decreases with increasing mass. Since $T=T_c$ when $\Omega_c=\Omega_K$ and $\Omega_K$ rapidly increases with mass and hence $T_c$ falls, vide Eq.(19). \begin{figure}[t] \vspace{0.0cm} \eject\centerline{\epsfig{file=fig09.eps,height=8cm,width=8cm}} \caption{Plots of time evolution of frequencies.} \label{fig9} \vspace{0.0cm} \end{figure} From Fig.-7 and Fig.-8 we see that the critical frequency and critical temperature decrease with mass and hence the r-mode instability window increases with the same. This means that for the same EoS and temperature, the massive configurations are more probable to r-mode instability and hence emission of gravitational waves than the less massive ones. This can be indirectly inferred from Fig.-6 where $\widetilde \tau_{GR}$ is much less than $\widetilde \tau_{ee}$ and $\widetilde \tau_{nn}$ for massive neutron stars and vice-versa for low mass neutron stars. Hence isolated young massive neutron stars have high probability for emission of gravitational waves through r-mode instability. It is worth noting that $\Omega_c$ is dependent on the density dependence of the symmetry energy and thus on $L$. Again, $R$, $R_c$, $I(R_c)$ and $\rho_t$ depend on $L$. Hence, for a fixed mass and temperature, $\Omega_c$ is dependent on the above parameters via the relation, \begin{equation} \Omega_c \sim \frac{R_c^{12/11}}{[I(R_c)]^{4/11}}\rho_t^{3/11} \label{eq47} \end{equation} In our case $L$, $\rho_t$ and $R_c$ are constants for a fixed neutron star mass and temperature. As a neutron star enters into the instability region due to accretion of mass from its companion, the amplitude of the r-mode $\alpha_r$ increases till reaching a saturation value. At this point the neutron star emits gravitational wave and releases its angular momentum and energy and spins down to the region of stability. Using the ideal condition that the decrease in temperature due to emission of gravitational wave is compensated by the heat produced due to viscous effects, the time evolution of spin angular velocity and spin down rate can be calculated for a neutron star from Eq.(23) and Eq.(25), respectively, provided M, T, $\Omega_{in}$ and $\alpha_r$ of the star is known. For the schematic values $\nu_{in}=\frac{\Omega_{in}}{2\pi}=700$ Hz and $\alpha_r=2 \times 10^{-7}$ used by Moustakidis \cite{Moustakidis2015}, the evolutions of spin are calculated for various neutron star masses and shown in Fig.-9. In Fig.-10 the spin down rates has been shown for these masses. In Fig.-11 the spin down rates as functions of spin frequency are shown. Some mention is to be made about the dependency of the critical frequency $\Omega_c$ on the symmetry energy slope parameter $L$. Although the slope $L$ depends on the strengths and ranges of the Yukawas for the DDM3Y EoS, it does not depend on the power of the density dependence $n$ and has a constant value of 45.1066 MeV. In a recent work, the critical frequency as a function of $L$ of the pulsar 4U 1608-52 was plotted using an estimated core temperature $\sim 4.55\times 10^8$ K and with different models of the EoS. In accordance with the Fig.-6 of \cite{Vidana2012}, using the measured spin frequency and the estimated core temperature, if the mass of 4U 1608-52 is $1.4M_\odot$ then it should marginally be unstable ($\Omega_c$ is smaller than its spin frequency), since the radius obtained from our mass-radius relation (Fig.-5) is $\sim 11.55$ kms and higher than 11.5 kms. In case of the highest mass configuration of 1.9227 M$_\odot$ with a radius of $\sim$9.75 kms, it is also likely to be in the instability region as $L<50$ MeV for our EoS. Thus we stress the fact that the r-mode instability window is enlarged for isolated neutron stars with a rigid crust if we consider the dissipation to be at the crust-core interface in agreement with \cite{We12}. \begin{figure}[t] \vspace{0.0cm} \eject\centerline{\epsfig{file=fig010.eps,height=8cm,width=8cm}} \caption{Plots of time evolution of spin-down rates.} \label{fig10} \vspace{0.0cm} \end{figure} \begin{figure}[t] \vspace{0.0cm} \eject\centerline{\epsfig{file=fig011.eps,height=8cm,width=8cm}} \caption{Plots of spin-down rates versus frequencies.} \label{fig11} \vspace{-0.07cm} \end{figure} The calculations are performed for five different $n$ values that correspond to SNM incompressibility ranging from $\sim$180-330 MeV. For each case, the constants $C$ and $\beta$ obtained by reproducing the ground state properties of SNM become different leading to five different sets of these three parameters. We certainly can not change strengths and ranges of the M3Y interaction. In Table-IV, the variations of the core-crust transition density, pressure and proton fraction for $\beta$-equilibrated neutron star matter, symmetric nuclear matter incompressibility $K_\infty$, isospin dependent part $K_\tau$ of isobaric incompressibility, neutron star's maximum mass with corresponding radius and crustal thickness with parameter $n$ are listed. It is important to mention here that recent observations of the binary millisecond pulsar J1614-2230 by P. B. Demorest et al. \cite{De10} suggest that the masses lie within $1.97\pm0.04$ M$_\odot$. Recently the radio timing measurements of the pulsar PSR J0348 + 0432 and its white dwarf companion have confirmed the mass of the pulsar to be in the range 1.97-2.05 M$_\odot$ at 68.27$\%$ or 1.90-2.18 M$_\odot$ at 99.73$\%$ confidence \cite{An13}. The observed $1.97\pm0.04$ M$_\odot$ neutron star rotates with 3.1 ms and results quoted in Table-5.2 are for non-rotating case. For rotating stars \cite{BCM14} present EoS predict masses higher than the lower limit of 1.93 M$_\odot$ for maximum mass of neutron stars. We used the same value of $\rho_0=0.1533$ fm$^{-3}$ since we wanted to keep consistency with all our previous works on nuclear matter. We would like to mention that if instead we would have used the value of 0.16 fm$^{-3}$ for $\rho_0$, the value of $K_\infty$ would have been slightly higher by $\sim$2 MeV and correspondingly maximum mass of neutron stars by $\sim$ 0.01 M$_\odot$. \noindent \section{Summary and conclusions} \label{Section 7} In the present work we have studied the r-mode instability of slowly rotating neutron stars with rigid crusts with their EoS obtained from the DDM3Y effective nucleon-nucleon interaction. This EoS provides good descriptions for proton, $\alpha$ and cluster radioactivities, elastic and inelastic scattering, symmetric and isospin asymmetric nuclear matter and neutron star masses and radii \cite{BCM14}. We have calculated the fiducial gravitational radiation and shear viscosity timescales within the DDM3Y framework for a wide range of neutron star masses. It is observed that the gravitational radiation timescale decreases rapidly with increasing neutron star mass while the viscous damping timescales exhibit an approximate linear increase. Next, we have studied the temperature dependence of the critical angular frequency for different neutron star masses. The implication is that for neutron stars rotating with frequencies greater than their corresponding critical frequencies have unstable r-modes leading to the emission of gravitational waves. Further, our study of the variation of the critical temperature as a function of mass shows that both the critical frequency and temperature decrease with increasing mass. The conclusion is that massive hot neutron stars are more susceptible to r-mode instability through gravitational radiation. Finally we have calculated the spin down rates and angular frequency evolution of the neutron stars through r-mode instability. We have also pointed out the fact that the critical frequency depends on the EoS through the radius and the symmetry energy slope parameter $L$. If the dissipation of r-modes from shear viscosity acts along the boundary layer of the crust-core interface then the r-mode instability region is enlarged to lower values of $L$.

\section{Introduction} Please follow the steps outlined below when submitting your manuscript to the IEEE Computer Society Press. This style guide now has several important modifications (for example, you are no longer warned against the use of sticky tape to attach your artwork to the paper), so all authors should read this new version. \subsection{Language} All manuscripts must be in English. \subsection{Dual submission} Please refer to the author guidelines on the CVPR 2017 web page for a discussion of the policy on dual submissions. \subsection{Paper length} Papers, excluding the references section, must be no longer than eight pages in length. The references section will not be included in the page count, and there is no limit on the length of the references section. For example, a paper of eight pages with two pages of references would have a total length of 10 pages. {\bf There will be no extra page charges for CVPR 2017.} Overlength papers will simply not be reviewed. This includes papers where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide. Note that this \LaTeX\ guide already sets figure captions and references in a smaller font. The reason such papers will not be reviewed is that there is no provision for supervised revisions of manuscripts. The reviewing process cannot determine the suitability of the paper for presentation in eight pages if it is reviewed in eleven. \subsection{The ruler} The \LaTeX\ style defines a printed ruler which should be present in the version submitted for review. The ruler is provided in order that reviewers may comment on particular lines in the paper without circumlocution. If you are preparing a document using a non-\LaTeX\ document preparation system, please arrange for an equivalent ruler to appear on the final output pages. The presence or absence of the ruler should not change the appearance of any other content on the page. The camera ready copy should not contain a ruler. (\LaTeX\ users may uncomment the \verb'\cvprfinalcopy' command in the document preamble.) Reviewers: note that the ruler measurements do not align well with lines in the paper --- this turns out to be very difficult to do well when the paper contains many figures and equations, and, when done, looks ugly. Just use fractional references (e.g.\ this line is $095.5$), although in most cases one would expect that the approximate location will be adequate. \subsection{Mathematics} Please number all of your sections and displayed equations. It is important for readers to be able to refer to any particular equation. Just because you didn't refer to it in the text doesn't mean some future reader might not need to refer to it. It is cumbersome to have to use circumlocutions like ``the equation second from the top of page 3 column 1''. (Note that the ruler will not be present in the final copy, so is not an alternative to equation numbers). All authors will benefit from reading Mermin's description of how to write mathematics: \url{}. \subsection{Blind review} Many authors misunderstand the concept of anonymizing for blind review. Blind review does not mean that one must remove citations to one's own work---in fact it is often impossible to review a paper unless the previous citations are known and available. Blind review means that you do not use the words ``my'' or ``our'' when citing previous work. That is all. (But see below for techreports.) Saying ``this builds on the work of Lucy Smith [1]'' does not say that you are Lucy Smith; it says that you are building on her work. If you are Smith and Jones, do not say ``as we show in [7]'', say ``as Smith and Jones show in [7]'' and at the end of the paper, include reference 7 as you would any other cited work. An example of a bad paper just asking to be rejected: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of our previous paper [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Removed for blind review \end{quote} An example of an acceptable paper: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of the paper of Smith \etal [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Smith, L and Jones, C. ``The frobnicatable foo filter, a fundamental contribution to human knowledge''. Nature 381(12), 1-213. \end{quote} If you are making a submission to another conference at the same time, which covers similar or overlapping material, you may need to refer to that submission in order to explain the differences, just as you would if you had previously published related work. In such cases, include the anonymized parallel submission~\cite{Authors14} as additional material and cite it as \begin{quote} [1] Authors. ``The frobnicatable foo filter'', F\&G 2014 Submission ID 324, Supplied as additional material {\tt fg324.pdf}. \end{quote} Finally, you may feel you need to tell the reader that more details can be found elsewhere, and refer them to a technical report. For conference submissions, the paper must stand on its own, and not {\em require} the reviewer to go to a techreport for further details. Thus, you may say in the body of the paper ``further details may be found in~\cite{Authors14b}''. Then submit the techreport as additional material. Again, you may not assume the reviewers will read this material. Sometimes your paper is about a problem which you tested using a tool which is widely known to be restricted to a single institution. For example, let's say it's 1969, you have solved a key problem on the Apollo lander, and you believe that the CVPR70 audience would like to hear about your solution. The work is a development of your celebrated 1968 paper entitled ``Zero-g frobnication: How being the only people in the world with access to the Apollo lander source code makes us a wow at parties'', by Zeus \etal. You can handle this paper like any other. Don't write ``We show how to improve our previous work [Anonymous, 1968]. This time we tested the algorithm on a lunar lander [name of lander removed for blind review]''. That would be silly, and would immediately identify the authors. Instead write the following: \begin{quotation} \noindent We describe a system for zero-g frobnication. This system is new because it handles the following cases: A, B. Previous systems [Zeus et al. 1968] didn't handle case B properly. Ours handles it by including a foo term in the bar integral. ... The proposed system was integrated with the Apollo lunar lander, and went all the way to the moon, don't you know. It displayed the following behaviours which show how well we solved cases A and B: ... \end{quotation} As you can see, the above text follows standard scientific convention, reads better than the first version, and does not explicitly name you as the authors. A reviewer might think it likely that the new paper was written by Zeus \etal, but cannot make any decision based on that guess. He or she would have to be sure that no other authors could have been contracted to solve problem B. FAQ: Are acknowledgements OK? No. Leave them for the final copy. \begin{figure}[t] \begin{center} \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}} \end{center} \caption{Example of caption. It is set in Roman so that mathematics (always set in Roman: $B \sin A = A \sin B$) may be included without an ugly clash.} \label{} \label{} \end{figure} \subsection{Miscellaneous} \noindent Compare the following:\\ \begin{tabular}{ll} \verb'$conf_a$' & $conf_a$ \\ \verb'$\mathit{conf}_a$' & $\mathit{conf}_a$ \end{tabular}\\ See The \TeX book, p165. The space after \eg, meaning ``for example'', should not be a sentence-ending space. So \eg is correct, {\em e.g.} is not. The provided \verb'\eg' macro takes care of this. When citing a multi-author paper, you may save space by using ``et alia'', shortened to ``\etal'' (not ``{\em et.\ al.}'' as ``{\em et}'' is a complete word.) However, use it only when there are three or more authors. Thus, the following is correct: `` Frobnication has been trendy lately. It was introduced by Alpher~\cite{Alpher02}, and subsequently developed by Alpher and Fotheringham-Smythe~\cite{Alpher03}, and Alpher \etal~\cite{Alpher04}.'' This is incorrect: ``... subsequently developed by Alpher \etal~\cite{Alpher03} ...'' because reference~\cite{Alpher03} has just two authors. If you use the \verb'\etal' macro provided, then you need not worry about double periods when used at the end of a sentence as in Alpher \etal. For this citation style, keep multiple citations in numerical (not chronological) order, so prefer \cite{Alpher03,Alpher02,Authors14} to \cite{Alpher02,Alpher03,Authors14}. \begin{figure*} \begin{center} \fbox{\rule{0pt}{2in} \rule{.9\linewidth}{0pt}} \end{center} \caption{Example of a short caption, which should be centered.} \label{} \end{figure*} \section{Formatting your paper} All text must be in a two-column format. The total allowable width of the text area is $6\frac78$ inches (17.5 cm) wide by $8\frac78$ inches (22.54 cm) high. Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch (0.8 cm) space between them. The main title (on the first page) should begin 1.0 inch (2.54 cm) from the top edge of the page. The second and following pages should begin 1.0 inch (2.54 cm) from the top edge. On all pages, the bottom margin should be 1-1/8 inches (2.86 cm) from the bottom edge of the page for $8.5 \times 11$-inch paper; for A4 paper, approximately 1-5/8 inches (4.13 cm) from the bottom edge of the page. \subsection{Margins and page numbering} All printed material, including text, illustrations, and charts, must be kept within a print area 6-7/8 inches (17.5 cm) wide by 8-7/8 inches (22.54 cm) high. \subsection{Type-style and fonts} Wherever Times is specified, Times Roman may also be used. If neither is available on your word processor, please use the font closest in appearance to Times to which you have access. MAIN TITLE. Center the title 1-3/8 inches (3.49 cm) from the top edge of the first page. The title should be in Times 14-point, boldface type. Capitalize the first letter of nouns, pronouns, verbs, adjectives, and adverbs; do not capitalize articles, coordinate conjunctions, or prepositions (unless the title begins with such a word). Leave two blank lines after the title. AUTHOR NAME(s) and AFFILIATION(s) are to be centered beneath the title and printed in Times 12-point, non-boldface type. This information is to be followed by two blank lines. The ABSTRACT and MAIN TEXT are to be in a two-column format. MAIN TEXT. Type main text in 10-point Times, single-spaced. Do NOT use double-spacing. All paragraphs should be indented 1 pica (approx. 1/6 inch or 0.422 cm). Make sure your text is fully justified---that is, flush left and flush right. Please do not place any additional blank lines between paragraphs. Figure and table captions should be 9-point Roman type as in Figures~\ref{} and~\ref{}. Short captions should be centred. \noindent Callouts should be 9-point Helvetica, non-boldface type. Initially capitalize only the first word of section titles and first-, second-, and third-order headings. FIRST-ORDER HEADINGS. (For example, {\large \bf 1. Introduction}) should be Times 12-point boldface, initially capitalized, flush left, with one blank line before, and one blank line after. SECOND-ORDER HEADINGS. (For example, { \bf 1.1. Database elements}) should be Times 11-point boldface, initially capitalized, flush left, with one blank line before, and one after. If you require a third-order heading (we discourage it), use 10-point Times, boldface, initially capitalized, flush left, preceded by one blank line, followed by a period and your text on the same line. \subsection{Footnotes} Please use footnotes\footnote {This is what a footnote looks like. It often distracts the reader from the main flow of the argument.} sparingly. Indeed, try to avoid footnotes altogether and include necessary peripheral observations in the text (within parentheses, if you prefer, as in this sentence). If you wish to use a footnote, place it at the bottom of the column on the page on which it is referenced. Use Times 8-point type, single-spaced. \subsection{References} List and number all bibliographical references in 9-point Times, single-spaced, at the end of your paper. When referenced in the text, enclose the citation number in square brackets, for example~\cite{Authors14}. Where appropriate, include the name(s) of editors of referenced books. \begin{table} \begin{center} \begin{tabular}{|l|c|} \hline Method & Frobnability \\ \hline\hline Theirs & Frumpy \\ Yours & Frobbly \\ Ours & Makes one's heart Frob\\ \hline \end{tabular} \end{center} \caption{Results. Ours is better.} \end{table} \subsection{Illustrations, graphs, and photographs} All graphics should be centered. Please ensure that any point you wish to make is resolvable in a printed copy of the paper. Resize fonts in figures to match the font in the body text, and choose line widths which render effectively in print. Many readers (and reviewers), even of an electronic copy, will choose to print your paper in order to read it. You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic. When placing figures in \LaTeX, it's almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below {\small\begin{verbatim} \usepackage[dvips]{graphicx} ... \includegraphics[width=0.8\linewidth] {myfile.eps} \end{verbatim} } \subsection{Color} Please refer to the author guidelines on the CVPR 2017 web page for a discussion of the use of color in your document. \section{Final copy} You must include your signed IEEE copyright release form when you submit your finished paper. We MUST have this form before your paper can be published in the proceedings. Please direct any questions to the production editor in charge of these proceedings at the IEEE Computer Society Press: Phone (714) 821-8380, or Fax (714) 761-1784. {\small \bibliographystyle{ieee} \section*{Acknowledgments} \noindent This publication has emanated from research conducted with the financial support of Science Foundation Ireland (SFI) under grant number SFI/12/RC/2289 and SFI/15/SIRG/3283. This research was supported by contract SGR1421 by the Catalan AGAUR office. The work has been developed in the framework of project TEC2016-75976-R, funded by the Spanish Ministerio de Economia y Competitividad and the European Regional Development Fund (ERDF). The authors also thank NVIDIA for generous hardware donations. \section{Introduction} Content-based image retrieval has been mostly tackled as the problem of instance-level image retrieval~\cite{Oxford5k,tembedding,vlad}. The notion of an ``instance'' specifically limits the search to ``one'' instance of a semantic class (i.e to retrieve instances of one specific object, person or location as in~\cite{Awad2017}). This contrasts with less specific retrieval pipelines in which any instance of any member of a class will satisfy the search. \begin{figure}[t!] \begin{center} \includegraphics[width=0.48\textwidth]{Figures/figure1__H} \caption{Top 5 retrieved images for a query in INSTRE dataset using different spatial weighting schemes with BLCF. Correct results are highlighted in green, irrelevant images in red.} \label{} \end{center} \end{figure} Retrieval methods based on convolutional neural networks (CNN) have become increasingly popular in recent years. Off-the-shelf CNN representations extracted from convolution layers can be directly aggregated via spatial max/sum-pooling generating a global image representation that achieves a good precision and recall performance for a reasonable footprint~\cite{BabenkoL15,visualInstanceRazavianSMC14}. Moreover, when suitable training data is available, CNN models can be trained for similarity learning to adapt the learned representations to the end task of retrieval as well as to the target image domain (e.g. landmarks as in~\cite{gordo2016deep,Radenović2016}). However, the fine-tuning process involves the collection, annotation, and cleaning of a suitable training dataset, which is not always feasible. Furthermore, in generic instance search scenarios the target instances are unknown, which potentially makes off-the-shelf CNN representations more adequate for the task. In this work we propose using state-of-the-art saliency models to weight the contribution of local convolutional features prior to aggregation. The notion of saliency in recent CNN-based image retrieval works has referred to the most active regions obtained from a specific filter within a CNN such as in~\cite{CaoLWHSS16,Jimenez_2017_BMVC}, or to a 2D feature saliency map derived from the convolutional layers~\cite{QAwrmac17,CroW,}. In these cases, salient regions are detected with the same CNN from which image representations are extracted. On the other hand, human-based saliency models, such as those derived from the popular Itti and Koch model~\cite{itti1998model}, have been applied only in image retrieval pipelines based in local handcrafted features~\cite{Wu2017}. To the best of our knowledge, it has not yet been explored how the most recent CNN-based retrieval pipelines can benefit from modern saliency models, and how different models affect the retrieval performance. The contributions of this paper are the following: \begin{itemize} \item We propose a novel approach to instance search combining saliency weighting over off-the-shelf convolutional features which are aggregated using a large vocabulary with a bag of words model. The bags of local convolutional features (BLCF) scheme has the advantage of retaining the sparsity and scalability of the original bag of words model while giving improved performance across multiple benchmarks. \item We demonstrate that this weighting scheme gives outperforms all other state of the art on the challenging INSTRE benchmark, without requiring any feature fine tuning. Furthermore, it also offers comparable or better performance to more complicated encoding schemes for off-the-shelf features on Oxford and Paris benchmarks. \item We investigate the impact of the specific visual attention model used, and show that, perhaps surprisingly, higher performance on saliency benchmarks does not necessary equate to improved performance when used in the instance search task. \end{itemize} The remainder of the paper is organized as follows: Section~\ref{} presents recent CNN-based retrieval methods, Section~\ref{} describes the proposed retrieval pipeline, Section~\ref{} introduces the different saliency models evaluated, and finally, Section~\ref{} presents the datasets and retrieval results obtained. \section{Related Work} \label{} Several works have investigated the usage of convolutional layers for retrieval. Early approaches focused on the usage of fully connected layers as global image representation~\cite{neuralcodes}, that when combined with spatial search and data augmentation strategies~\cite{cnnofftheshelf} achieved a significant boost in performance in several retrieval benchmarks. This improvement came at the cost of significantly increasing the memory and computational requirements, meaning that the the system was not scalable to larger more realistic scenarios. A second generation of works explored the convolutional layers to derive image representations for retrieval. In these approaches, the activations of the different neuron arrays across all feature maps in a convolutional layer are treated as local features. Babenko~\cite{BabenkoL15} have shown that a simple spatial sum pooling on the convolutional layer outperformed fully connected layers using pre-trained CNN models for retrieval for a reasonable footprint of 512D. Furthermore, the power of this representation could be enhanced by applying a simple Gaussian center prior scheme to spatially weight the contribution of the activations on a convolutional layer prior to aggregation. Following a similar line, Kalantidis~\cite{CroW} proposed a cross dimensional weighting scheme (CroW) that consisted of the channel sparsity of a convolutional layer and a spatial weighting based in the $L^2$-norm of the local convolutional features. The spatial weighting scheme directly depended on the image content, which was shown to be a more effective approach than the fixed center prior approach. A different line are the works performing region analysis using convolutional features. Tolias \textit{et al.}~\cite{rmac-ToliasSJ15} proposed the popular regional of maximum activation (R-MAC). This is an aggregation method for convolutional features that, given a convolutional layer, generates a set of regional vectors by performing spatial max-pooling within a particular region. The regions are individually post-processed (by $L^2$-normalization, followed by a PCA whitening, and a second $L^2$-normalization), and aggregated via sum pooling into a single compact representation. One of the main limitations of the approach is using a fixed grid of locations. Several works has been proposed to overcome the R-MAC limitations. Jimenez \textit{et al.}~\cite{Jimenez_2017_BMVC} explored the Class Activation Maps (CAMs) to improve over the fix region strategy of R-MAC. CAMs generates a set of spatial maps highlighting the contribution of the areas within an image that are most relevant to classify an image as a particular class. In their approach, each of the maps is used to weight the convolutional features and generate a set class vectors, which are post-processed and aggregated as the region vectors in R-MAC. Similarly, Cao \textit{et al.}~\cite{CaoLWHSS16} proposed a method to derive a set of base regions directly from the convolutional layer, followed with a query adaptive re-ranking strategy. Lanskar and Kannala~\cite{QAwrmac17} used a saliency measure directly derived from the convolutional features to weight the contribution of the regions of R-MAC prior to aggregation. Their proposed saliency measure consists in sum aggregating the feature maps over the channel dimension. Similarly, the very recent work of Simeoni \textit{et al} ~\cite{} propose using a saliency measure also derived from a convolutional layer that consists in a weighted sum across channels where weights represent the channel sparsity. They propose a method to detect a set of rectangular regions directly from the feature saliency map, improving over the uniform sampling of R-MAC. The current state-of-the-art retrieval approaches~\cite{gordo2016deep,Radenović2016} use models fine-tuned with a ranking loss. In particular, Gordo~\emph{et al.}~\cite{gordo2016deep} improves over the original R-MAC encoding by learning a region proposal network~\cite{FastERGirshick15} for the task of landmark retrieval. Region proposal and feature learning are optimized end-to-end achieving excellent performance in the popular Paris, Oxford, and Holidays datasets. This approach, however, requires the construction of a suitable training dataset, which is usually domain specific, time consuming to construct and it is unclear how those models generalize in more generic scenarios such as the INSTRE dataset. Intuitively, one can see that pooling descriptors from different regions and then subsequently pooling all of the resulting together is similar to applying a weighting scheme to the original convolutional features since region descriptors and globally pooled descriptors are built from the same set of local features. Figure~\ref{} visualizes the region sampling of the original R-MAC~\cite{rmac-ToliasSJ15} as a spatial weighting strategy (without considering the region post-processing). The approach provides more emphasis to the center located CNN features. In contrast to the approaches in the literature, we propose a simple yet effective CNN-based pipeline that does not require any region analysis. We demonstrate the generalization of our approach across different retrieval domains, resulting in a scalable retrieval system that does not require any additional feature finetuning. \begin{figure}[t!] \begin{center} \includegraphics[width=.48\textwidth]{Figures/state_of_the} \caption{Spatial weighting schemes used in the literature on three INSTRE images. R-MAC visualizes the window location strategy used in~\cite{rmac-ToliasSJ15}, Gaussian center prior is used in~\cite{BabenkoL15}, saliency measure derived from the local convolutional features used in~\cite{CroW,QAwrmac17,} (we visualize the $L^2$-norm of each local feature), and human-based saliency prediction used in this work from SalGAN model~\cite{SalGAN}.} \label{} \end{center} \end{figure} \section{Saliency models} \label{} \begin{figure*}[t] \begin{center} \includegraphics[width=1.0\textwidth]{Figures/diagram2} \caption{Three sample query images from Oxford (top row), Paris (middle row), and INSTRE (bottom row) with their associated spatial weighting schemes: From left to right weighting schemes based on Gaussian center prior, $L^2$-norm of convolutional features, IttiKoch model, BMS, SalNet, SalGAN, SAM-VGG, and SAM-ResNet.} \label{} \end{center} \end{figure*} One of the main features of the human vision system is its capacity to actively focus on the most salient regions and movements. Visual attention models detect regions that stand out from their surroundings, producing saliency maps that highlight the most prominent regions within an image. Saliency has been shown beneficial in content-based image retrieval when using traditional the Bag of Words model and SIFT features~\cite{Wu2017}. Some works have investigated the usage of saliency as a mechanism to reduce the number of local features and reduce the computational complexity of SIFT-based BoW frameworks~\cite{awad2012saliency,liang2010salient,nakamoto2011combination}. Other works, rather than completely discard the background information, have used saliency to weight the contribution of the foreground and the background simultaneously~\cite{de2012spatial,wu2017visual}. However, the usage of saliency models in the task of image retrieval has thus far been restricted to handcrafted CBIR features and handcrafted saliency models~\cite{itti1998model}. With the emergence of challenges such as the MIT saliency benchmark~\cite{MIT300} and the Large-Scale Scene Understanding Challenge~\cite{LSUNchallenge} (LSUN), and the appearance of large-scale public annotated datasets such as iSUN~\cite{iSUN}, SALICON~\cite{SALICON}, MIT300~\cite{MIT300}, or CAT2000~\cite{CAT2000}, data-driven approaches based on CNN models trained end-to-end have become dominant, generating more accurate models each year~\cite{Deepgaze,SALICON,Deepfix,SalGAN,SalNet}. Having access to large datasets has not only allowed the development of deep learning based saliency models, but has also enabled a better comparison of the performance between models. Our goals are to first evaluate whether saliency is useful in a CNN-based retrieval system, and second, to examine how the performance of different saliency models on saliency benchmarks relates to performance of instance search. To this end, we select two handcrafted models and four deep learning based models. The handcrafted models selected are the classic approach proposed by Itti and Koch~\cite{itti1998model}, and the Boolean Map based saliency (BMS)~\cite{zhang2013saliency}, which represents the best performing non-deep learning algorithm on the MIT300 benchmark. The four deep learning based models differ in CNN architecture, the loss functions used in training, and the training strategy~\cite{SalNet,SalGAN,cornia2016predicting}, providing a reasonable cross section of the current state-of-the-art. Table~\ref{} summarizes the saliency prediction performance of the selected models, where we report the the AUC-Judd value (which is variant of the Area Under the Curve), the Pearson's linear correlation coefficient (CC), the Normalized Scanpath Saliency (NSS), and the Kullback-Leibler divergence (KL) to evaluate how accurate the saliency prediction is when comparing with the eye-tracking ground truth in the MIT300 benchmark. \begin{table}[b] \centering \caption{Performance of CNN saliency models on the MIT300 saliency benchmark~\cite{MIT300}.\vspace{0.1cm}} \label{} \begin{tabular}{@{}p{3.9cm}llll@{}} \toprule Saliency model & \begin{tabular}[c]{@{}l@{}}AUC\\ Judd\end{tabular} & CC & NSS & KL \\ \midrule Human & 0.92 & 1.00 & 3.29 & 0.00 \\ SAM-ResNet~\cite{cornia2016predicting} & 0.87 & 0.78 & 2.34 & 1.27 \\ SAM-VGG~\cite{cornia2016predicting} & 0.87 & 0.77 & 2.30 & 1.13 \\ SalGAN~\cite{SalGAN} & 0.86 & 0.73 & 2.04 & 1.07 \\ SalNet~\cite{SalNet} & 0.83 & 0.58 & 1.51 & 0.81 \\ BMS~\cite{zhang2013saliency} & 0.83 & 0.55 & 1.41 & 0.81 \\ IttiKoch~\cite{walther2006modeling} & 0.60 & 0.14 & 0.43 & 2.30 \\ \bottomrule \end{tabular} \end{table} The \textbf{Itti-Koch model}~\cite{walther2006modeling} model is inspired by how the primary visual cortex processes visual information. It explicitly models local variations on three early visual features to generate the final saliency maps. A salient region represents a region presenting a clear contrast with its surroundings in terms of color, intensity, and/or orientation. The proposed algorithm analyzes the image at eight different scales and computes center surround difference between fine and coarse levels. Three ``conspicuity'' maps, one for each visual feature, are obtained by normalizing and performing across-scale addition. The final saliency map is generated by a linear combination of the three normalized conspicuity maps. The final step of the model includes a winner-take-all (WTA) network that simulates a scan path over the image in order of decreasing saliency of attended locations. We use the MATLAB implementation from Walther and Koch\footnote{}, which extends the work of Itti-Koch~\cite{itti1998model} to generate saliency regions with the extent of the objects contained in the image. \textbf{Boolean Map based Saliency}~\cite{zhang2013saliency} generates a set of Boolean maps by randomly thresholding the different color channels of an image. Then, an attention map for each Boolean map is computed by assigning 1 to the union of surrounded regions, and 0 to the rest of the map. The attention maps are normalized (with a dilatation operation and a $L^2$-normalization) and linearly combined into a full-resolution mean attention map. A Gaussian blur is applied on the obtained map generating the final saliency prediction. Despite its simplicity, this method achieves the best performance among other hand-crafted approaches in the saliency MIT300 benchmark. \textbf{Deep Salnet}~\cite{SalNet} is a end-to-end deep learning model trained to predict human eye fixations. It uses a fully convolutional network consisting of ten layers: one input layer, eight convolution layers, and one deconvolution layer~\cite{deconvolution}. The ReLU is used as an activation function and pooling layers follow the first two convolutional layers, effectively reducing the width and height of the feature maps in the intermediate layers by a factor of four. The final deconvolution generates the final prediction, up-sampling the resolution to the original input size. A transfer learning strategy is applied on this network by re-using and adapting the weights from the first three convolutional layers from a pre-trained VGG architecture~\cite{vgg}. This acts as a regularizer and improves the final network result. The loss function used during training is the mean squared error computed pixel-wise between the saliency prediction and the ground truth map. \begin{figure*}[t] \begin{center} \includegraphics[width=1.0\textwidth]{Figures/las2} \caption{Effect on performance for five examples from the INSTRE dataset after applying saliency weighting (SalGAN). The first column depicts the query with its associated average precision (AP). On the left, AP when performing unweighted BLCF, and on the right when performing saliency weighted BLCF. Retrieved images are ranked in decreasing order of ranking difference between unweighted BLCF and saliency weighted BLCF. For each image, precision at the position where each image is retrieved is reported for unweighted BLCF (left) and saliency weighted BLCF (right).} \label{} \end{center} \end{figure*} The \textbf{SalGAN}~\cite{SalGAN} model introduces the use of adversarial training~\cite{goodfellow2014generative} for the task of visual saliency prediction. The model is the \textit{generator} module in the adversarial setup. It follows the VGG-16~\cite{vgg} architecture without the fully connected layers, and it is initially trained using binary cross entropy loss to predict eye fixations, similarly to~\cite{SalNet}. The \textit{decoder} architecture is structured in the same way as the encoder, but in reverse order and with pooling layers replaced by upsampling layers. The weights of the decoder are randomly initialized. The final loss used for the adversarial training is a combination of the a content loss (binary-cross entropy of predicted saliency maps and the saliency ground truth) and an adversarial loss quantifying the error of the discriminator. The \textbf{Saliency Attentive Model}~\cite{cornia2016predicting} is a CNN model whose distinctive feature is to include an attentive mechanism based on Long-Short Term Memory (LSTM) to iteratively refine predicted saliency maps. The authors also explicitly model the tendency of looking towards the center of images~\cite{tatler2007central} by including a set of trainable Gaussian priors on their model. The base CNN model uses dilated convolutions~\cite{yu2015multi} to better preserve the spatial information across layers, and its weights are initialized on ImageNet. Parameters of the attention mechanism and center-priors are fit for saliency prediction using a loss function that consists of a combination of three popular salience estimation measurements (NSS, CC, and KL). Two models are proposed, one based on a VGG16-like architecture~\cite{vgg} (SAM-VGG), and a second based on ResNet-50~\cite{He2016DeepRL} (SAM-ResNet). Figure~\ref{} shows the different spatial weights evaluated in this work. Additionally, we evaluate a fixed spatial weighting approach (Gaussian center-prior~\cite{BabenkoL15}) and a weighting scheme derived from the activations of the convolutional layer (by measuring the $L^2$-norm of each local CNN representation similar to~\cite{CroW}). \section{Bags of Local Convolutional Features} \label{} Our pipeline extends the work proposed in~\cite{mohedano2016bags}, that consists of using the traditional bag of visual words encoding on local pre-trained convolutional features. The bags of local convolutional features (BLCF) has the advantage of generating high-dimensional and sparse representations. The high-dimensionality makes the aggregated local information more likely to be linearly separable, while relatively few non-zero elements means that they are efficient in terms of storage and computation, as shown in Table~\ref{}. \begin{table}[t!] \center \caption{Average number of visual words per image in 25k BoW representation on different benchmarks. We also report the query time (seconds).\vspace{0.1cm}} \label{} \begin{tabular}{@{}p{3.7cm}cccc@{}} \toprule & INSTRE & Oxford & Paris \\ \midrule words/image & 289 & 174 & 163 \\ query/time & 0.019 & 0.002 & 0.003 \\ \bottomrule \end{tabular} \end{table} The volume of activations from a convolutional layer $\mathcal{X}$ in $\mathbb{R}^{M \times N \times D}$, is re-interpreted as a set of $M \times N$ local descriptors of $D$ dimensions. A visual vocabulary is learned using $k$-means on the local CNN features. With this procedure each image is represented with an \textit{assignment map}, which consists of a compact representation of a 2D array of size $H \times W$ relating each pixel of the original image with its visual work with precision of ($W/N$, $H/M$); being $W$ and $H$ the the with and height of the original images. This is represented in the first stream of Figure~\ref{}, which involves the semantic representation of the image. One of the main advantages of this representation is to preserve the spatial layout of the image, so it is possible to apply a spatial weighting scheme prior to constructing the final BoW representation. In particular, a saliency weighting scheme $\alpha(i,j)$ with resolution $W \times H$ is down-sampled to match the spatial resolution of the assignment maps $M \times N$, and normalized to have values between 0 and 1, illustrated in the second stream in Figure~\ref{}. Then, the final representation consists in an histogram $\mathbf{f}_{\text{BOW}}=(h_{1}, ... h_{K})$, where each component is the sum of the spatial weight $w_{ik}$ assigned to a particular visual word $k$, $h_{k}=\sum_{i=1}^N{w_{ik}}$. The final vector is $L^2$-normalized and cosine similarity is used to rank the images. \begin{figure}[b!] \begin{center} \includegraphics[width=.48\textwidth]{Figures/BLCF_saliency6} \caption{The Bag of Local Convolutional Framework (BLCF) pipeline with saliency weighting.} \label{} \end{center} \end{figure} We use descriptors from \textit{conv5\_1} from a pre-trained VGG16~\cite{vgg} without fully connected layers. The VGG16 network was chosen following Jimenez et al. ~\cite{Jimenez_2017_BMVC}, where it was found that the features from this network outperform those from ResNet50 using the CAM approach. We confirmed this result in our own experiments, finding a similar result for BCLF encoding when using the local search scheme: there was a consistent reduction ($>$ 10 points) in mAP when compared with VGG16 regardless of the weighting scheme used. Images are resized to have its maximum size of 340 pixels ($1/3$ of the original resolution) before performing mean subtraction prior to being forwarded through the network. This resolution is also used to extract the different saliency maps. Local features are post-processed with $L^2$-normalization, followed by a PCA whitening (512D), and $L^2$-normalized again prior to being clustered. Clustering is performed using approximate $k$-means with $k=25,000$ words. The visual vocabulary and PCA models are fit on the target dataset. The obtained saliency maps are down-sampled by computing the maximum across $16\times 16$ non-overlapped blocks. For the query images, we interpolate the volume of activations to $(2H,2W)$, as this interpolation of the feature maps has been shown to improve performance~\cite{mohedano2016bags}. However, we only apply this strategy to the query images, and not to the entire dataset, so as not to increase dataset memory consumption. The bounding box of the queries is mapped to the assignment maps, and only the information relevant to the instance is used to create $\mathbf{f}_{\text{BOW}}$. This procedure is similar to the one used in~\cite{Radenović2016}, where a bounding box is mapped to the convolutional feature maps instead of cropping the original image, which was shown to provide better results. \section{Experiments} \label{} We experiment with three different datasets: the challenging INSTRE dataset for object retrieval ~\cite{Instre}, and two well-known landmark-related benchmarks, the Oxford~\cite{Oxford5k} and Paris~\cite{paris6k} datasets. The former dataset was specifically designed to evaluate instance retrieval approaches. It consists in 23,070 manually annotated images, including 200 different instances from diverse domains, including logos, 3D objects, buildings, and sculptures. Performance is evaluated using mean average precision (mAP) following the same standard protocols as in Oxford and Paris benchmarks, and evaluating mAP over 1,200 images as described in~\cite{difussion}. \subsection{Weighting schemes} \label{} \begin{table}[b!] \caption{Performance (mAP) of different spatial weighting schemes using the BLCF approach.\vspace{0.1cm}} \centering \begin{tabular}{@{}p{3.9cm}lll@{}} \toprule Weighting & INSTRE & Oxford & Paris \\ \midrule None & 0.636 & 0.722 & 0.798 \\ Gaussian & 0.656 & 0.728 & 0.809 \\ $L^2$-norm & 0.674 & 0.740 & \textbf{0.817} \\ Itti-Koch~\cite{itti1998model} & 0.633 & 0.693 & 0.785 \\ BMS~\cite{zhang2013saliency} & 0.688 & 0.729 & 0.806 \\ SalNet~\cite{SalNet} & 0.688 & 0.746 & 0.814 \\ SalGAN~\cite{SalGAN} & \textbf{0.698} & \textbf{0.746} & 0.812 \\ SAM-VGG~\cite{cornia2016predicting} & 0.688 & 0.686 & 0.785 \\ SAM-ResNet~\cite{cornia2016predicting} & 0.688 & 0.673 & 0.780 \\ \bottomrule \end{tabular} \label{} \end{table} In this section we evaluate different spatial weighting schemes with the BLCF framework. We find that saliency weighting schemes (BMS, SalNet, and SalGAN) improve retrieval performance across the evaluated datasets with respect to the unweighted assignment maps. Figure~\ref{} shows some of the cases where saliency (SalGAN model) is most beneficial, allowing the efficient localisation of the target instances in most of the cases, despite of the high variability of the backgrounds and relative positions within the images in the INSTRE dataset. Table~\ref{} contains the performance of different weighting schemes on the BLCF approach. The simple Gaussian weighting achieves a boost in performance with respect to the baseline, which indicates a natural tendency of the instances to appear in the center of the images in all the datasets. The saliency measure derived from the convolutional features ($L^2$-norm weighting) allows to roughly localize the most relevant parts of the image, which represents a boost in performance with respect the Gaussian center prior weighting. However, the $L^2$ weighting appears in general to be much more noisy than the human attention-based saliency models, as depicted in Figure~\ref{} and Figure~\ref{}, which leads to lower mAP when compared with the BMS, SalNet, and SalGAN models. Saliency models achieve the best performing results in the INSTRE dataset, with the exception of the Itti-Koch model, which decreases performance with respect the baseline in all datasets. This result is consistent with the quality of the saliency prediction achieved in the MIT300 benchmark, where it is rated in the second lowest rank of the evaluated models~\footnote{\url{}}. The more accurate saliency models (BMS, SalNet, SalGAN, SAM-VGG, and SAM-ResNet), however, achieve almost equivalent performance on the INSTRE dataset. The relatively small size of the instances within this dataset and the high variance in their relative position within the images makes saliency a very efficient spatial weighting strategy for instance search, with no substantial difference between saliency models. This contrasts with results for the Oxford and Paris datasets, where a coarser saliency prediction (i.e the one provided by the SalNet model) achieves better results than the one obtained with the more accurate models of SAM-VGG and SAM-ResNet. Figure~\ref{} illustrates this effect on three different instances of the Notre Dame cathedral from the Paris dataset. The most accurate saliency models (i.e. SAM-VGG and SAM-ResNet) detect saliency regions ``within'' the building, instead of detecting the full building as relevant. Also, SAM-VGG, SAM-ResNet, and SalGAN appear to be more sensitive to detecting people, omitting any other salient region of the image and thus decreasing the overall retrieval performance. \begin{figure}[t!] \begin{center} \includegraphics[width=.48\textwidth]{Figures/notre_dame} \caption{\small Sample of saliency predictions for three examples of Notre Dame cathedral in the Paris dataset. Average precision (AP) for this query is 0.868, which is improved by BMS and SalNet models (achieving 0.880 and 0.874 AP respectively). More accurate saliency models decrease performance with respect the baseline, achieving an AP of 0.862 in the case of SalGAN, 0.857 SAM-VGG, 0.856 SAM-ResNet and 0.853 Itti-Koch models.} \label{} \end{center} \end{figure} \subsection{Aggregation methods} We also tested saliency weighting in combination with classic sum pooling over the spatial dimension to compare with the BCLF weighting. Here we report the best performing saliency models (SalNet and SalGAN) compared with a no weighting baseline using the VGG-16 network \textit{pool5} layer at full image resolution. Although combining sum pooling with saliency weighting did give a performance improvement on the INSTRE dataset (mAP 0.527 for SalGAN vs 0.405 for no pooling), saliency weighting combined with BLCF gives substantially better mAP (see Table~\ref{}). Furthermore, sum pooling with saliency weighting gave little or no improvement on the Oxford and Paris datasets (mAP Oxford: 0.681 for SalNet vs 0.686 for no weighting, and Paris: 0.766 for SalNet vs 0.765 for no weighting). This result is perhaps unsurprising. Weighting the feature assignments prior to pooling in the BCLF framework can be interpreted as a simple importance weighting on each discrete feature. However, the interpretation for sum pooling is less clear, since the feature vectors at each spatial location usually have multiple activations of varying magnitude, and weighting alters the potentially semantically meaningful magnitudes of the activations. The BCLF-saliency approach also has the advantage of having two to three times fewer non-zero elements in the representation, giving substantially improved query times. \begin{table}[t!] \caption{Performance (mAP) of different spatial weighting schemes using the sum pooling aggregation on (SUM) and the BLCF approach.\vspace{0.1cm}} \centering \begin{adjustbox}{width=0.48\textwidth} \begin{threeparttable} \begin{tabular}{@{}lllllll@{}} \toprule Method & \multicolumn{2}{c}{Instre} & \multicolumn{2}{c}{Oxford} & \multicolumn{2}{c}{Paris} \\ \midrule Weighting & SUM & BLCF & SUM & BLCF & SUM & BLCF \\ None & 0.405 & 0.636 & 0.686 & 0.722 & 0.765 & 0.798 \\ SalNet & 0.519 & 0.688 & 0.681 & 0.746 & 0.766 & 0.814 \\ SalGAN & 0.527 & 0.698 & 0.612 & 0.746 & 0.749 & 0.812 \\ \end{tabular} \end{threeparttable} \end{adjustbox} \label{} \end{table} \subsection{Comparison with the state-of-the-art} Our approach compares favorably with other methods exploiting pre-trained convolutional features, as shown in Table~\ref{}. In particular, we achieve state-of-the-art performance using pre-trained features in Oxford and INSTRE datasets, when combining BLCF with saliency prediction from SalGAN. CAMs~\cite{Jimenez_2017_BMVC}, and R-MAC~\cite{rmac-ToliasSJ15} slightly outperform our approach in the Paris dataset, where they achieve $0.855$, and $0.835$ mAP respectively, whereas our method achieves $0.812$ mAP. It is worth noting, however, that our system is much less complex (we do not perform region analysis), and more scalable, since the sparse representation only needs a maximum of $336$ non-zero elements out of the 25K dimensions. Fine-tuned CNN models~\cite{Radenović2016,gordo2016deep} significantly outperform our approach in Oxford and Paris. However, saliency weighted BLCF achieves state-of-the art performance in the INSTRE dataset with $0.698$ mAP, outperforming the fined-tuned models. This is probably a consequence of restricting the training domain to landmarks images in~\cite{Radenović2016,gordo2016deep}, since R-MAC with pre-trained features~\cite{rmac-ToliasSJ15} ($0.523$ mAP) already achieves a higher performance in INSTRE than the fine-tuned R-MAC~\cite{gordo2016deep} ($0.477$ mAP). This provides evidence that similarity learning is a strategy to greatly improve performance on CNN representations where the target instance domain is restricted. In more generic scenarios, such as the INSTRE dataset, pre-trained ImageNet models achieve more general image representations. One common technique to improve retrieval results is the average query expansion (AQE)~\cite{chum2007total}. Given a list of retrieved images, a new query is generated by sum aggregating the image representations of the top $N$ images for a particular query. We select the top 10 images, and $L^2$-normalization on the new query representation. Table~\ref{} With this simple post processing we substantially improve the results in the INSTRE dataset (mAP increased from 0.698 to 0.757), achieving the best performance in comparison with other methods using the same post-processing strategy. \begin{table}[t!] \centering \caption{Performance comparison with the state-of-the-art.\vspace{0.1cm}} \label{} \begin{adjustbox}{width=0.48\textwidth} \begin{threeparttable} \begin{tabular}{lccccc} \hline Method & \multicolumn{1}{l}{Off-the-shelf} & dim & INSTRE & Oxford & Paris \\ \hline CroW~\cite{CroW} & yes & 512 & 0.416 & 0.698 & 0.797 \\ CAM~\cite{Jimenez_2017_BMVC}\tnote{*} & yes & 512 & 0.325 & 0.736 & 0.855 \\ R-MAC~\cite{rmac-ToliasSJ15} & yes & 512 & 0.523 & 0.691 & 0.835 \\ R-MAC~\cite{Radenović2016}\tnote{\dag} & No & 512 & 0.477 & 0.777 & 0.841 \\ R-MAC-ResNet~\cite{gordo2016deep}\tnote{\dag} & No & 2048 & 0.626 & \textbf{0.839} & \textbf{0.938} \\ (our) BLCF & yes & 336 & 0.636 & 0.722 & 0.798 \\ (our) BLCF-Gaussian & yes & 336 & 0.656 & 0.728 & 0.809 \\ (our) BLCF-SalGAN & yes & 336 & \textbf{0.698} & 0.746 & 0.812 \\ \end{tabular} \begin{tablenotes} \centering \footnotesize \item Results marked with (*) are provided by the authors. Those marked with (\dag) are reported in Iscen et al.~\cite{difussion}. Otherwise they are based on our own implementation. \end{tablenotes} \end{threeparttable} \end{adjustbox} \end{table} \begin{table}[t!] \centering \caption{Performance comparison with the state-of-the-art with average query expansion.\vspace{0.1cm}} \label{} \begin{adjustbox}{width=0.48\textwidth} \begin{threeparttable} \begin{tabular}{@{}lccccc@{}} \toprule Method & \multicolumn{1}{l}{Off-the-shelf} & dim & INSTRE & Oxford & Paris \\ \midrule CroW~\cite{CroW} & yes & 512 & 0.613 & 0.741 & 0.855 \\ CAM~\cite{Jimenez_2017_BMVC}\tnote{*} & yes & 512 & & 0.760 & 0.873 \\ R-MAC~\cite{rmac-ToliasSJ15} & yes & 512 & 0.706 & 0.770 & 0.884 \\ R-MAC~\cite{Radenović2016}\tnote{\dag} & No & 512 & 0.573 & 0.854 & 0.884 \\ R-MAC-ResNet~\cite{gordo2016deep}\tnote{\dag} & No & 2048 & 0.705 & \textbf{0.896} & \textbf{0.953} \\ (ours) BLCF & yes & 336 & 0.679 & 0.751 & 0.788 \\ (ours) BLCF-Gaussian & yes & 336 & 0.731 & 0.778 & 0.838 \\ (ours) BLCF-SalGAN & yes & 336 & \textbf{0.757} & 0.778 & 0.830 \\ \end{tabular} \begin{tablenotes} \centering \footnotesize \item Results marked with (*) are provided by the original publications. Those marked with (\dag) are reported in Iscen et al.~\cite{difussion}. Otherwise they are based on our own implementation. \end{tablenotes} \end{threeparttable} \end{adjustbox} \end{table} \section{Conclusions and future work} \label{} We have proposed a generic method for instance search that relies on BoW encoding of local convolutional features and attention models. We have demonstrated that saliency models are useful for the instance search task using a recently proposed CNN-based retrieval framework. Results indicate that different state-of-the-art saliency prediction models are equally beneficial for this task in the challenging INSTRE dataset. In landmark related datasets such as Oxford an Paris, however, a coarse saliency is more beneficial than highly accurate saliency models such as SAM-VGG or SAM-ResNet. Better query expansion strategies can be applied to further improve results, such as the diffusion strategy proposed in~\cite{difussion}, where global diffusion using $2,048$ dimensional descriptors from fine tuned R-MAC~\cite{gordo2016deep} achieves $0.805$ mAP in INSTRE, and a regional diffusion $0.896$ mAP. This strategy can be also applied to the proposed saliency weighted BLCF representation, potentially increasing retrieval performance of the proposed saliency weighted BLCF representations.

\section{Introduction} A prevalant family of human pose estimation works generally fall into two groups: 2D human pose estimation and 3D human pose estimation. Last year has witnessed the revolution of 2D human pose estimation, thanks to the development of heatmap-based network and the availability of deep residual network \cite{he2016deep}. However, the research on 3D human pose estimation has been significantly lagging behind its counterpart. Prior works on 3D human pose estimation can be roughly categorized into two families: \emph{regression methods} and \emph{reconstruction methods}. \emph{Regression methods} directly learn a mapping function from input image to the target 3D joint locations\cite{li20143d, li2014heterogeneous, li2015maximum, moreno20163d, park20163d, sun2017compositional, tekin2016structured, zhou2017towards, tekin2017learning, mehta2016monocular}. As popularized in 3D human pose estimation, the performance is not as excellent as expected. \emph{Reconstruction methods} typically follow a two-stage schema where in the first stage a 2D pose estimator is employed and then 3D pose is reconstructed aiming to minimize the reprojection error via \emph{optimization} \cite{akhter2015pose, bogo2016keep, ramakrishna2012reconstructing, fan2014pose, zhou2017sparse} or \emph{matching} \cite{jahangiri2017generating, chen20163d}. While \emph{optimization} can generate only one 3D output, recent work of Jahangiri and Yuille \cite{jahangiri2017generating} has highlighted the importance of having multiple hypotheses under the widely known problem of depth ambiguity. A fundamental challenge limiting previous methods from attacking 3D human pose estimation is the insufficient training data. Most top-performing methods in 3D human pose estimation are restricted in laboratory environment where the appearance variation is far less than outdoor scene. On the other hand, there exists no accurate 3D ground truth for in-the-wild dataset to date. Fusing 2D and 3D data, dates back to at least \cite{yasin2016dual}, therefore has become an emerging trend \cite{sun2017compositional, zhou2017towards}. This naturally brings us to a question, \emph{ for accurate 3D human pose estimation, is mixing different data sources really indispensible?} In this work, we argue that combined training is not necessary by better exploiting the data that we already have. We put forth a novel expressive intermediate feature representation called \emph{skeleton map}. For an example see Figure ~\ref{}. \begin{figure}[t] \begin{center} \includegraphics[width=0.5\linewidth]{skeletonmap.png} \end{center} \caption{An example of \emph{skeleton map}. 15 body parts are drawn in different colors.} \label{} \end{figure} The core insight is that \emph{human being can easily reason about 3D human pose given skeleton map by our prior knowledge about human kinematics, anatomy and anthropometrics, rather than image observation}. This suggests that most of the image cues, \eg lightning, texture, and human-object interaction are useless for 3D pose inference. \emph{Skeleton map} is simple, compact and effective. Unlike full human body part segmentation map which requires pixel-wise segmentation, it only models the connection between adjacent joints in the human skeleton. It contains rich structural information and exhibits inherent occlusion-aware property for regression. \emph{Skeleton map} is general for both indoor and outdoor scenario. Researchers previously argued that 2D and 3D data complements each other and mixing them is central to better shared feature learning. But with our pure and succinct \emph{skeleton map} representation, we are able to achieve near state-of-the-art performance. We then take a leap forward to generate multiple hypotheses, sharing the same spirit of ``wisdom of the crowd" \cite{lifshitz2016human}. Different scales of \emph{skeleton maps} are conveniently generated by deconvolutional networks for subsequent regression. It can be intepretered as implicit data augmentation \emph{without data fusion or data synthesis}. This desirable property further resolves the depth ambiguity. To the best of our knowledge, this is the first work to \begin{itemize} \item[$\bullet$] Perform regression directly from \emph{skeleton map} alone. \item[$\bullet$] Generate multiple hypotheses \emph{without 3D library}. \item[$\bullet$] Unite segmentation, pose regression and heatmap detection into a framework we call \emph{DeepSkeleton} for 3D human pose estimation. \end{itemize} \emph{DeepSkeleton} achieves 86.5\emph{mm} average joint error on 3D Human3.6M dataset and delivers considerable 3D poses on 2D MPII dataset. \section{Related Work} For a comprehensive review on the literature of 2D human pose estimation we refer the readers to \cite{gong2016human, luvizon2017human, sun2017human}. Here we survey on previous works most relevant to our approach. \textbf{3D Human Pose Estimation} Li and Chan \cite{li20143d} pioneered the work of direct 3D human pose regression from 2D image. Several approaches have been proposed to learn the pose structure from data. Tekin \etal \cite{tekin2016structured} learn latent pose representation using an auto-encoder. Zhou \etal \cite{zhou2016deep} integrate a generative forward kinematics layer into the network to learn the joint angle parameters. Realizing the difficulty to minimize per-joint error, \cite{park20163d, mehta2016monocular, sun2017compositional} advocate to predict relative joint locations to multiple joints. More recently, Rogez \etal \cite{rogez2017lcr} simplify the problem by local residual regression in the classified pose class. Another research direction has been focused on inferring 3D pose from 2D estimates. The underlying premise is that 2D human pose estimation can be regarded as nearly solved\cite{wei2016convolutional, newell2016stacked, carreira2016human, bulat2016human, chu2016structured, chu2017multi, gkioxari2016chained, lifshitz2016human, yang2016end}. Consequently, the challenge of 3D pose estimation has been shifted from predicting accurate 2D towards predicting depth from RGB image. As an example, Moreno-Noguer \cite{moreno20163d} employs FCN\cite{long2015fully} to obtain pairwise 3D distance matrix from 2D observation for absolute 3D pose recovery. Tekin \etal \cite{tekin2017learning} fuse image space and heatmap space to combine the best of both worlds. Tome \etal \cite{tome2017lifting} adopt an iterative multi-stage architecture where 2D confidence is progressively lifted to 3D and projected back to 2D, ensuring the match between 2D observation and 3D prediction. Zhou \etal \cite{zhou2016sparseness} combine heatmap and 3D geometric prior in EM algorithm to reconstruct 3D skeleton from 2D joints. Zhou \etal \cite{zhou2017towards} output depth from 2D heatmap and intermediate feature maps in a weakly-supervised fashion. Nevertheless, the inherent depth ambiguity presents the foremost obstacle impeding the progress of 3D human pose estimation. Recent techniques ameliorate this issue through dual-source training. Examples include \cite{yasin2016dual, zhou2017towards, sun2017compositional}. Our proposed \emph{DeepSkeleton} stands in contrast to the latent assumption of these works that shared feature learning from both 2D and 3D data is essential. Our results suggest that 2D and 3D data sources may be completely indepedent, which is in line with recent observation \cite{tome2017lifting, martinez2017simple}. \textbf{Exemplar based pose estimation} Most previous work on 3D human pose estimation rely on generating only one single 3D pose. This is problematic as multiple 3D poses may have similar 2D projection. With early reference to \cite{cho2015accurate}, generating multiple hypotheses has been repurposed for 3D human pose estimation in \cite{jahangiri2017generating}. In their work, a generative model is responsible for generating multiple 3D poses from 3D MoCap library. Similarly, Chen and Ramanan \cite{chen20163d} retrieve 3D pose using neareast neighbour search. \emph{DeepSkeleton} shares some resemblence to these works expect for the need of large-scale offline MoCap dataset. \textbf{Semantic Segmentation} Fully convolutional network(FCN)\cite{long2015fully} made the first attempt to solve semantic segmentation using deep neural network. More recent works enjoy the benefits of residual connection \cite{lin2016refinenet, chen2016deeplab}. However, accurate pixel-wise annotation of segmentation map is time-intensive. Different from full human body part segmentation\cite{oliveira2016deep, xia2015zoom, liang2015human} , synthesizing ground truth of our proposed \emph{skeleton map} only requires 2D annotation. \textbf{Joint Semantic Segmentation and Pose Estimation} Segmentation and pose estimation are two intrinsically complementary tasks in the sense that pose esmation benefits from the topology of semantic parts, while the estimated pose skeleton offers natural cue for the alignment with part instance. One of the most significant works that jointly solve these two problems is Shotton \etal \cite{shotton2013real}, which derives an intermediate body part segmentation representation for 3D joint prediction from depth image. A large body of literature has been devoted to this field \cite{dong2014towards, ladicky2013human, yamaguchi2012parsing, kohli2008simultaneous, alahari2013pose}. Tripathi \etal \cite{tripathi2017pose2instance} demonstrate the impact of human pose prior on person instance segmentation. Perhaps the most related approach to ours is Xia \etal \cite{xia2017joint}. Their \emph{skeleton label map} serves as a prior to regularize the segmentation task, and is derived from 2D pose prediction. In contrast, our \emph{skeleton map} is independent of 2D detection and is directly taken as input to the subsequent regression module. Another major difference is that they aim to solve multi-person 2D pose estimation, while we target at single-person 3D pose estimation. \section{Methodology} Given a RGB image $\mathbf{I} \in {\mathcal{R}}^{224\times224}$ of a person, our goal is to output 3D joint locations $X\in {\mathcal{R}}^{K\times3} $ with joint number $K=16$. We break the problem down into three steps: \begin{itemize} \item Segmentation(Section \ref{}): For each configuration $p_i=\left\lbrace c_i, l_i \right\rbrace$ with crop scale $c_i$ and stick width $l_i$, \emph{foreground skeleton map} $S_i^f \in {\mathcal{R}}^{56\times56}$ and \emph{background skeleton map} $S_i^b \in {\mathcal{R}}^{56\times56} $ are generated via deconvolutional network ${Deconv}_i$($i=1,...,n$). \item Regression(Section \ref{}): $\emph{Skeleton maps}$ $\mathcal{S}=\left\lbrace (S_i^f,S_i^b) |i=1,...n\right\rbrace$ are individually fed into \emph{separate} regression networks $R=\left\lbrace {Regression}_i|i=1,...n\right\rbrace$, where ${Regression}_i$ takes \emph{skeleton map} $S_i^f, S_i^b$ as input and outputs 3D pose hypothesis $X^i$, resulting in multiple 3D hypotheses $\mathcal{H}=\left\lbrace X^i | i=1,...n \right\rbrace $. \item Matching(Section \ref{}): To match with 2D observation $x \in \mathcal{R}^{K\times2}$, the hypothesis $X^{\ast} \in \mathcal{H}$ with minimum projection error to 2D joint detection is selected as final output. \end{itemize} \subsection{Skeleton Map} \emph{Skeleton map} $S_i^f$($S_i^g$) draws a stick with width $l_i$ between neighbouring joints in the human skeleton and assigns different colors to distinguish body parts, an example of which is depicted in Figure ~\ref{}. The human skeleton we use follows the paradigm of \cite{sun2017compositional} expect that we define \emph{thorax} as root in all our experiments, in which 15 body parts are defined. \emph{Skeleton map}, like body part segmentation map, encodes part relationship in the body segments and imposes strong prior on human pose. However, training networks for full human body semantic segmentation needs labor-intensive dense and precise annotations, which is impractical for large human pose dataset \eg Human3.6M \cite{ionescu2014human3}. The simplicity of \emph{skeleton map} naturally addresses this issue. But what is a good architecture for \emph{skeleton map} generation? \subsection{Deconvolution for Generating Skeleton Map} \label{} \textbf{Deconvolutional Network Design} A simple choice is to employ the encoder-decoder architecture. In practice, we apply ResNet-50\cite{he2016deep} in a fully convolutional manner, producing pixel-level real values. We replace the fully connected layer after \emph{pool5} with deconvolutional layers. The network structure shown in Figure ~\ref{} starts with a 224$\times$224 image and extracts features along the downsampling process. Herein only \emph{res2c}, \emph{res3d}, \emph{res4f} and \emph{res5c} are sketched for brevity. The last fully connected layer is simply discarded. The devolutional module built upon \emph{pool5} gradually processes feature maps to the final output: a three-channel 56$\times$56 \emph{skeleton map}. It comprises of repeated blocks of upsampling layer (with initial weights set to bilinear upsampling) followed by residual module. To encourage the learning of local and global context, high-level feature maps are combined with low-level feature maps through skip connections, analogous to those used in deep networks\cite{long2015fully, newell2016stacked}. The output composes of three channels representing the \emph{skeleton map}. Rather than performing per-pixel classification into body part classes, we found that per-pixel regression results in better segmentation accuracy. We opt for sigmoid cross entropy loss in training. However, one common problem in training deep network is the notorious \emph{vanishing gradient}. To remedy this issue, each blob feeding into the pixel-wise summation layer branches off and connects to a residual block. Intermediate supervision is then applied on the output of each residual block, allowing for learning \emph{skeleton map} at multiple resolutions. \begin{figure*} \begin{center} \includegraphics[width=1.0\linewidth]{diagram.png} \end{center} \caption{Deconvolutional Network Architecture. The blocks from input image up to \emph{res2c} are not drawn for simplicity. Intermediate supervision is not plotted. See Section \ref{} for details.} \label{} \end{figure*} After the initial performance gain brought by common pratice \ie skip connection and intermediate supervision, investigations on prevalant \emph{conv-deconv} network architectures \eg \emph{RefineNet}\cite{lin2016refinenet} and \emph{Stacked Hourglass}\cite{newell2016stacked} did little to further improve the segmentation accuracy. \textbf{Deal with trunction} Truncation, that is, partial visibility of human body joints caused by image boundary, poses a noticeable challenge especially for in-the-wild 3D human pose estimation. In our case, this is supported by the fact that the deconvolutional network is uncertain about whether to plot the segments associated with cropped endpoint joint \ie \emph{wrist} or \emph{ankle}, due to lack of image evidence. A standard way to deal with this is to use multiple image crop, which is made possible by multiplying the provided rough person scale in the dataset with the rescaling factor \ie $c_i \in \left\lbrace 1.0, 1.25, 1.5 \right\rbrace$. The 2D joint ground truth is rescaled in the cropped window accordingly. Note that indoor dataset faces no truncation problem, and \emph{crop scale $c_i$} is always set to 1.0. \textbf{Deal with stick width} Several endeavours have been made to effectively excavate features for human pose estimation, most of which focus on \emph{multi-stream} learning\cite{xie2015holistically}. \emph{DeepSkeleton} differs in that it directly modifies the target output by changing stick width. Our design of \emph{multi-scale skeleton map} is motivated by the claim that each convolutional layer is responsible for skeleton pixel whose scale is less than the receptive field\cite{shen2016object}. In concrete terms, only convolutional layers with receptive field size larger than stick width can capture features of body parts. Hence, coarse segmentation(\emph{large stick width}) and fine segmentation(\emph{small stick width}) feature varying combinations of low-level and high-level features. Note that care has to be taken when deciding the body segment width of ground truth \emph{skeleton map}. We consider two practial concerns:(1)\emph{the parts should be small enough to localize semantic body parts.} (2)\emph{the parts should be large enough for convolution and deconvolution.} The stick width \ie $l_i$ is empirically set to be in the range $[5,15]$. \textbf{Deal with occlusion} Severe occlusion hinders accurate human pose estimation. Answers to simple queries such as \emph{``Is the left upper leg in front of right upper leg?"} can actually provide important occlusion cues for regression. Motivated by \emph{inside/outside score map} \cite{li2016fully}, in this work, \emph{foreground skeleton map} $S_i^f$ displays body parts that are occluding others, while \emph{background skeleton map} $S_i^b$ displays those occluded by others. That said, \emph{skeleton map} explicitly models the occlusion relationship of two overlapping body parts. As far as we know, this straightforward formulation has never been done in the literature of human pose estimation. In more detail, recall that each 2D endpoint joint on the bone results in a ray oriented towards the camera optical center. Assume that 3D point $\left\lbrace \mathbf{X}_u,\mathbf{Y}_u,\mathbf{Z}_u\right\rbrace$ on bone $\mathbf{B}_u$ and $\left\lbrace \mathbf{X}_v,\mathbf{Y}_v,\mathbf{Z}_v \right\rbrace$ on bone $\mathbf{B}_v$ yield the same 2D projection $\left\lbrace \mathbf{x},\mathbf{y} \right\rbrace$. Denote the point with smaller depth(closer to camera in $Z$ direction) as $id$, and the other as $\widetilde{id}$. \emph{Foreground skeleton map} assigns color of bone $\mathbf{B}_{id}$ to the pixel $\left\lbrace \mathbf{x},\mathbf{y} \right\rbrace$. In contrast, \emph{background skeleton map} assigns color of bone $\mathbf{B}_{\widetilde{id}}$, pretending bone $\mathbf{B}_{\widetilde{id}}$ is occluding bone $\mathbf{B}_{id}$. See Figure \ref{} for an example. This inherent \emph{occlusion-aware} property of \emph{skeleton map} is important for regression. \begin{figure} \begin{center} \includegraphics[width=0.19\linewidth]{28433.png} \includegraphics[width=0.19\linewidth]{28433_front.png} \includegraphics[width=0.19\linewidth]{28433_back.png} \includegraphics[width=0.19\linewidth]{28433_2d.png} \includegraphics[width=0.19\linewidth]{28433_3d.png} \end{center} \caption{A example of occlusion handling. From left to right: raw image, predicted \emph{foreground skeleton map}, predicted \emph{background skeleton map}, inferred 2D landmark, inferred 3D joint position. Note how the left/right body parts are distinguishably segmented in \emph{foreground/background skeleton map}.} \label{} \end{figure} \subsection{Regression} \label{} Plagued by the cluttered background of RGB image, a longstanding research direction in 3D human pose estimation has been exploiting better features from raw RGB input. We show that regression from \emph{skeleton map} alone is feasible. We employ state-of-the-art ResNet-50\cite{he2016deep} as the backbone network. Since \emph{skeleton maps} $(S_i^b, S_i^f)$ generated by deconvolutional network ${Deconv}_i$ are $56\times56$, they are rescaled to $224\times224$ first and concatenated together, which is then taken as input to ${Deconv}_i$ and processed along the downsampling path. The last fully connected layer is repurposed to output 3D position $X^i=\left\lbrace \mathbf{X}_k^i | k=1,2,..K \right\rbrace$ of all $K$ joints. Euclidean distance loss is applied for back propogation. Multiple 3D predictions $\mathcal{H}=\left\lbrace X^i | i=1,...n \right\rbrace $ are made by training independent regression networks for different \emph{skeleton maps} input. We want to emphasize that na\"{i}vely concatenating all the \emph{skeleton maps} $\mathcal{S}$ as input (\emph{early fusion}) or learning parallel streams for multiple input and fusing the feature responses subsequently (\emph{late fusion}) will not help. As an alternative option, one might consider to concatenate \emph{skeleton map} along with raw RGB image, which however, does not boost the performance in our observation. Therefore, we stick with the original design \ie learning 3D solely from intermediate \emph{skeleton map} feature representation. \subsection{Matching} \label{} Now we have multiple 3D pose hypotheses $\mathcal{H}$, the problem boils down to select the optimal hypothesis $X^{\ast} \in \mathcal{H}$ as final 3D output. The simpliest way is to choose the candidate whose projection best matches the 2D pose detection results. Writing $\mathit{Proj}$ as the camera projection matrix and $x$ as 2D pose detection, we seek to find the optimal 3D pose $X^{\ast} \in \mathcal{H}$ via minimizing the reprojection error: \begin{equation} X^{\ast}=\underset{X^i}{\operatorname{arg\,min}}||Proj(X^i)-x||^2 \end{equation} We use the pre-trained state-of-the-art 2D detector \emph{Stacked Hourglass}\cite{newell2016stacked} for generating $x$. No finetuning is employed. We remark that our 3D hypotheses are completely independent of 2D pose detection, rather they are learnt from multi-level discriminative \emph{skeleton maps}. \section{Discussions} In principle, any expressive intermediate representation can be used for regression. Heatmap, for example, has been explorered in \cite{wu2016single} to bridge the gap between real and synthetic dataset. Carreira \etal \cite{carreira2016human} stack heatmap with RGB image for coarse-to-fine regression. Yet, an obvious drawback of heatmap is different body joints are encoded in \emph{discrete gaussian peaks}, thus the dependence between joints is not well exploited. \emph{Skeleton map} overcomes this problem by explicitly connecting adjacent joints of each bone by a stick. The colored semantic body part offers a strong cue for regression learning. For generating \emph{skeleton map}, perhaps the easiest way is to firstly detect 2D joints and then draw lines between neighbouring joints. However, this introduces two disadvantages: \emph{1. Occlusion relationship information is completely discarded. 2. The inaccurate 2D detection has impact on the following regression.} Our initial exploration shows that this has no apparent benefits. \section{Experiments} Our approach is evaluated on the largest human pose benchmarks MPII and Human3.6M. \emph{MPII} \cite{andriluka20142d} is a 2D real-world human pose dataset. It contains around $25k$ natural images collected from YouTube with a variety of poses and complicated image appearances. Cluttered background, multiple people, severe truncation and occlusion make it the most challenging 2D human pose dataset. For 2D pose evaluation in MPII, we use PCKh\cite{andriluka20142d} metric which measures the percentage of joints with distance to ground truth below a certain threshold. \emph{Human3.6M}\cite{ionescu2014human3} is a large-scale 3D human pose dataset consisting of 3.6M video frames captured in controlled laboratory environment. 5 male and 6 female actors performing 17 daily activities are captured by motion capture system in 4 different camera views. The image appearance in clothes and background is limited compared to \emph{MPII}. Following standard practice in \cite{zhou2017towards, li20143d, zhou2016deep, zhou2016sparseness}, five subjects(S1, S5, S6, S7, S8) are used in training. Every $64^{th}$ frame of the two subjects(S9, S11) is used in testing. MPJPE(mean per joint position error)\cite{ionescu2014human3} is used as evaluation metric after aligning 3D poses to root joint. We represent 3D pose in local camera coordinate system following the methodology of Zhou \etal \cite{zhou2017towards}. \subsection{Implementation Detail} For training the network we use Caffe\cite{jia2014caffe} with 15 GPUs. Deconvolutional network training starts with base learning rate 0.00001 and mini-batch size 12. For training regression network, we set base learning rate to 0.01 and mini-batch size to 32. Learning rate is dropped by a factor of 10 after error plateu on the validation set. The network is trained until convergence. For optimization, stochastic gradient descent is adopted. Weight decay is 0.0002, and momentum is 0.9. No data augmentation or data fusion is used. \subsection{Baseline Settings} To validate the effectiveness of \emph{skeleton map} and \emph{multiple hypothesis}, we test two baselines: \begin{itemize} \item[$\bullet$] \emph{Direct RGB} It performs regression directly from raw RGB input. \item[$\bullet$] \emph{Ours w/o Mul-Hyp} It performs regression from only one \emph{skeleton map}. \end{itemize} Unless otherwise specified, we set $c_i=1.0,l_i=10$ for all the experiments of one hypothesis(\emph{skeleton map}) both on Human3.6M and MPII. Our final system is denoted as \emph{Ours w Mul-Hyp} (equivalent to \emph{DeepSkeleton}). \subsection{Experiments on 3D dataset Human3.6M} \label{} \textbf{Comparison with state-of-the-art} For fair comparison, we compare with state-of-the-art methods without mixed 2D and 3D data for training in Table \ref{}. Note that \emph{Compositional Pose Regression} \cite{sun2017compositional} provides results with and without 2D data. We therefore denote $\emph{CompBone}^* $ as \emph{Compositional Pose Regression} without extra 2D training data and report the results from the original paper. Table \ref{} shows that our final system \emph{ours w Mul-Hyp} outperforms the main competitor Tome \etal \cite{tome2017lifting}. Notably, it surpasses competeting methods in actions \emph{Sit}, \emph{SitDown}, \emph{Photo} by a large margin. The improvement comes from our novel \emph{skeleton map} representation and the expressiveness of multiple hypotheses. Visualized 3D poses are displayed in Figure \ref{}. \begin{table*} \scriptsize \begin{center} \begin{tabular}{lcccccccc} \toprule \textbf{Method} & \textbf{Direction} & \textbf{Discuss} & \textbf{Eat} & \textbf{Greet} & \textbf{Phone} & \textbf{Pose} & \textbf{Purchase} & \textbf{Sit} \\ \hline\ Tekin\cite{tekin2016direct} & 102.4 & 147.7 & 88.8 & 125.3 & 118.0 & 112.4 & 129.2 & 138.9\\ Chen\cite{chen20163d} & 89.9 & 97.6 & 90.0 & 107.9 & 107.3 & 93.6 & 136.1 & 133.1\\ Zhou\cite{zhou2016sparseness} & 87.4 & 109.3 & 87.1 & 103.2 & 116.2 & 106.9 & 99.8 & 124.5\\ Xingyi\cite{zhou2016deep} & 91.8 & 102.4 & 97.0 & 98.8 & 113.4 & 90.0 & 93.8 & 132.2\\ ${CompBone}^{*}$\cite{sun2017compositional} & 90.2 & 95.5 & 82.3 & 85.0 & 87.1 & 87.9 & 93.4 & 100.3 \\ Tome\cite{tome2017lifting} & \textbf{65.0} & \textbf{73.5} & \textbf{76.8} & 86.4 & \textbf{86.3} & \textbf{68.9} & \textbf{74.8} & 110.2\\ Moreno-Noguer\cite{moreno20163d} & 69.5 & 80.2 & 78.2 & 87.0 & 100.8 & 76.0 & 69.7 & 104.7\\ Ours w Mul-Hyp & 75.6 & 75.0 & 94.9 & \textbf{82.4} & 107.7 & 91.2 & 86.6 & \textbf{73.3}\\ \hline \toprule \textbf{Method} & \textbf{SitDown} & \textbf{Smoke} & \textbf{Photo} & \textbf{Wait} & \textbf{Walk} & \textbf{WalkDog} & \textbf{WalkTogether} & \textbf{Avg} \\ \hline\ Tekin\cite{tekin2016direct} & 224.9 & 118.4 & 182.7 & 138.8 & 55.1 & 126.3 & 65.8 & 125.0\\ Chen\cite{chen20163d} & 240.1 & 106.7 & 139.2 & 106.2 & 87.0 & 114.1 & 90.6 & 114.2\\ Zhou\cite{zhou2016sparseness} & 199.2 & 107.4 & 143.3 & 118.1 & 79.4 & 114.2 & 97.7 & 113.0\\ Xingyi\cite{zhou2016deep} & 159.0 & 106.9 & 125.2 & 94.4 & 79.0 & 126.0 & 99.0 & 107.3\\ ${CompBone}^{*}$\cite{sun2017compositional} & 135.4 & 91.4 & 94.5 & 87.3 & 78.0 & 90.4 & 86.5 & 92.4 \\ Tome\cite{tome2017lifting} & 173.9 & 85.0 & 110.7 & \textbf{85.8} & \textbf{71.4} & 86.3 & \textbf{73.1} & 88.4\\ Moreno-Noguer\cite{moreno20163d} & 113.9 & 89.7 & 102.7 & 98.5 & 79.2 & 82.4 & 77.2 & 87.3\\ Ours w Mul-Hyp & \textbf{80.5} & \textbf{83.9} & \textbf{81.5} & 97.1 & 99.4 & \textbf{78.9} & 87.0 & \textbf{86.5}\\ \bottomrule \end{tabular} \end{center} \caption{Comparison with state-of-the-art on Human3.6M. No mixed 2D and 3D data training is used in all the methods. MPJPE(mean per joint position error) is used as evaluation metric.} \label{} \end{table*} \textbf{Comparison with Regression from RGB} Table \ref{} shows that \emph{Ours w/o Mul-Hyp} significantly improves baseline \emph{Direct RGB} by 12.3\emph{mm}(relative 10.8\%), demonstrating the strength of \emph{skeleton map}. \begin{table} \scriptsize \begin{center} \begin{tabular}{ll} \toprule \textbf{Method} & \textbf{Avg MPJPE} \\ \hline\ Direct RGB & 114.2\\ Ours w/o Mul-Hyp & $101.9_{\downarrow 12.3}$\\ Ours w Mul-Hyp & ${\textbf{86.5}}_{\downarrow 27.7}$\\ \bottomrule \end{tabular} \end{center} \caption{Comparison of ours with regression from raw RGB on Human3.6M. Mul-Hyp=multiple hypotheses. MPJPE metric is used.} \label{} \end{table} \emph{Does skeleton map brings better 2D estimation, or better depth estimation?} In order to answer this question, we evaluate the average joint error given ground truth depth and ground truth 2D respectively in the following. Without loss of generality, we restrict ourselves to generate one hypothesis. \textbf{Impact of Skeleton Map on 2D Estimation} We make use of ground truth depth and predicted 2D to recover 3D pose in the camera coordinate system. Table \ref{} reports the result of average joint error using different input sources for regression network. One can see that 25.9\emph{mm}(relative 27.9\%) error reduction is obtained after feeding predicted \emph{skeleton map} to regression network. Further 17.4\emph{mm} decrease is achieved by using ground truth \emph{skeleton map} for regression. This can be interpreted as \emph{skeleton map} simplifies the 2D learning procedure and prevents overfitting. Strong shape prior serves as important regularization cue for learning 2D location. \begin{table} \scriptsize \begin{center} \begin{tabular}{ll} \toprule \textbf{Method} & \textbf{Avg MPJPE} \\ \hline\ Direct RGB w GT Depth, Pred 2D & 92.8\\ Pred Ske w GT Depth, Pred 2D & $66.9_{\downarrow 25.9}$\\ GT Ske w GT Depth, Pred 2D & ${\textbf{49.5}}_{\downarrow 43.3}$\\ \bottomrule \end{tabular} \end{center} \caption{Performance given ground truth depth of different regression input on Human3.6M. Pred(GT) Ske=Use predicted(ground truth) \emph{skeleton map} for regression. Direct RGB=Use RGB for regression. Pred 2D=Use predicted 2D. GT Depth=Use ground truth depth. MPJPE metric is used.} \label{} \end{table} \textbf{Impact of Skeleton Map on Depth Estimation} To gain insight into the importance of \emph{skeleton map} for depth estimation, we use ground truth 2D and predicted depth to acquire 3D joints. We see in Table \ref{} that depth regression from predicted \emph{skeleton map} shows evident superiority over RGB image, yielding 21.5\emph{mm}(relative 22.7\%) error reduction. This indicates that \emph{skeleton map} is more favorable for depth prediction. \textbf{Impact of Multiple Hypotheses} Next we elaborate on the effect of using multiple hypotheses. Here we use $k=11$ hypotheses.\footnote{$c_i=1.0, l_i \in \lbrace 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 \rbrace$} We first assume that the ground truth \emph{skeleton map} is provided. In Table \ref{}, multiple hypotheses slightly improves the accuracy, but to a lower extent than expected. This implies that ground truth \emph{skeleton map} is sufficiently powerful to reduce ambiguity. We then move to a realistic scenario where ground truth \emph{skeleton map} is unavailable. Quite surprisingly, using multiple hypotheses reduces the average MPJPE from 101.9 \emph{mm} to 86.5\emph{mm} in Table \ref{}, which largely narrows down the performance gap between ground truth and predicted \emph{skeleton map}. Generated multiple hypotheses are illustrated in Fig \ref{}. The third hypothesis is chosen as final output based on simple matching. One could argue that similar performance might be accomplished by ensembling multiple runs of the same regression network ${Regression}_i$. To examine this, we take the regression outputs of $k=11$ different runs from single \emph{skeleton map}, denoted as \emph{Ensemble}. The result in Table \ref{} suggests that our multi-level \emph{skeleton map} provides more information than single \emph{skeleton map}. A natural problem arises: \emph{What is the performance upper bound of multiple hypotheses?} To investigate this problem, we select the optimal 3D hypothesis with minimum 3D error to ground truth 3D pose, producing an error of 68.3$mm$. This is promising as we are able to excel most state-of-the-art works \emph{without offline 3D pose library}. However, how to select the optimal 3D hypothesis remains unclear. \begin{table} \scriptsize \begin{center} \begin{tabular}{ll} \toprule \textbf{Method} & \textbf{Avg MPJPE} \\ \hline\ Direct RGB w GT 2D, Pred Depth & 94.6\\ Pred Ske w GT 2D, Pred Depth & $73.1_{\downarrow 21.5}$\\ GT Ske w GT 2D, Pred Depth & $\textbf{58.7}_{\downarrow 35.9}$\\ \hline \end{tabular} \end{center} \caption{Performance given ground truth 2D of different regression input on Human3.6M. Pred(GT) Ske=Use predicted(ground truth) \emph{skeleton map} for regression. Direct RGB=Use RGB for regression. GT 2D=Use ground truth 2D. Pred Depth=Use predicted depth. MPJPE metric is used.} \label{} \end{table} \begin{figure} \begin{center} \includegraphics[width=1.0\linewidth]{multihyp.png} \end{center} \caption{Visualization of multiple hypotheses on Human3.6M. Bottom shows the predicted hypotheses and ground truth hypotheses(white) from a novel viewpoint. Top shows the projection of 3D hypotheses and raw image. The third hypothesis is the final output.} \label{} \end{figure} \begin{figure*} \begin{center} \includegraphics[width=0.075\linewidth]{0.png} \includegraphics[width=0.075\linewidth]{1.png} \includegraphics[width=0.075\linewidth]{12.png} \includegraphics[width=0.075\linewidth]{13.png} \includegraphics[width=0.075\linewidth]{4.png} \includegraphics[width=0.075\linewidth]{5.png} \includegraphics[width=0.075\linewidth]{6.png} \includegraphics[width=0.075\linewidth]{7.png} \includegraphics[width=0.075\linewidth]{8.png} \includegraphics[width=0.075\linewidth]{9.png} \includegraphics[width=0.075\linewidth]{10.png} \includegraphics[width=0.075\linewidth]{11.png} \includegraphics[width=0.075\linewidth]{14.png} \includegraphics[width=0.075\linewidth]{15.png} \includegraphics[width=0.075\linewidth]{16.png} \includegraphics[width=0.075\linewidth]{17.png} \includegraphics[width=0.075\linewidth]{18.png} \includegraphics[width=0.075\linewidth]{19.png} \includegraphics[width=0.075\linewidth]{20.png} \includegraphics[width=0.075\linewidth]{21.png} \includegraphics[width=0.075\linewidth]{22.png} \includegraphics[width=0.075\linewidth]{23.png} \includegraphics[width=0.075\linewidth]{24.png} \includegraphics[width=0.075\linewidth]{25.png} \includegraphics[width=0.075\linewidth]{26.png} \includegraphics[width=0.075\linewidth]{27.png} \includegraphics[width=0.075\linewidth]{28.png} \includegraphics[width=0.075\linewidth]{29.png} \includegraphics[width=0.075\linewidth]{30.png} \includegraphics[width=0.075\linewidth]{31.png} \includegraphics[width=0.075\linewidth]{32.png} \includegraphics[width=0.075\linewidth]{33.png} \includegraphics[width=0.075\linewidth]{34.png} \includegraphics[width=0.075\linewidth]{35.png} \includegraphics[width=0.075\linewidth]{36.png} \includegraphics[width=0.075\linewidth]{37.png} \includegraphics[width=0.075\linewidth]{38.png} \includegraphics[width=0.075\linewidth]{39.png} \includegraphics[width=0.075\linewidth]{40.png} \includegraphics[width=0.075\linewidth]{41.png} \includegraphics[width=0.075\linewidth]{42.png} \includegraphics[width=0.075\linewidth]{43.png} \includegraphics[width=0.075\linewidth]{44.png} \includegraphics[width=0.075\linewidth]{45.png} \includegraphics[width=0.075\linewidth]{46.png} \includegraphics[width=0.075\linewidth]{47.png} \includegraphics[width=0.075\linewidth]{48.png} \includegraphics[width=0.075\linewidth]{49.png} \end{center} \caption{Qualitative results on Human3.6M(First row) and MPII(Second to fourth row). 3D poses are illustrated from a novel viewpoint. Note that 3D pose results for natural images(MPII) are quite plausible. Different colors are used to differentiate MPII from Human3.6M.} \label{} \end{figure*} \begin{table} \scriptsize \begin{center} \begin{tabular}{ll} \toprule \textbf{Method} & \textbf{Avg MPJPE} \\ \hline\ GT Ske w/o Mul-Hyp & 76.2\\ GT Ske w Mul-Hyp & $\textbf{69.7}_{\downarrow 6.5}$\\ \bottomrule \end{tabular} \end{center} \caption{Performance gain from multiple hypotheses given ground truth skeleton map on Human3.6M. MPJPE metric is used.} \label{} \end{table} \begin{table*} \scriptsize \begin{center} \begin{tabular}{lllllllll} \toprule \textbf{Method} & \textbf{Direction} & \textbf{Discuss} & \textbf{Eat} & \textbf{Greet} & \textbf{Phone} & \textbf{Pose} & \textbf{Purchase} & \textbf{Sit} \\ \hline\ Ours w/o Mul-Hyp & 94.5 & 89.1 & 103.8 & 101.7 & 131.4 & 97.0 & 107.7 & 84.4\\ Ensemble & $89.7_{\downarrow 4.8}$ & $85.2_{\downarrow 3.9}$ & $100.3_{\downarrow 3.5}$ & $97.7_{\downarrow 4.0}$ & $126.5_{\downarrow 4.9}$ & $94.0_{\downarrow 3.0}$ & $102.2_{\downarrow 5.5}$ & $80.7_{\downarrow 3.7}$\\ Ours w Mul-Hyp & $\textbf{75.6}_{\downarrow 18.9}$ & $\textbf{75.0}_{\downarrow 14.1}$ & $\textbf{94.9}_{\downarrow 8.9}$ & $\textbf{82.4}_{\downarrow 19.3}$ & $\textbf{107.7}_{\downarrow 23.7}$ & $\textbf{91.2}_{\downarrow 5.8}$ & $\textbf{86.6}_{\downarrow 21.1}$ & $\textbf{73.3}_{\downarrow 11.1}$\\ \hline \toprule \textbf{Method} & \textbf{SitDown} & \textbf{Smoke} & \textbf{Photo} & \textbf{Wait} & \textbf{Walk} & \textbf{WalkDog} & \textbf{WalkTogether} & \textbf{Avg} \\ \hline\ Ours w/o Mul-Hyp & 97.7 & 93.8 & 98.0 & 111.6 & 110.9 & 97.6 & 111.0 & 101.9\\ Ensemble & $95.0_{\downarrow 2.7}$ & $90.9_{\downarrow 2.9}$ & $94.0_{\downarrow 4.0}$ & $104.6_{\downarrow 7.0}$ & $106.8_{\downarrow 4.1}$ & $92.6_{\downarrow 5.0}$ & $106.6_{\downarrow 4.4}$ & $97.7_{\downarrow 4.2}$\\ Ours w Mul-Hyp & $\textbf{80.5}_{\downarrow 17.2}$ & $\textbf{83.9}_{\downarrow 9.9}$ & $\textbf{81.5}_{\downarrow 16.5}$ & $\textbf{97.1}_{\downarrow 14.5}$ & $\textbf{99.4}_{\downarrow 11.5}$ & $\textbf{78.9}_{\downarrow 18.7}$ & $\textbf{87.0}_{\downarrow 24.0}$ & $\textbf{86.5}_{\downarrow 15.4}$ \\ \bottomrule \end{tabular} \end{center} \caption{Performance gain from multiple hypotheses given predicted \emph{skeleton map} on Human3.6M. MPJPE metric is used. } \label{} \end{table*} \subsection{Experiments on 2D dataset MPII} We present 2D and 3D pose estimation for in-the-wild dataset MPII. We use MPII validation set \cite{tompson2014joint} including 2958 images for ablation study. \textbf{Pseudo 3D Ground Truth} MPII only provides 2D annotation, but training our network requires 3D pose ground truth. We use state-of-the-art 3D reconstruction approach \cite{zhou2017sparse} to initialize 3D pose from 2D landmark. Note that most of the reconstructed poses are already reasonable despite occasional incorrect inference. We then introduce human assistance, where a human expert is presented with the initialized 3D pose along with input image and asked to manually adjust wrong limb orientation. We stress that the goal of semi-automatic annotation is to resolve the depth ambiguity as far as possible by aligning 3D pose with image observation. Since accurate 3D MoCap pose is impratical for natural images, we call this \emph{pseudo 3D ground truth}. \begin{table} \scriptsize \begin{center} \begin{tabular}{lcccccccc} \toprule \textbf{Method} & \textbf{All}\\ \hline\ Wei\cite{wei2016convolutional} & 88.5\\ Newell\cite{newell2016stacked} & 90.9\\ Chu\cite{chu2017multi} & \textbf{91.5}\\ \toprule Carreira(IEF)\cite{carreira2016human} & 81.3\\ \toprule Sun(CompBone)\cite{sun2017compositional} & \textbf{86.4}\\ ${CompBone}^{**}$ & 79.6\\ Rogez(LCR-Net)\cite{rogez2017lcr} & 74.2\\ Ours w Mul-Hyp & 73.1 \\ \bottomrule \end{tabular} \end{center} \caption{Comparison with state-of-the-art on MPII test set. [email protected] is used as evaluation metric. All denotes [email protected] of all joints. Top section: 2D detection based. Middle section: 2D regression based. Bottom section: 3D regression based.} \label{} \end{table} \textbf{Comparison with state-of-the-art} Previous approaches generally fall into three families: 2D detection based, 2D regression based and 3D regression based. Our method belongs to 3D regression based. In this family, our closest competitors are \cite{rogez2017lcr, sun2017compositional}. The comparison is not completely fair as they both use additional 3D data. Sun \etal \cite{sun2017compositional} integrate a two-stage state-of-the-art 2D regression based method IEF\cite{carreira2016human} into their network. For the completeness of this work, we also report Stage 0 result provided in their paper, denoted as $\emph{CompBone}^{**} $. This amounts to direct regression without ad-hoc stage. We observe in Table \ref{} \emph{ours w Mul-Hyp} is on par with state-of-the-art 3D regression methods without 3D data or post processing. Qualitative results are shown in Figure \ref{}. \textbf{Comparison with Regression from RGB} Table \ref{} compares \emph{Ours w/o Mul-Hyp} with \emph{Direct RGB}. We observe that each joint gains tremendous improvement. For instance, elbow [email protected] is improved by 12.0\%(relative 20.7\%) and ankle [email protected] is improved by 5.7\%(relative 17.2\%). This again demonstrates the remarkable merit of \emph{skeleton map}. \begin{table} \scriptsize \begin{center} \begin{tabular}{lllll} \toprule \textbf{Method} & \textbf{Head} & \textbf{Sho.} & \textbf{Elb.} & \textbf{Wri.}\\ \hline\ Direct RGB & 79.1 & 75.1 & 58.0 & 46.9 \\ Ours w/o Mul-Hyp & $\textbf{90.6}_{\uparrow 11.5}$ & $\textbf{83.4}_{\uparrow 8.3}$ & $\textbf{70.0}_{\uparrow 12.0}$ & $\textbf{54.5}_{\uparrow 7.6}$ \\ \toprule \textbf{Method} & \textbf{Hip} & \textbf{Knee} & \textbf{Ank.} & \textbf{Mean} \\ \hline\ Direct RGB & 64.5 & 49.0 & 33.1 & 61.8\\ Ours w/o Mul-Hyp & $\textbf{74.2}_{\uparrow 9.7}$ & $\textbf{59.4}_{\uparrow 10.4}$ & $\textbf{38.8}_{\uparrow 5.7}$ & $\textbf{71.2}_{\uparrow 9.4}$\\ \bottomrule \end{tabular} \end{center} \caption{Comparison to direct regression from RGB on MPII validation set. Regression from only one \emph{skeleton map} increases mean [email protected] by 9.4$\%$.} \label{} \end{table} \textbf{Impact of Multiple Hypotheses} Table \ref{} shows the effect of multi-scale and multi-crop \emph{skeleton map}. We observe the same conclusion as in Table \ref{}. Using multi-scale \emph{skeleton map} results in 9.4\%(relative 24.2\%) improvement of ankle [email protected]. Multi-crop \emph{skeleton map} yields extra 7.0\% improvement. It is noteworthy that ensemble of the same regression network falls behind our final system, indicating multi-level \emph{skeleton map} is able to capture diverse semantic features from input image. \begin{table} \scriptsize \begin{center} \begin{tabular}{lllll} \toprule \textbf{Method} & \textbf{Head} & \textbf{Sho.} & \textbf{Elb.} & \textbf{Wri.} \\ \hline\ Base & 90.6 & 83.4 & 70.0 & 54.5 \\ Ensemble & $89.4_{\downarrow 1.2}$ & $83.7_{\uparrow 0.3}$ & $66.9_{\downarrow 3.1}$ & $54.8_{\uparrow 0.3}$ \\ Base+Mul-S & $90.1_{\downarrow 0.5}$ & $85.6_{\uparrow 2.2}$ & $\textbf{70.7}_{\uparrow 0.7}$ & $57.2_{\uparrow 2.7}$ \\ Base+Mul-S+Mul-C & $\textbf{90.5}_{\uparrow 0.2}$ & $\textbf{86.0}_{\uparrow 2.6}$ & ${70.5}_{\uparrow 0.5}$ & $\textbf{58.0}_{\uparrow 3.5}$ \\ \toprule \textbf{Method} & \textbf{Hip} & \textbf{Knee} & \textbf{Ank.} & \textbf{Mean} \\ \hline\ Base & 74.2 & 59.4 & 38.8 & 71.2\\ Ensemble & $73.5_{\downarrow 0.7}$ & $63.3_{\uparrow 3.9}$ & $48.9_{\uparrow 10.1}$ & $71.9_{\uparrow 0.7}$\\ Base+Mul-S & $\textbf{75.6}_{\uparrow 1.4}$ & $65.3_{\uparrow 5.9}$ & $48.2_{\uparrow 9.4}$ & $73.7_{\uparrow 2.5}$\\ Base+Mul-S+Mul-C & ${75.1}_{\uparrow 0.9}$ & $\textbf{68.4}_{\uparrow 9.0}$ & $\textbf{55.2}_{\uparrow 16.4}$ & $\textbf{74.8}_{\uparrow 3.6}$\\ \bottomrule \end{tabular} \end{center} \caption{Comparison to ours with single hypothesis on MPII validation set. Base: \emph{Ours w/o Mul-Hyp}. Mul-S: Vary stick width of \emph{skeleton map} $l_i$ in $\lbrace 5, 6, 7, 8, 9, 10\rbrace $. Mul-C: Vary crop size of raw image $c_i$ in $\lbrace 1.0, 1.25, 1.5\rbrace $. Ensemble: 18 different runs of regression from one \emph{skeleton map}. [email protected] metric is used.} \label{} \end{table} \textbf{Performance Upperbound} One remaining question is \emph{what is the limit of skeleton map applied in natural unconstrained scenario?} To assess the upper bound, we perform regression from one single ground truth \emph{skeleton map} on MPII. We see in Table \ref{} regression from single ground truth \emph{skeleton map} achieves 94.5\% overall [email protected]. This validates the effectiveness of \emph{skeleton map} representation. \begin{table} \scriptsize \begin{center} \begin{tabular}{lllll} \toprule \textbf{Method} & \textbf{Head} & \textbf{Sho.} & \textbf{Elb.} & \textbf{Wri.}\\ \hline\ Ours w/o Mul-Hyp & 90.6 & 83.4 & 70.0 & 54.5\\ GT Ske w/o Mul-Hyp & $\textbf{96.9}_{\uparrow 6.3}$ & $\textbf{99.4}_{\uparrow 16.0}$ & $\textbf{97.2}_{\uparrow 27.2}$ & $\textbf{93.4}_{\uparrow 38.9}$\\ \toprule \textbf{Method} & \textbf{Hip} & \textbf{Knee} & \textbf{Ank.} & \textbf{Mean} \\ \hline\ Ours w/o Mul-Hyp & 74.2 & 59.4 & 38.8 & 71.2\\ GT Ske w/o Mul-Hyp & $\textbf{97.7}_{\uparrow 23.5}$ & $\textbf{94.0}_{\uparrow 34.6}$ & $\textbf{70.1}_{\uparrow 31.3}$ & $\textbf{94.5}_{\uparrow 23.3}$\\ \bottomrule \end{tabular} \end{center} \caption{Comparison of regression from one predicted \emph{skeleton map}(\emph{Ours w/o Mul-Hyp}) and regression from one ground truth \emph{skeleton map}(\emph{GT Ske w/o Mul-Hyp}) on MPII validation set. [email protected] metric is used.} \label{} \end{table} \section{Conclusion} We have sucessfully shown how to push the limit of 3D human pose estimation using \emph{skeleton map} without fusing different data sources. \emph{Skeleton map} is an impressive abstraction of input, which when combined with multiple hypotheses generation is able to achieve compelling results on both indoor and in-the-wild dataset. We also carry out exhaustive experimental evaluation to understand the performance upper bound of our novel intermediate representation. We expect to further narrow down the performance gap between ground truth and predicted \emph{skeleton map} by better segmentation network. We hope the idea of combining semantic segmentation and pose estimation inspire a new research direction in 3D human pose estimation. {\small \bibliographystyle{ieee}

\section{Introduction} Progress on quantum field theory (QFT) hinges on understanding non-perturbative effects such as instantons, monopoles, vortices and so on. One of much less understood non-perturbative effects is monopole bubbling effect, which is screening of magnetic charges of singular Dirac monopoles \cite{} by regular 't Hooft-Polyakov monopoles \cite{} \footnote{See \cite{} for an explicit example.}. Aim of this paper is to understand properties of weak coupling perturbative series in the presence of monopole bubbling effects as well as instantons. The monopole bubbling effects typically appear in 't Hooft loop \cite{}, which is magnetic version of Wilson loop and detects Higgs phase by area law, while area law of Wilson loop implies confinement \cite{}. The 't Hooft loop can be defined as a partition function with singular boundary conditions for gauge fields given by Dirac monopoles \cite{}. For example, a straight 't Hooft line along the Euclidean time at spatial origin in $\mathbb{R}^4$ is described by \begin{\eq} F = \frac{B}{4} \epsilon_{\mu\nu\rho} \frac{x^\mu}{|x|^3} dx^\nu \wedge dx^\rho , \end{\eq} where $B$ is a flux on $S^2$ surrounding the loop $B=\frac{1}{2\pi} \int_{S^2} F$. It is known that 't Hooft loop receives not only instanton corrections but also monopole bubbling effects. In this paper we study perturbative series of half-BPS 't Hooft loops in 4d $\mathcal{N}=2$ supersymmetric (SUSY) gauge theories. To preserve SUSY, we need an additional boundary condition for the adjoint scalar $\Phi$ in the $\mathcal{N}=2$ vector multiplet: $\Phi = \frac{B}{2|x|}$. Here we analyze the SUSY 't Hooft loops put on a curved space. Before going to our setup in detail, let us recall general expectations on perturbative series in QFT. Perturbative series in QFT is typically not convergent \cite{}. In mathematics, there is a standard way to resum non-convergent series called Borel resummation. Given a perturbative series $\sum_{\ell =0}^\infty c_\ell g^{a+\ell} $ of a quantity $I(g)$, its Borel resummation along $\mathbb{R}_+$ is defined by \begin{\eq} \mathcal{S}_0 I (g) =\int_0^{\infty } dt\ e^{-\frac{t}{g}} \mathcal{B}I(t) , \label{} \end{\eq} where $\mathcal{B}I(t)$ is an analytic continuation of the formal Borel transformation $\sum_{\ell =0}^\infty \frac{c_\ell}{\Gamma (a+\ell )} t^{a+\ell -1}$ after the summation. Perturbative series in typical QFT is expected to be non-Borel summable due to singularities along $\mathbb{R}_+$ of $\mathcal{B}I(t)$ and Borel resummation formula has ambiguities since the integral depends on how to avoid the singularities. However, it is generally unclear when perturbative series in QFT are Borel summable. Another important question is in the case that we do not have Borel ambiguities, how the resummation is related to exact results. In \cite{} one of the authors initiated to address these questions. It has been proven that perturbative series in 4d $\mathcal{N}=2$ and 5d $\mathcal{N}=1$ SUSY gauge theories with Lagrangians are Borel summable along $\mathbb{R}_+$ for various observables in sectors with fixed instanton numbers \footnote{ See \cite{} for earlier checks in some examples. }. This result for the 4d $\mathcal{N}=2$ case is expected from a recent proposal \cite{} on a semi-classical realization of IR renormalons \cite{}, where the semi-classical solution does not exist in the 4d $\mathcal{N}=2$ theories \cite{} (see also \cite{}). In this paper we study the 't Hooft loop in the 4d $\mathcal{N}=2$ theories, which receives monopole bubbling effects in addition to instanton effects. Now we explain details of our setup. Let us consider 4d $\mathcal{N}=2$ theories with Lagrangians and non-positive beta functions. In this class of theories we study the supersymmetric 't Hooft loop placed at a circle in a squashed sphere $S^4_b$, which is defined as the hypersurface in $\mathbb{R}^5$ \begin{\eq} X_0^2 +b^{-2}(X_1^2 +X_2^2 ) +b^2 (X_3^2 +X_4^2 ) = r^2 . \end{\eq} In this geometry, we can place half-BPS line operators on the circles $S^1_b$ and $S^1_{b^{-1}}$, which are given by $X_3 =X_4 =0$ and $X_1 =X_2=0$ at fixed $X_0$, respectively. Though we focus on $S^1_b$ at $X_0 =0$, the result for $S^1_{b^{-1}}$ is simply obtained by the replacement $b\rightarrow b^{-1}$, which does not change our main result. The 't Hooft loop receives both instantons and monopole bubbling effects \cite{} \footnote{ Note that because of $|v|\leq |B|$, the contributions from the sector $|v|\neq |B|$ are relatively exponentially suppressed corrections. }: \begin{\eq} T(B) = \sum_{k,\bar{k},v} e^{-kS_{\rm inst}-\bar{k}\bar{S}_{\rm inst} -{\rm tr}(v^2 ) S_{\rm mono}} T^{(k,\bar{k},v)} (g,\theta ) , \label{} \end{\eq} where $v$ denotes a screened monopole charge and \begin{\eq} S_{\rm inst}=-2\pi i\tau ,\quad \bar{S}_{\rm inst}=2\pi i\bar{\tau},\quad S_{\rm mono}= -\frac{\pi b^2 }{2g} , \end{\eq} with a complex gauge coupling $\tau =\theta /2\pi +i/g$. Eq.~\eqref{} is schematic in the sense that theories with product gauge group have multiple couplings and multiple $(k,\bar{k},v)$. Here we are interested in weak coupling expansion by (square) Yang-Mills coupling $g$: \begin{\eq} T^{(k,\bar{k},v)} (g,\theta ) \simeq \sum_\ell c^{(k,\bar{k},v)}_\ell (\theta ) g^{\sharp +\ell} , \end{\eq} where $\sharp$ is a leading order exponent. In this paper we show that the perturbative series in the sector with fixed $(k,\bar{k},v )$ is Borel summable along $\mathbb{R}_+$ \footnote{ \cite{} studied the 't Hooft loop in $SU(2)$ $\mathcal{N}=2^\ast$ Yang-Mills theory on round $S^4$ in perturbative sector. } in the case that explicit expressions for instantons and monopole bubbling effects in SUSY localization formula \cite{} are available. Our main result is \begin{\eq} T^{(k,\bar{k},v)} (g,\theta ) =\mathcal{S}_0 T^{(k,\bar{k},v)} (g,\theta ) . \end{\eq} Thus we can rewrite the whole exact result in terms of the Borel resummation: \begin{\eq} T(B) = \sum_{k,\bar{k},v} e^{-kS_{\rm inst}-\bar{k}\bar{S}_{\rm inst} -{\rm tr}(v^2 ) S_{\rm mono}} \mathcal{S}_0 T^{(k,\bar{k},v)} (g,\theta ) . \end{\eq} This equation can be understood as ``semi-classical decoding" \cite{decoding} of the 't Hooft loop in 4d $\mathcal{N}=2$ theories though the resurgence structure itself is trivial for this case. \section{'t Hooft loop and Borel resummation} \label{} \subsubsection*{Exact result} Instead of a path integral expression, we use a conjectural finite dimensional integral representation for the 't Hooft loop \cite{} which has passed highly nontrivial tests. There are exact results for the 't Hooft loop in the 4d $\mathcal{N}=2$ theories on round $S^4$ \cite{}, $S^1 \times\mathbb{R}^3$ \cite{} and $S^1 \times S^3$ \cite{} by SUSY localization \cite{}. Although there is no explicit computation for our $S_b^4$ case, one can find the reasonable expression for the exact result by combining the results for round $S^4$ \cite{}, $S^1 \times\mathbb{R}^3$ \cite{} and the partition function on $S_b^4$ \cite{} (see also \cite{}). The expression is consistent with the exact results on round $S^4$ for $b=1$, the partition function on $S_b^4$ for $B=0$, and AGT relation \cite{} for some theories. Let us consider the 4d $\mathcal{N}=2$ SUSY gauge theories with a semi-simple gauge group $G=G_1 \times\cdots\times G_n$, which are coupled to hyper multiplets of representations $(\mathbf{R_1} ,\cdots ,\mathbf{R_{N_f}} )$. According to \cite{}, the 't Hooft loop is given by \footnote{We take $r=1$. We can recover $r$-dependence by taking $a\rightarrow ra$.} \begin{\eq} T(B) =\sum_v \int d^{|G|}a Z_{\rm cl}^{(v)} Z_{\rm 1loop}^{(v)} Z_{\rm NP}^{(v)} , \label{eq:MM} \end{\eq} where $v$ describes the screened magnetic charges. The classical contribution $Z_{\rm cl}^{(v)} (a)$ is \begin{\eq} Z_{\rm cl}^{(v)} (a) =\prod_{p=1}^n \exp\Biggl[ - \frac{{\rm tr}_p a^2}{g_p} -b \theta_p {\rm tr}_p (a\cdot v ) \Biggr] , \end{\eq} where $g$ is proportional to square of one-loop effective Yang-Mills coupling at scale $1/r$ and ${\rm tr}_p$ denotes trace in the gauge group $G_p$. The one-loop determinant $Z_{\rm 1loop}^{(v)}$ has contributions both from poles and equator of $S_b^4$ \footnote{ Note that $\alpha\cdot v$ and $\rho\cdot v$ are integers by Dirac quantization conditions \cite{}. }: \begin{\eqa} && Z_{\rm 1loop}^{(v)} (a) = \left| Z_N (a_N ) \right|^2 Z_{\rm eq}(a ) , \NN\\ && Z_N (a) =\Biggl[ \frac{\prod_{\alpha \in \Delta_+} \Upsilon (i \alpha\cdot a ) \Upsilon (-i \alpha\cdot a ) } {\prod_{I=1}^{N_f} \prod_{\rho_I \in \mathbf{R_m}} \Upsilon \left( i \rho_I \cdot a +\frac{Q}{2} \right) } \Biggr]^{1/2} , \NN\\ && Z_{\rm eq} (a) \NN\\ && = \frac{\prod_{I,\rho_I} \prod_{k=0}^{|\rho_I \cdot v|-1} \cosh^{\frac{1}{2}}{\Bigl[ \pi b \left(\rho_I \cdot a +i b \left( k-\frac{|\rho_I \cdot v|-1}{2}\right) \right) \Bigr]} } {\prod_{\alpha \in \Delta} \prod_{k=0}^{|\alpha \cdot v|-1} \sinh^{\frac{1}{2}}{\Bigl[ \pi b\left(\alpha \cdot a +ib \left( k-\frac{|\alpha \cdot v|}{2} \right) \right)\Bigr]} } ,\NN\\ \end{\eqa} where $Q=b+b^{-1}$, $a_N =a +\frac{ibv}{2}$, and $\Upsilon (x)$ is the Upsilon function defined by \begin{\eq} \Upsilon (x) =\prod_{m_1 ,m_2 \geq 0} (m_1 b +m_2 b^{-1} +x)(m_1 b +m_2 b^{-1} +Q-x) . \label{} \end{\eq} $Z_{\rm NP}^{(v)} $ expresses non-perturbative effects, which are more involved: \begin{\eqa} && Z_{\rm NP}^{(v)}(a) =Z_{\rm inst}^{(v)} (a) Z_{\rm mono}^{(v)} (a) ,\NN\\ && Z_{\rm inst}^{(v)} =|Z_{\rm Nek}(a_N )|^2 =\sum_{k,\bar{k}} e^{-kS_{\rm inst}-\bar{k}\bar{S}_{\rm inst} } Z_{\rm inst}^{(k,\bar{k},v)} , \end{\eqa} where $Z_{\rm Nek}^{(v)}$ is Nekrasov instanton partition function \cite{} with $\Omega$-background parameters $(\epsilon_1 ,\epsilon_2 )=(b,b^{-1})$, and $Z_{\rm mono}^{(v)}$ is a contribution from monopole bubbling effects. Note that $Z_{\rm mono}^{(v)}(a)$ for $v=B$ is trivial: \begin{\eq} Z_{\rm mono}^{(B)}(a) =1 , \label{} \end{\eq} which reflects the absence of the monopole screening. Now we are interested in properties of small-$g$ expansion of \begin{\eq} T^{(k,\bar{k},v)} (g,\theta ) = \int_{-\infty}^\infty d^{|G|}a \ Z_{\rm cl}^{(v)}Z_{\rm 1loop}^{(v)} Z_{\rm NP}^{(k,\bar{k},v)} , \end{\eq} where \begin{\eq} Z_{\rm NP}^{(k,\bar{k},v)} (a) =Z_{\rm inst}^{(k,\bar{k},v)}(a) Z_{\rm mono}^{(v)} (a) . \end{\eq} \subsubsection*{$SU(N)$ superconformal QCD} We first discuss 4d $\mathcal{N}=2$ $SU(N)$ superconformal QCD for simplicity of explanations which is SUSY QCD (SQCD) with $2N$ fundamental hyper multiplets. We will consider more general theories later. The 't Hooft loop of our SQCD in the sector $(k,\bar{k},v)$ is \begin{\eqa} && T_{\rm SQCD}^{(k,\bar{k},v)} (g,\theta ) = \int d^N a\ \delta \Bigl( \sum_j a_j \Bigr) e^{-\frac{1}{g}\sum_j a_j^2 -b\theta \sum_j v_j a_j} \NN\\ && \prod_{1\leq i< j \leq N} \frac{\left| \Upsilon (ia_{N,ij} ) \Upsilon (-ia_{N,ij} ) \right|} { \prod_{k=0}^{|v_{ij}|-1} \left| \sinh{\Bigl[ \pi b \left( a_{ij} +ib \left( k-\frac{|v_{ij}|}{2} \right) \right)\Bigr]} \right| } \NN\\ && \prod_{j =1}^N \frac{ \prod_{k=0}^{|v_j |-1} \cosh^{N}{\Bigl[ \pi b \left( a_j +ib \left( k-\frac{|v_j|-1}{2}\right) \right) \Bigr]} } { \left| \Upsilon \left( i a_{N,j} +\frac{Q}{2} \right) \right|^{2N} } Z_{\rm NP}^{(k,\bar{k},v)} , \NN\\ \label{} \end{\eqa} where $a_{ij}=a_i -a_j$ and $v_{ij}=v_i -v_j$. Let us study small-$g$ expansion of this object. Instead of explicitly computing perturbative coefficients, we explicitly find Borel transformation, somehow already hidden in eq.~\eqref{} as in \cite{}. To see this, first we make a change of the variables as $a_j =\sqrt{t}\hat{x}_j$, where $\hat{\mathbf{x}}=(\hat{x}_1 ,\cdots ,\hat{x}_N)$ is the unit vector in $\mathbb{R}^N$. Then, we rewrite the 't Hooft loop as \begin{\eq} T_{\rm SQCD}^{(k,\bar{k},v)} (g,\theta ) = \int_0^\infty dt\ e^{-\frac{t}{g}} f^{(k,\bar{k},v)} (t,\theta ) , \label{} \end{\eq} where \begin{\eqa} && f^{(k,\bar{k},v)} (t ) =t^{\sharp}\int_{S^{N-1}}d^{N-1}\hat{x} \delta \Bigl( \sum_j \hat{x}_j \Bigr) h^{(k,\bar{k},v)} (t,\hat{x} ) ,\NN\\ && h^{(k,\bar{k},v)} (t,\hat{x} ) =\left. e^{-b\theta \sum_j v_j a_j} Z_{\rm 1loop}^{(v)} Z_{\rm NP}^{(k,\bar{k},v)} \right|_{a_j =\sqrt{t}\hat{x}_j} . \end{\eqa} The exponent $\sharp$ is a constant depending on $N$ and $v_{ij}$, whose detail is not important for our purpose \footnote{ Explicitly, $\sharp =\frac{N-2}{2} +\sum_{i<j} \delta_{v_{ij} ,0}$. }. The functions $f^{(k,\bar{k},v)}$ and $h^{(k,\bar{k},v)}$ depend on $\theta$ but we do not often write $\theta$ explicitly in the arguments for simplicity. Since the expression \eqref{} is the form of the Laplace transformation and similar to the Borel resummation, one may wonder whether the function $f^{(k,\bar{k},v)} (t)$ of the perturbative series is the Borel transformation. Indeed, we can prove that this is the case as in \cite{}: \begin{\eq} f^{(k,\bar{k},v)} (t,\theta ) =\mathcal{B}T_{\rm SQCD}^{(k,\bar{k},v)} (t,\theta ) . \label{} \end{\eq} The proof takes the following three steps. (1) We show that the integrand $h^{(k,\bar{k},v)}(t,\hat{x})$ is identical to analytic continuation of a convergent power series of $t$. (2) We exchange the order of the power series expansion of $h^{(k,\bar{k},v)}(t,\hat{x})$ and integration over $\hat{x}$. This is rigorously justified by proving uniform convergence of the small-$t$ expansion. (3) It is easily seen that the coefficient of the perturbative series of $f^{(k,\bar{k},v)}(t)$ is given by $c_\ell^{(k)} /\Gamma (\frac{\sharp +\ell}{2})$ as guaranteed by the Laplace transformation \eqref{}. First, let us focus on perturbative sector $(k,\bar{k},v)=(0,0,B)$. We will consider non-perturbative effects later. To prove the uniform convergence of the small-$t$ expansion of $h^{(k,\bar{k},v)}(t,\hat{x})$, it is convenient to use Weierstrass M-test, i.e. we find a sequence $\{ M_\ell \}$ such that $| h^{(0)}_\ell (\hat{x})| < M_\ell $ and $\sum_{\ell =0}^\infty M_\ell t^{\sharp +\ell} <\infty $ for fixed $t$. We can easily find such a series as in \cite{} by replacing all the sources of negative contributions to the small-$t$ expansion by positive definite larger values \footnote{ For example, it is convenient to make the replacements $\pm \hat{x}_i \rightarrow 1$, $\pm\hat{x}_{ij}\rightarrow 2$, $\zeta (\ell >1 )\rightarrow 2$ and $1/[(m_1 b +m_2 b^{-1})^2 +x^2 ]$ $\rightarrow$ $1/[b^{\pm 2}(m_1 +m_2 )^2 +x^2 ]$. }. As an example, the following function generates $M_\ell$: \begin{\eqa} \frac{P(t) e^{b\theta \sqrt{t}\sum_j |v_j | -4(5N^2 -4N) b^{2}t }}{(1 - b^{2}t )^{4N^2} \prod_\pm \left( 1-2b^{2}(2t \pm \sqrt{t}Q) \right)^{2(N^2 -N)} } , \end{\eqa} where $P(t)$ is an appropriate finite order polynomial. Thus, $f^{(0,0,B)} (t )$ is actually the Borel transformation. Now we can explicitly study analytic properties of the Borel transformation. Since we have expressed the Borel transformation in terms of the one-loop determinant $Z_{\rm 1loop}^{(v)}$, this problem boils down to analytic properties of the one-loop determinant, whose details are studied in app.~\ref{} for general case. As a result, the one-loop determinant in our SQCD has degree-$2N$ poles at \begin{\eq} a_j = \pm i \left[ \left( m_1 +\frac{|v_j |+1}{2} \right) b +\left(m_2 +\frac{1}{2} \right) b^{-1} \right] , \end{\eq} where $m_1 ,m_2 \in \mathbb{Z}_{\geq 0}$. We emphasize that all the apparent branch cuts are canceled in the whole expression. Most important point here is that the one-loop determinant does not have singularities for $a\in\mathbb{R}$. Thus, the Borel transformation has singularities (poles) only along $\mathbb{R}_-$ for $b\in\mathbb{R}$ and the perturbative series is Borel summable along $\mathbb{R}_+$. \subsubsection*{General $\mathcal{N}=2$ theories with Lagrangian} We can easily generalize the above analysis to general $\mathcal{N}=2$ theories with Lagrangian. Taking the polar coordinate $a_i^{(p)}=\sqrt{t_p} \hat{x}_i^{(p)} $ with $\hat{x}_i^{(p)}\in S^{|G_p |-1}$, we find \footnote{If $G_p$ is $SU$, then we insert the traceless constraint by the delta function into $h^{(k,\bar{k},v)}$.} \begin{\eq} T^{(k ,\bar{k}, v )} (g,\theta ) = \int_0^\infty d^n t\ e^{-\sum_{p=1}^n \frac{t_p}{g_p}} f^{(k ,\bar{k}, v )} (t) , \label{} \end{\eq} where \footnote{ $(k,\bar{k},v)$ stands for $k=\{ k_1 ,\cdots ,k_p \}$, $\bar{k}=\{ \bar{k}_1 ,\cdots ,\bar{k}_p \}$ and $v=\{ v^{(1)}, \cdots ,v^{(p)} \}$. } \begin{\eqa} && f^{(k,\bar{k},v )} (t ) = t^{\sharp} \int_{\rm sphere}d\hat{x}\ h^{( k,\bar{k},v)} (t,\hat{x} ) ,\NN\\ && \left. h^{(k,\bar{k},v )} (t,\hat{x} ) = e^{-b\sum_p \theta_p {\rm tr}_p v\cdot a} Z_{\rm 1loop}^{(v )} Z_{\rm NP}^{(k,\bar{k},v )} \right|_{a_i^{(p)}=\sqrt{t_p} \hat{x}_i^{^{(p)}} } , \NN\\ \label{} \end{\eqa} where $t^\sharp =\prod_{p=1}^n t_p^{\sharp_p} $ with a constant $\sharp_p$ unimportant for our purpose \footnote{ Explicitly, $\sharp_p =\frac{|G_p|-1}{2} +\sum_{\alpha_p \in \Delta_+^{(p)}} \delta_{\alpha_p \cdot v^{(p)} ,0}$. }. Again, eq.~\eqref{} is similar to the form of Borel resummation generalized to multiple couplings. Indeed, one can show that the function $f^{( k,\bar{k},v )} (t )$ is nothing but the Borel transformation of the perturbative series. Let us focus on the sector with $(k,\bar{k},v)=(0,0,B)$ again. As in the previous case, we can always show that $f^{( k,\bar{k},v )} (t )$ is the Borel transformation: \begin{\eq} \mathcal{B}T^{(\{k ,\bar{k}, v \})} (t,\theta ) = f^{(\{k ,\bar{k}, v \})} (t,\theta ) . \label{} \end{\eq} Let us look at analytic properties of the Borel transformation. According to app.~\ref{}, the one-loop determinant has singularities at \begin{\eq} \rho\cdot a = \pm i \left[ \left( m_1 +\frac{|\rho\cdot v|+1}{2} \right) b +\left(m_2 +\frac{1}{2} \right) b^{-1} \right] . \end{\eq} This shows that the one-loop determinant does not have singularities for real $a$ and therefore the perturbative series is Borel summable along positive real axis. \subsubsection*{Instanton corrections} Generalization to non-zero instanton sector is also straightforward because $Z_{\rm Nek}(a )$ with fixed instanton number and omega background parameters $(\epsilon_1 ,\epsilon_2 )=(b,b^{-1})$ is a rational function, which does not have poles for real $a$ unless $m_1 b +m_2 b^{-1}$ ($m_{1,2} \in\mathbb{Z}$) can be purely imaginary. Therefore, eq.~\eqref{} still holds in nonzero instanton sector. For example, Nekrasov partition function for $U(N)$ SQCD with $N_f$ fundamental hypers is given by \begin{\eq} Z_{\rm Nek}(a) =\sum_Y e^{-|Y| S_{\rm inst}} \frac{\left( \prod_{j=1}^N n^f_j (a,Y) \right)^{N_f} } {\prod_{i,j =1}^N n^V_{i,j} (a ,Y ) } , \end{\eq} where $Y=(Y_1 ,\cdots ,Y_N )$ denotes a set of Young tableau in $U(N)$, $|Y|$ is the total number of boxes of $Y$, and \begin{\eqa} && n^V_{i,j}(a, Y ) = \prod_{ s\in Y_i} E_{ij}(a ,s) ( Q -E_{ij}(a ,s) ) ,\NN\\ && E_{ij} (a ,s) = -bA_{Y_j} (s ) +b^{-1}( L_{Y_i} (s) +1 ) -ia_{ij} ,\NN\\ && n^f_j (a,Y ) = \prod_{s\in Y_j } \phi_j ( a,s ) (\phi_j (a,s) +Q ) ,\NN\\ && \phi_j (a,s) = -ia_j +b ( s_h -1) +b^{-1}( s_v -1 ) . \end{\eqa} The indices $s=(s_h ,s_v )$ label a box in Young diagram at $s_h$-th column and $s_v$-th row, and $L_Y (s)$ ($A_Y (s)$) is leg (arm) length of Young tableau $Y$ at $s$. The vector contribution gives poles at \begin{\eq} a_{ij} = -ib \left( \frac{v_{ij}}{2} -A_{Y_j} (s) \right) -ib^{-1} \left( L_{Y_i}(s)+1\right) , \end{\eq} and \begin{\eq} a_{ij} = ib \left( 1-\frac{v_{ij}}{2} +A_{Y_j} (s) \right) -ib^{-1}L_{Y_i}(s) . \end{\eq} Although the contribution from each Young diagram may have poles at $a_{ij} =0$, this type of poles are canceled after summing over Young diagrams with the same instanton number \cite{}. Thus the Borel transformation has singularities only along $\mathbb{R}_-$ and especially perturbative series is Borel summable along $\mathbb{R}_+$ even if we include the instanton corrections. \subsubsection*{Monopole bubbling effects} Now let us add monopole bubbling effects. As in the instanton corrections, generalization is straightforward as long as we know explicit expressions of $Z_{\rm mono}^{(v)}(a)$ because the monopole bubbling effect is described by a ratio of finite products of hyperbolic functions, whose poles are not on real axis. Hence, \eqref{} still holds in the presence of the monopole bubbling effects. For example, for $SU(N)$ or $U(N)$ gauge theory with fundamentals and adjoints, the contribution from the monopole bubbling effects is given by \begin{\eq} Z_{\rm mono}^{(v)}(a) = \hat{\sum_{Y}} Z_Y^{\rm vec} (a) \prod_I Z_Y^{R_I} (a) , \label{} \end{\eq} where \begin{\eqa} && Z_Y^{\rm vec} (a) = \frac{1} {\hat{\prod}_{i, j, s\in Y_i, \pm} \sinh{\frac{2\pi ba_{ij} +i\pi b^2(A_{Y_i}(s) -L_{Y_j}(s) \pm 1 )}{2}} } ,\NN\\ && Z_Y^{\rm adj} (a) = \hat{\prod}_{i,j,s\in Y_i} \cosh^2{\frac{2\pi b a_{ij}+i\pi b^2 (A_{Y_i}(s) -L_{Y_j}(s) )}{2} } ,\NN\\ && Z_Y^{\rm fund} (a) = \hat{\prod}_{j, s\in Y_j} \cosh{\frac{2\pi ba_j +i\pi b^2 ( s_v +s_h -1)}{2} } . \label{} \end{\eqa} The sum in \eqref{} is over a set of Young tableau with the total number of boxes $\frac{1}{2}{\rm tr}(B^2 -v^2)$. Note also that we have put the symbol $\hat{}\ $ for the sum and product which means the sum over $Y$ and product over $s$ are constrained. Namely, we include only $Y$ and $s$ satisfying the constraints in contrast to the instanton corrections. We explain details on this in app.~\ref{} because the constraints are quite complicated and nevertheless their details are not so important for our purpose. The most important thing for us is singularity structure of $Z_{\rm mono}^{(v)}(a)$. For the contribution from single set of Young diagram $Y$, the vector contribution gives poles at \begin{\eq} a_{ij} = ib^{-1}m -\frac{ib}{2} \left( A_{Y_i} (s) -L_{Y_j}(s) \pm 1 \right) . \end{\eq} When $A_{Y_i}(s) -L_{Y_j}(s) \pm 1 \neq 0$ or $m\neq 0$, the poles are not located along real axis unless $b^2$ is purely imaginary. When $A_{Y_i}(s) -L_{Y_j}(s) \pm 1 = 0$, we have the poles at $a_{ij} =0$ but this type of poles are canceled from other Young diagrams as long as we consider well-defined 't Hooft loops. Thus, $Z_{\rm mono}^{(v)}(a)$ does not have singularities along real axis and the perturbative series is Borel summable along $\mathbb{R}_+$. Furthermore, the Borel transformation has singularities only along $\mathbb{R}_-$ for $b\in\mathbb{R}$ including all the non-perturbative corrections. \subsubsection*{Dyonic loop} Supersymmetric dyonic loop can be computed by putting SUSY Wilson loop on $S_b^1$ \cite{} in the setup of the 't Hooft loop. If we assume the conjectural expression \eqref{eq:MM} for the 't Hooft loop, then the exact result for the dyonic loop should be given by \begin{\eq} D=\langle {\rm tr}_R e^{ba} \rangle_{M.M.}, \end{\eq} where $\langle \cdots \rangle_{M.M.}$ denotes unnormalized expectation value in the matrix model \eqref{eq:MM}. Since this is just insertion of the function of $a$ without singularities, this does not prevent us from application of the above technique to this case. Then the Borel transformation is simply given by $f^{(k,\bar{k},v)}$ in \eqref{} with the extra insertion of ${\rm tr}_R e^{ba}$ to the integrand. Thus, the analytic properties of $h^{(k,\bar{k},v)}$ do not change and the perturbative series is still Borel summable along $\mathbb{R}_+$ including the instantons and the monopole bubbling effects. \section{Conclusion and Discussions} \label{} In this paper we have studied weak coupling perturbative series in the presence of monopole bubbling effects as well as instanton effects. We have shown that the perturbative series of the 't Hooft loop in the 4d $\mathcal{N}=2$ supersymmetric gauge theories are Borel summable along $\mathbb{R}_+$. It has turned out that the exact result is the same as the sum of the Borel resummations over instanton-anti-instanton numbers $(k,\bar{k})$ and screened magnetic charges $v$. Our result is also non-trivially consistent with the conjecture that 4d $\mathcal{N}=2$ theories do not have IR renormalon type singularities \cite{}. There is a confusing point on our results. According to Lipatov's argument \cite{} and topological selection rule \cite{} (see also \cite{}), it is expected that Borel singularities come from saddle points with the same topological number. Hence, one may expect that there are Borel singularities coming from instantons-anti-instantons with the same $k-\bar{k}$ in our setup. However, we have seen that we do not have such singularities. Absence of this type of singularities as well as IR renormalons lead the Borel summability along $\mathbb{R}_+$ and makes the perturbative series in every sector isolated in contrast to usual resurgence scenario in quantum mechanics \cite{} and QFT \cite{}. It is interesting to find physical interpretations for this point. One of possible scenarios would be ``Chesire cat resurgence", which has recently appeared in some SUSY theories \cite{}. Another possibility is that there is something like a generalization of topological selection rule particular for our setup. We have found that the Borel transformation has infinitely many singularities in Borel plane. It is interesting to find their physical interpretations. Technically, these singularities come from singularities of the one-loop determinant, Nekrasov partition functions, and monopole bubbling effects. At least for those from the one-loop determinant, we expect that they can be explained by complexified SUSY solutions as in 3d $\mathcal{N}=2$ case \cite{}, which formally satisfy SUSY conditions but may not be on the original path integral contour. The formula for the 't Hooft loop, which we have used, has not been explicitly computed by the localization despite it has the passed nontrivial checks. It is nice to derive the localization formula directly. It is known that the 't Hooft loop operator is S-dual to SUSY Wilson loop, whose perturbative expansion is also Borel summable along $\mathbb{R}_+$ \cite{}. It would be illuminating to study whether there are implications of the $S$-duality for structures of Borel singularities. \subsection*{Acknowledgments} We thank Takuya Okuda for his correspondence to our questions on \cite{}. M.~H. would like to thank Centro de Ciencias de Benasque, CERN, Fudan University, KITP, RIKEN, and YITP for hospitality. D.~Y. would like to thank Tokyo Institute of Technology for hospitality. This research was supported in part by the National Science Foundation under Grant No. NSF PHY-1125915. D.~Y. was supported by the ERC Starting Grant N. 304806, ``The Gauge/Gravity Duality and Geometry in String Theory''.

\section{Introduction} Synthetic microswimmers or active particles (APs) \cite{Paxton,BechingerReview,Dreyfus,howse,palacci,jiang,buttinoni,romanczuk,zoettl} currently receive considerable attention because their properties resemble the behavior of living microorganisms \cite{Cates2012,elgeti2015,Schwarz,Polin}. Similar to their biological counterparts, APs can form clusters \cite{Theurkauff2012,Palacci2013,Buttinoni2013}, respond to external chemical \cite{Liebchen} and light \cite{Lozano2016} gradients or to gravity \cite{Borge2014} and external flows \cite{Palacci2016}. In both cases, the translational mean square displacement (MSD) displays a short-time ballistic and a long-time effective diffusive behavior. Despite such common features between biological and synthetic microswimmers, notable differences occur. For instance, the trajectories of \textit{E. Coli} are comprised of alternate periods of rather straight segments with almost constant velocity $v$ (\textit{runs}), and abrupt short-lived reorientation events (\textit{tumbles}) which result in rather sudden directional changes \cite{BergRT,Rupprecht}. In contrast, the angular dynamics of synthetic microswimmers is determined by their rotational diffusion which - for micron-sized particles - is on the order of several tens up to hundreds of seconds \cite{catesRT,Schwarz}. Because orientational and positional changes are superimposed, directional changes of active particles are typically rather smooth compared to their biological counterparts. Here we demonstrate that APs can also perform a swimming motion where translational and orientational changes are decoupled in time. Such run-and-tumble-like propulsion scheme can be induced in a suspension of APs being immersed in a viscoelastic medium when their self-propulsion is periodically turned on and off with modulation time $T_{mod}$. While the particles perform a rather straight forward swimming motion during the on-period, they experience a pronounced elastic recoil when the propulsion is turned off. Similar recoils have been previously observed in microrheological experiments where a (passive) colloidal particle was externally driven by an optical laser tweezer through a viscoelastic liquid ~\cite{RubenNJP}. When the optical trap was suddenly turned off, the relaxing microstructure of the fluid caused a transient backward motion of the particle, i.e. an elastic recoil. In case of an AP with time-modulated self-propulsion as studied here, such recoils are found to strongly affect both, its translational and orientational dynamics. We observe that the shape of the trajectories strongly depends on $T_{mod}$ and that particle reorientation is most rapid when it is close to the relaxation time $\tau$ of the fluid. Our experimental findings suggest novel steering strategies for active colloids in e.g. crowded systems \cite{Li} or within patterned confinements \cite{Yazdi}. \begin{figure} \includegraphics[width=0.9\columnwidth]{Fig1.pdf} \caption{ (a) Schematic illustration of the sample cell containing a dilute suspension of light-activated Janus particles in a fluid. (b) Phase diagram of a Newtonian binary critical mixture of water and PnP (blue) with the critical point at $T_c$ = $31.4^{\circ}$C and 0.4 PnP mass fraction. The green line corresponds to the phase diagram after addition of $0.05\%$ polyacrylamide (PAAm), which renders the fluid viscoelastic. Inset: snapshot of the spinodal demixing of the viscoelastic fluid at $T$ = $32^{\circ}$C. The nature of the demixing process is not affected by the presence of PAAm. (c) Self-propulsion velocity of a Janus particle with $2R = 7.75$~$\mu$m as a function of the light intensity under steady illumination conditions for Newtonian (blue circles) and viscoelastic mixture (green squares). (d) Intensity protocol to realize a periodic modulation of particle self-propulsion. (e),(f) Positional and orientational configuration of an active particle whose self-propulsion is modulated in time with $T_{mod}= 1.5s$ and $I_{max}=4$~$\mu$W/$\mu$m$^2$. (g),(h) Translational and angular mean-square displacements for periodical modulation (same conditions as in (e,f)) (blue) and homogeneous illumination with $I_{max}=2$~$\mu$W/$\mu$m$^2$ to ensure identical conditions (gray).} \label{} \end{figure} \section{Experimental description} \begin{figure} \includegraphics[width=\columnwidth]{Fig2.pdf} \caption{ (a) Typical trajectory of a self-propelled particle in a viscoelastic fluid at constant illumination with $I=2$~$\mu$W/$\mu$m$^2$. (b) Trajectory of the same particle for $T_{mod}= 1.5s$ and $I=4$~$\mu$W/$\mu$m$^2$. The arrows indicate events where the self-propulsion is turned off. (c),(d) Positional and orientational configuration of an active particle with periodically modulated self-propulsion for the same conditions as in (b).} \label{} \end{figure} As active colloids, we used spherical silica particles with diameter $2R = 7.75$~$\mu$m, which are half-coated by a 50 nm sputtered carbon layer (Fig.1a). Such Janus-particles are immersed in a critical mixture of water and propyleneglycol n-propyl ether (PnP) with 0.4 PNP mass fraction. To render the fluid viscoelastic, we added $0.05\%$ polyacrylamide (PAAm) ($M_w=18 \times 10^{6}$). The solution exhibits a lower critical point as shown by the green line in Fig.1b. In the semidilute regime, such solutions are viscoelastic with a stress-relaxation time $\tau=1.50 \pm 0.10$s (for details, see~\cite{RubenPRL}). As experimentally confirmed, the solution displays shear thinning behavior for Weissenberg numbers $\mathrm{Wi} > 1$, where Wi$=v\tau/2R$, with $v$ the particle propulsion velocity \cite{gomezsolano1}. It should be emphasized that in all experiments shown below, $\mathrm{Wi} < 0.3$ where the fluid has a constant (zero) shear viscosity $\eta_0$=$0.100 \pm 0.015$ Pa s. In this regime, the mean viscous friction force on a particle has been confirmed to be proportional to its velocity ~\cite{RubenNJP}. A small amount of APs is immersed in the mixture and contained in a thin sample cell with height $h$ $\approx 4R $ (Fig.1a). Under such spatial confinement translational and rotational dynamics is limited to two dimensions ~\cite{das,popescu,rubenSPP}. In the absence of self-propulsion, the translational and rotational diffusion coefficients determined from our experiments are $D^0_t = (3.69 \pm 0.60)\times 10^{-4}\,\mu \mathrm{m}^2\,\mathrm{s}^{-1}$ and $D^0_r = (1.84 \pm 0.40)\times 10^{-5}\,\mathrm{s}^{-1}$, respectively. This value is about 50\% below the corresponding Stokes-Einstein values and caused by hydrodynamic coupling to the walls \cite{brenner}. The sample cell is thermally coupled to a water bath which is kept constant below the critical temperature at $T=25 \pm 0.1^{\circ}$C. When the cell is illuminated with laser light ($\lambda$ = 532 nm), it is absorbed by the carbon caps, which results in local demixing of the fluid (for our intensities this demixing region is at least one order of magnitude smaller than the particle size \cite{rubenSPP}). This leads to a self-diffusiophoretic motion whose propulsion velocity is controlled by the incident light intensity as shown by the green symbols in Fig.1b \cite{buttinoni,samin,rubenSPP}. To rule out a possible temperature dependence of the rheological properties of the viscoelastic mixture, which would influence the viscous drag experienced by APs, we have measured the shear-rate dependent viscosity of our system between $25^{\circ}$C (bath temperature) and $31^{\circ}$C, the latter being close to $T_C$ (for further details see appendix A). Within our experimental resolution, only a negligible temperature-dependence is observed. Time-dependent propulsion forces were created by periodic variation of the laser intensity with $T_{mod}$ by a mechanical shutter, which periodically switches the illumination intensity between $I_{max}$ and zero (Fig. 1d). In our protocol, we have chosen identical \emph{on} and \emph{off} duration times which guarantees that the time-averaged incident light intensity is independent of $T_{mod}$. Particle positons and orientations were obtained by digital video microscopy with a frame rate of 10 Hz and 50nm of spatial resolution. The particle orientation was determined directly from the optical contrast due to the carbon cap ~\cite{RubenPRL}. \section{Experimental Results} \begin{figure} \includegraphics[width=0.7\columnwidth]{Fig3.pdf} \caption{Comparison of the recoil motion of an AP when the illumination is suddenly turned off (green) and that of a colloidal particle which is suddenly released from a moving optical trap (red). The particle velocity prior to the recoil (t=0) was in both cases $v \approx 0.4\mu m/s$ and the particle size $2R = 7.75 \mu m$. The solid black line corresponds to an exponential fit $x \propto \exp(-t/\tau)$, where $\tau = 1,5$~s. Inset: semilog representation of the main figure.} \label{} \end{figure} Before discussing the behavior of active particles in viscoelastic fluids, we will briefly discuss their motion in a Newtonian binary mixture, i.e. in absence of the PAAm. As shown by the blue lines in Fig.1b,c phase diagram of a water-PNP mixtures is little affected by the presence of PAAm and only leads to small changes in the critical temperature and the light-induced particle propulsion. Figs.1e,f show the in-plane positional and orientational component of an active particle which is subjected to a periodic laser illumination. In contrast to the translational motion, which alternates between self-propulsion (light on) and Brownian diffusion (light off), no differences in the angular component are found. The latter is expected, because for homogeneous illumination, the particle orientation is entirely determined by its rotational diffusion and independent of self-propulsion. To compare the particle's motion in Newtonian liquids for continuous and periodic illumination, we have analyzed the positional and orientational component of their trajectories. To allow for a direct comparison, the time-averaged illumination intensity was identical in both cases. Figs.1g,h show the positional and angular mean square displacements (MSDs) for periodic, $T_{mod}=1.5 s$ (blue), and steady (gray) illumination, respectively. Apart from deviations in the short time-behavior of the translational MSD \cite{hartmut}, the diffusive long-time behavior is identical for periodic and steady illumination. This demonstrates that a steady state in the particle motion is reached within the illumination period. Such behavior is consistent with the fact that the temperature profile around a micron-sized capped colloid suspended in a liquid with thermal diffusivity $\alpha \sim 10^{-7}m^2/s$ reaches its steady state after a time scale $R^2/\alpha \sim10^{-3}$s \cite{rubenSPP}. Note that this is at least two orders of magnitude shorter than the time-scales of our experiments, namely $\tau$ = 1.5 s and the smallest modulation period $T_{mod}$ = 0.2 s in our experiments. In the following, we discuss our observations after adding PAAm to the critical solution, i.e. in a viscoelastic fluid. Because the viscosity is enhanced by a factor of 25 after adding the polymer ($\eta_0$=$0.100 \pm 0.015$ Pa s), the particle's diffusive reorientation dynamics becomes much slower compared to above. This is seen in Fig.2a, which shows the rather straight particle trajectory under constant illumination over 400s \cite{footnote1}. When the active motion is periodically turned on and off, drastic directional changes appear. These reorientation events mainly occur when the illumination is turned off, as shown by the arrows in Fig.2b. The effect of the modulated propulsion on the particle's trajectory is seen in more detail in Fig.2c. During each illumination period the particle is translated by several hundreds of nanometres. In the absence of illumination, a much smaller reverse motion occurs. Contrary to the Newtonian case, here pronounced differences between periods with and without illumination are also observed in the orientational particle dynamics (Fig.2d): while only slow orientational changes occur during self-propulsion, abrupt directional changes arise when the illumination is turned off. The above observations can be explained by an elastic recoil, which is experienced by the particle after illumination is turned off. The presence of such recoils has been recently demonstrated in microrheological experiments, where a colloidal particle was driven by a moving optical tweezer through a viscoelastic solution \cite{RubenNJP}. When the driving force was suddenly removed, a back motion due to the recovery of the deformed fluid's microstructure was observed. Such behavior, which is absent in Newtonian liquids, is caused by the rather long (seconds) structural relaxation times being characteristic for viscoelastic fluids \cite{Bird}. Fig.3 (red curve) shows the result of such a recoil experiment in the same viscoelastic solvent and for the same bath temperature $T=25^{\circ}$C and particle size $2R = 7.75 \mu m$ as used in the present study. The velocity before release was set to $v \approx 0.4 \mu m/s$ which results in an exponential decay with a decay time of $\tau = 1.5s$ (black line) \cite{footnote2}. The green curve shows the transient of an equally sized active particle, which was originally self-propelling with velocity $v \approx 0.4 \mu m/s$ and whose active motion was suddenly stopped by turning the illumination off. In both cases, we observe the same exponential decay, which suggests the occurrence of elastic recoil in APs in viscoelastic fluids with modulated self-propulsion. In analogy with the Deborah number (De), which is used in rheology to characterize flows of viscoelastic materials under time-dependent deformations, in the following we will quantify the motion of active particles with periodic propulsion modulation by \begin{equation}\label{eq:De} De=\frac{\tau}{T_{mod}}. \end{equation} \begin{figure} \includegraphics[width=\columnwidth]{Fig4.pdf} \caption{(a) Particle trajectories (400 s each) for different modulation times and for $I_{max}=4$~$\mu$W/$\mu$m$^2$. The local particle orientations are indicated by arrows. (b) Normalized probability distribution functions of orientational changes $\Delta \theta$ during $\Delta t=T_{mod}$ for different De. (c) Corresponding orientational mean square displacements which indicate an effective diffusive behavior at long times. (d) Normalized orientational diffusion coefficient vs. De for $I_{max}= 2$~$\mu$W/$\mu$m$^2$ (black circles), $4$~$\mu$W/$\mu$m$^2$ (red squares). (e) Normalized orientational diffusion coefficient vs. intensity $I$ at De=1.} \label{} \end{figure} Fig.4a shows typical APs trajectories measured over 400 s duration and for $T_{mod}= 0.2, 1.5, 2.5, 4.0 s$ (corresponding to De = 7.5, 1, 0.6, 0.375) and for $I_{max}$=$4$~$\mu$W/$\mu$m$^2$ , where the arrows show the corresponding particle orientation determined from the video images by detecting the cap orientation. At low and high De, the trajectories are rather straight, but become quite wiggly at De =1. To quantify the change of the particle orientation $\Delta \theta$ during the off-period, we have calculated the normalized probability distributions of $\Delta \theta$ (Fig.4b). Compared to constant illumination (De=0), the distributions become wider, which suggests that the particle reorientation is strongly modified by $T_{mod}$. Around De=1, the width of the distribution is the largest and exhibits orientational changes of several tens of degrees after a single modulation. We have also calculated the mean-square displacement of the angular particle orientation $\theta$ , $\langle \Delta \theta (t)^2 \rangle = \langle [\theta(t_0 + t) - \theta(t_0)]^2 \rangle $, where the brackets denote time averaging over $t_0$. We plot the behavior of $\langle \Delta \theta (t)^2 \rangle$ for different De in Fig.4c. Towards long times, the growth of $\langle \Delta \theta (t)^2 \rangle $ becomes linear with $t$. Interestingly, the slope changes with De. Therefore, in the long-time limit we can determine an effective rotational diffusion coefficient $D^{(\mathrm{De})}_r$, which depends on the Deborah number, as \begin{equation}\label{} \langle \Delta \theta (t)^2 \rangle = 2D^{(\mathrm{De})}_r t. \end{equation} From the slope we obtain a De-dependent $D_r$ which is plotted (normalized to $D_r^0$) in Fig.4d. The graph displays a maximum at De=1 with an enhancement factor around 800. We also investigated the dependence of $D_r/D_r^0$ on the propulsion velocity. The data, which are plotted in Fig.4e for De=1, show a remarkably strong and nonlinear increase of the rotational diffusion on $I_{max}$. Qualitatively, this can be understood by considering that the elastic stress in the fluid increases when the particle is faster. Accordingly, the deformation of the fluid and, thus, the recoil (being responsible for the particle's reorientation dynamics) becomes enhanced. As mentioned above, the elastic recoil also affects the translational particle dynamics. This is shown in Fig.5a, where we show that the translational MSD $\langle \Delta r (t)^2 \rangle$, (calculated in the same manner as the angular MSD) obtained from trajectories of at least 3600 s for different De. The deviations from a linear behavior (oscillations) at short times are due to the propulsion modulation. Similar to the situation of Newtonian liquids (Fig.1g), the long-time behavior is linear in time, i.e. entirely diffusive ($\langle \Delta r (t)^2 \rangle$ = $4D_t t$ ). Finally, we also have investigated the dependence of $D_t/D_t^0$ vs. De for $I_{max}= 4$~$\mu$W/$\mu$m$^2$ (red squares) and $2$~$\mu$W/$\mu$m$^2$ (black circles). Independent of the illumination intensity, we observe a strong enhancement of $D_t/D_t^0$ around De=1. A dependence of the translational diffusion coefficient for APs with time modulated propulsion force has been theoretically predicted also for viscous fluids \cite{hartmut}. The magnitude of the enhancement observed by us is, however, considerably larger and, obviously, results from the periodic elastic recoils which are experienced by the AP when its self-propulsion is modulated in time. \begin{figure} \includegraphics[width=0.6\columnwidth]{Fig5.pdf} \caption{(a) Translational mean square displacement for different De = 7.5, 1.5, 1, 0.6, 0.375, for $I_{max}=4.0$ $\mu$W$/$$\mu$$m^2$. The solid line with slope equal to 1 shows the long-time translational diffusivity. (b) Normalized translational diffusion coefficient vs. De for $I_{max}= 2$~$\mu$W/$\mu$m$^2$ (black circles) and $4$~$\mu$W/$\mu$m$^2$ (red squares). } \label{} \end{figure} \section{Conclusions} In contrast to Newtonian fluids, where orientational changes of APs are usually determined by their slow rotational diffusion, we have shown that much faster directional changes can be induced in viscoelastic environments when the self-propulsion of APs is periodically modulated in time. As a result of the modulated propulsion strength, the translational and directional dynamics becomes disentangled in time. Our observations are explained by the elastic stress, which builds up during self-propulsion and which is suddenly released when the propulsion is turned off. Because this relaxation process couples to the particle dynamics, it results in an increased translational and rotational particle diffusion, which peaks when modulation time is identical to the microstructural relaxation time of the fluid. In addition to a variation of the modulation time, we expect that the particle's dynamics should also depend on the functional shape of the modulation (rectangular, sinusoidal, sawtooth, etc.) and thus allows to optimize the locomotion to different types of environments. Because the particle reorientation dynamics is key for the stability of clusters and the motion of APs through topographical landscapes, the possibility to control the rotational motion of APs may suggest novel strategies for particle steering under conditions where rapid directional changes are required. \section*{Acknowledgments} We thank H-J. K\"ummerer, C. Mayer and U. Rau for their technical support. C. B. acknowledges financial support from the German Research Foundation (DFG) through the priority programme SPP 1726 on microswimmers and by the ERC Advanced Grant ASCIR (Grant No.693683). J-R.G-S. was supported by Deutsche Forschungsgemeinschaft (DFG) grant No. GO 2797/1-1.

\section{End-to-End Framework for learning Cellular Automata rules} In this section, we discuss the details of our architecture, illustrated in Fig. \ref{} (a), for generating a transition function $f_T$ for the cellular automata model described in eqn (\ref{}). The architecture comprises of the following components. \begin{itemize} \item Two data repositories consisting of built-up rasters and remotely sensed rasters which will be required during model building. \item Data representation module is for representing the data obtained from the data repositories in such a way so as to facilitate learning and knowledge representation. \item The learning algorithm stage is expected to learn patterns from the data which is received from the previous stage. \item The knowledge representation module stores the knowledge learned by the learning algorithm for further querying. \item Finally, the prediction stage predicts future built-up conditions. \end{itemize} \subsection{Data Representation} \label{} The data representation component consists of procedures to build a \textit{data matrix} and a \textit{label matrix} from the raster and built-up maps. The \textit{data matrix} is formed from the raster and \textit{built-up} maps in form of a matrix of size $n\times (1+|N(l_p^t)|+|\phi_{raster}^{len}(R_p, N(R_p))|)$, where $n$ is the number of data points in the raster, $N()$ is a neighborhood function and $|.|$ is the cardinality of a set. The neighborhood set $N()$ plays a significant role in the framework, as it represents the effect of neighboring pixels. Increase in the neighborhood set size increases the model complexity by increasing the dimensions of the \textit{data matrix} (\cite{sante2010cellular}). \begin{figure} \centering \subfloat[]{\includegraphics[scale=0.3]{\locimage{ca_decision_tree}}} \subfloat[]{\includegraphics[scale=0.3]{\locimage{transition_classes}}} \caption{(a) \textit{End-to-End} framework for prediction of urban growth (b) Venn diagram of transition classes relevant to Urban growth prediction} \label{} \end{figure} A crucial step at this stage is the design of the encoder function $\phi_{raster}^{len}$. We can build an encoder function by training an autoencoder\cite{vincent2010stacked} with the vectors of form $<R_p, N(R_p)>$ to form a $len$ length representation. An autoencoder is a neural network consisting of an encoder which encodes the data vector of a certain size and a decoder which reconstructs the data vector back from the encoding. The network is trained by minimizing the error between the decoded output and the original output. Finally the encoder from the autoencoder is used to encode the input which is known as a representation. Since the method is unsupervised, therefore there is no requirement to label the data for encoder design. The encoder design for our framework is depicted in Fig. \ref{}. If a vector is represented by $X = <R_p, N(R_p)>$, then $\phi_{raster}^{len}$ can be expressed as \begin{eqnarray} \phi_{raster}^{len}(X) = \sigma(W_k^e...\sigma(W_1^eX)) \label{} \end{eqnarray} , where $W_k^e$ is a weight matrix and $\sigma$ is an activation function (eg. sigmoid, relu). The eqn. \ref{} represents a $k$ layer feed forward neural network which encodes the information in a raster position and its neighborhood. Similar to the encoder function, we can express the decoder function (say $\phi'$) as \begin{eqnarray} \phi'(X) = \sigma(W_k^d..\sigma(W_1^d \phi_{raster}^{len}(X))) \label{} \end{eqnarray} , where $W_k^d$ is a weight matrix and $\sigma$ is an activation function. The only constraint in eqn. \ref{} is that the output vector $\phi'(X)$ must have the equal dimension to that of $X$. Then the optimization of the parameters can be done using the following minimization equation. \begin{eqnarray} E = \frac{1}{n} \sum{(X-\phi'(X))^2} \label{} \end{eqnarray} Subsequent to the optimization, the representations of length $len$ can be generated from the vectors $<R_p, N(R_p)>$ using the function $\phi_{raster}^{len}$ as in eqn. \ref{}. \begin{figure} \centering \includegraphics[scale=0.6]{\locimage{rasterrep}} \caption{Learning Representation from Neighborhood using an Autoencoder.} \label{} \end{figure} The \textit{data matrix} can be obtained by concatenating the \textit{built-up} vectors ($<l_p^t,N(l_p^t)>$) and \textit{raster representation} ($<\phi_{raster}^{len}(R_p, N(R_p))>$) columnwise. The \textit{raster representation} variables are non-categorical with values ranging from $[-1,1]$, whereas the \textit{built-up} variables are categorical with value either $-1$ (Non Built-up) or $1$ (Built-up). The rows of the \textit{data matrix} can be directly supplied to the function given in eqn. (\ref{}). The \textit{label matrix} is an $n\times 1$ matrix which consists of labels of $4$ transition classes namely, \textit{Non Built-up} to \textit{Non Built-up}, \textit{Non Built-up} to \textit{Built-up}, \textit{Built-up} to \textit{Non Built-up} and \textit{Built-up} to \textit{Built-up}. The transition classes and their class labels are illustrated as a Venn diagram in Fig \ref{} (b). Classes $2$ and $3$ are within classes $0$ and $1$ because in the dataset usually number of transition cells are lesser in number than number of persistent cells. The rows of the \textit{data matrix} form the input vector to the function $f_T$, whereas rows of the label matrix form output values of the function $f_T$. The procedure for building the \textit{data and label matrix} is given in Algorithm \ref{}. Contrary to other works (\cite{moghadam2013spatiotemporal, shafizadeh2015performance, yang2008cellular}), where two classes \{\textit{Built-up, Non Built-up}\} have been used for training classifiers, we have used four classes in our \textit{label matrix}. This resolves two following issues. \begin{itemize} \item With the two class approach, the framework turns into a built-up classification problem because the raster images consist of information regarding built-up. In our case, we have four classes out of which two are transition and two are persistence classes. This tweak divides the \textit{data matrix} into four parts and patterns of transition and persistence classes get separated. The classifier still does classification, but the pattern classification problem turns into, ``Which patterns lead to transition and which ones lead to persistence?". \item Comparing the \textit{Built-up} maps of two close time steps, we have observed that transition from \textit{Non Built-up} to \textit{Built-up} and \textit{Built-up} to \textit{Non Built-up} occurs in a small fraction of cells when compared to cells which are persistent. In the $2$-class approach of modeling, certain learning algorithms tend to ignore patterns that are less fraction of the dataset considering them as noise. For instance, if there are $10\%$ transition pixels and $90\%$ persistent pixels, then a default choice for high predictive accuracy is to give an accuracy of $90\%$ (\cite{chawla2002smote}). The imbalance in the dataset is unavoidable in this case as it is a natural property of the problem at hand, which gets covered up when not considered separately. Imbalanced datasets are common in many practical problems and have been addressed multiple times (\cite{chawla2005data}). In our case, we have used three separate metrics, namely \textit{Figure of Merit (FoM), Producer's Accuracy (PA)} and \textit{User's Accuracy (UA)} other than classification accuracy to track whether this problem is tackled by the knowledge representation or not (\cite{pontius2008comparing}). These three metrics are designed to check whether a landuse change model can predict the transitions properly. Therefore, it is crucial to have high FoM, PA and UA along with OA to guarantee a good simulation technique. \end{itemize} \subsection{Knowledge Representation} \textit{Knowledge as defined by Fischler and Firschein 1987, refers to stored information or models used by person or machine to interpret, predict, and appropriately respond to the outside world}. By the very nature of the definition, knowledge is task specific (\cite{haykin2009neural}). There are various ways in which knowledge can be represented such as rules (Decision Trees), decision surfaces (SVM), computation nodes (ANN), probability distributions (Probabilistic Graphical models) and nearest neighbors (KNN). Table \ref{} provides a description of the various kinds of supervised classification methods, which can be considered in the \textit{End-to-End} framework for knowledge representation (Only short mathematical descriptions are provided in the table). The reason behind referring these as knowledge representation techniques is as follows. A classifier is primarily comprised of a mathematical form and a parameter estimation (training) technique. The mathematical form consists of a set of parameters which represents a family of functions. The parameter estimation technique is an algorithm for finding a suitable set of parameters corresponding to the data, that optimizes a particular objective. When a classifier is trained using data, it builds a knowledge structure based on its corresponding mathematical representation. The developed knowledge represented by a classifier is the set of parameters, which have been obtained after parameter estimation. The performance of the knowledge representation technique on a particular problem depends on how precise the set of parameters is estimated from the data. To the best of our knowledge, the selection of a classifier requires a profound understanding of the data, which may not be always available for complex multivariate problems. Since any of the classifiers in Table \ref{} can represent the transition function $f_T$ (given in eqn. (\ref{})), therefore one needs to select a classifier based on certain criterion. In our case, the selection of the classifier as a knowledge representation technique is based on eight metrics, namely \textit{Cross validation, Training time, Prediction time, Figure of Merit (FoM), Producer's accuracy (PA), User's accuracy (UA)} and \textit{Overall accuracy (OA)}. Each of these metrics is a different window that provides a certain view of the framework --- \textit{Cross validation} measures overfitting; \textit{FoM, PA, UA} and \textit{OA} measure framework's performance in learning patterns and simulation; \textit{Training} and \textit{Prediction time} measure swiftness of the knowledge representation technique. Based on all the views, we have selected our suitable knowledge representation technique as a Decision tree or Ensemble of Decision trees (Random Forests) for generation of cellular automata rules. Hence we call the framework a \textit{Rule based End-to-End} framework. \begin{table} \centering \resizebox{\textwidth}{!} { \begin{tabular}{|c|c|c|c|} \hline {\bf Classifier} & {\bf Knowledge representation} & \specialcell{\bf Mathematical representation} & {\bf Training procedure}\\ \hline Logistic regression & \specialcell{Stores Knowledge\\ in terms of\\ weights ($W$ and $b$)\\ of a hyperplane} & $Y = \sigma(W*X+b)$ & Stochastic gradient descent \\ \hline Gaussian Naive Bayes & \specialcell{Stores knowledge\\ in terms of\\ Probability distribution} & $Y = \operatorname*{arg\,max}_{k\in \{1,2,..K\}} p(C_k) \prod_{i=1}^{N} P(x_i|C_k)$ & Maximum likelihood estimation\\ \hline Support vector machines & \specialcell{Stores knowledge in\\ terms of weights\\ ($W$ and $b$) and kernel $\phi$} & $W*\phi(X)-b=0$ & Sequential minimal optimization\\ \hline Multi Layer Perceptrons & \specialcell{Stores knowledge\\ in terms of\\ computation nodes} & $\sigma(W_1*\sigma(..\sigma(W_n * X)))$ & Backpropagation\\ \hline K-Nearest Neighbors & \specialcell{Stores feature vectors \\and class labels \\as knowledge} & \specialcell{feature vectors and\\ class labels} & \specialcell{instance based learning\\ or non-generalizing learning}\\ \hline Decision Tree & \specialcell{Stores knowledge\\ in terms\\ of rules} & $(X_i = a_i) \wedge ... \wedge (\theta_j = a_j)$ & CART, ID3, C4.5\\ \hline \specialcell{Ensemble methods\\ (Random forest, AdaBoost)} & \specialcell{Stores knowledge\\ in terms of the\\ unit classifier \\in the ensemble} & Depends on the unit classifier & Bagging/Boosting\\ \hline \end{tabular} } \caption{Features of popular Classifiers which are also knowledge representation technique.} \label{} \end{table} \subsection{Methodology} As discussed in Section \ref{}, the Urban Growth Cellular Automata model consists of three components namely, cell state $S_t^p$, transition function $f_T$ and update rule (eq. (\ref{})). The cell state and update rule needs to be predefined and is not flexible, while the transition function is capable of taking any form during the training. Hence the framework is generic and does not depend on manually designed features. The procedures for dataset preparation, training and prediction are given in the Algorithms \ref{}, \ref{} and \ref{} respectively. Algorithm \ref{} consists of the following steps. \begin{itemize} \item Gathering raster and built-up values of a point and its neighborhood in a matrix $X_R$ with row vector ($x_r \gets \{R(p), N(R(p)\}$) and in a matrix $X_B$ with row vector ($x_b \gets \{B_p^t, N(B_p^t)\}$) respectively. The matrix $X_R$ is used to build an autoencoder using a custom function $build\_autoencoder()$. This is followed by generation of representations ($\phi_{raster}^{len}(x_r)$) and collection of all points in $X_B$ and representations $\phi_{raster}^{len}(x_r)$ to form a data matrix. It should be noted that the neighborhood criterion needs to be decided before the Data generation process. In this paper, we have used a standard Moore neighborhood criterion. \item Preparation of label matrix which is composed of transition and persistent classes. To prepare the matrix, we have declared \textit{isUrban($B_p^t$)} function, which returns \textit{true} if the point $p$ at time $t$ is urban and \textit{false} otherwise. It is required to determine whether at point $p$ and time interval $\{t, t+1\}$ transition has occurred or not. \end{itemize} Algorithm \ref{} consists of the following steps. \begin{itemize} \item Retrieve the \textit{Data matrix} and \textit{Label matrix} from Algorithm \ref{}. \item Select a set of classifiers $S_C$ as in Table \ref{} and train these classifiers using their corresponding training algorithm. \item The resultant is a set of trained classifiers, each of which can be considered as a transition function $f_T$. \item Then, depending on the Figure of Merit, Producer's accuracy, User's accuracy, Overall accuracy, cross-validation, training time and prediction time, one can choose which one to use. In certain cases, it may be convenient to use cross-validation, training time and prediction time to remove certain classifiers from the list initially. \end{itemize} Algorithm \ref{} consists of the following steps. \begin{itemize} \item Gathering raster and built-up values of a point and its neighborhood in row vectors $x_r$ and $x_b$ respectively. \item Using $\phi_{raster}^{len}$ from Algorithm \ref{} to generate representations. This is followed by concatenation of the vectors $x_b$ and representations $\phi_{raster}^{len}(x_r)$ to form $x$. \item Use a trained classifier $f_T$ to predict on $x$. \item If $f_T$ predicts a transition from \textit{Non Built-up} to \textit{Built-up} or a persistence from \textit{Built-up} to \textit{Built-up}, then the final class at $t+1$ is \textit{Built-up} and \textit{Non Built-up} otherwise. \end{itemize} It may be noted that at least one built-up image is necessary as an initial point to start prediction procedure. This is due to the fact that we have modeled urban growth as a cellular automaton for which the framework is recurrent in nature (see Fig. \ref{}). Figure \ref{} represents the flowchart of the simulation procedure. It shows that the framework can simulate urban growth up to as many years as possible in the future each time utilizing the last predicted image. It has been proved in (\cite{omohundro1984modelling}) that \textit{an arbitrary $9$ neighbor, two-dimensional cellular automata can be simulated in terms of a set of ten partial differential equations}. The PDEs indicate the relation between the dynamic variable and space-time, which can be a theoretical form of the urban growth cellular automata. Without the cellular automata, the procedure would be merely a built-up/landuse classification rather than prediction. \begin{algorithm} \begin{algorithmic} \STATE {\bf Input:} \STATE $B^t \gets \textit{Built-up raster at time t}$ \STATE $B^{t+1} \gets \textit{Built-up raster at time t+1}$ \STATE $R \gets \textit{Raster image}$ \STATE $isUrban()\gets$ function which returns True if point is Urban. \STATE $build\_autoencoder() \gets$ function which builds a function $\phi_{raster}^{len}$ after training on a dataset. \STATE {\bf Output:} \STATE $X, Y, \phi_{raster}^{len} \gets$ \textit{Data matrix}, \textit{Label matrix} and \textit{Encoder function}. \STATE {\bf Procedure:} \STATE $X, Y, X_R, X_B\gets \phi$ \FORALL {pixel $p$} \STATE $x_b \gets \{B_p^t, N(B_p^t)\}$ \STATE $x_r \gets \{R(p), N(R(p))\}$ \STATE $X_B \gets X_B\cup x_b$ \STATE $X_R \gets X_R\cup x_r$ \IF{!isUrban($B_p^{t+1}$) and !isUrban($B_p^{t}$)} \STATE $Y \gets Y\cup C_{NB}^{NB}$ \ELSIF{isUrban($B_p^{t+1}$) and isUrban($B_p^{t}$)} \STATE $Y \gets Y\cup C_{B}^{B}$ \ELSIF{isUrban($B_p^{t+1}$) and !isUrban($B_p^{t}$)} \STATE $Y \gets Y\cup C_{B}^{NB}$ \ELSE \STATE $Y \gets Y\cup C_{NB}^{B}$ \ENDIF \ENDFOR \STATE $\phi_{raster}^{len} \gets build\_autoencoder(X_R, len)$ \STATE $X \gets X \cup \{x_b,\phi_{raster}^{len}(x_r)\}; \forall x_r,x_b\in X_R,X_B$ \STATE return $X,Y,\phi_{raster}^{len}$ \end{algorithmic} \caption{Data Representation procedure} \label{} \end{algorithm} \begin{algorithm} \begin{algorithmic} \STATE {\bf Input:} \STATE $X \gets$ \textit{Data Matrix} \STATE $Y \gets$ \textit{Label Matrix} \STATE $S_C \gets$ Set of classifiers as in Table \ref{} \STATE $C.train() \gets$ Trains $C$ with corresponding training algorithm \STATE {\bf Output:} \STATE $S_C^T \gets$ Set of trained classifiers on $X$ and $Y$ \STATE {\bf Procedure:} \STATE $S_C^T \gets \phi$ \FORALL {$C\in S_C$} \STATE $C.train(X,Y)$ \STATE $S_C^T = S_C^T\cup C$ \ENDFOR \STATE return $S_C^T$ \end{algorithmic} \caption{Parameter estimation procedure} \label{} \end{algorithm} \begin{algorithm} \begin{algorithmic} \STATE {\bf Input:} \STATE $B^t \gets \textit{Built-up raster at time t}$ \STATE $R \gets \textit{Raster image}$ \STATE $f_T \gets C \in S_C^T$ (from Procedure \ref{}) \STATE $\phi_{raster}^{len} \gets $ Encoder function from the output of Algorithm \ref{}. \STATE {\bf Output:} \STATE $B_{t+1} \gets$ Future Built-up map. \STATE {\bf Procedure:} \FORALL {pixel $p$} \STATE $x_b \gets \{B_p^t, N(B_p^t)\}$ \STATE $x_r \gets \{R(p), N(R(p)\}$ \STATE $x \gets \{x_b, \phi_{raster}^{len} (x_r)\}$ \IF{$f_T(x) \in \{C_{NB}^B, C_B^B\}$} \STATE $B_p^{t+1} \gets$ Built-up \ELSE \STATE $B_p^{t+1} \gets$ Non Built-up \ENDIF \ENDFOR \STATE return $B^{t+1}$ \end{algorithmic} \caption{Prediction procedure} \label{} \end{algorithm} \begin{figure}[t] \centering \includegraphics[scale=0.25]{\locimage{recurrent_framework}} \caption{Recurrence of the Rule based Framework due to the cellular automata} \label{} \end{figure} \subsection{Features of the End-to-End Framework} The proposed \textit{End-to-End} architecture has certain key features which differentiates it from earlier proposed urban growth prediction frameworks (\cite{shafizadeh2015performance, lin2011predictive, liu2008modelling, van2009modeling}). \begin{itemize} \item The key improvement is the removal of a manual feature extraction module, which is inherent in the existing architectures and replacing it with a data representation module. The data representation module generates representations from the raw satellite images in an unsupervised manner in a form as described in Section \ref{} and does not require any separate database consisting of manually selected explanatory features. \item The knowledge representation layer stores knowledge that directly relates built up and raster representations with the transition classes. This is a significant distinction from the earlier models where learning models were used to establish relationships between manually selected explanatory features and \textit{Built-up}. The removal of manual selection processes reduces the bias on models and knowledge structures, thereby creating an opportunity to develop new theories and explain the results of a complex process as urban growth. \item Furtheremore, since representation learning is unsupervised, therefore it can be done without taking into account the final objective (in this case urban growth). Thus we can say that the representations are generic in nature and can be used for other applications as well. \end{itemize} Finally, these qualities come with a drawback, as implementation becomes easier but knowledge representation turns incomprehensible. Hence theoretically, it becomes a challenging task to extract meaning from these rules. This is due to the fact that the variables in the \textit{Data matrix} are frequency/band values which can have multiple semantics. Therefore, the rules generated can have multiple interpretations among which the most appropriate interpretation needs to be found out. \section{Experiments and Results} We have conducted our experiments on the region of Mumbai, which is the capital city of Maharashtra, India (latitude: $19.0760\degree$ N, longitude: $72.8777\degree$ E). The city lies on the west coast of India and has a natural harbor. Mumbai is one of the mega cities of India also often colloquially known as the City of Dreams. According to Census India, the population of Mumbai has been steadily rising from approximately $9$ million in $1991$ to more than $12$ million in $2011$. The region under consideration is shown in Fig. \ref{. The experiments are conducted in a Virtual Machine (VM) of an OpenStack based cloud infrastructure\footnote{} with the \textit{8 VCPUs}, \textit{16 GB RAM} and operating system \textit{Ubuntu 14.04}. \subsection{Data Collection and Preprocessing} We have collected remotely sensed natural color ($3$ bands) Landsat images from the United States Geological Survey (USGS) website\footnote{} for the years $1991$, $2001$ and $2011$ for the Mumbai region. The Mumbai region has been extracted from the raster and the final image consisted of $972280$ data pixels. The images have been segmented to generate the \textit{built-up} raster images that are binary images with white pixels representing \textit{built-up} and black pixels representing \textit{non built-up}. The segmentation has been carried out using maximum likelihood classification implemented in semi-automatic classification plugin\footnote{} of QGIS\footnote{}. The semi-automatic classification method is essentially a semi-manual labeling method where initial labels are to be provided by a human and then segmented maps are generated using the raster values of the Landsat image. Since it is manual and inaccurate it needs to be verified to some source or reference maps which in our case have been Google Earth and Mumbai maps from a previously published work on Mumbai \cite{shafizadeh2015performance}. The total number of pixels that transformed from \textit{non built-up} to \textit{built-up} during the $20$ years considered in our study are given in Table \ref{}. The percentage of pixels that changed from non-urban to urban is approximately $10\%$, which is similar to other studies conducted on the region of Mumbai (\cite{shafizadeh2015performance, moghadam2013spatiotemporal}). The slight aberrations are due to the classification inaccuracy of the classifiers used for performing the landcover classification. \begin{table}[h] \centering \begin{tabular}{|c|c|c|} \hline Time step & Pixels transformed & Pixels persistent\\ \hline $1991-2001$ & $149746$ & $822534$\\ \hline $2001-2011$ & $103360$ & $868920$\\ \hline \end{tabular} \caption{Number of pixels transformed vs number of pixels persistent} \label{} \end{table} \begin{figure*} \centering \subfloat[][]{\includegraphics[scale = 0.07]{\locimage{mumbai1991}}}\hspace{0.1cm} \subfloat[][]{\includegraphics[scale = 0.07]{\locimage{cimg1991}}}\hspace{0.1cm} \subfloat[][]{\includegraphics[scale = 0.07]{\locimage{cimg2001}}}\hspace{0.1cm} \subfloat[][]{\includegraphics[scale = 0.07]{\locimage{cimg2011}}} \caption{Rasters showing (a) Mumbai year 1991 and \textit{built-up} conditions of the year (b) 1991 (c) 2001 (d) 2011. (White represents built-up while black represents non built-up)} \label{} \end{figure*} The rasters showing the urban growth conditions of the years $1991$, $2001$ and $2011$ are shown in Fig \ref{}. In our case, we have considered $3$ classes i.e. $C_{NB}^{NB}, C_{NB}^{B}$ and $C_B^B$. The $4^{th}$ class $C_{B}^{NB}$ have had $0.02$ per cent contribution in the training set, probably as there have been no recent massive de-urbanization in the area. Hence we have considered all such instances to also fall in the class $C_{B}^{B}$. The \textit{data matrix} and the \textit{label matrix} are generated from the data considering a Moore neighborhood of radius $1$. The length of the encoded representation is varied in multiples of $5$. The autoencoder is trained using the Adaptive Gradient Descent Algorithm \cite{duchi2011adaptive} with batch size of $1000$. Subsequent to the training, the encoder present in the autoencoder is used to generate representations for creating the \textit{data matrix}. \subsection{Training and Validation} The \textit{data matrix} and \textit{label matrix} generated in the previous step have been used to develop knowledge structures using various kinds of knowledge representation techniques as given in Table \ref{}. The classifiers have been trained in a multi-parameter setting in the following ways. \begin{itemize} \item \texttt{Logistic Regression}: For this classifier, we have executed the training for $10$ times each with a values of L2 regularization in the range $[0.01,100]$. \item \texttt{Gaussian Naive Bayes}: No special prior probabilities have been set for this classifier. \item \texttt{Support Vector Machine}: Different kinds of kernel functions like linear, polynomial and radial basis function have been used to test performance metrics. For polynomial kernels, the degree of polynomial have been varied from $1$ to $10$. \item \texttt{Multi Layer Perceptron}: Parameters like number of hidden layers, hidden layer sizes, batch size, number of iterations, learning rate, momentum have been varied in different ranges to predict performance. The hidden layer configurations which have been tested are $(10,)$, $(20,15)$, $(20,15,10)$, $(20,15,10,5)$, $(20,15,10,5,3)$. The number of iterations and batch size have been taken in multiples of $100$. Learning rate have been taken with constant, adaptive and inverse scaled and in the range $[0.001, 0.1]$. Momentum have been varied in the range $[0.5,0.9]$. \item \texttt{Single and Ensemble of decision trees}: For this classifier, we have used Gini impurity as the measure to separate the datasets. The maximum height of the tree is set to values in the range $\{10,100,200,\infty\}$. The algorithm used to build the decision tree is an optimized version of CART (Classification and Regression Trees), which can handle both categorical and non-categorical data\footnote{}. For ensemble method of knowledge representation, we have considered $\{10,100,1000\}$ decision trees/estimators. The implementations of the learning methods that we have used are available at scikit learn\footnote{} library in python. \end{itemize} The average of all the results from all the different kinds of parameter settings are taken as the final performance of a knowledge structure. According to \cite{pontius2008comparing}, only comparing classification accuracy is not enough for validating a landuse change model. They argued that in order to validate a change model, there needs to be a validation of $4$ metrics, namely Figure of Merit (FoM), Producer's accuracy (PA), User's Accuracy (UA) and Overall Accuracy (OA). Since urban growth is a land use change model, therefore we have used these $4$ metrics for model validation. The former $3$ metrics assist in comparing the original and predicted maps of urban growth, while Overall Accuracy (OA) can be thought of as classification accuracy. According to (\cite{pontius2008comparing}), the $4$ validation measures depend on $5$ variables which are \begin{itemize} \item $A$ = Area of error due to observed change predicted as persistence. \item $B$ = Area correct due to observed change predicted as change. \item $C$ = Area of error due to observed change predicted in the wrong gaining category. \item $D$ = Area of error due to observed persistence predicted as change. \item $E$ = Area correct due to observed persistence predicted as persistence. \end{itemize} Figure of Merit ($FoM$) provides us the amount of overlap between the observed and predicted change. Producer's accuracy ($PA$) gives the proportion of pixels that model accurately predicted as change, given that the reference maps indicate observed change. User's accuracy ($UA$) gives the proportion of pixels that the model predicts accurately as change, given that the model predicts change. The equations of the metrics are given as follows. \begin{eqnarray} FoM = \frac{B}{A+B+C+D} \label{} \end{eqnarray} \begin{eqnarray} PA = \frac{B}{A+B+C} \label{eqn:PA} \end{eqnarray} \begin{eqnarray} UA = \frac{B}{B+C+D} \label{eqn:UA} \end{eqnarray} \begin{eqnarray} OA = \frac{B+E}{A+B+C+D+E} \label{eqn:OA} \end{eqnarray} \begin{figure*} \centering \subfloat[][1991-2001]{\includegraphics[scale = 0.34]{\locimage{mumbai-pred-1991-2001-DT}}} \subfloat[][2001-2011]{\includegraphics[scale = 0.34]{\locimage{mumbai-pred-2001-2011-DT}}} \caption{Predicted maps of Mumbai using Decision Tree (a) $1991-2001$ (b) $2001-2011$} \label{} \end{figure*} \begin{figure*} \centering \subfloat[][1991-2001]{\includegraphics[scale = 0.34]{\locimage{mumbai-pred-1991-2001-RF}}} \subfloat[][2001-2011]{\includegraphics[scale = 0.34]{\locimage{mumbai-pred-2001-2011-RF}}} \caption{Predicted maps of Mumbai using Random Forest (a) $1991-2001$ (b) $2001-2011$} \label{} \end{figure*} The training and validation of \textit{End-to-End} framework have been done using \textit{cross validation}. \textit{Cross validation} by dividing the dataset (generated from the dataset built from the years $1991-2001$) into $10$ parts and randomly selecting a part as validation set while others are used for training. The reason for selecting this is because the number of pixels which changed during this period is more than the other periods (Table \ref{}). The validation results in terms of classification accuracy mean and variance is shown in Table \ref{}. Performance comparison of existing methods with \textit{End-to-End} approach is presented in Fig \ref{} (a) and (b). The first $4$ set of bars presents the results of existing methods applied to our dataset, whereas the last two set of bars provides the results of \textit{End-to-End} approach with single and ensemble of decision trees. The resultant built up maps representing transition classes predicted by both our approaches are displayed in Fig. \ref{} and \ref{}. Comparision of the training and prediction time taken by the knowledge structures is provided in Fig. \ref{} (a) and (b). We have compared our framework with four existing methods (\cite{shafizadeh2015performance}, \cite{lin2011predictive}, \cite{feng2016modeling} and \cite{kamusoko2015simulating}) in terms of $FoM, PA, UA$ and $OA$. Some of the distance based features which have been used in these works are distance to roads, built-up, river, urban planning area, central business district, railway, wetlands, forests, agricultural lands etc. Some of the non-distance based factors are digital elevation maps, slope, population density, land/crop density etc. During experimentation, we have manually generated each of these feature maps to model and compare the results of these works with our \textit{End-to-End} framework. \subsection{Discussion}\label{} We defend our hypothesis regarding \textit{End-to-End} learning for prediction of urban growth with the results of our experiments. The comparison of our framework with existing frameworks based on the four parameters ($FoM, PA, UA, OA$) as in Fig. \ref{} (a) and (b) reveals that \textit{End-to-End} learning performs significantly better than the existing learning based methods developed for urban growth prediction. We argue based on the results that this is possible due to the superior representation and robustness of encoded representations combined with an ensemble of decision trees. An approximate summary of enhancements provided by our \textit{End-to-End} framework on the dataset is given as follows. \begin{itemize} \item $18\%$ improvement on Figure of Merit $FoM$. \item $17\%$ improvement on Producer's accuracy $PA$. \item $2\%$ improvement on User's Accuracy $UA$. \item $3\%$ improvement of Overall Accuracy $OA$. \end{itemize} The cross validation accuracies and the training, as well as prediction time is given in Table \ref{}. It is evident from the table that both Decision trees and Random Forests (ensemble of decision trees) provide optimal results in terms of cross validation accuracy, training time, prediction time, $FoM, PA, UA$ and $OA$ as compared to other prediction models. However, Decision tree has a problem of overfitting which can be fixed by using a Random Forest. Hence we conclude that Random Forest (ensemble of decision trees) is an optimal choice for knowledge representation for our proposed \textit{End-to-End} framework. \begin{figure} \centering \subfloat[][Training time]{\includegraphics[scale=0.4]{\locimage{time_training}}} \subfloat[][Prediction time]{\includegraphics[scale=0.4]{\locimage{time_test}}} \caption{Comparison of (a) training and (b) prediction time with respect to the knowledge representation technique.} \label{} \end{figure} Figure \ref{} depicts the change of performance metrics with respect to the size of the encoding. It shows a sharp rise in the begining followed by saturation in the performance metrics as the encoding size is increased. This implies that with increase in encoding size, information is more precisely encoded by the autoencoder and hence performance of the simulation improves. However, the tradeoff is if encoding size is increased arbitrarily then time required to train the autoencoder increases more than performance metrics. Furthermore, increasing the encoding size increases size of the feature vector in the \textit{data matrix}, which brings in the \textit{curse of dimensionality} issue\footnote{When dimensionality of a feature vector is increased without increase in size of the data, then data tends to become sparse. This problem is referred to as \textit{curse of dimensionality}}. In this case, we believe that saturation is caused by increase in dimensions of the feature vector in the \textit{data matrix}. \begin{figure} \centering \subfloat[][1991-2001]{\includegraphics[scale=0.5]{\locimage{ugmetrics_1991_2001}}} \subfloat[][2001-2011]{\includegraphics[scale=0.5]{\locimage{ugmetrics_2001_2011}}} \caption{Performance metrics with respect to the encoding size.} \label{} \end{figure} From the comparison in Fig. \ref{} (a) and (b), we have observed that the Overall Accuracy (OA) and the User's Accuracy (UA) metric is much higher than the Figure of merit (FoM) and Producer's accuracy (PA) for certain existing methods. This is indeed the case that has also been reflected in results shown by experiments in some of the other works, for instance, \cite{shafizadeh2015performance} have claimed $0.435, 0.6062$ and $0.8513$ as FoM, PA and OA using MLP for the region of Mumbai. One of the reasons behind this peculiarity is due to the imbalance in the datasets problem that we discussed earlier in the paper (section III A). Since Figure of Merit and Producer's accuracy provides the performance of the model in terms of transitions and the fraction of transition pixels is low, therefore a training algorithm might not have learned them correctly. Furthermore, we can also see that existing models have comparable Overall Accuracy (OA) of about $90\%$. This can be due to the fact that only $10\%$ of the pixels fall in the transition class and the simple default strategy to give maximum accuracy is to give $90\%$ \cite{chawla2002smote}. In the \textit{End-to-End} approach, we have seen that the $3$ metrics $FoM,PA$ and $UA$ are comparable, which indicates that the imbalance in datasets have been handled. The user's accuracy metric is high for certain models because of the fact that there is only one direction of change i.e. \textit{non built-up} to \textit{built-up}. Therefore, predicted pixels in the transition category can be only one category and the model that predicted a higher number of pixels in the transition category have higher user's accuracy. \begin{table*} \centering \resizebox{\textwidth}{!}{\begin{tabular}{|c|c|c|c|} \hline {\bf Classifier} & {\bf Cross-Validation (Mean +/- variance)} & {\bf Training Time (in seconds)} & {\bf Prediction Time (in seconds)}\\ \hline MLP \cite{shafizadeh2015performance} & $0.899112 (+/- 0.041553)$ & $244.40$ & $0.46$\\ \hline Logistic Regression \cite{lin2011predictive} & $0.891636 (+/- 0.039668)$ & $3.79$ & $0.03$\\ \hline SVM \cite{feng2016modeling} & $ 0.628495 (+/- 0.111121)$ & $58284.61$ & $2184.21$\\ \hline Random Forests \cite{kamusoko2015simulating} & $ 0.851369 (+/- 0.056350)$ & $3.13$ & $0.09$\\ \hline \textit{End-to-End} approach (MLP single layer) & $0.911805 (+/- 0.029091)$ & $233.98$ & $3.7$\\ \hline \textit{End-to-End} approach (SGD) & $0.902009 (+/- 0.025873)$ & $4.84$ & $0.14$\\ \hline \textit{End-to-End} approach (Naive Bayes) & $0.853874 (+/- 0.120939)$ & $0.87$ & $1.58$\\ \hline \textit{End-to-End} approach (KNN) & $0.826471 (+/- 0.051649)$ & $5603.82$ & $22524.38$\\ \hline \textit{End-to-End} approach (AdaBoost) & $0.850093 (+/- 0.016626)$ & $138.40$ & $12.59$\\ \hline \textit{End-to-End} approach (Decision Tree) & $0.849341 (+/- 0.049533)$ & $31.44$ & $0.36$\\ \hline {\bf \textit{End-to-End} approach (Random Forests)} & {\bf 0.900792 (+/- 0.043120)} & {\bf 53.06} & {\bf 2.9}\\ \hline \end{tabular}} \caption{Performance of other classifiers in comparison to Decision tree and Random Forests} \label{} \end{table*} \begin{table*} \centering \resizebox{\textwidth}{!}{\begin{tabular}{|c|c|} \hline {\bf Model} & {\bf Remarks}\\ \hline MLP (\cite{shafizadeh2015performance}) & \specialcell{High Training time. Poor FoM, PA and UA than \\Proposed Method. Uses feature engineering.}\\ \hline Logistic Regression (\cite{lin2011predictive}) & \specialcell{Poor FoM, PA and UA than Proposed Method.\\ Uses feature engineering.}\\ \hline SVM (\cite{feng2016modeling}) & \specialcell{High training times. Poor FoM, PA and UA than \\Proposed Method. Uses feature engineering.}\\ \hline Random Forests (\cite{kamusoko2015simulating}) & \specialcell{Poor FoM, PA and UA than Proposed Method. \\Uses feature engineering.}\\ \hline \textit{End-to-End} approach (MLP single layer) & High Training time. Poor FoM, PA and UA than Proposed Method.\\ \hline \textit{End-to-End} approach (SGD) & Poor FoM, PA and UA than Proposed Method.\\ \hline \textit{End-to-End} approach (Naive Bayes) & Poor FoM, PA and UA than Proposed Method.\\ \hline \textit{End-to-End} approach (Decision Tree) & Optimal but has Problem of overfitting.\\ \hline {\bf \textit{End-to-End} approach (Random Forests)} & {\bf Optimal}\\ \hline \end{tabular}} \caption{Limitations of the other models to the proposed model.} \label{} \end{table*} \begin{figure*} \centering \subfloat[][1991-2001]{\includegraphics[scale=0.5]{\locimage{performance_1991_2001}}} \subfloat[][2001-2011]{\includegraphics[scale=0.5]{\locimage{performance_2001_2011}}} \caption{Performance comparison of our approach with existing methods for years (a) $1991-2001$ (b) $2001-2011$} \label{} \end{figure*} \subsection{Future Urban Growth Prediction} Figures \ref{ and \ref{ show future urban growth prediction for year $2051$ starting from $1991$ using our \textit{End-to-End} framework and knowledge structure as Decision Tree as well as Random Forest respectively. The raster over which the built-up is displayed is the year $1991$ natural color image of Mumbai. The white pixels represent the \textit{built-up} regions. It can be seen that the framework in case of Decision tree does not encroach upon water bodies and swamps, whereas in case of Random Forest, few encroachments are present. It may be noted that despite providing no explicit region information like water bodies, swamps, forests etc to the framework during training, the framework has been able to capture them. Hence, we can say that the rule based framework automatically divides the regions in the satellite image in some way to determine in which places growth can happen. This knowledge is currently encoded in the decision trees which can be extracted only if we discover the rules. \begin{figure*} \centering \subfloat[][]{\includegraphics[scale=0.10]{\locimage{mumbai_DT_2051}}}\hspace{0.1cm} \subfloat[][]{\includegraphics[scale=0.10]{\locimage{mumbai_RF_2051}}} \caption{Simulated urban growth of 2051 on mumbai region with knowledge structure as (a) Decision Tree and (b) Random Forests} \label{} \end{figure*} \section{Introduction} Urban regions are fast growing entities in the modern world. Areas which have been outskirts of a city/town twenty years back are now habituated and developed into urban form. Research has shown that uncontrolled sprawl of urban areas has potentially adverse effects on the environment in terms of biodiversity, loss of habitat and change of landscape (\cite{basawaraja2011analysis, moghadam2013spatiotemporal}). United nations have predicted based on past data that urban population will rise from a current proportion of $54$ per cent to a proportion of $66$ percent by the year $2050$\footnote{}. With the uncontrolled expansion of cities, it has become important to monitor and control changes in the urban landuse. A fundamental difficulty in monitoring urban areas is due to its huge spread, which makes manual tracking of global landuse/built-up changes intractable. However, there are several remote sensing satellites periodically capturing images of the earth surface, which contains plenty of real time information regarding change of urban landscape. Hence, satellite imagery has been considered as an effective component for real time monitoring of urban growth. Urban growth monitoring using satellite derived products has been addressed by several researchers (\cite{moghadam2013spatiotemporal, shafizadeh2015performance, lin2011predictive}). Models using cellular automaton, agent based methods, neural networks, logistic regression, fractals have been developed for robust and accurate prediction of urban growth (\cite{liu2008modelling, bhatta2012urban, musa2016review, aburas2016simulation}). Cellular automata is a widely preferred model for urban growth prediction due to its simplicity. Several CA models and corresponding case studies have been done on various major cities for detecting urban change, for instance, Vancouver Britain, Wuhan China etc (\cite{aburas2016simulation,rienow2015supporting, ke2016partitioned, van2009modeling, feng2016modeling}). Recently, artificial neural networks have also been quite popular in the land use change prediction (\cite{shafizadeh2015performance, razavi2014predicting}) and studies using the combination of neural networks and cellular automata models have been proposed (\cite{omrani2017integrating}). On the other hand, support vector machines (SVM), Decision trees and Random Forest based models have also been proposed (\cite{rienow2015supporting, ke2016partitioned, ahmadlou2016modeling, kamusoko2015simulating, li2004data}). Besides, logistic regression has been quite suitable for urban growth prediction in order to assess the driving forces behind urban growth (\cite{hu2007modeling}). Overall, among the learning techniques used for prediction, there have been broadly four categories which have dominated the urban growth community, namely logistic regression, artificial neural networks, support vector machines and inductive learning approaches (\cite{musa2016review}). Each of the models developed using learning methodologies mentioned above has utilized spatial and non-spatial explanatory features for predicting urban growth (\cite{musa2016review}). Spatial features such as distance from urban areas, water bodies, swamps etc. are derived from satellite imagery manually using supervised land use classification methods (\cite{shafizadeh2015performance, thapa2011urban}). Other kinds of spatial factors such as elevation, distance from roads and railways are derived from other open spatial data repositories like Open Street Maps\footnote{} and US Shuttle Radar Topography Mission\footnote{}. Three issues of this method are discussed as follows. \begin{itemize} \item It has been evident from practical experience (\cite{bengio2013representation}) that feature engineering significantly affects results of machine learning algorithms. At present, most of the urban growth prediction models use distance based features, for instance, distance from roads, city centers, railways, water bodies etc. and the distance metric is generally euclidean distance. Although the distance from certain regions does determine built-up, that may not be always euclidean distance. Moreover, there are other distance metrics, for instance, Minkowski's distance, Mahalanobis distance, etc which have not been sufficiently used in urban growth prediction modeling. It may be possible to achieve better performance using other distance metrics. The issue, in this case, is a selection of suitable distance based criterion which requires manual effort. \item The spectral properties can also be a crucial feature as different spectral properties indicate different landuse and different landuse may have different growth patterns. Hence another limitation of distance based features is that they are unable to capture the spectral properties of landuse. \item Due to the disparity in the geography of our earth surface and noise in satellite imagery, it is often excessively labor-intensive to derive all the different kinds of landuse feature maps from satellite imagery. \end{itemize} We intend to address these issues by designing an intelligent framework which has minimal human bias and build models in an \textit{End-to-End} fashion. The idea of \textit{End-to-End} learning is to bypass intermediate feature engineering steps in model building and design a model that can directly learn robust representations from input data (\cite{bengio2013representation, graves2014towards}). \textit{End-to-End} learning is a paradigm of machine learning that has recently gained importance due to the advent of \textit{Representation} learning and high-performance computing infrastructures. It has delivered exceptional performance in domains like image classification, speech processing, image segmentation etc (\cite{krizhevsky2012imagenet, badrinarayanan2015segnet, graves2013speech}). Majority of these techniques demonstrate the existence of complex relationships between various real world objects/phenomenon and conclusions that we draw from it. These complex relations are stored in our memory in terms of intricate knowledge structures which are hard to decode but makes human beings quite efficient in motor skills. Inspired by the above works, we put forward our hypothesis: \textit{Automatically extracted information from remotely sensed data can be an important feature in modeling urban growth.} The primary objective of this paper is to utilize the concept of \textit{Representation learning} in developing a novel \textit{End-to-End} framework for prediction of urban growth. \textit{The proposed framework utilizes unsupervised representation learning and supervised classification methods to learn rules of an urban growth cellular automata model from raw satellite data to form various knowledge structures.} Our proposed framework gives superior prediction performance than existing learning based methodologies in terms of various metrics which evaluates urban growth simulation performance. The key contributions of this paper are as follows. \begin{itemize} \item Eliminating feature engineering module from prediction framework. \item Generating representations from spectral information present in remote sensing satellite imagery in improving performance of urban growth prediction. \end{itemize} The rest of the sections are organized as follows. In section $2$, we define the cellular automata model for urban growth prediction. Section $3$ describes the framework that has been used to discover and store rules. Section $4$ presents results and discussions of the experiments conducted in the region of Mumbai, India. Finally, section $5$ provides conclusion and future research directions. \section{Urban Growth modeled using Cellular Automata (CA)} \label{} Urban growth, being a spatiotemporal dynamic process, can be modeled using the theory of cellular automata. A typical CA model is composed of an infinite array of cells having a finite number of states, which transform at discrete time steps using certain transition rules. The transition rules of a CA model signify the relationship between a cell and its neighborhood. The neighborhood criterion can be Von Neumann ($4$ neighbors) or Moore Neighborhood ($8$ neighbors). Transition rules are iteratively applied to the cells to simulate dynamic processes over multiple time steps. We define a cell state at point $p$ and time $t$ as $S_p^t = <l_p^t, \tau_p^t>$, where $l_p^t$ is a binary label variable representing \textit{Built-up} and \textit{Non Built-up} and $\tau_p^t$ is a transition indicator. The range of label variable $l_p^t$ is $\{-1,1\}$, where $-1$ represents \textit{Non Built-up} and $1$ represents \textit{Built-up}. Let us represent the set of \textit{Built-up} and \textit{Non Built-up} pixels originally as $S_{B}$ and $S_{NB}$. The transition indicator indicates whether the cell under consideration has undergone a transformation in the current time step. The range of the transition indicator is $\{0,1,2,3\}$, where $0,1,2,3$ represents transition from \textit{Non Built-up} to \textit{Non Built-up} ($C_{NB}^{NB}$), \textit{Built-up} to \textit{Built-up} ($C_{B}^{B}$), \textit{Non Built-up} to \textit{Built-up} ($C_{NB}^{B}$) and \textit{Built-up} to \textit{Non Built-up} ($C_{B}^{NB}$). A pictorial representation of the transition classes are displayed in Fig. \ref{} (b). We define the update rule for the transition indicator as \begin{eqnarray} \tau_p^{t+1} = f_T(l_p^t, N(l_p^t)) \label{} \end{eqnarray} , where $N$ is the neighborhood criterion and $f_T$ is a update function to be modeled. The transition function for update of the state $S_p^{t}$ to $S_p^{t+1}$ given as is, \begin{eqnarray} S_p^{t+1} = \begin{cases} (S_B,\tau_p^{t+1}) & \tau_p^{t+1}\in \{C_{NB}^B,C_B^B\} \\ (S_{NB},\tau_p^{t+1}) & \tau_p^{t+1}\in \{C_{NB}^{NB},C_{B}^{NB}\} \end{cases} \label{} \end{eqnarray} The theory of cellular automata is convenient in this scenario because it considers the space as a discrete array of cells. Since in our case, the data is of raster type, it is already in form of discrete matrix of pixels. So discretization of space is not required to be done explicitly. If the data was in vector form, then we would have to define the discretization of the space separately. \subsection{Extension of the CA model} The urban growth CA model can be extended to include other variables which are driving factors in urban growth (\cite{yang2008cellular, feng2016modeling}). In order to insert the raster information in the CA model, we have modified our CA model in the following way. \begin{itemize} \item The state variable $S_p^t$ is extended to include a raster data variable $R_p$ from the images to form $<l_p^t, \tau_p^t, R_p>$. The raster image can consist of multiple bands in which case, $R_p$ represents a vector of values of all bands at point $p$. \item The update equation (\ref{}) gets transformed to include the raster variable $R_p$ in the following way. \begin{eqnarray} \tau_p^{t+1} = f_T(l_p^t, N(l_p^t), \phi_{raster}^{len}(R_p, N(R_p))) \label{} \end{eqnarray} \end{itemize} The above modification allows us to reflect upon the relationship between the \textit{built-up} variables and the raster variables through the function $f_T$ and $\phi_{raster}$. The function $\phi_{raster}^{len}$ is an encoding function which is responsible for generating a desired length $len$ encoding of the information present in $<R_p, N(R_p)>$. Thus, we have reduced the problem of modeling urban growth from raster data to modeling the function $f_T$ and $\phi_{raster}^{len}$. \input{endtoend.tex} \input{exp.tex} \section{Conclusion and Future Work} We have introduced the concept of \textit{End-to-End} learning in urban growth prediction by developing a framework for learning rules of a cellular automata model directly by representing spectral information from remote sensing data. We have empirically verified our framework on predicting urban growth for the region of Mumbai over a time frame of $20$ years. Our \textit{End-to-End} framework has outperformed existing learning based methodologies with a simpler implementation than the existing frameworks. Future work can be based on challenges which we have encountered in this work. Since spatial resolution of the satellite is fixed, therefore as temporal resolution reduces, the number of cells in which built-up happens reduces. Due to this, the number of points in the transition classes reduces and persistence classes increases. Therefore, the dataset gets heavily imbalanced and the ability of the classifier to learn patterns of transition classes reduces. The problem of imbalanced datasets can be alleviated if both the temporal and spatial resolution of the satellite images are increased. This is possible if data from high-resolution satellites like IKONOS/ QuickBird are used. However, the data load of high-resolution satellites is high, hence more infrastructure and sophisticated algorithms would be necessary. Despite superior performance of the \textit{End-to-End} approach, one of the drawbacks of \textit{End-to-End} learning is essentially difficulty in understanding the rules. Hence, uncovering automatically generated rules in an \textit{End-to-End} learning framework can be considered as a challenge which needs to be resolved in the subsequent studies. Furthermore, the \textit{End-to-End} framework can be extended to incorporate other data resources which are vector form, which may further improve the performance of the framework. \input{main.bbl} \end{document}

\section{Introduction} \label{intro} Approximation by convolution involves selecting a suitable integrable function $K(x)$ (\textit{the convolution kernel}) satisfying $\int_{\mathbb{R}}K(x)dx=1.$ A parameterized family of convolution kernels is generated from $K$ by setting $K_{h}(x)=h^{-1}K(x/h)$ where $h>0.$ Then, for a given target function $f,$ its convolution approximation $f \ast K_{h}$ converges to $f$, as $h \to 0.$ The rate of convergence depends upon the smoothness of $f$ and the polynomial reproduction properties of the underlying convolution kernel. In this paper we consider the approximation of 1-periodic continuous functions by convolution with the Gaussian kernel. In this case it is only possible to reproduce the constant function and so, as we will see, convergence is limited to $\mathcal{O}(h^{2}),$ regardless of additional smoothness requirements. However, if we employ a multilevel iterative refinement scheme we see that we get very rapid convergence. If the width of the Gaussian is halved at each iteration, then at the $n$th level of refinement we have essentially a $d=2^n$ scaling in the Gaussian. For a discrete scheme we would apply a quadrature at $d=2^n$ equally spaced points, giving $d$ degrees of freedom. If the target function is taken from a certain periodic Sobolev space, whose order prescribes the smoothness of the functions, then we see improved but saturated convergence rates. If we consider functions in the native space of the Gaussian, a space of infinitely differentiable functions, then the order of convergence is $d^{-\frac{\ln(d)}{\ln(2)}}$. This work is motivated by the desire to prove convergence of the multilevel sparse grid quasi-interpolation introduced by Levesley and Usta \cite{usta}. Multilevel sparse grid algorithms using smooth functions were introduced by the author and collaborators \cite{georgoulis}.\\ Our focus is on the Gaussian kernel \[ \psi(x)={1 \over\sqrt{2\pi}} \exp(-x^2/2) \] and we define $\psi_h(x) = h^{-1} \psi(x/h)$. We note the Fourier transform of $\psi$ is \begin{eqnarray*} \widehat{ \psi} (x) & = & \int_{-\infty}^\infty \psi(y) \exp(-2\pi i x y) dy = \exp(-2\pi^2 x^2). \end{eqnarray*} We wish to approximate a continuous 1-periodic function \[f= \sum_{k\in \mathbb{Z}}^\infty f_k e_k,\,\,\,\,\,\,{\rm{where}}\,\,\,\,\,\,e_{k}(x) = \exp(2\pi i k x),\,\,\,\, (k\in \mathbb{Z}), \] by \begin{eqnarray*} f \ast \psi_h(x) & = & \int_{-\infty}^\infty f(t) \psi_h(x-t) dt \; = \; \sum_{j=-\infty}^\infty \int_{j}^{j+1} f(t) \psi_h(x-t) dt \nonumber \\ & = & \sum_{j=-\infty}^\infty \int_{0}^{1} f(t) \psi_h(x-t-j) dt \; = \; \int_{0}^{1} f(t) \phi_h(x-t) dt, \label{conv1} \end{eqnarray*} where $$ \phi_h(x) = \sum_{j=-\infty}^\infty \psi_h(x-j). $$ We note that $\phi_h$ is 1-periodic and a straightforward computation shows that \begin{eqnarray} \phi_h(x) & = & \sum_{k=-\infty}^\infty \widehat{\psi}(hk) e_k(x),\label{phirep} \end{eqnarray} In view of this and the well-known convolution formula for periodic functions (see e.g. \cite{bracewell}) we have that \begin{eqnarray*} f \ast \psi_h = \sum_{k=-\infty}^\infty f_k \widehat{\psi}(hk) e_k, \end{eqnarray*} so that the error in convolution approximation is \begin{equation} E_hf = f-f \ast \psi_h = \sum_{k=-\infty}^\infty f_k (1- \widehat{\psi}(hk)) e_k.\label{errhcon} \end{equation} This representation immediately shows that the convolution reproduces the constant, but not any other trigonometric polynomial.\\ In Section~\ref{secconconv} we will examine the convergence of the convolution approximation to target functions taken from certain periodic Sobolev spaces. We will show that no non-constant periodic function can have an approximation order smaller than $\mathcal{O}(h^{2})$. In view of this we explore a multilevel iterative refinement, halving $h$ in the convolution approximation at each level. In Section~\ref{secitref} we examine the error for this scheme and show that we can improve upon $\mathcal{O}(h^{2})$ for functions with additional but finite smoothness. Specifically, we see that rapid improvements in the accuracy are exhibited in the early iterations but once the number of iterations passes a certain level, relating to the smoothness of the function, the algorithm settles to converge at a polynomial rate. In Section~\ref{convnat} we introduce the {\sl native space} for Gaussian approximation. The native space is a subspace of infinitely smooth functions and, for such functions, we show that the algorithm exhibits rapid improvements in accuracy at every iteration. \section{Convergence of the convolution approximation} \label{secconconv} The functions we wish to approximate are continuous 1-periodic and taken from a periodic Sobolev space $$ \mathcal{N}_\beta = \left \{ f=\sum_{k=-\infty}^\infty f_k e_k : \| f \|_{\beta} = \left ( |f_{0}|^{2}+\sum_{k\in \mathbb{Z}} k^{2\beta}|f_{k}| \right )^{1/2} < \infty \right \}. $$ The Sobolev embedding theorem \cite{adams} ensures that if $\beta>\frac{1}{2}$ then all functions in $\mathcal{N}_\beta$ will be continuous. The following result gives error bounds for Gaussian convolution approximation of such functions. \begin{proposition}\label{prop1} Let $f \in \mathcal{N}_\beta,$ where $\beta >\frac{1}{2}.$ Then \[ \| E_h f \|_{\infty}\le\begin{cases} C_{1}h^2 \,\, & \,\, \textrm{for} \,\,\,\, \beta>\frac{5}{2}; \\ h^2(C_{2}\sqrt{\ln\left(\frac{1}{h}\right)}+C_{3}) \,\, & \,\, \textrm{for} \,\,\,\, \beta=\frac{5}{2}; \\ C_{4}h^{\beta-\frac{1}{2}} \,\, & \,\, \textrm{for} \,\,\,\, \frac{1}{2}< \beta < \frac{5}{2}, \end{cases} \] where $C_{i}$ $i=1,2,3,4$, are positive constants independent of $h.$ \end{proposition} \begin{proof} Since $\widehat{\psi}(0)=1$ and $\widehat{\psi}(-k)=\widehat{\psi}(k)$, $(k\in \mathbb{Z}_{+}$), we have, from (\ref{errhcon}), that \begin{equation} \| E_h f \| \le \sum_{k=1}^\infty (1-\widehat{\psi} (hk))(|f_k|+|f_{-k}|).\label{base-error} \end{equation} Suppose that $\beta = \frac{5}{2}+\alpha,$ where $\alpha>0.$ Using the elementary bound \begin{equation} 1-\exp(-x) \le x\quad {\rm{for}}\quad x>0, \label{elem} \end{equation} we have that $1-\widehat{\psi}(hk) \le 2\pi^{2}h^{2}k^{2}$. This yields \[ \begin{aligned} \| E_h f \|_{\infty } &\le 2\pi^{2}h^{2}\left(\sum_{k=1}^\infty k^{2}|f_k|+\sum_{k=1}^\infty k^{2}|f_{-k}|\right)\\ &=2\pi^{2}h^{2}\left(\sum_{k=1}^\infty k^{\frac{5}{2}+\alpha}|f_k|\frac{1}{k^{\alpha+\frac{1}{2}}}+\sum_{k=1}^\infty k^{\frac{5}{2}+\alpha}|f_{-k}|\frac{1}{k^{\alpha+\frac{1}{2}}}\right). \end{aligned} \] Applying the Cauchy Schwarz inequality we have \[ \begin{aligned} & \| E_h f \|_{\infty}\le\\ & 2\pi^{2}h^{2}\left(\sum_{k=1}^\infty \frac{1}{k^{2\alpha+1}}\right)^{\frac{1}{2}}\left[\left(\sum_{k=1}^\infty \left(k^{2}\right)^{\frac{5}{2}+\alpha}|f_k|^{2}\right)^{\frac{1}{2}}+\left(\sum_{k=1}^\infty \left(k^{2}\right)^{\frac{5}{2}+\alpha}|f_{-k}|^{2}\right)^{\frac{1}{2}}\right]\\ &\le 2\pi^{2}h^{2}C\|f\|_{\frac{5}{2}+\alpha}. \end{aligned} \] Now assume that $\beta = \frac{5}{2}-\alpha$ where $0\le \alpha < 2.$ In the following development we will partition the right hand side of (\ref{base-error}) as follows \[ \| E_h f \| \le \sum_{k=1}^{m_{h}} (1-\widehat{\psi}(hk))(|f_k|+|f_{-k}|)+\sum_{k=m_{h}+1}^\infty (1-\widehat{\psi}(hk))(|f_k|+|f_{-k}|), \] where $m_{h}$ denotes the integer satisfying $m_{h}\le \frac{1}{h} \le m_{h}+1.$ For $k \le m_h$ we will, as before, use $1-\widehat{\psi}(hk)\le 2\pi^{2}h^{2}k^{2}$ whereas for $k \ge m_h +1$ we use $1-\widehat{\psi}(hk)\le 1.$ This leads to \[ \begin{aligned} \| E_h f \|& \le \underbrace{2\pi^{2}h^{2}\sum_{k=1}^{m_{h}}k^{2}|f_k|+\sum_{k=m_{h}+1}^\infty |f_k|}_{E_{+}}+\underbrace{2\pi^{2}h^{2}\sum_{k=1}^{m_{h}}k^{2}|f_{-k}|+\sum_{k=m_{h}+1}^\infty |f_{-k}|}_{E_{-}}. \end{aligned} \] Focusing on $E_{+}$ we can write: \[ \begin{aligned} E_{+}=2\pi^{2}h^{2}\sum_{k=1}^{m_{h}}k^{\frac{5}{2}-\alpha}|f_k|k^{\alpha-\frac{1}{2}}+\sum_{k=m_{h}+1}^\infty k^{\frac{5}{2}-\alpha}|f_k|\frac{1}{k^{\frac{5}{2}-\alpha}}.\\ \end{aligned} \] Applying the Cauchy-Schwarz inequality to the first sum, in the case where $0 <\alpha <2,$ we have \[ \begin{aligned} 2\pi^{2}h^{2}\sum_{k=1}^{m_{h}}k^{\frac{5}{2}-\alpha}|f_k|k^{\alpha-\frac{1}{2}}&\le 2\pi^{2}h^{2}\left(\sum_{k=1}^{m_{h}}\left(k^{2}\right)^{\frac{5}{2}-\alpha}|f_k|^{2}\right)^{\frac{1}{2}} \left(\sum_{k=1}^{m_{h}}k^{2\alpha -1}\right)^{\frac{1}{2}}\\ &\le 2\pi^{2}h^{2}C\|f\|_{\frac{5}{2}-\alpha}m_{h}^{\alpha}\le 2\pi^{2}Ch^{2-\alpha}\|f\|_{\frac{5}{2}-\alpha}. \end{aligned} \] For $\alpha = 0$ we have \[ \begin{aligned} 2\pi^{2}h^{2}\sum_{k=1}^{m_{h}}k^{\frac{5}{2}}|f_k|k^{-\frac{1}{2}}&\le 2\pi^{2}h^{2}\left(\sum_{k=1}^{m_{h}}\left(k^{2}\right)^{\frac{5}{2}}|f_k|^{2}\right)^{\frac{1}{2}} \left(\sum_{k=1}^{m_{h}}\frac{1}{k}\right)^{\frac{1}{2}}\\ &\le 2\pi^{2}h^{2}\|f\|_{\frac{5}{2}}(2\ln(m_{h}))^{\frac{1}{2}}\le \sqrt{2}\pi^{2}h^{2}\left(\ln\left(\frac{1}{h}\right)\right)^{\frac{1}{2}}\|f\|_{\frac{5}{2}}. \end{aligned} \] Applying the Cauchy-Schwarz inequality to the second sum we have \[ \begin{aligned} \sum_{k=m_{h}+1}^\infty k^{\frac{5}{2}-\alpha}|f_k|\frac{1}{k^{\frac{5}{2}-\alpha}}& \le \left(\sum_{k=m_{h}+1}^\infty\left(k^{2}\right)^{\frac{5}{2}-\alpha}|f_k|^{2}\right)^{\frac{1}{2}} \left(\sum_{k=m_{h}+1}^\infty\frac{1}{k^{5-2\alpha}}\right)^{\frac{1}{2}}\\ &\le \|f\|_{\frac{5}{2}-\alpha}\mathcal{C}\left(\frac{1}{(m_{h}+1)^{4-2\alpha}}\right)^{\frac{1}{2}}\le \mathcal{C}\|f\|_{\frac{5}{2}-\alpha}h^{2-\alpha}. \end{aligned} \] These two bounds allow us to deduce that $E_{+}\le \mathcal{C}h^{2-\alpha}$ for $0<\alpha<2$ and $E_{+}\le h^{2}(A\sqrt{\ln(1/h)}+B).$ The equivalent bound for $E_{-},$ the contribution involving Fourier coefficients of negative index, follows in the same fashion. Hence we conclude that \[ \|E_{h}f\| \le Ch^{2-\alpha}\|f\|_{\frac{5}{2}-\alpha}=Ch^{\beta-\frac{1}{2}}\|f\|_{\beta}\,\,\,\, {\rm{where}}\,\,\,\,\frac{1}{2}<\beta < \frac{5}{2}, \] and \[ \|E_{h}f\| \le \|f\|_{\frac{5}{2}}h^{2}((A\sqrt{\ln(1/h)}+B). \] \end{proof} \section{Iterative refinement} \label{secitref} The result from the previous section suggests that we cannot improve convergence by imposing additional smoothness conditions, beyond $f \in \mathcal{N}_{\frac{5}{2}}.$ In view of this we now consider using a multilevel iterative refinement where, at each level, we halve $h$ in the convolution approximation. In this setting the the multilevel error in convolution is defined by \begin{equation} M_1 f = f-f\ast \psi_h , \,\,\,\,\,\,\, {\rm and}\,\,\,\,\,\,\,M_j f := M_{j-1} f - M_{j-1} f \ast \psi_{\frac{h}{2^{j-1}}}, \quad (j \ge 2). \label{itdef} \end{equation} Our analysis here will follow in the same fashion as in the previous section. We begin by noting that if $f=\sum_{k\in \mathbb{Z}}f_{k}e_{k}$ is a continuous 1-periodic function then the Fourier series representation of the multilevel error is given by \begin{equation} \begin{aligned} M_j f & = & \sum_{k=-\infty}^\infty b_{j}(k) \widehat{f} (k) e_k, \quad j=1,2,\cdots, \label{mjrep} \end{aligned} \end{equation} where, for $k \in \mathbb{Z}$ and \begin{equation} b_{j}(k) = \prod_{\ell=1}^j \left(1-\widehat{\psi}\left( \frac{hk}{2^{\ell-1}}\right)\right)\quad j=1,2,\ldots. \label{multdef} \end{equation} For a given $f \in \mathcal{N}_{\beta},$ the following result quantifies the accuracy of the multilevel convolution approximation after $j$ steps of iterative refinement. \begin{proposition} \label{corr} Let $f \in \mathcal{N}_{\beta},$ $\beta >\frac{1}{2}$ and $0<h<\frac{1}{2\pi}.$ Then the multilevel error after the $j^{th}$ iterative refinement satisfies: $$ \| M_j f \|_{\infty} \le \begin{cases} C_{1}\left(\frac{2\pi h }{2^{\frac{j}{2}}}\right)^{2j} \,\, & \,\, \textrm{if} \,\,\,\, j<\frac{2\beta-1}{4}; \\ \left(\frac{2\pi h }{2^{\frac{j}{2}}}\right)^{2j}\left(C_{2}\sqrt{\ln\left(\frac{2^{\frac{j}{2}}}{2\pi h}\right)}+C_{3}\right) \,\, & \,\, \textrm{if} \,\,\,\, j=\frac{2\beta-1}{4}; \\ C_{4}\left(2\pi h \right)^{\beta-\frac{1}{2}}\left(\frac{1}{2^{j}}\right)^{\frac{2\beta-1}{4}} \,\, & \,\, \textrm{if} \,\,\,\, j>\frac{2\beta-1}{4}, \end{cases}$$ where $C_{i}$ $i=1,2,3,4$, are positive constants independent of $h.$ \end{proposition} \begin{proof} The proof follows the same pattern as that of Proposition \ref{prop1} and so rather than repeat the details we briefly outline the steps. We begin by noting that, since $\widehat{\psi}(0)=1$ and $\widehat{\psi}(-k)=\widehat{\psi}(k)$, $(k\in \mathbb{Z}_{+}$), we have \begin{equation} \| M_j f \|_{\infty} \le \sum_{k=1}^\infty b_{j}(k)(|f_k|+|f_{-k}|). \end{equation}\label{base-multierror} To reach the bound for the case where $j<\frac{2\beta-1}{4}$ (or equivalently $\beta>2j+1$) we note that (\ref{elem}) allows us to deduce that \begin{equation} b_{j}(k)\le \prod_{\ell=1}^j \frac{2\pi^{2}(2k)^{2}h^{2}}{2^{2\ell}}=\prod_{\ell=1}^j \frac{2^{3}\pi^{2}k^{2}h^{2}}{2^{2\ell}}= \left(\frac{2\pi h}{2^{\frac{j}{2}}}\right)^{2j}k^{2j}.\label{bj-bound} \end{equation} Using this we replicate the steps from Proposition 1 (associated to the $\beta>\frac{5}{2}$ case) to reach the required bound. For the remaining cases captured by $j\ge \frac{2\beta-1}{4}$ (or equivalently $\beta \le 2j+\frac{1}{2}$) we follow Proposition 1 again and define $m_{j}$ to be the integer satisfying \[ m_{j}\le\frac{2^{\frac{j}{2}}}{2\pi h}\le m_{j}+1. \] and, using this, the error expression is split into a finite sum (including the first $m_{j}$ terms) and the remaining infinite sum. For bounding purposes we employ (\ref{bj-bound}) for the finite sum and $b_{j}(k) \le 1$ for the infinite sum. Once again by mimicking the steps from Proposition 1 (associated to the $\beta\le\frac{5}{2}$ case) one can establish the stated bounds. \end{proof} We comment here that for functions of finite smoothness the speed at which the multilevel iterative refinement converges is restricted by the smoothness of the function. The error is reduced significantly in the early iterations, when $j < \frac{2\beta-1}{4}$ but beyond this point the error decays asymptotically at a polynomial rate. \section{Native space for Gaussian approximation} \label{convnat} It is more natural to approximate functions from the so called {\sl native space} for Gaussians: $$ \mathcal{N}_\psi = \left \{ f : \| f \|_\psi = \left ( \sum_{k=-\infty}^\infty \widehat{\psi}(k)^{-1} |f_k|^2 \right )^{1/2} < \infty \right \}. $$ If we approximate such very smooth functions we have the following result. \begin{theorem} \label{nat} Let $f \in \mathcal{N}_\psi$ and $0 <h<1$. Then $$ \| M_j f \|_{\infty} \le C\left(\frac{4h^{2}j}{e}\right)^{j}\left(\frac{1}{2^{j}}\right)^{j} \| f \|_\psi. $$ \end{theorem} \begin{proof} As in Proposition 2 we can use (\ref{bj-bound}) to develop the multilevel error as follows \begin{equation} \begin{aligned} \| M_j f \|_{\infty} &\le \sum_{k=1}^\infty b_{j}(k)(|f_k|+|f_{-k}|)\\ &\le\left(\frac{4\pi^{2}h^{2}}{2^{j}}\right)^{j}\sum_{k=1}^{\infty}k^{2j}\left(|f_{k}|+|f_{-k}|\right)\\ &=\left(\frac{4\pi^{2}h^{2}}{2^{j}}\right)^{j}\sum_{k=1}^{\infty}\sqrt{\widehat{\psi}(k)}k^{2j} \left(\frac{|f_{k}|}{\sqrt{\widehat{\psi}(k)}}+\frac{|f_{-k}|}{\sqrt{\widehat{\psi}(k)}}\right)\\ &\le \left(\frac{4\pi^{2}h^{2}}{2^{j}}\right)^{j}\sqrt{\sum_{k=1}^\infty\widehat{\psi}(k)k^{4j}}\left( \sqrt{\sum_{k=1}^\infty \frac{|f_{k}|^{2}}{\widehat{\psi}(k)}}+\sqrt{\sum_{k=1}^\infty \frac{|f_{k}|^{2}}{\widehat{\psi}(k)}}\right)\\ &\le 2\left(4\pi^{2}h^{2}\right)^{j}\left(\frac{1}{2^{j}}\right)^{j}\sqrt{\sum_{k=1}^\infty\widehat{\psi}(k)k^{4j}}\|f\|_{\psi}. \end{aligned}\label{base-multierror-nat} \end{equation} Concerning the infinite sum appearing in the bound above, we observe that for each $j=1,2,\ldots$ the non-negative function $x\mapsto\widehat{\psi}(x)x^{4j}$ is increasing for $0\le x \le \frac{\sqrt{j}}{\pi}$ and decreasing for $x\ge \frac{\sqrt{j}}{\pi}.$ Let $m_{j}$ denote the integer satisfying $m_{j} \le \frac{\sqrt{j}}{\pi} \le m_{j}+1,$ then we can write \[ \begin{aligned} &\sum_{k=1}^\infty\widehat{\psi}(k)k^{4j}=\sum_{k=1}^{m_{j}}\widehat{\psi}(k)k^{4j}+ \sum_{k=m_{j}+1}^{\infty}\widehat{\psi}(k)k^{4j}\\ &\le \int_{1}^{m_{j}}\widehat{\psi}(x)x^{4j}dx +\widehat{\psi}(m_{j})(m_{j})^{4j} +\widehat{\psi}(m_{j}+1)(m_{j}+1)^{4j}+\int_{m_{j}+1}^{\infty}\widehat{\psi}(x)x^{4j}dx\\ &\le 2\widehat{\psi}\left(\frac{\sqrt{j}}{\pi}\right) \left(\frac{\sqrt{j}}{\pi}\right)^{4j}+\int_{0}^{\infty}\widehat{\psi}(x)x^{4j}dx=2\left(\frac{j}{\pi^{2}e}\right)^{2j} +\int_{0}^{\infty}e^{-2\pi^{2}x^{2}}x^{4j}dx. \end{aligned} \] Examining the integral we have \[ \begin{aligned} \int_{0}^{\infty}e^{-2\pi^{2}x^{2}}x^{4j}dx &=\int_{0}^{\infty}e^{-s}\left(\frac{s}{2\pi}\right)^{2j}\frac{ds}{2\sqrt{2s}}\\ &= \frac{1}{\sqrt{2}(2\pi)(2\pi^{2})^{2j}}\int_{0}^{\infty}e^{-s}s^{2j-\frac{1}{2}}ds=\frac{\Gamma\left(2j+\frac{1}{2}\right)}{\sqrt{2}(2\pi)(2\pi^{2})^{2j}}. \end{aligned} \] Using Stirling's formula for the Gamma function \cite{AS} Formula 6.1.39 we have \[ \Gamma\left(2j+\frac{1}{2}\right)\le C\sqrt{2\pi}\left(\frac{2j}{e}\right)^{2j}. \] Substituting this into the bound above we see that \[ \sum_{k=1}^\infty\widehat{\psi}(k)k^{4j}\le C\left(\frac{j}{e\pi^{2}}\right)^{2j}. \] Taking the square root and substituting into (\ref{base-multierror-nat}) we conclude that \[ \begin{aligned} \| M_j f \|_{\infty}&\le 2\left(4\pi^{2}h^{2}\right)^{j}\left(\frac{1}{2^{j}}\right)^{j}\left(\frac{j}{e\pi^{2}}\right)^{2j}\|f\|_{\psi}\\ &\le C\left(\frac{4h^{2}j}{e}\right)^{j}\left(\frac{1}{2^{j}}\right)^{j}\|f\|_{\psi}. \end{aligned} \] \end{proof} We remark that, setting $d=2^j$, we get a convergence rate of $\mathcal{O}\left(\frac{4 h^2 \alpha \ln d}{e d}\right)^{\alpha \ln d},$ where $\alpha = 1/\ln(2)$ which is faster than any polynomial. \ifx\undefined\leavevmode\hbox to3em{\hrulefill}\, \newcommand{\leavevmode\hbox to3em{\hrulefill}\,}{\leavevmode\hbox to3em{\hrulefill}\,} \fi

\section{Introduction} \label{} A W-algebra is, generally speaking, a vertex operator algebra{} whose set of generating fields includes conformal primaries of conformal weights greater than $1$. Aside from the Virasoro minimal models, such algebras were first considered by Zamolodchikov \cite{ZamInf85} shortly after the genesis of conformal field theory \cite{BPZCFT84}. We refer to \cite{BSWsym93} for an overview of W-algebras and their role in conformal field theory. Here, we focus on the so-called \Walg{N} algebras, where \(N\ge 2\), which were first considered in \cite{BaiExt88}. They form a distinguished family of W-algebras that are generated by \(N-1\) fields of conformal weights \(2,3,\dots, N\), where the generating field of conformal weight $2$ is the energy momentum tensor and all the remaining generators are conformal primaries. While these algebras have received a lot of attention, there is much that remains poorly understood. Even for the case $N=3$ there are significant difficulties, see \cite{BSW3} for example. The \Walg{N} algebras were first proposed in the context of statistical mechanics, where they were linked to the continuum limits of certain \(\fld{Z}_N\)-symmetric lattice models \cite{FatZn85}. In particular, $\wvoa$ was first introduced to describe the extended symmetry of the $3$ state Potts model \cite{ZamInf85}. They enjoyed a period of intense popularity shortly thereafter as physicists explored the possibilities in their largely unsuccessful quest to classify all rational conformal field theories{}. More recently, under the name of minimal model holography, \Walg{N} algebras (and their supersymmetric generalisations) have been the focus of intense scrutiny because of their role in constructing the AdS/CFT duals of Vasiliev higher spin theories of gravity on \(\grp{AdS}_3\) \cite{VasHS96}, see \cite{GabMMH13} for an in depth review. The \Walg{N} algebras are also of great mathematical interest, since they are closely related to vertex operator algebras{} constructed from affine Lie algebras through either cosets \cite{GKO85,BaiCos88} of the form \begin{equation} \frac{\AKMA{sl}{N}_k\otimes\AKMA{sl}{N}_1}{\AKMA{sl}{N}_{k+1}} \end{equation} or through quantum hamiltonian reductions associated with principal nilpotents of \(\SLA{sl}{N}\) \cite{FeiQDS90,deBRel94}. Although the equivalence of the coset and principal reduction pictures is only now being established rigorously \cite{AraCos17,ACL18}, this has the happy consequence that even though \Walg{N} algebras are quite different, structurally, to affine vertex operator algebras{}, Lie-theoretic ideas can still be used to analyse them. The purpose of this paper is to derive explicit formulae for singular vectors of the \Walg{N} algebras. Recall that the (universal) \Walg{N} algebra may be realised as a subalgebra of the rank \(N-1\) Heisenberg vertex algebra{} and thus, by restriction, the Fock spaces of the Heisenberg algebra become modules over the \Walg{N} algebras. It is within these Fock spaces that we derive explicit formulae for \Walg{N} singular vectors. The derivation uses the simple, but far-reaching, fact that the universal enveloping algebra of the creation operators of a Heisenberg algebra is, as an associative algebra, isomorphic to the complexification of a ring of symmetric functions.\footnote{We adhere to the convention that a symmetric polynomial in a countably infinite number of variables is called a symmetric function.} The appeal of this fact is that it allows one to express complicated singular vectors, consisting of large linear combinations of products of Heisenberg creation operators, as images of symmetric functions that have a very simple form under the aforementioned isomorphism. This identification of singular vectors with symmetric functions can be traced back to Wakimoto and Yamada \cite{WakIrr83}, who discovered that Virasoro singular vectors in Fock spaces for the rank $1$ Heisenberg algebra of central charge \(c=1\) can be elegantly expressed in terms of Schur functions, a much studied basis of the ring of symmetric functions. This work was later generalised to free field realisations of the Virasoro algebra at arbitrary central charge by Mimachi and Yamada \cite{MimJa95}, with Schur functions being replaced by Jack symmetric functions \cite{JacCla70}. We recall that the Jack functions form a one-parameter family of bases of the ring of symmetric functions and that the Schur functions correspond to the parameter being set to $1$. This work relied heavily on a construction by Tsuchiya and Kanie \cite{TKScr286,TKScr86} of Virasoro module homomorphisms called screening operators. With these, singular vectors may be realised as images of Fock space \hwvs{}. These screening operators are also the key ingredient in Dotsenko and Fateev's Coulomb gas formalism \cite{DotScr84}. Similar singular vector{} constructions using screening operators have recently been detailed for the free field realisations of the vertex operator algebras{} associated with \(\AKMA{sl}{2}\) and the \(\mathcal{N}=1\) superconformal algebras, with singular vectors being evaluated using Jack symmetric functions, their supersymmetric generalisations and related families \cite{KatMis92,DesSup01,DesSJa12,RWsl215,YanUgl15,OPDS16}. The Jack symmetric function basis is indexed by partitions of integers and, interestingly, the singular vectors of Fock spaces for the Virasoro algebra and \(\AKMA{sl}{2}\) are always associated to a single Jack symmetric function indexed by a rectangular partition, that is, a partition whose parts are all equal. It is therefore natural to try and find an interpretation for the Jack symmetric functions indexed by non-rectangular partitions. One such interpretation was found by Awata, Odake, Matsuo and Shiraishi \cite{AwaCS95} while studying a connection between the Calogero-Sutherland model, a model for a system of non-relativistic particles on a circle with an inverse square potential, and the \Walg{N} algebras. In \cite{AwaCS95}, it is noted that \Walg{N} singular vectors{} can be realised using screening operators acting on rank $N-1$ Fock spaces. However, these singular vectors were not explicitly evaluated, but were instead projected onto a rank $1$ Fock space, where they were identified as eigenstates of the hamiltonian of the Calogero-Sutherland model. As the eigenstates of the Calogero-Sutherland model may be expressed as products of a ground state wavefunction and Jack polynomials, the conclusion of \cite{AwaCS95} is then that \Walg{N} singular vectors{} can be used to construct arbitrary Jack polynomials. Here, in contrast to \cite{AwaCS95}, we do not project, but stay in the rank \(N-1\) Fock spaces, using symmetric function theory to find explicit formulae for \Walg{N} singular vectors. Our main result is that certain singular vectors in rank $N-1$ Fock spaces may be identified with linear combinations of Jack symmetric functions and their skew variants, parametrised by sequences of partitions. The motivation for this study of singular vectors is that they have proven invaluable in finding short elegant proofs of the classification theorems for modules of rank $1$ vertex operator algebras{} \cite{TsuExt13,RWVir15,RWsl215,BloN117}. We believe that a solid understanding of higher rank singular vectors will pave the way to similar classification theorems for higher rank vertex operator algebras{}. This paper is organised as follows. \cref{} form an overview of known results and serve to fix notation. In \cref{}, we review the rank \(r\) Heisenberg vertex algebra for later use. We show how energy-momentum tensors can be constructed for arbitrary choices of central charge and discuss vertex operators and their compositions. In \cref{sec:W3}, we give the standard free field realisation of the \Walg{3} algebras by explicitly constructing them as subalgebras of the rank $2$ Heisenberg vertex algebra. We then determine their screening operators and identify the Fock spaces on which we can consistently evaluate their action. \cref{} gives a brief overview of the ring of symmetric functions, with particular emphasis on the Jack symmetric function bases, these being the bases required to evaluate the action of screening operators. The main results, explicit formulae for \Walg{N} singular vectors in certain Fock spaces in terms of the Jack symmetric functions, are given in \cref{} --- see \cref{} for the \Walg{3} singular vector formula and Equation \eqref{} for that of \Walg{N}. In \cref{}, we give a detailed derivation of these formulae for the \Walg{3} algebras. This case already illustrates the complexity of the general computations while keeping the formulae reasonably brief. This is followed by a discussion of simple examples in \cref{} featuring low-grade singular vectors{}. Although we do not report the results of investigating any really complicated examples, symbolic algebra packages incorporating symmetric functions, \textsc{SageMath} for example, allow such investigations to be straightforwardly performed to quite high grades. In \cref{sec:WN}, we generalise the singular vector formulae of \cref{} to the \Walg{N} algebras for all \(N\ge3\). \section{The rank $r$ Heisenberg algebra} \label{} The theory of \(r\) free chiral bosons, also known as the rank \(r\) Heisenberg vertex algebra \(\hvoa{r}\), is a staple of conformal field theory and vertex operator algebra{} theory. Here, we construct the Heisenberg algebras by affinising abelian Lie algebras. The rank $r$ Heisenberg algebra is constructed from an \(r\)-dimensional complex vector space \(\hvs\) together with a non-degenerate symmetric bilinear form \(\kil{\blank}{\blank}\). We pick a basis \(\{\scrt{1},\dots,\scrt{r}\}\) of \(\hvs\) such that the Gram matrix of \(\kil{\blank}{\blank}\) is the Cartan matrix of \(\SLA{sl}{r+1}\): \begin{equation} \kil{\scrt{i}}{a^j}=2\delta_{i,j}-\delta_{i+1,j}-\delta_{i,j+1},\quad i,j=1,\dots,r. \end{equation} Since \(\kil{\blank}{\blank}\) is non-degenerate, it defines a vector space isomorphism \(\iota:\hvs\to\dvs\) by \(a\mapsto\kil{a}{\blank}\). The induced non-degenerate symmetric bilinear form will also be denoted by \(\kil{\blank}{\blank}\). We denote the images of the basis vectors \(\scrt{i}\) by \(\srt{i}=\iota(\scrt{i})\) and the elements of the basis of \(\dvs\) dual to \(\set{\scrt{i}}\) by \(\fwt{i}\). Thus, $\fwt{i}(\scrt{j}) = \delta_i^j$. The \(\srt{i}\) and \(\fwt{i}\) may therefore be identified as simple roots and fundamental weights, respectively, of \(\SLA{sl}{r+1}\). In this picture, the basis vectors \(\scrt{i}\in\hvs\) are the simple coroots of \(\SLA{sl}{r+1}\). To any vector \(a\in\hvs\), one assigns a field \(a(z)\) whose defining operator product expansions are \begin{equation} a(z)b(w)\sim \frac{\kil{a}{b}}{(z-w)^2}, \quad a,b\in\hvs. \end{equation} These fields admit Fourier expansions of the form \begin{equation} a(z)=\sum_{n\in\fld{Z}} a_{n} z^{-n-1},\quad a\in\hvs, \end{equation} whose modes satisfy the following commutation relations: \begin{equation} \comm[\big]{a_m}{b_n}=m\kil{a}{b} \delta_{m,-n}\mathsf{1}. \end{equation} The Heisenberg Lie algebra \(\heis\) is the infinite-dimensional Lie algebra spanned by the central element \(\mathsf{1}\) and the generators \(a_m\), for all \(a\in\hvs\) and \(m\in\fld{Z}\). We have chosen to denote the central element by \(\mathsf{1}\), since we assume that it will act as the identity on any \(\heis\)-module. A basis of $\heis$ is then given by $\mathsf{1}$ and the $\scrt{i}_m$, with $i=1,\dots,r$ and $m\in\fld{Z}$. The Heisenberg Lie algebra admits a triangular decomposition \begin{align} \heis=\heis_{-}\oplus \heis_0\oplus\heis_+, \quad \heis_\pm=\bigoplus_{i=1}^r\bigoplus_{m\ge 1}\fld{C} \scrt{i}_{\pm m}, \quad \heis_0=\bigoplus_{i=1}^r\fld{C} \scrt{i}_0\oplus\fld{C}\mathsf{1}. \end{align} Verma modules over $\heis$ are commonly referred to as Fock spaces. These are induced from the one-dimensional modules $\fld{C}\ket{\zeta}$, $\zeta\in\dvs$, over $\heis_{\ge}=\heis_0\oplus\heis_+$ that are defined by \begin{align} \mathsf{1}\ket{\zeta}=\ket{\zeta}, \qquad a_n\ket{\zeta}=\delta_{n,0} \zeta(a) \ket{\zeta},\quad a\in\hvs,\quad n\ge0. \end{align} The Fock spaces \begin{align} \Fock{\zeta}=\UEA{\heis}\otimes_{\UEA{\heis_\ge}}\fld{C}\ket{\zeta} \end{align} are well known to be simple $\heis$-modules, for all $\zeta \in \dvs$. As a module over itself, the Heisenberg vertex algebra $\hvoa{r}$ is identified with the Fock space \(\Fock{0}\) and the state-field correspondence is given by \begin{equation} \ket{0}\longleftrightarrow \mathsf{1}, \quad b^1_{-n_1-1}\cdots b^k_{-n_k-1}\ket{0}\longleftrightarrow \normord{\frac{\partial^{n_1}}{n_1!}b^1(z)\cdots\frac{\partial^{n_k}}{n_k!}b^k(z)}, \end{equation} where \(b^1,\dots,b^k\in \hvs\) and normal ordering is defined in the usual way. The Heisenberg vertex algebra \(\hvoa{r}\) can be endowed with the structure of a vertex operator algebra{} by choosing an energy-momentum tensor. This choice is not unique. For the purposes of this note, we shall restrict our attention to the following one-parameter family of energy-momentum tensors: \begin{equation}\label{} T(z) = \sum_{i=1}^r \sqbrac*{\frac{1}{2} \normord{\scrt{i}(z) \dscrt{i}(z)} + \alpha_0 \partial \dscrt{i}(z)}, \quad \alpha_0 \in \fld{C}. \end{equation} Here, the \(\dscrt{i} \in \hvs\) are dual to the coroots \(\scrt{i}\) in the sense that $\iota(\dscrt{i})=\fwt{i}$. We note that while the quadratic summand in the above energy-momentum tensor is basis independent, the linear summand is not. The central charge corresponding to this choice of energy-momentum tensor depends on the parameter \(\alpha_0\): \begin{equation}\label{} c = r - r(r+1)(r+2) \alpha_0^2. \end{equation} By definition, the coefficients of the Fourier expansion of the energy-momentum tensor satisfy the commutation relations of the Virasoro algebra. Thus, formula \eqref{} realises the Virasoro generators \(L_n\), $n \in \fld{Z}$, as infinite sums of products of Heisenberg generators: \begin{align} T(z)=\sum_{n\in \fld{Z}}L_n z^{-n-2}, \quad L_n=\sum_{i=1}^r\sqbrac*{\frac{1}{2}\sum_{m\in\fld{Z}} \normord{\scrt{i}_m\dscrt{i}_{n-m}}-\alpha_0(n+1)\dscrt{i}_n}. \end{align} This identification yields an action of the Virasoro algebra on the Fock spaces \(\Fock{\zeta}\), \(\zeta\in\dvs\). In this way, any \hwv{} \(\ket{\zeta} \in \Fock{\zeta}\) is also a Virasoro \hwv{}: \begin{align} \label{} L_n\ket{\zeta}=h_\zeta\delta_{n,0}\ket{\zeta},\quad n\ge 0,\quad h_\zeta=\frac{1}{2} \kil{\zeta}{\zeta-2\alpha_0\weyl}. \end{align} Here, $\weyl = \sum_i\fwt{i}$ is the Weyl vector of $\SLA{sl}{r+1}$. We note that while the Fock spaces are simple as Heisenberg modules, they need not be as Virasoro modules. The primary fields of the free boson theory are called vertex operators (not to be confused with elements of the Heisenberg vertex operator algebra). To define them, we first need to extend $\heis$ by \(\fld{C}[\dvs]\), the group algebra of \(\dvs\), treating \(\dvs\) as an abelian group under vector addition and $\fld{C}[\dvs]$ as an abelian Lie algebra. We denote the group algebra basis element corresponding to \(\eta\in\dvs\) by \(\gae{\eta}\) and define the commutation relations between the generators \(a_m\) and \(\gae{\eta}\) by \begin{align} \comm*{a_m}{\gae{\eta}}=\delta_{m,0} \eta(a)\gae{\eta}, \quad a\in\hvs, \quad \eta\in\dvs, \quad m\in\fld{Z}. \end{align} It is easy to check that this extension of $\heis$ by $\fld{C}[\dvs]$ is a semidirect sum of Lie algebras. A standard computation now shows that $\gae{\eta}$ maps the \hwv{} $\ket{\zeta} \in \Fock{\zeta}$ to a \hwv{} of $a_0$-eigenvalue $\zeta(a) + \eta(a)=(\zeta+\eta)(a)$. Following usual practice, we shall identify $\gae{\eta} \ket{\zeta}$ with $\ket{\zeta + \eta}$. The vertex operator corresponding to $\ket{\zeta} = \gae{\zeta} \ket{0}$ is \begin{align} \label{} \vop{\zeta}{z}=\gae{\zeta}z^{a_0}\prod_{m\ge1}\exp\brac*{\frac{a_{-m}}{m}z^{m}} \exp\brac*{-\frac{a_m}{m}z^{-m}}, \quad \zeta = \iota(a) \in \dvs. \end{align} These primary fields therefore define linear maps between Fock spaces: \begin{align} \vop{\zeta}{z}\colon\Fock{\eta}\to z^{(\zeta,\eta)} \Fock{\zeta+\eta}\llbracket z,z^{-1}\rrbracket. \end{align} It is easy to check from the $\hvoa{r}$-primary operator product expansion{} \begin{align}\label{ope:av} a(z)\vop{\zeta}{w}&\sim \frac{\zeta(a)\vop{\zeta}{w}}{z-w} \end{align} that $a(z)$ and $\vop{\zeta}{w}$ are mutually local for all $a \in \hvs$ and $\zeta \in \dvs$. The same is therefore true for an arbitrary field of $\hvoa{r}$ and any vertex operator, by Dong's lemma. Finally, suppose that $\zeta_i = \iota(b^i) \in \dvs$, for $i=1,\dots,k$. Then, a standard computation allows one to write the composition of the $k$ vertex operators $\vop{\zeta_i}{z_i}$ as \begin{align}\label{} \vop{\zeta_1}{z_1}\cdots \vop{\zeta_k}{z_k}=\prod_{i=1}^k \gae{\zeta_i} \cdot \prod_{\mathclap{1\le i<j\le k}} \: (z_i-z_j)^{(\zeta_i,\zeta_j)} \cdot \prod^k_{i=1} z_i^{b^i_0} \cdot \prod_{m\ge 1}\exp\brac*{\frac{1}{m}\sum^k_{i=1}b^i_{-m}z^m_i} \exp\brac*{-\frac{1}{m}\sum^k_{i=1}b^i_{m}z^{-m}_i}. \end{align} This explicit formula will be used many times in what follows. \section{The Universal \Walg{3} Vertex Operator Algebra} \label{sec:W3} In this section, we restrict ourselves to the rank $2$ Heisenberg vertex algebra \(\hvoa{2}\) and, in the vein of \cite{ZamZ85}, define a family of subalgebras, each denoted by \(\Wvoa{3}\), called the \Walg{3} vertex operator algebras, or \Walg{3} algebras for short. These algebras are parametrised by $\alpha_0 \in \fld{C}$ and are strongly generated by the energy-momentum tensor defined in \eqref{} and an additional primary field \(W(z)\) of conformal weight $3$. In the basis \(\set{\scrt{1},\scrt{2}}\) defined above, for which the Gram matrix of the inner product \(\kil{\blank}{\blank}\) is the Cartan matrix of \(\SLA{sl}{3}\), the energy-momentum tensor \(T(z)\) is \begin{subequations} \label{} \begin{align} T(z)&=\frac{1}{3}\normord{\scrt{1}(z)\scrt{1}(z)}+\frac{1}{3}\normord{\scrt{1}(z)\scrt{2}(z)}+\frac{1}{3}\normord{\scrt{2}(z)\scrt{2}(z)} +\alpha_0\partial_z \scrt{1}(z)+\alpha_0\partial_z \scrt{2}(z) \intertext{and the central charge is $c=2-24\alpha_0^2$. The conformal primary of weight $3$ is then} W(z)&=\frac{\sqrt{\beta}}{18\sqrt{3}}\Bigl[ 2\normord{\brac*{\scrt{2}(z)-\scrt{1}(z)}\brac*{\scrt{1}(z)+2\scrt{2}(z)}\brac*{2\scrt{1}(z)+\scrt{2}(z)}} \Bigr.\nonumber\\ &\quad\Bigl. +9\alpha_0\brac*{\normord{\partial \scrt{2}(z)\brac*{\scrt{1}(z)+2\scrt{2}(z)}} -\normord{\partial \scrt{1}(z)\brac*{2\scrt{1}(z)+\scrt{2}(z)}}} +9\alpha_0^2\brac*{\partial^2 \scrt{2}(z)-\partial^2\scrt{1}(z)}\Bigr], \end{align} \end{subequations} where \begin{equation} \beta=\frac{16}{22+5c} = \frac{2}{4-15\alpha_0^2} \end{equation} in the conventional normalisation, appropriate for $c \neq -\frac{22}{5}$ ($\alpha_0 \neq \pm \frac{2}{\sqrt{15}}$). A somewhat involved computation now determines the operator product expansion of \(W(z)\) with itself to be \begin{align} W(z)W(w)&\sim\frac{c/3}{(z-w)^6}+\frac{2T(w)}{(z-w)^4}+\frac{\partial T(w)}{(z-w)^3}+\frac{\frac{3}{10}\partial^2 T(w)+2\beta\Lambda(w)}{(z-w)^2}+\frac{\frac{1}{15}\partial^3T(w)+\beta\partial\Lambda(w)}{z-w}, \end{align} where \(\Lambda(z)=\normord{T(z)T(z)}-\frac{3}{10}\partial^2 T(z)\). This, along with the primary nature of $W(z)$, implies the commutation relations \begin{subequations} \label{} \begin{align} \comm[\big]{L_m}{W_n}&=(2m-n)W_{m+n},\\ \comm[\big]{W_m}{W_n}&=(m-n)\sqbrac*{\frac{1}{15}(m+n+3)(m+n+2)-\frac{1}{6}(m+2)(n+2)}L_{m+n}\nonumber\\ &\quad+\beta(m-n)\Lambda_{m+n}+\frac{c}{360}m(m^2-1)(m^2-4)\delta_{m+n,0}, \end{align} \end{subequations} where $W(z) = \sum_{n \in \fld{Z}} W_n z^{-n-3}$. Since Fock spaces are modules over the Heisenberg vertex operator algebra{} \(\hvoa{2}\) and we have defined the \Walg{3} algebra as a subalgebra of \(\hvoa{2}\), each Fock space is a \(\Wvoa{3}\)-module, by restriction. In particular, the \hwv{} \(\ket{\zeta}\in\Fock{\zeta}\), \(\zeta\in\dvs\), is also a \hwv{} for \(\Wvoa{3}\): \begin{align} L_n\ket{\zeta}= \delta_{n,0}h_\zeta \ket{\zeta},\quad W_n\ket{\zeta}= \delta_{n,0}w_\zeta \ket{\zeta},\quad n\ge0. \end{align} Here, $h_\zeta$ was given in \eqref{} and the $W_0$-eigenvalue is given by \begin{align} \label{} w_\zeta &= \sqrt{3\beta} \kil{\zeta}{\fwt{2}-\fwt{1}} \brac[\big]{\kil{\zeta}{\fwt{1}} - \alpha_0} \brac[\big]{\kil{\zeta}{\fwt{2}} - \alpha_0}. \end{align} Our main reason for introducing vertex operators in \eqref{} is to construct linear maps between Fock spaces that commute with the action of an appropriate subalgebra of the Heisenberg vertex algebra. Here, we wish to construct maps that commute with \(\Wvoa{3}\), that is, \(\Wvoa{3}\)-module homomorphisms. Such module homomorphisms are called \emph{screening operators} and they are constructed from \emph{screening fields}, these being vertex operators whose operator product expansions with the fields of \(\Wvoa{3}\) are total derivatives. For this, it clearly suffices to find fields whose operator product expansions with the generating fields \(T(z)\) and \(W(z)\) are total derivatives. As the vertex operator $\vop{\zeta}{w}$ is a conformal primary of weight $h_{\zeta}$, its operator product expansion{} with $T(z)$ will be a total derivative if and only if \(h_{\zeta}=1\). Unsurprisingly, the analogous computation for \(W(z)\) is more involved (we used Thielemans' \textsc{OPEdefs} package for \textsc{Mathematica}. We shall not give the unpleasant details, noting instead that a necessary condition for the operator product expansion \(W(z) \vop{\zeta}{w}\) to be a total derivative is that the coefficient of $(z-w)^{-1}$ in this expansion is a total derivative. Analysing this explicitly, for general \(\zeta\in\dvs\), and recalling that $h_{\zeta} = 1$, we conclude that this coefficient will be a total derivative if \(\zeta_1=0\), \(\zeta_2=0\) or if \(\zeta_1=\zeta_2\) and \(\alpha_0=0\) (where the $\zeta_i$ denote Dynkin labels: $\zeta = \sum_i \zeta_i \fwt{i}$). As we are only interested in screening operators that exist for all values of \(\alpha_0\), it follows that there are exactly four possible weights $\zeta$ that can be used to construct screening operators: $\zeta = \alpha_{\pm} \srt{1}, \alpha_{\pm} \srt{2}$. Here, we define \begin{align}\label{} \alpha_+=\frac{1}{2} \brac*{\alpha_0+\sqrt{\alpha_0^2+4}},\quad \alpha_-=\frac{1}{2} \brac*{\alpha_0-\sqrt{\alpha_0^2+4}} \end{align} to be the solutions of the quadratic equations \(h_{\zeta_i\srt{i}}=1\), for \(i=1,2\). It only remains to confirm that the full operator product expansion \(W(z) \vop{\zeta}{w}\), when $\zeta$ is one of the above weights, is indeed a total derivative: \begin{align} W(z)\vop{\alpha_\pm\srt{i}}{w}\sim-(-1)^i\sqrt{\frac{6}{4-15\alpha_0^2}}\partial_w \brac*{ \frac{\alpha_0\vop{\alpha_\pm\srt{i}}{w}}{2(z-w)^2}-\frac{\normord{\dscrt{1}(w)\vop{\alpha_\pm\srt{i}}{w}}}{z-w} }, \quad i=1,2. \end{align} Having identified screening fields for \(\Wvoa{3}\), we construct screening operators by taking residues: \begin{equation} \scr{\pm i}=\oint_0 \vop{\alpha_\pm\srt{i}}{w} \,\dd w. \end{equation} Here, the residue is indicated using a simple anticlockwise contour that encircles $0$ once (we absorb the usual factor of $2\pi\ii$ into the definition of the contour integral). These screening operators define \(\Wvoa{3}\)-module homomorphisms since \begin{align} \comm[\big]{T(z)}{\scr{\pm i}} =-\oint_z T(z)\vop{\alpha_\pm \srt{i}}{w} \,\dd w=0, \quad \comm[\big]{W(z)}{\scr{\pm i}} =-\oint_z W(z)\vop{\alpha_\pm \srt{i}}{w} \,\dd w=0. \end{align} These identities follow from the mutual locality of Heisenberg fields and vertex operators, see \eqref{ope:av}. Taking the residue of a screening field \(\vop{\alpha_\pm\srt{i}}{z}\) is of course only well defined when it is acting on a $\hvoa{2}$-module for which the exponents of \(z\) in the Fourier expansion of \(\vop{\alpha_\pm\srt{i}}{z}\) are all integers. In case the $\hvoa{2}$-module is the Fock space $\Fock{\eta}$, this is satisfied if and only if \(\alpha_{\pm}\kil{\srt{i}}{\eta}\in\fld{Z}\). Fortunately, one can also construct screening operators by integrating compositions \eqref{} of multiple screening fields. In particular, composing \(r_2\) copies of \(\vop{\alpha_\pm \srt{2}}{w}\) with \(r_1\) copies of \(\vop{\alpha_\pm\srt{1}}{z}\) and then acting on $\Fock{\eta}$ gives \begin{multline}\label{} \Bigl.\vop{\alpha_\pm\srt{1}}{z_1}\cdots\vop{\alpha_\pm\srt{1}}{z_{r_1}} \vop{\alpha_\pm \srt{2}}{w_1}\cdots\vop{\alpha_\pm\srt{2}}{w_{r_2}}\Bigr\rvert_{\Fock{\eta}}\\ =\prod_{1\le i<j\le r_1}\brac*{z_i-z_j}^{2\alpha_\pm^2} \cdot\prod_{1\le i<j\le r_2}\brac*{w_i-w_j}^{2\alpha_\pm^2} \cdot\prod_{i=1}^{r_1}\prod_{j=1}^{r_2}\brac*{z_i-w_j}^{-\alpha\pm^2} \cdot\prod_{i=1}^{r_1}z_i^{\alpha_\pm\kil{\srt{1}}{\eta}} \cdot\prod_{j=1}^{r_2}w_j^{\alpha_\pm\kil{\srt{2}}{\eta}}\\ \quad\cdot \gae{r_1\alpha_\pm\srt{1}+r_2\alpha_\pm\srt{2}} \prod_{m\ge1} \exp\sqbrac*{\alpha_\pm\brac*{ \scrt{1}_{-m}\sum_{i=1}^{r_1}\frac{z_i^m}{m} +\scrt{2}_{-m}\sum_{i=1}^{r_2}\frac{w_i^m}{m} }} \exp\sqbrac*{-\alpha_\pm\brac*{ \scrt{1}_{m}\sum_{i=1}^{r_1}\frac{z_i^{-m}}{m} +\scrt{2}_{m}\sum_{i=1}^{r_2}\frac{w_i^{-m}}{m} }}. \end{multline} Up to a complex phase, which we suppress, the first five multivalued factors in this expression can be rewritten in the form \begin{multline}\label{} \prod_{1\le i\neq j\le r_1}\brac*{1-\frac{z_i}{z_j}}^{\alpha_\pm^2} \cdot\prod_{1\le i\neq j\le r_2}\brac*{1-\frac{w_i}{w_j}}^{\alpha_\pm^2} \cdot\prod_{i=1}^{r_1}\prod_{j=1}^{r_2}\brac*{1-\frac{w_j}{z_i}}^{-\alpha_\pm^2}\\ \cdot\prod_{i=1}^{r_1}z_i^{\alpha_\pm\kil{\srt{1}}{\eta}+\alpha_\pm^2\brac*{r_1-r_2-1}} \cdot\prod_{j=1}^{r_2}w_j^{\alpha_\pm\kil{\srt{2}}{\eta}+\alpha_\pm^2\brac*{r_2-1}}, \end{multline} thereby isolating the non-integer exponents of the $z_i$ and $w_j$ in the last two factors. Finding closed (multivariable) contours over which multivalued functions such as \eqref{} can be integrated (to obtain $\Wvoa{3}$-module homomorphisms) is a highly non-trivial problem. Fortunately, Tsuchiya and Kanie solved this problem for the rank $1$ Heisenberg vertex algebra \cite{TKScr86} by constructing cycles with non-trivial homology classes over which screening operators can be integrated. These cycles, which we shall denote by \(\cyc{m;t}\) for \(m\in \fld{Z}_{\ge0}\) and \(t\in\fld{C}\setminus \fld{Q}_{\le0}\),\footnote{ The range of the parameter \(t\) could in principle be extended to \(\fld{C}\setminus\{0\}\). However, to avoid singularities in certain coefficients, this would require one to use a different normalisation of the Jack symmetric function basis presented in \cref{}. Moreover, some linear independence arguments would become more complicated. For simplicity, we therefore avoid non-positive rational values of the parameter \(t\).} allow one to integrate expressions of the form \begin{equation} \int_{\cyc{m;t}}\prod_{1\le i\neq j\le m}\brac*{1-\frac{z_i}{z_j}}^{1/t} \cdot f(z) \,\dd z_1\cdots \dd z_m, \end{equation} where \(f(z)\) is a Laurent polynomial in \(z_1,\dots,z_m\) which is invariant with respect to permuting the indices of its variables. We shall not describe the construction of these cycles in any detail. It will, however, be convenient to normalise them by requiring that \begin{equation}\label{} \int_{\cyc{m;t}}\prod_{1\le i\neq j\le m}\brac*{1-\frac{z_i}{z_j}}^{1/t} \frac{\dd z_1\cdots \dd z_m}{z_1 \cdots z_m}=1. \end{equation} The cycles \(\cyc{m;t}\) can be used to construct screening operators from the compositions \eqref{} whenever the exponents of the \(z_i\) and \(w_j\) are integers. If this is the case, then the screening operators are defined as \begin{equation}\label{} \scrs{\pm}{r_1,r_2}=\int_{\cyc{r_1;1/\alpha_\pm^2}}\int_{\cyc{r_2;1/\alpha_\pm^2}} \vop{\alpha_\pm\srt{1}}{z_1}\cdots\vop{\alpha_\pm\srt{1}}{z_{r_1}} \vop{\alpha_\pm \srt{2}}{w_1}\cdots\vop{\alpha_\pm \srt{2}}{w_{r_2}} \,\dd z_1\cdots \dd z_{r_1}\dd w_1\cdots\dd w_{r_2}. \end{equation} By construction, these screening operators are \(\Wvoa{3}\)-module homomorphisms when acting on appropriate Fock spaces. We parametrise the Fock space weights for which the screening operators \eqref{} are defined as follows: \begin{equation} \label{} \svwts{u_1,v_1}{u_2,v_2}=\brac[\big]{\brac{1-u_1}\alpha_+ + \brac{1-v_1}\alpha_-}\fwt{1} +\brac[\big]{\brac{1-u_2}\alpha_+ + \brac{1-v_2}\alpha_-}\fwt{2}, \quad u_1,u_2,v_1,v_2\in\fld{Z}. \end{equation} Considering the exponents of the last two factors of \eqref{}, we conclude that the screening operators define $\Wvoa{3}$-module homomorphisms between the following Fock spaces: \begin{equation} \label{} \begin{aligned} \scrs{+}{r_1,r_2}\colon\Fock{\svwts{r_1-r_2,s_1}{r_2,s_2}}&\to\Fock{\svwts{-r_1,s_1}{r_1-r_2,s_2}}, & r_1,r_2&\in\fld{Z}_{\ge0}, & s_1,s_2&\in\fld{Z},\\ \scrs{-}{s_1,s_2}\colon\Fock{\svwts{r_1,s_1-s_2}{r_2,s_2}}&\to\Fock{\svwts{r_1,-s_1}{r_2,s_1-s_2}}, & r_1,r_2&\in\fld{Z}, & s_1,s_2&\in\fld{Z}_{\ge0}. \end{aligned} \end{equation} Evaluating the action of these screening operators initially appears rather daunting. However, we know from \eqref{} that compositions of screening fields factorise into a product of a multivalued function and certain power series in the \(z_i\) and \(w_j\) that are \emph{symmetric} with respect to permuting the \(z_i\) among themselves and, separately, the \(w_j\) among themselves. The theory of symmetric functions provides the tools that allow us to evaluate the action of these screening operators on certain Fock spaces. We therefore turn to a discussion of these tools. \section{The ring of symmetric functions} \label{} The purpose of this section is to review various results from the theory of symmetric functions that will be used to evaluate the action of screening operators on certain Fock spaces. The standard reference for symmetric functions and their myriad properties is Macdonald's book \cite{MacSym95} to which we refer the reader for more details. Let $\fsym{n}$ denote the ring of symmetric polynomials in the \(n\) variables $z_1, \dots, z_n$. This is the subring of \(\fld{C}[z_1,...,z_n]\) that consists of the polynomials that are invariant with respect to permuting the indices of the $z_i$. It admits numerous interesting generators such as the power sums \begin{align} \powsum{k}=\sum_{i=1}^n z^k_i, \quad k\ge1. \end{align} For \(1\le k\le n\), the \(\powsum{k}\) are algebraically independent and freely generate \(\fsym{n}\), that is, \begin{align} \fsym{n}=\fld{C}[\powsum{1},\cdots,\powsum{n}]. \end{align} We can therefore use partitions \(\lambda=\partn{\lambda_1,\lambda_2,\dots}\), whose parts $\lambda_i$ are bounded by \(n\), to define \begin{equation} \powsum{\lambda}=\powsum{\lambda_1}\cdots\powsum{\lambda_k}. \end{equation} These power sums, labelled by partitions whose parts do not exceed $n$, thus form a basis of \(\fsym{n}\): \begin{equation}\label{} \fsym{n}=\bigoplus_{\lambda, \lambda_1\le n}\fld{C} \powsum{\lambda}. \end{equation} Another family of symmetric polynomials is given by the monomial symmetric polynomials \begin{equation} \monsym{\lambda}=\sum_{\sigma}z_1^{\lambda_{\sigma(1)}}\cdots z_n^{\lambda_{\sigma(n)}}, \end{equation} where \(\sigma\) runs over all distinct permutations of the partition \(\lambda\). In this case, \(\lambda\) is not constrained by a bound on its individual parts, but by their number \(\len{\lambda}\) (the length of \(\lambda\)) which is at most $n$. Note that each monomial summand of $\monsym{\lambda}$ has coefficient $1$. For example, \begin{equation} \fmonsym{[2,2]}{z_1,z_2}=z_1^2z_2^2,\quad \fmonsym{[2,2]}{z_1,z_2,z_3}=z_1^2z_2^2+z_1^2z_3^2+z_2^2z_3^2. \end{equation} The monomial symmetric polynomials also form a basis of \(\fsym{n}\): \begin{equation}\label{} \fsym{n}=\bigoplus_{\lambda,\len{\lambda}\le n}\fld{C}\monsym{\lambda}. \end{equation} The respective restrictions on parts and lengths of partitions in the definitions of these symmetric polynomials can be avoided by taking a formal limit to infinitely many variables. The resulting ring \(\sym\) is called the ring of symmetric functions and, unsurprisingly, its elements are called symmetric functions. The ring \(\fsym{n}\) of symmetric polynomials in \(n\) variables can then be easily recovered from \(\sym\) by setting all but the first \(n\) variables to $0$. This amounts to a projection \begin{align} \pi_n:\sym\to\fsym{n}, \quad f(x_1,x_2,\dots)\mapsto f(x_1,\dots,x_n,0,0,\dots). \end{align} In \(\sym\), the power sums \(\powsum{k}\) are algebraically independent for all \(k\ge1\) and they freely generate \(\sym\), that is, \begin{equation} \label{} \sym=\fld{C}[\powsum{1},\powsum{2},\dots]. \end{equation} Similarly, the restrictions on the sizes of the parts and the lengths of the partitions labelling power sums and monomial symmetric functions, respectively, no longer apply. Both classes of symmetric functions give bases of \(\sym\): \begin{equation} \sym = \bigoplus_{\lambda}\fld{C} \powsum{\lambda}=\bigoplus_{\lambda}\fld{C}\monsym{\lambda}. \end{equation} We note that \(\pi_n\brac*{\monsym{\lambda}}=0\) if and only if \(\len{\lambda}>n\), but that no such truncations exist for the power sums \(\powsum{k}\): their images under \(\pi_n\) are all non-zero. There exists another family of bases of \(\sym\) and \(\fsym{n}\) labelled by partitions, called the Jack symmetric functions and Jack symmetric polynomials (or just Jack functions or polynomials for short), respectively. These are defined using the dominance partial ordering of partitions: if $\lambda$ and $\mu$ are both partitions of the same non-negative integer, then we write \(\lambda\ge \mu\) (and say that $\lambda$ dominates $\mu$) if \begin{equation} \lambda_1+\cdots+\lambda_i\ge\mu_1+\cdots+\mu_i, \end{equation} for all \(i\ge1\). For each $t\in\fld{C}\setminus\fld{Q}_{\le0}$ (the non-positive rationals are excluded to avoid certain normalisation problems), the Jack functions \(\jack{\lambda}{t}\) are uniquely defined by the following two properties: \begin{enumerate} \item For any partition \(\lambda\), \(\jack{\lambda}{t}\) admits an upper triangular decomposition of the form \begin{align}\label{upptri} \jack{\lambda}{t}=\monsym{\lambda}+\sum_{\lambda>\mu} v_{\lambda,\mu}(t)\monsym{\mu}, \qquad v_{\lambda,\mu}(t) \in \fld{C}. \end{align} \item The Jack functions form an orthogonal basis of $\sym$ with respect to the inner product defined by \begin{align}\label{innprod} \cjprod{\powsum{\lambda},\powsum{\mu}}{t}=t^{\len{\lambda}} \delta_{\lambda\mu} \prod_{i\ge1}i^{m_i}m_i!, \end{align} where $m_i$ denotes of number of parts of $\lambda$ equal to $i$. \end{enumerate} For each $n\ge1$, the Jack polynomials in \(\fsym{n}\) may be defined as the images of the corresponding Jack functions in $\sym$ under the projection \(\pi_n\). As with monomial symmetric polynomials, we have \(\pi_n\brac*{\jack{\lambda}{t}}=0\) if and only if \(\len{\lambda}>n\). For \(\len{\lambda}\le n\), the Jack polynomials \begin{align} \fjack{\lambda}{t}{z_1,\dots,z_n}=\pi_n(\jack{\lambda}{t}) \end{align} are linearly independent and form a basis of \(\fsym{n}\). For the application to follow, we mention the following important examples in $\fsym{n}$ called the \emph{rectangular} Jack polynomials. In these, the partition has the form $\lambda = [m^n]$ in which all $n$ parts are equal to $m$. Rectangular Jack polynomials have a very simple form: \begin{equation} \label{} \fjack{[m^n]}{t}{z_1,\dots,z_n} = \fmonsym{[m^n]}{z_1,\dots,z_n} = \prod_{i=1}^n z_i^m. \end{equation} This follows because all partitions of $mn$ that are strictly dominated by $[m^n]$ have length greater than $n$. They also have extremely simple products with other Jacks. For $\len{\lambda} \le n$, denote by $\lambda + [m^n]$ the partition with parts $\lambda_i + m$. Then, \begin{equation} \label{} \fjack{[m^n]}{t}{z_1,\dots,z_n} \fjack{\lambda}{t}{z_1,\dots,z_n} = \fjack{\lambda + [m^n]}{t}{z_1,\dots,z_n}. \end{equation} We emphasise that rectangular Jack polynomials are independent of the parameter $t$. The Jack functions and polynomials satisfy many properties that shall be essential for what follows. We list some of them here for convenience. \begin{enumerate} \item We denote by $\djack{\lambda}{t}$ the elements of the basis dual to the $\jack{\lambda}{t}$ with respect to the inner product \eqref{innprod}. Since the Jack functions form an orthogonal basis, $\jack{\lambda}{t}$ is proportional to $\djack{\lambda}{t}$: \begin{align} \djack{\lambda}{t}=\nc{\lambda}{t}\jack{\lambda}{t}, \quad \nc{\lambda}{t} =\frac{1}{\cjprod{\jack{\lambda}{t},\jack{\lambda}{t}}{t}}. \end{align} The proportionality constant \(\nc{\lambda}{t}\) is given explicitly by \begin{align}\label{} \nc{\lambda}{t}=\prod_{s\in\lambda}\frac{a(s)t+l(s)+1}{(a(s)+1)t+l(s)}, \end{align} where \(a(s)\) and \(l(s)\) denote the arm and leg lengths, respectively, of the box \(s\) in the Young diagram of \(\lambda\). \item The Jack functions and their duals admit a kind of generating function called the \emph{Cauchy kernel}: \begin{align}\label{cauchykernel} \prod_{i,j}(1-y_iz_j)^{-1/t}= \prod_{m\ge1}\exp\brac*{\frac{1}{t}\frac{\fpowsum{m}{y}\fpowsum{m}{z}}{m}}= \sum_{\lambda}\fjack{\lambda}{t}{y}\fdjack{\lambda}{t}{z}. \end{align} In this identity, the two alphabets $\set{y_i}$ and $\set{z_j}$ may be finite or infinite. \item Given partitions \(\lambda\) and \(\mu\), the skew Jack functions \(\jack{\lambda/\mu}{t}\) and \(\djack{\lambda/\mu}{t}\) are defined to be the unique symmetric functions satisfying \begin{equation} \cjprod{\jack{\lambda/\mu}{t},\djack{\nu}{t}}{t}=\cjprod{\jack{\lambda}{t},\djack{\mu}{t}\djack{\nu}{t}}{t} \qquad\text{and}\qquad \cjprod{\djack{\lambda/\mu}{t},\jack{\nu}{t}}{t}=\cjprod{\djack{\lambda}{t},\jack{\mu}{t}\jack{\nu}{t}}{t} \end{equation} for all partitions \(\nu\). Let us write \(\mu\subseteq\lambda\) if the Young diagram of \(\mu\) is contained in that of \(\lambda\). Then, \(\jack{\lambda/\mu}{t}= \djack{\lambda/\mu}{t}=0\) unless \(\mu\subseteq\lambda\). Finally, the ordinary and dual skew Jack functions are proportional: \begin{align}\label{skewdjack} \djack{\lambda/\mu}{t}=\frac{\nc{\lambda}{t}}{\nc{\mu}{t}}\jack{\lambda/\mu}{t}. \end{align} \item Consider an alphabet \(z=(z_1,z_2,\dots)\), partitioned into two subsets \(x=(x_1,x_2,\dots)\) and \(y=(y_1,y_2,\dots)\). Any symmetric function in \(z\) may obviously be decomposed into symmetric functions in \(x\) and \(y\). For Jack functions, this decomposition is \begin{align} \label{unionjack} \fjack{\lambda}{t}{z}=\fjack{\lambda}{t}{x \cup y}=\sum_\nu \fjack{\nu}{t}{x}\fjack{\lambda/\nu}{t}{y}, \quad \fdjack{\lambda}{t}{z}=\fdjack{\lambda}{t}{x \cup y}=\sum_\nu \fdjack{\nu}{t}{x}\fdjack{\lambda/\nu}{t}{y}. \end{align} Both sums may clearly be restricted to partitions satisfying $\nu \subseteq \lambda$. \item The Jack polynomials $\fjack{\lambda}{t}{z_1,\dots,z_n}$ are orthogonal with respect to the inner product \begin{align} \label{} \jprod{f,g}{n}{t}=\int_{\cyc{n;t}} \fintker{n}{t}{x} f(x)\overline{g(x)}\,\frac{\dd x_1 \cdots \dd x_n}{x_1 \cdots x_n}, \end{align} where \(\cyc{n;t}\) is the cycle normalised in \eqref{}, $\overline{g(x_1,x_2,\dots)}=g(x_1^{-1},x_2^{-1},\dots)$ and \begin{align} \fintker{n}{t}{x}=\prod_{1\le i\neq j\le n} \brac*{1-\frac{x_i}{x_j}}^{1/t} \end{align} is called the integrating kernel. With respect to this integral inner product, the Jack polynomials satisfy \begin{align} \label{} \jprod{\fjack{\lambda}{t}{x},\fdjack{\mu}{t}{x}}{n}{t}=\delta_{\lambda,\mu} \nc{\lambda}{t}(n), \quad \nc{\lambda}{t}(n) = \prod_{s\in\lambda}\frac{n+a^\prime(s)t-l^\prime(s)}{n+(a^\prime(s)+1)t-l^\prime(s)-1}, \end{align} where \(a^\prime(s)\) and \(l^\prime(s)\) denote the arm and leg colengths, respectively, of the box \(s\) in the Young diagram of \(\lambda\). \end{enumerate} \section{Explicit evaluation of $\Wvoa{3}$ singular vectors} \label{} With this symmetric function technology under our belts, we now turn to the computation of singular vectors in Fock spaces, the idea being to realise them as images of \hwvs{} under a $\Wvoa{3}$-module homomorphism (screening operator). For definiteness, we shall choose the screening operator $\scrs{+}{r_1,r_2}$ defined in \eqref{} that was constructed from $r_1$ copies of $\vo{\alpha_+ \alpha_1}$ and $r_2$ copies of $\vo{\alpha_+ \alpha_2}$. The computation for $\scrs{-}{r_1,r_2}$ is exactly the same and will be omitted. By \eqref{}, $\scrs{+}{r_1,r_2}$ has a well defined action on the Fock space $\Fock{\eta}$, where $\eta = \svwts{r_1-r_2,s_1}{r_2,s_2}$, sending it into $\Fock{\theta}$, where $\theta = \svwts{-r_1,s_1}{r_1-r_2,s_2}$. We can now explicitly evaluate the action of the screening operator $\scrs{+}{r_1,r_2}$ on the \hwv{} $\ket{\eta} \in \Fock{\eta}$. Using \eqref{}, \eqref{} and \eqref{}, this action is \begin{align}\label{integral} \scrs{+}{r_1,r_2} \ket{\eta} &= \int_{\Delta} \vop{\alpha_+\srt{1}}{z^1_1} \cdots \vop{\alpha_+\srt{1}}{z^1_{r_1}} \vop{\alpha_+ \srt{2}}{z^2_1} \cdots \vop{\alpha_+ \srt{2}}{z^2_{r_2}} \ket{\eta} \diff \nonumber \\ &= \int_{\Delta} \dprod{k=1}{2}{1\le i\neq j\le r_k}{} \brac[\Big]{1 - \frac{z^k_i}{z^k_j}}^{\alpha_+^2} \cdot \dprod{i=1}{r_1}{j=1}{r_2} \brac[\Big]{1 - \frac{z^2_j}{z^1_i}}^{-\alpha_+^2} \cdot \dprod{k=1}{2}{i=1}{r_k} \brac[\big]{z^k_i}^{\alpha_+ \kil{\srt{k}}{\eta} + \alpha_+^2 (r_k-1) + 1} \nonumber\\ &\qquad \cdot \prod_{i=1}^{r_1} \brac[\big]{z^1_i}^{-\alpha_+^2 r_2} \cdot \dprod{k=1}{2}{m\ge1}{} \exp \brac*{\frac{\alpha_+ a^k_{-m}}{m} \sum_{i=1}^{r_k} \brac[\big]{z^k_i}^m} \cdot \ket{\theta} \diffz \nonumber \\ &= \int_{\Delta} \dprod{k=1}{2}{1\le i\neq j\le r_k}{} \brac[\Big]{1 - \frac{z^k_i}{z^k_j}}^{\alpha_+^2} \cdot \dprod{i=1}{r_1}{j=1}{r_2} \brac[\Big]{1 - \frac{z^2_j}{z^1_i}}^{-\alpha_+^2} \nonumber\\ &\qquad \cdot \dprod{k=1}{2}{i=1}{r_k} \brac[\big]{z^k_i}^{s_k} \cdot \dprod{k=1}{2}{m\ge1}{} \exp \brac*{\frac{\alpha_+ a^k_{-m} \fpowsum{m}{z^k}}{m}} \cdot \ket{\theta} \diffz. \end{align} Here, the integrals are over the product cycle $\Delta = \cyc{r_1;\alpha_+^{-2}} \times \cyc{r_2;\alpha_+^{-2}}$, see \cref{sec:W3}. To proceed, we note that the tensor product $\sym \otimes_{\fld{C}} \sym$ is isomorphic to $\UEA{\heis_{-}} = \fld{C}[a^k_{-m} \st k=1,2,\ m \in \fld{Z}_{>0}]$ as an algebra, by \eqref{}. Concretely, let $y^1_i$ and $y^2_i$ denote the variables for the two factors of $\sym\otimes_{\fld{C}}\sym$ and consider the isomorphism \begin{align}\label{homom} \rho_+ \colon \Lambda \otimes_{\fld{C}} \Lambda \longrightarrow \UEA{\heis_-}, \quad \powsum{m}(y^k) \longmapsto \frac{1}{\alpha_+} a^k_{-m}, \quad k=1,2,\ m \in \fld{Z}_{>0}. \end{align} Then, we may write \begin{align} \prod_{m\ge1} \exp \brac*{\frac{\alpha_+ a^k_{-m} \fpowsum{m}{z^k}}{m}} = \rho_+ \brac*{\prod_{m\ge1} \exp \brac*{\alpha_+^2 \frac{\powsum{m}(y^k) \powsum{m}(z^k)}{m}}} = \rho_+ \brac*{\prod_{i\ge1} \prod_{j=1}^{r_k} (1-y^k_i z^k_j)^{-\alpha_+^2}}, \end{align} recognising the Cauchy kernel \eqref{cauchykernel} with parameter $t = \alpha_+^{-2}$. For $k=1$, we expand this Cauchy kernel in terms of Jacks and their duals as in \eqref{cauchykernel}: \begin{equation} \label{} \prod_{m\ge1} \exp \brac*{\frac{\alpha_+ a^1_{-m} \fpowsum{m}{z^1}}{m}} = \rho_+ \brac*{\sum_\lambda \fjack{\lambda}{t}{y^1} \fdjack{\lambda}{t}{z^1}}. \end{equation} For $k=2$, we first combine the Cauchy kernel with that appearing in the second factor of the integrand of \eqref{integral}: \begin{align} \label{} \rho_+ \brac*{\prod_{i\ge1} \prod_{j=1}^{r_2} (1-y^2_i z^2_j)^{-\alpha_+^2}} \prod_{i=1}^{r_1} \prod_{j=1}^{r_2} \brac*{1 - \brac{z^1_i}^{-1} z^2_j}^{-\alpha_+^2} = \rho_+ \brac*{\sum_\mu \fjack{\mu}{t}{y^2 \cup (z^1)^{-1}} \fdjack{\mu}{t}{z^2}}. \end{align} Here, we have noted that the product is a Cauchy kernel in the alphabets $\set{y^2_i} \cup \set{(z^1_\ell)^{-1}}$ and $\set{z^2_j}$. This may be further simplified using skew-Jacks as in \eqref{unionjack}: \begin{equation} \fjack{\mu}{t}{y^2 \cup (z^1)^{-1}} = \sum_{\nu} \overline{\fjack{\nu}{t}{z^1}} \fjack{\mu / \nu}{t}{y^2}. \end{equation} We recall that the skew-Jack $\jack{\mu / \nu}{t}$ is $0$ unless $\nu \subseteq \mu$. Next, note that we also have the integrating kernels \begin{equation} \dprod{k=1}{2}{1\le i\neq j\le r_k}{} \brac[\Big]{1 - \frac{z^k_i}{z^k_j}}^{\alpha_+^2} = \fintker{r_1}{t}{z^1} \fintker{r_2}{t}{z^2} \end{equation} of the symmetric polynomial inner product \eqref{}. Finally, the product $\dprod{k=1}{2}{i=1}{r_k} \brac[\big]{z^k_i}^{s_k}$ is a product of rectangular Jack polynomials. However, here we have to be careful with the signs of the $s_k$. Indeed, \eqref{} and \eqref{} show that the conformal weights of the \hwvs{} $\ket{\eta}$ and $\ket{\theta}$ differ by \begin{equation} \label{} h_{\eta} - h_{\theta} = -r_1 s_1 - r_2 s_2. \end{equation} This must be non-negative if the screening operator $\scrs{+}{r_1,r_2}$ is to map $\ket{\eta}$ to a singular descendant of $\ket{\theta}$. We shall therefore assume from here on that $s_1, s_2 \in \fld{Z}_{\le 0}$. Thus, \begin{equation} \dprod{k=1}{2}{i=1}{r_k} \brac[\big]{z^k_i}^{s_k} = \overline{\fjack{[-s_1^{r_1}]}{t}{z^1}} \, \overline{\fjack{[-s_2^{r_2}]}{t}{z^2}}, \end{equation} using \eqref{}. Putting all this back in \eqref{integral}, the integrand factorises and we get \begin{align}\label{} \scrs{+}{r_1,r_2} \ket{\eta} &= \sum_{\lambda, \mu, \nu} \rho_+ \brac*{\fjack{\lambda}{t}{y^1}} \int_{\cyc{r_1;t}} \fintker{r_1}{t}{z^1} \overline{\fjack{[-s_1^{r_1}]}{t}{z^1}} \, \overline{\fjack{\nu}{t}{z^1}} \fdjack{\lambda}{t}{z^1} \prod_{i=1}^{r_1} \frac{\dd z^1_i}{z^1_i} \notag \\ &\mspace{80mu} \cdot \rho_+ \brac*{\fjack{\mu / \nu}{t}{y^2}} \int_{\cyc{r_2;t}} \fintker{r_2}{t}{z^2} \overline{\fjack{[-s_2^{r_2}]}{t}{z^2}} \fdjack{\mu}{t}{z^2} \prod_{i=1}^{r_2} \frac{\dd z^2_i}{z^2_i} \cdot \ket{\theta} \notag \\ &= \sum_{\lambda, \mu, \nu} \jprod{\djack{\lambda}{t}, \jack{\nu + [-s_1^{r_1}]}{t}}{r_1}{t} \jprod{\djack{\mu}{t}, \jack{[-s_2^{r_2}]}{t}}{r_2}{t} \rho_+ \brac*{\fjack{\lambda}{t}{y^1} \fjack{\mu / \nu}{t}{y^2}} \ket{\theta} \notag \\ &= \sum_{\substack{\nu \subseteq [-s_2^{r_2}]\\\ell(\nu)\le r_1}} \nc{\nu + [-s_1^{r_1}]}{t}(r_1) \nc{[-s_2^{r_2}]}{t}(r_2) \rho_+ \brac*{\fjack{\nu + [-s_1^{r_1}]}{t}{y^1} \fjack{[-s_2^{r_2}] / \nu}{t}{y^2}} \ket{\theta}, \end{align} by using \eqref{}, \eqref{} and \eqref{}. As the second factor is independent of $\nu$ (and non-zero), it may be absorbed into the normalisation of the singular vector{}. Our final result is therefore \begin{equation} \label{} \scrs{+}{r_1,r_2} \ket{\eta} = \sum_{\substack{\nu \subseteq [-s_2^{r_2}]\\\ell(\nu)\le r_1}} \nc{\nu + [-s_1^{r_1}]}{t}(r_1) \rho_+ \brac*{\fjack{\nu + [-s_1^{r_1}]}{t}{y^1} \fjack{[-s_2^{r_2}] / \nu}{t}{y^2}} \ket{\theta}. \end{equation} This form is now easily implemented in computer algebra packages. The right-hand side{} of \eqref{} is easily seen to be manifestly non-zero by noting that the total degree, with respect to the \(a^2_{-m}\), of the summand corresponding to the empty partition \(\nu=\sqbrac*{\ }\) is maximal and that all other summands have strictly lesser degrees. Since \(\nc{[-s_1^{r_1}]}{t}(r_1)\), \(\fjack{[-s_1^{r_1}]}{t}{y^1}\) and \(\fjack{[-s_2^{r_2}]}{t}{y^2}\) are all non-zero, this summand is therefore linearly independent of all others. The conclusion is that \eqref{} defines a singular vector{} for every $r_1, r_2 \in \fld{Z}_{\ge 0}$, $s_1, s_2 \in \fld{Z}_{\le0}$ and $t \in \fld{C} \setminus \fld{Q}_{\le 0}$. \section{Examples} \label{} We now illustrate the $\wvoa{}$ singular vector{} formula \eqref{} with three examples. \subsection{Example 1} \label{ex:1} For our first example, we compute a singular vector for the case when $t=\frac{4}{5}$, so that \begin{equation} \alpha_+ = \frac{\sqrt{5}}{2}, \quad \alpha_- = -\frac{2}{\sqrt{5}}, \quad \alpha_0 = \frac{1}{2 \sqrt{5}}, \quad c = \frac{4}{5}. \end{equation} This central charge corresponds to that of the $3$-state Potts model, described by the $\wvoa$ minimal model $\wminmod{4,5}$ (the parameters here are the numerator and denominator of $t$ in reduced form). Take $r_1=r_2=-s_1=-s_2=1$ for simplicity. Then, the map $\scrs{+}{1,1}$ sends $\Fock{\eta}$, where $\eta=\svwts{0,-1}{1,-1}$, into $\Fock{\theta}$, where $\theta=\svwts{-1,-1}{0,-1}$. We note that \begin{equation} \label{} h_{\eta} = \frac{13}{6}, \quad h_{\theta} = \frac{1}{6}, \quad w_{\eta} = \frac{187}{9 \sqrt{390}}, \quad w_{\theta} = -\frac{7}{9 \sqrt{390}}, \end{equation} by \eqref{} and \eqref{}. The conformal weight $h_{\theta}$ is not one of those associated with the $3$-state Potts model. Nevertheless, the Fock space $\Fock{\theta}$ has a singular vector at grade 2 in accordance with \eqref{}. \cref{} writes it in the form \begin{align} \scrs{+}{1,1} \ket{\eta} &= \sum_{\nu \subseteq [1]} \nc{\nu + [1]}{4/5}(1) \rho_+ \brac*{\fjack{\nu + [1]}{4/5}{y^1} \fjack{[1] / \nu}{4/5}{y^2}} \ket{\theta}. \end{align} There are only two partitions $\nu$ to consider. Using \eqref{}, \eqref{homom} and \textsc{SageMath} to write Jacks and skew-Jacks in terms of power sums, we have \begin{equation} \begin{aligned} \nu &= [0]: & \nc{[0]+[1]}{4/5}(1) &= \frac{5}{4}, & \rho_+ \brac*{\fjack{[0]+[1]}{4/5}{y^1}} &= \rho_+ \brac*{\fpowsum{[1]}{y^1}}=\frac{2}{\sqrt{5}}a^{1}_{-1}, \\ &&&& \rho_+ \brac*{\fjack{[1]/[0]}{4/5}{y^2}} &= \rho_+ \brac*{\fpowsum{[1]}{y^2}}=\frac{2}{\sqrt{5}}a^{2}_{-1}. \\ \nu &= [1]: & \nc{[1]+[1]}{4/5}(1) &= \frac{5}{4} \frac{9}{8}, & \rho_+ \brac*{\fjack{[1]+[1]}{4/5}{y^1}} &= \rho_+ \brac*{\frac{5}{9}\fpowsum{[1,1]}{y^1}+\frac{4}{9}\fpowsum{[2]}{y^1}} \\ &&&&&= \frac{4}{5} \frac{5}{9} a^1_{-1} a^1_{-1} + \frac{2}{\sqrt{5}} \frac{4}{9} a^1_{-2}, \\ &&&& \rho_+ \brac*{\fjack{[1]/[1]}{4/5}{y^1}} &= \rho_+ \brac*{\fpowsum{[0]}{y^1}}=1. \end{aligned} \end{equation} The singular vector is therefore explicitly identified as \begin{align} \label{} \scrs{+}{1,1} \ket{\eta}=\brac*{a^1_{-1}a^2_{-1}+\frac{5}{8}a^1_{-1}a^1_{-1}+\frac{\sqrt{5}}{4}a^1_{-2}}\ket{\theta}. \end{align} Consider the $\wvoa$ Verma module $\Verma{\vartheta}$ whose \hwv{} $\ket{\vartheta}$ has $L_0$- and $W_0$-eigenvalue $h_{\theta}$ and $w_{\theta}$, as given in \eqref{}. By direct calculation, $\Verma{\vartheta}$ has a singular vector $\ket{\chi}$, unique up to normalisation, at grade 2: \begin{align} \ket{\chi} = \brac*{\frac{390}{119} W_{-1}W_{-1} - \frac{\sqrt{390}}{17} W_{-2} + \frac{10\sqrt{390}}{119} L_{-1}W_{-1} + L_{-1}L_{-1}} \ket{\vartheta}. \end{align} The free field realisation $f \colon \wvoa \ira \hvoa{2}$ defined by \eqref{} induces a $\wvoa{}$-module homomorphism \begin{equation} f_{\vartheta} \colon \Verma{\vartheta} \longrightarrow \Fock{\theta}, \quad f_{\vartheta}(U \ket{\vartheta}) = f(U) \ket{\theta}. \end{equation} Here, $U$ is an arbitrary element of the $\wvoa$ mode algebra, this being the (unital) associative algebra generated by the $L_m$ and $W_n$ subject to \eqref{}. Explicit calculation now verifies that the image of the singular vector $\ket{\chi}$ under $f_{\vartheta}$ is, of course, that constructed in \eqref{}: \begin{align} f_{\vartheta} \brac*{\ket{\chi}} = \brac*{\frac{5}{4} a^1_{-1} a^2_{-1} + \frac{25}{32} a^1_{-1} a^1_{-1} + \frac{5 \sqrt{5}}{16} a^1_{-2}} \ket{\theta} =\frac{5}{4} \scrs{+}{1,1} \ket{\eta}. \end{align} \subsection{Example 2} For our second example, we compute a grade three singular vector for general central charges. Let $r_1=2$ and $r_2=-s_1=-s_2=1$, so that $\eta = \svwts{1,-1}{1,-1}$ and $\theta = \svwts{-2,-1}{1,-1}$. In order to evaluate the singular vector formula \eqref{}, we need to compute \begin{equation} \begin{gathered} \nc{[1,1]}{t}(2) = \frac{2}{t+1} \frac{1}{t}, \quad \nc{[2,1]}{t}(2) = \frac{2}{t+1} \frac{1}{t} \frac{t+2}{2t+1}, \\ \jack{[1,1]}{t} = \frac{1}{2} \powsum{[1,1]} - \frac{1}{2} \powsum{[2]}, \quad \jack{[1]/[0]}{t} = \powsum{[1]}, \\ \jack{[2,1]}{t} = \frac{1}{t+2} \powsum{[1,1,1]} + \frac{t-1}{t+2} \powsum{[2,1]} - \frac{t}{t+2} \powsum{[3]}, \quad \jack{[1]/[1]}{t} = 1, \end{gathered} \end{equation} again using \eqref{}, \eqref{homom} and \textsc{SageMath}. The singular vector is thus \begin{align} \scrs{+}{2,1} \ket{\eta} &= \sqbrac*{ \nc{[1,1]}{t}(2) \rho_+ \brac*{\fjack{[1,1]}{t}{y^1} \fjack{[1]/[0]}{t}{y^2}} + \nc{[2,1]}{t}(2) \rho_+ \brac*{\fjack{[2,1]}{t}{y^1} \fjack{[1]/[1]}{t}{y^2}} } \ket{\theta} \notag \\ &= \left[ \frac{1}{t(t+1)} \brac*{\frac{1}{\alpha_+^2} a^1_{-1} a^1_{-1} - \frac{1}{\alpha_+} a^1_{-2}} \frac{1}{\alpha_+} a^2_{-1} \right. \notag \\ &\qquad \left. + \frac{2}{t(t+1)(2t+1)} \brac*{\frac{1}{\alpha_+^3} a^1_{-1} a^1_{-1} a^1_{-1} + \frac{t-1}{\alpha_+^2} a^1_{-2} a^1_{-1} - \frac{t}{\alpha_+} a^1_{-3}} \right] \ket{\theta} \notag \\ &= \left[ \frac{2/\alpha_+}{(t+1)(2t+1)} a^1_{-1} a^1_{-1} a^1_{-1} + \frac{1/\alpha_+}{t+1} a^1_{-1} a^1_{-1} a^2_{-1} \right. \notag \\ &\qquad \left. + \frac{2(t-1)}{(t+1)(2t+1)} a^1_{-2} a^1_{-1} - \frac{1}{t+1} a^1_{-2} a^2_{-1} - \frac{1/\alpha_+}{(t+1)(2t+1)} a^1_{-3} \right] \ket{\theta}. \end{align} We note that the result is manifestly well defined and non-zero for all $\alpha_+$ such that $t = \alpha_+^{-2} \in \fld{C} \setminus \fld{Q}_{\le 0}$, as expected. This region includes all central charges less than $98$. \subsection{Example 3} Our final example concerns singular vectors{} for quite arbitrary central charges (including all $c < 98$). This time, we fix $\theta = 0$ and use \eqref{} to construct singular vectors{} in the Fock space $\Fock{0}$. First, we note that $\theta = \svwts{-r_1,s_1}{r_1-r_2,s_2} = 0$ may be solved for $r_1$ and $s_1$: \begin{equation} r_1 = -1 + (-s_1 + 1) t, \quad r_2 = -2 + (-s_1 - s_2 + 2) t. \end{equation} Since $r_1, r_2, s_1, s_2 \in \fld{Z}$, we will only find singular vectors{} when $t \in \fld{Q}_{>0}$. Writing $t = \frac{u}{v}$, where $u$ and $v$ are coprime integers, it follows that \begin{equation} r_1 = mu-1, \quad -s_1 = mv-1, \quad r_2 = nu-2, \quad -s_2 = (n-m)v-1, \end{equation} for some $m,n \in \fld{Z}$. Given that $r_1, r_2 \in \fld{Z}_{\ge 0}$ and $s_1, s_2 \in \fld{Z}_{<0}$, we conclude that $m$, $n$ and $n-m$ must be positive integers. We thereby obtain, for each fixed $t \in \fld{Q}_{>0}$, an infinite sequence of singular vectors{}, generically indexed by integers $n>m>0$, of the form $\scrs{+}{mu-1,nu-2} \ket{\svwts{(m-n)u+1,-mv+1}{nu-2,-(n-m)v+1}}$. Among these, the singular vector{} of lowest grade corresponds, assuming that $u>1$, to $(m,n) = (1,2)$. Moreover, the grade of $\scrs{+}{u-1,2(u-1)} \ket{\svwts{-(u-1),-(v-1)}{2(u-1),-(v-1)}}$ is $3(u-1)(v-1)$, by \eqref{}. It is not clear if these singular vectors{} of the Fock space $\Fock{0}$ correspond, in the sense of Example~\ref{ex:1} to singular vectors{} in the $\wvoa$ vacuum Verma module $\Verma{0}$ or not. However, there are five other Fock spaces $\Fock{\zeta}$ whose \hwvs{} $\ket{\zeta}$ have $h_{\zeta} = w_{\zeta} = 0$. This follows from the easily verified fact that both $h_{\zeta}$ and $w_{\zeta}$ are left invariant by the following shifted action of the Weyl group $\symgp{3}$: \begin{equation} \sigma \cdot \zeta = \sigma(\zeta - \alpha_0 \weyl) + \alpha_0 \weyl, \quad \sigma \in \symgp{3}. \end{equation} Each of these five other Fock spaces has an infinite sequence of singular vectors{} given by \eqref{} and it is interesting to ask whether these also correspond to singular vectors{} in $\Verma{0}$ or not. We shall not investigate this question here. We only note the following observation: $\Fock{2 \alpha_0 \weyl}$ has such a singular vector{} at grade $3$ and it corresponds to just one of the \emph{two} linearly independent grade $3$ singular vectors{} of $\Verma{0}$. Which one is obtained depends on the branch of the square root of $\beta$ chosen in \eqref{}. We conclude by remarking that the question of whether the Fock space singular vectors{} constructed here exhaust the singular vectors{} of $\Verma{0}$ is much easier to answer. They do not. We cannot obtain the two linearly independent singular vectors{} at grade $1$ using \eqref{} (for $c \neq 2$; when $c=2$, $W_0$ acts non-diagonalisably). Nor can we obtain, when $t = \frac{u}{v} \in \fld{Q}_{>0}$, the grade $(u-2)(v-2)$ singular vector{} whose image is non-zero in the universal $\wvoa$ vacuum module. This singular vector{} can be constructed formally using screening operators, but we do not know how to actually evaluate the integral in this case. What is needed is a certain $\SLA{sl}{3}$ analogue of the theory of Jack functions, something which does not appear to have yet been developed (see \cite{TarSel03,WarSel09,WarSL310} for work in this direction). We hope to return to this important point in the future. \section{Explicit evaluation of $\Wvoa{N}$ singular vectors} \label{sec:WN} In the previous section, we computed explicit formulae for $\Wvoa{3}$ singular vectors in Fock spaces. In this section we generalise the results of \cref{} and derive explicit formulae for $\Wvoa{N}$ singular vectors in Fock spaces. Continuing the pattern of ranks $1$ and $2$, the rank $N-1$ Heisenberg vertex operator algebra, with choice of energy-momentum tensor \eqref{}, has $2(N-1)$ screening operators, $\vop{\zeta}{w}$ for $\zeta = \alpha_\pm \srt{i}$, where $\alpha_\pm$ was defined in \eqref{} and the $\srt{i}$ are the simple roots of $\alg{sl}(N)$. The \Walg{N} vertex operator algebra $\Wvoa{N}$ is usually described as being generated by the Virasoro field $T(z)$ and $N-2$ Virasoro primary fields $W^3(z),\dots,W^{N}(z)$ of conformal weights $3,\dots,N$, respectively. Unfortunately, explicit formulae for these primaries, for example in terms of Heisenberg fields, rapidly increase in complexity as \(N\) increases and there are no known closed formulae for general \(N\). Fortunately our computations do not require explicit expressions for the $W^k(z)$, only the fact that they commute with the screening operators. We therefore turn to the definition of the \Walg{N} vertex operator algebra \cite{Luk88} in terms of a generating function called the quantum Miura transform. This constructs a different set of generators of $\Wvoa{N}$ that are not conformal primaries in general, but which are easily verified to commute with screening operators. We denote these new generating fields by $U^2(z)=T(z)$, $U^3(z),\dots,U^{N}(z)$ and their generating function by \begin{align} \label{} R_{N}(z)&=-\sum^{N}_{k=0}U_k(z)(\alpha_0\partial)^{N-k} =\normord{(\alpha_0\partial_z-\epsilon^1(z))\cdots(\alpha_0\partial_z-\epsilon^{N}(z))}, \end{align} where the \(\epsilon^i\) are the weights of the defining representation of \(\alg{sl}(N)\) so that $\epsilon^1 + \cdots + \epsilon^N = 0$ and \(\alpha^i=\epsilon^{i}-\epsilon^{i+1}\), for \(i=1,\dots,N-1\). With the \Walg{N} algebra now defined explicitly as the algebra generated by the \(U^i(z)\), \(i=2,\dots N\), we construct screening fields in a manner similar to $\Wvoa{3}$. As mentioned above, the vertex operators $\vop{\zeta}{w}$ with Heisenberg weights $\zeta=\alpha_{\pm}\alpha_1,\dots,\alpha_{\pm}\alpha_{N-1}$ are screening fields, because their operator product expansions with $R_N(z)$ are total derivatives: \begin{equation} \label{} R_{N}(z)\vop{\alpha_\pm\alpha_i}{w}\sim \partial_w\brac*{\frac{\normord{R^i_{N}(w)\vop{\alpha_\pm\alpha_i}{w}}}{z-w}}. \end{equation} Here, \(R^i_{N}(z)\) is defined as the product in \eqref{}, but without the factors involving \(\epsilon^i\) and \(\epsilon^{i+1}\). As in the rank $2$ case, the residues of the screening fields, when defined, commute with the \Walg{N} algebra, because their operator product expansions with the generating \(U^i\) fields are total derivatives, and therefore define module homomorphisms. Also as in the rank 2 case, one can compose screening fields and integrate them over suitable contours to construct yet more module homomorphisms. Note that it is sufficient to only compose screening fields whose weights are all rescalings of simple \(\alg{sl}(N)\) roots by either \(\alpha_+\) or \(\alpha_-\). This is because the two screening operators corresponding to the residues of $\vop{\alpha_+\alpha^i}{w}$ and $\vop{\alpha_-\alpha^j}{w}$ commute and can thus be considered independently. We shall therefore only present calculations involving the $\vop{\alpha_+\alpha^i}{w}$; those involving the $\vop{\alpha_-\alpha^j}{w}$ work in exactly the same way. We therefore compose $r_1$ copies of $\vop{\alpha_+\alpha^1}{z^1}$ with $r_{2}$ copies of $\vop{\alpha_+\alpha^{2}}{z^{2}}$ and so on, evaluating this composition on a Fock space of weight $\eta$, to obtain \begin{align} &\Bigl.\prod_{i=1}^{r_1} \vop{\alpha_+ \srt{1}}{z^1_i} \cdots \prod_{i=1}^{r_{N-1}} \vop{\alpha_+ \srt{N-1}}{z^{N-1}_i} \Bigr\rvert_{\Fock{\eta}}\nonumber\\ &\quad = \prod^{N-1}_{k=1} \prod_{1\le i<j\le r_k} (z^k_i-z^k_j)^{2\alpha_+^2} \cdot \prod^{N-2}_{k=1} \prod_{i=1}^{r_k} \prod_{j=1}^{r_{k+1}} (z^k_i-z^{k+1}_j)^{-\alpha_+^2} \cdot \prod_{k=1}^{N-1} \prod_{i=1}^{r_k} (z^k_i)^{\alpha_+\kil{\srt{k}}{\eta}}\nonumber\\ &\quad\quad \cdot \prod_{k=1}^{N-1} \gae{r_k \alpha_+ \srt{k}} \cdot \prod_{k=1}^{N-1} \prod_{m\ge1} \exp \brac*{\frac{\alpha_+ \scrt{k}_{-m}}{m} \sum_{i=1}^{r_k} (z_i^k)^m} \exp \brac*{-\frac{\alpha_+ \scrt{k}_{m}}{m} \sum_{i=1}^{r_k} (z_i^k)^{-m}} \nonumber\\ &\quad = \prod^{N-1}_{k=1} \prod_{1\le i\neq j\le r_k} \brac[\Big]{1-\frac{z_i^k}{z^k_j}}^{\alpha_+^2} \cdot \prod^{N-1}_{k=2} \prod_{i=1}^{r_k} \prod_{j=1}^{r_{k+1}} \brac[\Big]{1-\frac{z^k_j}{z^{k-1}_i}}^{-\alpha_+^2} \cdot \prod_{k=1}^{N-1} \prod_{i=1}^{r_k} (z^k_i)^{\alpha_+^2 (r_k-r_{k+1}-1) + \alpha_+ \kil{\srt{k}}{\eta}}\nonumber\\ &\quad\quad\cdot \prod_{k=1}^{N-1} \gae{r_k \alpha_+ \srt{k}} \cdot \prod_{k=1}^{N-1} \prod_{m\ge1} \exp \brac*{\frac{\alpha_+ \scrt{k}_{-m}}{m} \sum_{i=1}^{r_k} (z_i^k)^m} \exp \brac*{-\frac{\alpha_+ \scrt{k}_{m}}{m} \sum_{i=1}^{r_k} (z_i^k)^{-m}}, \end{align} where we define \(r_{N}=0\). In analogy to the reasoning presented for the \(\Wvoa{3}\) algebra in \cref{sec:W3}, one can construct a $\Wvoa{N}$-module homomorphism by choosing an appropriate contour. Integrating over the contours of Tsuchiya and Kanie \cite{TKScr86} is well defined whenever \begin{equation} \label{} \alpha_+^2\brac*{r_{k}-r_{k+1}-1}+\alpha_+\kil{\srt{k}}{\eta}\in \fld{Z},\ \text{for all}\ k=1,\dots,N-1. \end{equation} To parametrise the weights satisfying these constraints, we define \begin{align} \zeta_{\vec{u}, \vec{v}} =\sum^{N-1}_{i=1}\brac*{(1-u_i)\alpha_++(1-v_i)\alpha_-}\fwt{i}, \quad \vec{u}=(u_1,\dots,u_{N-1}),\ \vec{v}=(v_1,\dots,v_{N-1})\in \fld{Z}^{N-1}, \end{align} and define screening operators \begin{align} \scrs{+}{\vec{r}} &= \int_{\cyc{r_1;1/\alpha_+^2}} \cdots \int_{\cyc{r_{N-1};1/\alpha_+^2}} \prod_{i=1}^{r_1}\vop{\alpha_\pm\srt{1}}{z^1_i} \cdots \prod_{i=1}^{r_{N-1}} \vop{\alpha_+ \srt{N-1}}{z^{N-1}_i} \cdot \dprod{k=1}{N-1}{i=1}{r_k} \dd z^k_i, \quad \vec{r} \in \fld{Z}_{\ge 0}^{N-1}. \end{align} These, in turn, induce $\Wvoa{N}$-module homomorphisms \begin{align} \scrs{+}{\vec{r}} \colon \Fock{\eta^+_{\vec{r},\vec{s}}} &\to \Fock{\theta^+_{\vec{r},\vec{s}}}, \quad \vec{r} \in \fld{Z}_{\ge 0}^{N-1},\ \vec{s} \in \fld{Z}^{N-1}, \end{align} where $\eta^+_{\vec{r},\vec{s}} = \zeta_{(r_1-r_2,\dots,r_{N-2}-r_{N-1},r_{N-1}), \vec{s}}$ and $\theta^+_{\vec{r},\vec{s}} = \zeta_{(-r_1,r_1-r_2,\dots,r_{N-2}-r_{N-1}), \vec{s}}$. Similar screening operators $\scrs{-}{\vec{s}}$ are obtained by swapping the roles of $\alpha_+$ and $\alpha_-$, as well as $\vec{r}$ and $\vec{s}$, in this development. If we apply the screening operator $\scrs{+}{\vec{r}}$ to the $\hvoa{N-1}$ \hwv{} $\ket{\eta^+_{\vec{r},\vec{s}}}$, we get \begin{align}\label{integralWN} \scrs{+}{\vec{r}} \ket{\eta^+_{\vec{r},\vec{s}}} &= \int_{\cyc{r_1;1/\alpha_+^2}} \cdots \int_{\cyc{r_{N-1};1/\alpha_+^2}} \prod_{i=1}^{r_1} \vop{\alpha_+ \srt{1}}{z^1_i} \cdots \prod_{i=1}^{r_{N-1}} \vop{\alpha_+ \srt{N-1}}{z^{N-1}_i} \cdot \ket{\eta^+_{\vec{r},\vec{s}}} \dprod{k=1}{N-1}{i=1}{r_k} \dd z^k_i \nonumber \\ &= \int_{\cyc{r_1;1/\alpha_+^2}} \cdots \int_{\cyc{r_{N-1};1/\alpha_+^2}} \dprod{k=1}{N-1}{1\le i\neq j\le r_k}{} \brac[\Big]{1 - \frac{z^k_i}{z^k_j}}^{\alpha_+^2} \cdot \prod_{k=2}^{N-1} \dprod{i=1}{r_{k-1}}{j=1}{r_{k}} \brac[\Big]{1 - \frac{z^k_j}{z^{k-1}_i}}^{-\alpha_+^2} \nonumber\\ &\quad \cdot \dprod{k=1}{N-1}{i=1}{r_k} \brac[\big]{z^k_i}^{\alpha_+ \kil{\srt{k}}{\eta^+_{\vec{r},\vec{s}}} + \alpha_+^2 (r_k-1) + 1}\cdot \dprod{k=2}{N-1}{i=1}{r_k} \brac[\big]{z^1_i}^{-\alpha_+^2 r_k}\nonumber\\ &\quad\quad\cdot \dprod{k=1}{N-1}{m\ge1}{} \exp \brac*{\frac{\alpha_+ a^k_{-m}}{m} \sum_{i=1}^{r_k} \brac[\big]{z^k_i}^m} \cdot \ket{\theta^+_{\vec{r},\vec{s}}} \dprod{k=1}{N-1}{i=1}{r_k} \frac{\dd z^k_i}{z^k_i} \nonumber \\ &= \int_{\cyc{r_1;1/\alpha_+^2}} \cdots \int_{\cyc{r_{N-1};1/\alpha_+^2}} \dprod{k=1}{N-1}{1\le i\neq j\le r_k}{} \brac[\Big]{1 - \frac{z^k_i}{z^k_j}}^{\alpha_+^2} \cdot \prod_{k=1}^{N-2}\dprod{i=1}{r_{k-1}}{j=1}{r_{k}} \brac[\Big]{1 - \frac{z^k_j}{z^{k-1}_i}}^{-\alpha_+^2} \nonumber\\ &\qquad \cdot \dprod{k=1}{N-1}{i=1}{r_k} \brac[\big]{z^k_i}^{s_k} \cdot \dprod{k=1}{N-1}{m\ge1}{} \exp \brac*{\frac{\alpha_+ a^k_{-m} \fpowsum{m}{z^k}}{m}} \cdot \ket{\theta^+_{\vec{r},\vec{s}}} \dprod{k=1}{N-1}{i=1}{r_k} \frac{\dd z^k_i}{z^k_i}. \end{align} As in the \(\Wvoa{3}\) case, we can evaluate these integral formulae for singular vectors in terms of symmetric functions. Recall that the tensor product of $N-1$ copies of the ring of symmetric functions $\sym^{\otimes N-1}$ is isomorphic to \begin{equation} \UEA{\heis_{-}} = \fld{C}[a^k_{-m} \st k=1,\dots,N-1,\ m \in \fld{Z}_{>0}] \end{equation} as an algebra, by \eqref{}. Distinguishing the alphabets of the tensor factors by superscripts, so that the alphabet in the \(i\)-th tensor factor is denoted by \(y^i\), we define the following algebra isomorphism generalising that of \eqref{homom}: \begin{align} \rho_+ \colon \sym^{\otimes N-1} \rightarrow \UEA{\heis_-^{\otimes N-1}}, \quad \fpowsum{n}{y^k} \mapsto \frac{1}{\alpha_+} a^k_{-n}. \end{align} This isomorphism allows us to write \begin{align} \prod_{m\ge1} \exp \brac*{\frac{\alpha_+ a^k_{-m} \fpowsum{m}{z^k}}{m}} = \rho_+ \brac*{\prod_{m\ge1} \exp \brac*{\alpha_+^2 \frac{\powsum{m}(y^k) \powsum{m}(z^k)}{m}}} = \rho_+ \brac*{\prod_{i\ge1} \prod_{j=1}^{r_k} (1-y^k_i z^k_j)^{-\alpha_+^2}}, \end{align} for $k=1,\dots,N-1$. We now identify, with $t = \alpha_+^{-2}$, \begin{align} \dprod{k=1}{N-1}{1\le i\neq j\le r_k}{} \brac[\Big]{1 - \frac{z^k_i}{z^k_j}}^{\alpha_+^2}=\prod_{k=1}^{N-1}\fintker{r_k}{t}{z^k} \end{align} as the product of the integrating kernels for the variables $z_k$. For $k=1$, as in \eqref{}, we write \begin{equation} \prod_{m\ge1} \exp \brac*{\frac{\alpha_+ a^1_{-m} \fpowsum{m}{z^1}}{m}} = \rho_+ \brac*{\sum_{\mu_1} \fjack{\mu_1}{t}{y^1} \fdjack{\mu_1}{t}{z^1}}. \end{equation} For $k=2,\dots,N-1$, similar to \eqref{}, we have instead \begin{align} \prod_{m\ge1} \exp \brac*{\frac{\alpha_+ a^k_{-m} \fpowsum{m}{z^k}}{m}} \cdot \dprod{i=1}{r_{k-1}}{j=1}{r_{k}} \brac[\Big]{1 - \frac{z^{k}_j}{z^{k-1}_i}}^{-\alpha_+^2} &=\rho_+ \brac*{\sum_{\mu_k} \fjack{\mu_k}{t}{y^k \cup (z^{k-1})^{-1}}\fdjack{\mu_k}{t}{z^k}}\nonumber\\ &= \rho_+ \brac*{\sum_{\mu_k,\nu_k}\overline{\fjack{\nu_k}{t}{z^{k-1}}}\fjack{\mu_k / \nu_k}{t}{y^k} \fdjack{\mu_k}{t}{z^k}}. \end{align} Putting everything together, we have \begin{align} \scrs{+}{\vec{r}} \ket{\eta^+_{\vec{r},\vec{s}}} &= \sum_{\substack{\mu_1,\mu_2,\dots, \mu_{N-1} \\ \hphantom{\mu_1,} \nu_2,\dots,\nu_{N-1}}} \rho_+ \brac*{\fjack{\mu_1}{t}{y^1}} \int_{\cyc{r_1;t}} \fintker{r_1}{t}{z^1} \overline{\fjack{[-s_1^{r_1}]}{t}{z^1}} \, \overline{\fjack{\nu_2}{t}{z^1}} \fdjack{\mu_1}{t}{z^1} \prod_{i=1}^{r_1} \frac{\dd z^1_i}{z^1_i} \notag \\ &\quad\cdot \rho_+ \brac*{\fjack{\mu_2 / \nu_2}{t}{y^2}} \int_{\cyc{r_2;t}} \fintker{r_2}{t}{z^2} \overline{\fjack{[-s_2^{r_2}]}{t}{z^2}} \, \overline{\fjack{\nu_3}{t}{z^2}} \fdjack{\mu_2}{t}{z^2} \prod_{i=1}^{r_2} \frac{\dd z^2_i}{z^2_i} \notag \\ &\mspace{250mu}\vdots\notag\\ &\quad \cdot \rho_+ \brac*{\fjack{\mu_{N-2} / \nu_{N-2}}{t}{y^{N-2}}} \int_{\cyc{r_{N-2};t}} \fintker{r_{N-2}}{t}{z^{N-2}} \overline{\fjack{[-s_{N-2}^{r_{N-2}}]}{t}{z^{N-2}}} \, \overline{\fjack{\nu_{N-1}}{t}{z^{N-2}}} \fdjack{\mu_{N-2}}{t}{z^{N-2}} \prod_{i=1}^{r_{N-2}} \frac{\dd z^{N-2}_i}{z^{N-2}_i} \notag \\ &\quad \cdot \rho_+ \brac*{\fjack{\mu_{N-1} / \nu_{N-1}}{t}{y^{N-1}}} \int_{\cyc{r_{N-1};t}} \fintker{r_{N-1}}{t}{z^{N-1}} \overline{\fjack{[-s_{N-1}^{r_{N-1}}]}{t}{z^{N-1}}} \fdjack{\mu_{N-1}}{t}{z^{N-1}} \prod_{i=1}^{r_{N-1}} \frac{\dd z^{N-1}_i}{z^{N-1}_i} \cdot \ket{\theta^+_{\vec{r},\vec{s}}} \notag \\ &=\sum_{\substack{\mu_1,\mu_2,\dots, \mu_{N-1} \\ \hphantom{\mu_1,} \nu_2,\dots,\nu_{N-1}}} \prod_{k=1}^{N-2} \jprod{\djack{\mu_k}{t}, \jack{\nu_{k+1}+[-s_k^{r_k}]}{t}}{r_k}{t} \cdot \jprod{\djack{\mu_{N-1}}{t}, \jack{[-s_{N-1}^{r_{N-1}}]}{t}}{r_{N-1}}{t} \rho_+ \brac*{\fjack{\mu_1}{t}{y^1}} \prod^{N-1}_{k=2} \rho_+ \brac*{\fjack{\mu_k / \nu_k}{t}{y^k}} \cdot \ket{\theta^+_{\vec{r},\vec{s}}} \notag \\ &= \sum_{\nu_2,\dots,\nu_{N-1}} \brac*{\prod_{k=1}^{N-2} \nc{\nu_{k+1} + [-s_k^{r_k}]}{t}(r_k)\cdot \nc{[-s_{N-1}^{r_{N-1}}]}{t}(r_{N-1})} \notag \\ &\quad \cdot \rho_+ \brac*{\fjack{\nu_2+[-s_1^{r_1}]}{t}{y^1}} \prod^{N-2}_{k=2} \rho_+ \brac*{\fjack{\brac*{\nu_{k+1}+[-s_k^{r_k}]} / \nu_k}{t}{y^k}} \cdot \rho_+ \brac*{\fjack{[-s_{N-1}^{r_{N-1}}] / \nu_{N-1}}{t}{y^{N-1}}} \ket{\theta^+_{\vec{r},\vec{s}}}. \end{align} As before, the factor $\nc{[-s_{N-1}^{r_{N-1}}]}{t}(r_{N-1})$ does not depend on the summation indices $\nu_2,\dots,\nu_{N-1}$, appears in every summand, and is non-zero, so it can be suppressed. Moreover, the skew-Jack polynomials vanish unless the summation indices $\nu_2,\dots,\nu_{N-1}$ satisfy the relations \begin{align} \nu_{k} \subseteq \nu_{k+1} + [-s_{k}^{r_{k}}],\quad k=2,\dots,N-2, \quad \nu_{N-1} \subseteq [-s_{N-1}^{r_{N-1}}]. \end{align} Thus, the singular vector \(\scrs{+}{\vec{r}} \ket{\eta^+_{\vec{r},\vec{s}}} \in \Fock{\theta^+_{\vec{r},\vec{s}}}\) is proportional to \begin{multline}\label{} \sum_{\nu_2,\dots,\nu_{N-1}} \brac*{\prod_{k=1}^{N-2} \nc{\nu_{k+1} + [-s_k^{r_k}]}{t}(r_k)} \rho_+ \brac*{\fjack{\nu_2+[-s_1^{r_1}]}{t}{y^1}} \\ \cdot \prod^{N-2}_{k=2} \rho_+ \brac*{\fjack{\brac*{\nu_{k+1}+[-s_k^{r_k}]} / \nu_k}{t}{y^k}} \cdot \rho_+ \brac*{\fjack{[-s_{N-1}^{r_{N-1}}] / \nu_{N-1}}{t}{y^{N-1}}} \ket{\theta^+_{\vec{r},\vec{s}}}. \end{multline} This is our final formula for $\Wvoa{n}$ singular vectors generalising the $n=3$ case in \eqref{}. As before, considering the summand with $\nu_2 = \cdots = \nu_{N-1} = \sqbrac*{\ }$ shows that the right-hand side{} is non-zero for every \(\vec{r}\in\fld{Z}_{\ge0}^{N-1}\), \(\vec{s}\in\fld{Z}_{\le0}^{N-1}\) and \(t\in\fld{C}\setminus \fld{Q}_{\le0}\). This singular vector formula also has the nice property of being comparatively easy to evaluate using computer algebra packages such as \textsc{SageMath}. \section*{Acknowledgements} We thank Peter Forrester and Ole Warnaar for discussions relating to the results presented here. DR's research is supported by the Australian Research Council Discovery Project DP160101520 and the Australian Research Council Centre of Excellence for Mathematical and Statistical Frontiers CE140100049. SW's research is supported by the Australian Research Council Discovery Early Career Researcher Award DE140101825 and the Australian Research Council Discovery Project DP160101520. \flushleft \singlespacing

\section{Critical configurations as energy vector maximizers}\label{} In this section, we prove that we never reach a stable configuration from a critical one by reverse-firing unconstrainedly any non-empty multi-subsets of vertices (Theorem \ref{T:script}). This is an analogue of Aval \emph{et al.}'s result~\cite{ADDL16} in the case of digraphs with a global sink. Then, by associating each configuration with an energy vector, we will show that among stable configurations in a same equivalence class, the critical configuration is the unique one with the greatest energy vector by the containment order (Theorem \ref{T:max2}). An affirmative answer for a question in \cite{PP16} about the maximum weight of critical configurations is also given using $G$-strongly positive scripts. Before presenting these results, we first introduce the concepts of $G$-positive scripts and $G$-strongly positive scripts. \medskip \begin{definition}[$G$-positive and $G$-strongly positive scripts]\mbox{} Let $\sigma$ be an $n$-script. We say that: \begin{enumerate}[label=\roman*)] \item $\sigma$ is \emph{$G$-positive} if $\sigma\Delta \succ \pmb 0$; \item $\sigma$ is \emph{$G$-strongly positive} if it is $G$-positive and $supp(\sigma\Delta)$ intersects each source component of $G$. \end{enumerate} \end{definition} \begin{remarks}\mbox{} \begin{itemize} \item[-] In case that $G\setminus \{n+1\}$ is strongly connected, the concepts of $G$-positiveness and $G$-strong positiveness coincide. \item[-] A $G$-strongly positive script is $G$-positive but the converse is not true. For instance, consider the graph in Figure \ref{F:critdigraph}. The scripts $(1,2,4)$ and $(0,0,1)$ both are $G$-positive but the former is $G$-strongly positive while the latter is not. \item[-] $G$-positive and $G$-strongly positive scripts always exist. Indeed, since the entries of $\Delta^{-1}$ are positive rational numbers, one can find a positive integer vector $u$ such that $u \Delta^{-1}$ is also a positive integer vector. Then $\sigma = u \Delta^{-1}$ is a $G$-strongly positive script. \item[-] Let $\sigma$ be a $G$-strongly positive script. The support of $\sigma\Delta$ may not intersect all strongly connected components of $G$. For example, consider the graph in Figure \ref{F:critdigraph}. The script $\sigma = (1, 2, 3)$ is $G$-strongly positive and $(\sigma \Delta)(v_3) = 0$. \end{itemize} \end{remarks} \begin{figure}[hbt] \centering \includegraphics[scale=0.8]{critdigraph2.pdf} \caption{A graph with global sink $s$ \label{F:critdigraph}} \end{figure} \medskip The following lemma is a crucial result which shows that for stable configurations, legal firing consumes less time than reverse firing. \begin{lemma}\label{L:prec} Let $\sigma=(\sigma_1,\dots,\sigma_n)$ be an $n$-script and let $\pmb a$ be a stable configuration of $\CFG(G)$ such that $\pmb a +\sigma \Delta$ is non-negative. Let $\tau$ be the firing script in the stabilizing process of $\pmb a+\sigma\Delta$. Then $\tau\preceq \sigma$. \end{lemma} \begin{proof} Let $\pmb b=\pmb a+\sigma\Delta$. Since $\pmb a$ is stable, we have, for all $i=1,2\dots,n$: \begin{equation} 0\leq a_i=b_i-\sigma_i\Delta_{ii}+\sum_{j \neq i}\sigma_j\Delta_{ji}< \Delta_{ii}\,. \label{E:1} \end{equation} Let $s=(s_1,s_2,\dots,s_k)$ be a legal firing sequence in the stabilizing process of $\pmb b$. Note that some of $s_i$ could appear more than once. Denote by $s^{\leq t}=(s_1, s_2,\dots,s_t)$ the legal firing sequence which is the prefix of $s$ of length $t$ and $\tau^{\leq t}$ its associated firing script, {\em i.e.} $\tau_j^{\le t}$ is the number of firings of the vertex $j$ until time $t$. Assume to the contrary that $\tau \npreceq \sigma$. Let $t_0$ be the first time in the process of firing $s$ that the number of occurrences of the vertex $i = s_{t_0}$ exceeds $\sigma_{s_{t_0}}$. We have: \begin{equation*} \begin{cases} \tau^{\leq t_0}_j=\sigma_j+1 &\text{ if } j=i,\\ \tau^{\leq t_0}_j\leq \sigma_j &\text{ if } j\ne i. \end{cases} \end{equation*} Combining this with the fact that $\Delta_{ij} \le 0$ for all $j \neq i$ and (\ref{E:1}), we have the following evaluation for the number of chips at vertex $i$ after the firing of $s^{\le t_0}$: \begin{align*} &b_i-\tau^{\leq t_0}_t\Delta_{ii}-\sum_{j\ne i} \tau^{\leq t_0}_j \Delta_{ji}\leq b_i-(\sigma_i+1)\Delta_{ii}-\sum_{j\ne i} \sigma_j\Delta_{ji} < 0\,, \end{align*} which contradicts the legality of $s$. \end{proof} Note that the proof of Lemma \ref{L:prec} does not require $\pmb a+\sigma\Delta$ to be non-negative. We keep this hypothesis, however, for ease of exposition and in fact, it does not affect our other proofs. The next lemma states that reverse-firing a $G$-positive script followed by the stabilization does not decrease the weight of a non-negative configuration. \begin{lemma}\label{L:gre} Let $\sigma$ be a $G$-positive script and let $\pmb a,\pmb b$ be stable configurations such that $\pmb b=(\pmb a+\sigma\Delta)^\circ$. Then $w(\pmb b)\geq w(\pmb a)$. \end{lemma} \begin{proof} Let $\tau$ be the firing script of the stabilization of $\pmb a+\sigma\Delta$. We have $\pmb b=\pmb a+\sigma\Delta-\tau\Delta$, and: \begin{align*} w(\pmb b)=\pmb b\cdot \pmb 1&=\pmb a \cdot \pmb 1+ (\sigma-\tau)\Delta \cdot \pmb 1\\ & =w(\pmb a)+ \sum_{i: (i,n+1)\in E} (\sigma_i-\tau_i) e_{i,n+1}. \end{align*} By Lemma \ref{L:prec}, $\sigma_{i}\geq \tau_{i}$ for all $i$. Hence, $w(\pmb b)\geq w(\pmb a)$. \end{proof} The next proposition shows the recurring nature of critical configurations. This result was mentioned in \cite[Theorem 2.27]{PPW11}, but we have not found a proof in the literature. We restate it here with a complete proof. \begin{proposition}\label{P:cri} Let $\pmb a$ be a stable configuration and let $\sigma$ be a $G$-strongly positive script. Then $\pmb a$ is critical if and only if $(\pmb a+\sigma\Delta)^\circ=\pmb a$. Moreover, if $\pmb a$ is critical then the firing script in the stabilizing process of $\pmb a+\sigma\Delta$ is $\sigma$. \end{proposition} \begin{proof} We first prove that $(\pmb c_{max}+\sigma\Delta)^\circ=\pmb c_{max}$. Since $\sigma\Delta \succeq 0$, $(\pmb c_{max}+\sigma\Delta)^\circ$ is stable, hence $(\pmb c_{max}+\sigma\Delta)^\circ\preceq \pmb c_{max}$. On the other hand, by Lemma \ref{L:gre}, $w(\pmb c_{max}+\sigma\Delta)^\circ\geq w(\pmb c_{max})$. Hence, $(\pmb c_{max}+\sigma\Delta)^\circ= \pmb c_{max}$. Now let $\pmb a$ be an arbitrary critical configuration. By Lemma~\ref{}, there exists $\pmb c\succeq 0$ such that $\pmb a=(\pmb c_{max}+\pmb c)^\circ$. Hence: \begin{align*} (\pmb a+\sigma\Delta)^\circ &=((\pmb c_{max}+\pmb c)^\circ+\sigma\Delta)^\circ\\ &=(\pmb c_{max}+\pmb c+\sigma\Delta)^\circ \\ &=((\pmb c_{max}+\sigma\Delta)^\circ+\pmb c)^\circ \\ &=(\pmb c_{max}+\pmb c)^\circ=\pmb a. \end{align*} Conversely, suppose that $(\pmb a+\sigma\Delta)^\circ=\pmb a$. Observe that if $u$, $v$ are vertices of $G$ and there is a directed path from $u$ to $v$, then for all $p > 0$, by putting sufficiently many chips in $u$ and repeatedly (and legally) firing the vertices along that path, one can increase the number of chips in $v$ by at least $p$ (for instance, if $u = v_0, v_1, \cdots, v_k = v$ is a directed path from $u$ to $v$, then $p \prod_{i = 0}^{k - 1} \mbox{d}^+(v_i)$ chips in $u$ are enough). Thus, for all $p > 0$, starting with sufficiently many chips in one vertex of each source component of $G$, one can redistribute chips in $G$ such that every non-sink vertex has at least $p$ chips. Since $\sigma$ is $G$-strongly positive, the support of $\sigma\Delta$ on each source component is non-empty. Therefore, for a sufficiently large number $m \in \mathbb N$, there exists a legal firing sequence leading from $\pmb a + m \sigma \Delta$ to a configuration $\pmb a' \succeq \pmb c_{max}$. In other words, there exist a positive integer $m$, a positive configuration $\pmb c$ and a legal firing sequence from $\pmb a + m \sigma \Delta$ with firing script $\tau$ such that: $$\pmb c_{max}+\pmb c=(\pmb a+m\sigma\Delta)-\tau\Delta\,,$$ whence \[ (\pmb c_{max} + \pmb c)^\circ = (\pmb a + m \sigma \Delta - \tau \Delta)^\circ = (\pmb a + m \sigma \Delta)^\circ\,. \] On the other hand, since $(\pmb a+\sigma\Delta)^\circ=\pmb a$ and $\sigma\Delta \succeq \pmb 0$, we have, by induction: \begin{align*} (\pmb a + m \sigma \Delta)^\circ &= (\pmb a +\sigma\Delta+ (m - 1) \sigma \Delta)^\circ=((\pmb a+\sigma\Delta)^\circ + (m - 1) \sigma \Delta)^\circ\\ &=(\pmb a + (m - 1) \sigma \Delta)^\circ = \dots = (\pmb a + \sigma \Delta)^\circ = \pmb a\,. \end{align*} So $(\pmb c_{max} + \pmb c)^\circ = \pmb a$, which means $\pmb a$ is critical. Finally, suppose $\pmb a$ is critical and let $\tau$ be the firing script of the stabilization of $\pmb a + \sigma \Delta$. We have $\pmb a + \sigma \Delta - \tau \Delta = \pmb a$, hence $(\tau - \sigma) \Delta = \pmb 0$. Since $\Delta$ is non-singular, $\tau = \sigma$. \end{proof} \begin{remark} The above proof shows that the ``only if'' part also holds when $\sigma$ is $G$-positive but not $G$-strongly positive. The ``if'' part, however, requires $G$-strong positiveness to be correct. Consider the graph in Figure \ref{F:critdigraph} and the scripts $\sigma_1=(0,0,1)$, $\sigma_2=(1,2,4)$. We can easily verify that $\sigma_1$ is $G$-positive but not $G$-strongly positive and $\sigma_2$ is $G$-strongly positive. Let $\pmb a=(1,1,1)$. Then $(\pmb a+\sigma_2\Delta)^\circ=(6,3,1)\ne \pmb a$, which implies that $\pmb a$ is not critical. Meanwhile, $(\pmb a+\sigma_1\Delta)^\circ=\pmb a$, which means that $\pmb a$ is ``recurrent'' when reverse-firing by the script $\sigma_1$. \end{remark} \medskip In \cite{PP16}, Perrot and Pham posed a question about the maximum weight of critical configurations in their equivalence classes and gave the answer for Eulerian digraphs. Here, we give the answer for general digraphs with a global sink. \begin{theorem} \label{T:max} Let $\pmb a$ be a critical configuration and $\pmb b$ be a stable configuration in the equivalence class of $\pmb a$. Then $w(\pmb a)\geq w(\pmb b)$. \end{theorem} \begin{proof} Let $\sigma$ be a $G$-strongly positive script. Consider the following sequence of stable configurations: \[ \pmb b_0 = \pmb b\,, \pmb b_{k + 1} = (\pmb b_k + \sigma \Delta)^\circ \mbox{ for all } k \ge 0\,. \] By Lemma \ref{L:gre}, $\{w(\pmb b_k)\}$ is non-decreasing. In particular, $w(\pmb b_k) \ge w(\pmb b)$ for all $k$. The stable configurations $(\pmb b_k)$ all belong to the equivalence class of $\pmb a$, which is finite, so there exist $k < \ell$ such that $\pmb b_k = \pmb b_{\ell}$, or: \[ \pmb b_k= \pmb b_\ell = (\pmb b_k + (\ell - k) \sigma \Delta)^\circ\,. \] Note that if $\sigma$ is $G$-strongly positive, then $m \sigma$ is also $G$-strongly positive for all positive integer $m$. So applying Proposition \ref{P:cri} with $\pmb b_k$ and $(\ell - k) \sigma$ yields that $\pmb b_k$ is critical. But since each equivalence class has only one critical configuration, $\pmb b_k = \pmb a$. Thus $w(\pmb a) \ge w(\pmb b)$. \end{proof} As mentioned at the end of section \ref{}, to decide whether a configuration is critical, one may use Dhar's burning algorithm~\cite{Dha90} for undirected graphs and Speer's script algorithm~\cite{Spe93} for strongly connected digraphs. Both algorithms are based on Proposition \ref{P:cri}: the idea is to use a minimal $G$-strongly positive script to test for the recurrence of a critical configuration. In Dhar's algorithm, the minimum script is always $(1,1,\dots,1)$. In Speer's algorithm, the script is not explicitly given, but one can construct it through a greedy iteration process: starting with the initial script $\sigma=(1,0,\dots,0)$, at each step, one chooses an index $i$ such that $(\sigma\Delta)_i$ is negative and increments $\sigma_i$. In 2011, Perkinson et al.~\cite{PPW11} developed the script algorithm for digraphs with a global sink by taking the initial script $\sigma=(1,1,\dots,1)$. This initial script indicates that every vertex has to fire at least once in a recurrent firing sequence. It is showed in \cite{Spe93} that the output of the script algorithm is uniquely determined and is in fact the unique minimum $G$-strongly positive script. We denote this script by $\sigma^M$. {\bf Example:} For the graph in Figure \ref{F:critdigraph}, the minimum $G$-strongly positive script is $\sigma^M=(1,2,3)$. \medskip The next theorem is an extension of Aval \emph{et al.}'s result~\cite{ADDL16} for undirected graphs to digraphs with a global sink. Essentially, it states that one cannot obtain a stable configuration from a critical one by reverse-firing a non-empty multi-subset of vertices. This result is instrumental in the proof of the energy maximizing characterization of critical configurations (Theorem \ref{T:max2}) and in the proof of the duality theorem (Theorem \ref{T:dua}). \begin{theorem}\label{T:script} Let $\sigma^M$ be the minimum $G$-strongly positive script and let $\pmb a$ be a stable configuration of $\CFG(G)$. The following statements are equivalent: \begin{enumerate}[label=\roman*)] \item $\pmb a$ is critical; \item $\pmb a+\tau\Delta$ is not stable for all $\tau \succ \pmb 0$; \item $\pmb a+\tau\Delta$ is not stable for all $\tau$ such that $\pmb 0 \prec \tau \prec \sigma^M$. \end{enumerate} \end{theorem} \begin{proof}\mbox{} $i)\Rightarrow ii)$: Let $\pmb a$ be critical and suppose that there exists a script $\tau \succ \pmb 0$ such that $\pmb b = \pmb a + \tau \Delta$ is stable. Let $m = \max\{\tau_i\} + 1$ and $\sigma = m \sigma^M$. Then $\sigma - \tau \succ \pmb 0$. Since $\sigma$ is $G$-strongly positive, we have, by Proposition \ref{P:cri}, $(\pmb a + \sigma \Delta)^\circ = \pmb a$ and that $\sigma$ is the stabilizing script of $\pmb a + \sigma \Delta$ which is also equal to $\pmb b + (\sigma - \tau) \Delta$. But by Lemma \ref{L:prec}, the stabilizing script of $\pmb b + (\sigma - \tau) \Delta$ cannot exceed $\sigma - \tau$, contradiction! $ii)\Rightarrow iii)$: Trivial. $iii)\Rightarrow i)$: Let $\pmb b = (\pmb a + \sigma^M \Delta)^\circ$. Suppose that $\pmb b \neq \pmb a$. Let $\tau$ be the stabilizing script of $\pmb a + \sigma^M \Delta$. By Lemma \ref{L:prec} and since $\pmb b \neq \pmb a$, $\tau \prec \sigma^M$. Now $\pmb b = \pmb a + (\sigma^M - \tau) \Delta$ is stable and $\pmb 0 \prec \sigma^M - \tau \prec \sigma^M$, contradiction! So $(\pmb a + \sigma^M \Delta)^\circ = \pmb a$, which means $\pmb a$ is critical. \end{proof} \medskip Theorem \ref{T:max} shows that a critical configuration has maximum weight among all of the stable configurations of its equivalence class. This property, however, does not characterize critical configurations, in the sense that there may be non-critical configurations with the same weight, see Figure \ref{F:sameweight} for an example. The notion of energy vector, defined hereafter, will provide a maximal characterization of critical configurations. \begin{definition}[$\CFG$ order and energy vector]\mbox{} Let $\pmb a$ and $\pmb b$ be two configurations. We say that $\pmb a \preceq_{\CFG} \pmb b$ if $\pmb a \Delta^{-1} \preceq \pmb b \Delta^{-1}$. The vector $\pmb a \Delta^{-1}$ is called the {\em energy vector} of the configuration $\pmb a$. \end{definition} \begin{figure}[hbt] \begin{center} \includegraphics[scale=0.7]{Fig-sameweight.pdf} \caption{Two stable configurations in the same equivalence class with the same weight of $5$. The left one is critical whereas the right one is not. \label{F:sameweight}} \end{center} \end{figure} The following properties are immediate. \begin{lemma}\label{L:evectorproperty}\mbox{} Let $\pmb a, \pmb b$ be two configurations of $\CFG(G)$. Then \begin{enumerate}[label=\roman*)] \item The binary relation $\preceq_{\CFG}$ is a partial order on $\mathbb Z^n$. \item If $\pmb a \sim \pmb b$ and $\pmb a \preceq_{\CFG} \pmb b$, then there exists a unique non-negative script $\sigma$ such that $\pmb b - \pmb a = \sigma \Delta\,.$ \end{enumerate} \end{lemma} \begin{remark} In the chip-firing game literature, $\pmb b$ is said to be {\em accessible} from $\pmb a$ if there is a legal firing sequence leading from $\pmb a$ to $\pmb b$. The accessibility relation induces a partial order on the set of non-negative configurations. It is easy to see that the accessibility order is a sub-order of $\preceq_{\CFG}$. \end{remark} \begin{comment} Now let $\pmb a, \pmb b\in \CFG(G)$. We say that $\pmb a\preceq_{\CFG} \pmb b$ if $\pmb a\Delta^{-1} \preceq \pmb b\Delta^{-1}$. The vector $pmb a \Delta^{-1}$ is called an \emph{energy vector} of $\pmb a$. It is easy to check that $\preceq_{\CFG}$ is a partial order on $\CFG(G)$. In case of $\pmb a\sim \pmb b$ and $\pmb a\preceq_{\CFG} \pmb b$, then $\pmb b-\pmb a$ is represented uniquely as a linear combination with non-negative coefficients of the rows of $\Delta$ . Note that in the literature ones consider the accessibility relation of a $\CFG$: $\pmb a$ is accessible from $\pmb b$ if $\pmb a$ is obtained from $\pmb b$ by a legal firing sequence. The set of configurations accessible from an initial configuration together with the accessibility relationship forms a partial order and more precisely a lattice~\cite{LP01}. It is easily seen that the order induced by the accessibility is in fact a suborder of the order $\preceq_{\CFG}$. We have the following \end{comment} \begin{theorem}\label{T:max2} In the set of stable configurations of each equivalence class, the critical configuration is the greatest with respect to the $\preceq_{\CFG}$ order. Equivalently, the critical configuration is of the greatest energy (w.r.t. the containment order). \end{theorem} \begin{proof} Let $\pmb a$ be an arbitrary stable configuration and $\pmb a^*$ be the critical configuration in its equivalence class. Assume that $\pmb a$ is not critical. We will prove that $\pmb a\prec_{\CFG} \pmb a^*$. Since $\pmb a$ is not critical, by Theorem \ref{T:script}, there exists a script $\tau_1\succ \pmb 0$ such that $\pmb a_1=\pmb a+\tau_1\Delta$ is stable. Then $\pmb a_1\sim \pmb a$ and $\pmb a_1\succ_{\CFG} \pmb a$ (since $\pmb a_1\Delta^{-1}-\pmb a\Delta^{-1}=\tau_1\succ \pmb 0$). Now, if $\pmb a_1$ is critical, then $\pmb a_1=\pmb a^*$, hence $\pmb a^*=\pmb a_1\succ_{\CFG} \pmb a$. If $\pmb a_1$ is not critical, by Theorem \ref{T:script}, there exists a script $\tau_2 \succ \pmb 0$ such that $\pmb a_2=\pmb a_1+\tau_2\Delta$ is stable, and so $\pmb a\prec_{\CFG} \pmb a_1\prec_{\CFG} \pmb a_2$. Repeating the above arguments, we obtain a sequence of stable configurations $\pmb a\prec_{\CFG} \pmb a_1\prec_{\CFG} \pmb a_2\prec_{\CFG}\cdots$ Since the number of stable configurations in each equivalence class is finite, this increasing sequence (w.r.t. the $\preceq_{\CFG}$ order) must stop at a critical configuration which in fact coincides with $\pmb a^*$. Thus, $\pmb a^*\succ_{\CFG} \pmb a$. \end{proof} \section{The duality between critical configurations and superstable configurations}\label{} In this section, we revisit the duality between critical configurations and superstable configurations (Theorem \ref{T:dua}). It has been shown, by Asadi and Backman~\cite[Theorem 3.15]{AB11} and Perkinson \emph{et al.}~\cite[Corollary 5.15]{PPW11} that for a digraph with a global sink, a configuration is superstable if and only if its complementary configuration is critical. The proof of Perkinson \emph{et al.} used advanced algebra techniques such as the coordinate ring induced by the Laplacian matrix and Gr\"obner bases, while Asadi and Backman's proof is combinatorial. We give another combinatorial proof which is different and independent from Asadi and Backman's one. \begin{definition}[Super-stable configuration] \mbox{} A non-negative configuration $\pmb a$ is \emph{superstable} if for all scripts $\sigma \succ \pmb 0$, $\pmb a - \sigma\Delta$ has a negative component. \end{definition} Super-stable configurations are also known as \emph{reduced divisors} or \emph{$z$-superstable configurations}~\cite{BS13, GK15}. It can be readily seen that if $\pmb a$ is superstable and $\pmb 0 \preceq \pmb b\preceq \pmb a$, then $\pmb b$ is also superstable. Moreover, superstable configurations are stable. Indeed, if $\pmb a$ is not stable and $\tau$ is the stablizing script of $\pmb a$, then $\tau \succ \pmb 0$ and $\pmb a-\tau\Delta = \pmb a^\circ$ is non-negative. \medskip We now state the duality theorem. \begin{theorem}[Duality Theorem]\label{T:dua} Let $\pmb a$ be a stable configuration. Then $\pmb a$ is critical if and only if $\pmb c_{max}-\pmb a$ is superstable. Consequently, the number of critical configurations is equal to the number of superstable configurations. \end{theorem} \begin{proof} Let $\pmb a$ be a critical configuration. Suppose that $\pmb c_{max} - \pmb a$ is not superstable. Then there exists a script $\sigma \succ \pmb 0$ such that $(\pmb c_{max} - \pmb a) - \sigma \Delta \succeq \pmb 0$. Let $\pmb b = (\pmb c_{max} - \pmb a - \sigma \Delta)^\circ$ and let $\tau$ be the stabilizing script of $\pmb c_{max} - \pmb a - \sigma \Delta$. Then $\pmb a + (\sigma + \tau) \Delta = \pmb c_{max} - \pmb b$, which contradicts Theorem \ref{T:script} since $\sigma + \tau \succ \pmb 0$ and $\pmb c_{max} - \pmb b$ is stable. So $\pmb c_{max} - \pmb a$ is superstable. \smallskip Conversely, suppose that $\pmb c_{max} - \pmb a$ is superstable for some non-critical stable configuration $\pmb a$. By Theorem \ref{T:script}, there exists some script $\sigma \succ \pmb 0$ such that $\pmb a + \sigma \Delta$ is stable. In particular, $\pmb a + \sigma \Delta \preceq \pmb c_{max}$, hence $(\pmb c_{max} - \pmb a) - \sigma \Delta \succeq \pmb 0$, contradiction! \end{proof} \begin{corollary}\label{C:dua} Let $\pmb a$ be a stable configuration and $\sigma^M$ be the minimum $G$-strongly positive script. The following statements are equivalent: \begin{enumerate}[label=\roman*)] \item $\pmb a$ is superstable; \item $\pmb a-\sigma\Delta$ has a negative component for all $\pmb 0 \prec \sigma \preceq \sigma^M$; \item $\pmb a-\sigma\Delta$ is not stable for all $\pmb 0\prec \sigma \preceq \sigma^M$. \end{enumerate} \end{corollary} \smallskip Recall that for undirected graphs, $\sigma^M=(1,1,\dots,1)$. By Corollary \ref{C:dua}, in order to check if a configuration $\pmb a$ is superstable, it suffices to check the non-negativeness or non-stableness of configurations obtained from $\pmb a$ by firing a \emph{subset} of vertices. This is not true for digraphs in general. Consider, for instance, the digraph in Figure \ref{F:Fig-sta-3}: the configuration $\pmb a=(0,3)$ is not superstable ($\pmb a - (2, 1) \Delta = (1, 1) \succ \pmb 0$) despite the fact that $\pmb a - \sigma \Delta$ has at least one negative component for all $\sigma \in \left\{ (0, 1)\,, (1, 0)\,, (1, 1) \right\}$. \medskip The energy minimizing characteristic of superstable configurations is a natural consequence of the duality and the energy maximizing characteristic of critical configurations. \begin{corollary}\label{C:sma} Among the non-negative configurations of each equivalence class, the superstable configuration is the smallest with respect to the $\preceq_{\CFG}$ order. \end{corollary} \begin{proof} The result follows Theorem \ref{T:max2}, Theorem \ref{T:dua}, and the following observations: \begin{enumerate}[label=\roman*)] \item Since $\pmb a \succeq_{\CFG} \pmb a^\circ$ for all non-negative configuration $\pmb a$, it suffices to show that superstable configurations are the smallest among the stable configurations of each equivalence class. \item If $\pmb a$ and $\pmb b$ are equivalent stable configurations, then $\pmb c_{max} - \pmb a$ and $\pmb c_{max} - \pmb b$ are also equivalent stable configurations. \item For any configurations $\pmb a$ and $\pmb b$, $\pmb a \succeq_{\CFG} \pmb b$ if and only if $\pmb c_{max} - \pmb a \preceq_{\CFG} \pmb c_{max} - \pmb b$. \end{enumerate} \end{proof} \medskip \begin{figure}[hbt] \begin{center} \includegraphics[scale=0.7]{Fig-sta-4.pdf} \caption{A non-superstable configuration which is non-negative when firing any subsets of vertices.} \label{F:Fig-sta-3} \end{center} \end{figure} We conclude with a conjecture on the $\preceq_{\CFG}$ order. Let $\pmb a_*$ be a superstable configuration and let $\pmb a^*$ be the critical configuration equivalent to $\pmb a_*$. Consider the following increasing sequence of stable configurations with respect to the $\preceq_{CFG}$ order: \begin{equation} \pmb a_* \preceq_{CFG} (\pmb a_* + \sigma^M \Delta)^\circ \preceq_{CFG} (\pmb a_* + \sigma^M \Delta)^\circ \preceq_{\CFG} \cdots \preceq_{\CFG} \pmb a^* \label{} \end{equation} Consider the graph $G$ shown in Figure \ref{F:Fig-crit-2} with $\pmb a_* = (1, 0, 0, 1))$ and $\pmb a^* = (3, 1, 1, 0)$. For this graph: \[ \Delta = \begin{bmatrix} 5 & -3 & 0 & -1 \\ -1 & 2 & -1 & 0 \\ 0 & -1 & 2 & -1 \\ 0 & 0 & 0 & 2 \end{bmatrix}\,, \sigma^M = (1, 2, 1, 1)\,, \] and we have the sequence: \[ (1, 0, 0, 1) \prec_{\CFG} (0, 1, 1, 0) \prec_{\CFG} (4, 0, 0, 1) \prec_{\CFG} (3, 1, 1, 0)\,. \] The above sequence contains all the stable configurations equivalent to $(1, 0, 0, 1)$. In fact, we have not found an example in which the sequence (\ref{}) does not cover all the stable configurations of the corresponding equivalence class. We conjecture that this is true for digraphs with a global sink in general. \begin{figure}[hbt] \begin{center} \includegraphics[scale=0.7]{Fig-crit-3.pdf} \caption{Example for the Conjecture} \label{F:Fig-crit-2} \end{center} \end{figure} \begin{conjecture} The $\preceq_{CFG}$ order is linear on the set of all stable configurations of each equivalence class. \end{conjecture} \section{Introduction}\label{} Originally developped in the early nineties of the last centuries in the context of self-organized criticality~\cite{Dha90} and as a ``balancing game'' on graphs~\cite{BL92,BLS91}, chip-firing games (CFG) have become an attractive mathematical model in combinatorics~\cite{Big99,Mer05,Mer97,PP16}. In recent years, new connections have been found between CFG and the Riemann-Roch theory on graphs~\cite{AB11,BN07} and potential theory on graphs~\cite{BS13,GK15}. This manuscript contributes some new results to the latter category and at the same time provides a combinatorial insight on existing results. \smallskip We consider a CFG on a digraph $G$ with a global sink. A configuration is a distribution of chips on the non-sink vertices. A transition, called a firing, consists in choosing a vertex and sending one of its chips along each out-going arc to its neighbors. A firing is legal if the chosen vertex has at least as many chips as its out-going arcs. Suppose we restrict the game to non-negative configurations and legal firings. Then, a stable configuration is a configuration in which all the non-sink vertices cannot fire. A critical configuration is a stable configuration which is ``attainable'' from a configuration with many chips. The detailed descriptions of the model will be given in Section \ref{}. Further discussions on critical configurations can be found in \cite{HLMPPW08}. In order to investigate the critical configurations, we introduce the notion of $G$-strongly positive scripts. The $G$-strongly positive scripts are closely related to Speer's algorithm~\cite{Spe93} (indeed, the minimum $G$-strongly positive script coincides with the output of Speer's algorithm) and are equivalent to Perkinson {\em et al.}'s ``burning configurations''~\cite{PPW11}. The key observation is that reverse-firing a $G$-strongly positive script ``increases'' the configurations. Thus, by repeatedly reverse-firing a $G$-strongly positive script then stabilizing, we obtain an ``increasing'' sequence of stable configurations. Since there are only finitely many stable configurations in each equivalence class, this process must eventually stop at a fixed point, which is the critical configuration. \smallskip Our contribution in this paper is threefold. First, we show that for CFG on digraphs with a global sink, among all the stable configurations of each equivalence class, the critical configuration has the maximum weight (Theorem \ref{T:max}). This has been proved in \cite{PP16} for Eulerian digraphs, but the question remained open for digraphs with a global sink. Secondly, we give an extension of Aval {\em et al.}~\cite{ADDL16}'s result from undirected graphs to digraphs with a global sink. Namely, one cannot obtain a stable configuration from a critical one by reverse-firing a non-empty multi-subset of vertices (Theorem \ref{T:script}). Using this result, we revisit the duality between critical configurations and superstable configurations (Theorem \ref{T:dua}). It should be noted that the duality is well-known in the literature: the result for Eulerian digraphs was given by Holroyd \emph{et al.}~\cite{HLMPPW08}, for strongly connected digraphs by Asadi and Backman~\cite{AB11}, for digraphs with a global sink by Perkinson \emph{et al.}~\cite{PPW11}. There are remarkable differences between the techniques used in those proofs. For the case of digraphs with a global sink, Perkinson {\em et al.} used advanced algebra techniques such as the coordinate ring and Gr\"obner bases, while Asadi and Backman's proof is a combinatorial one. Our proof is combinatorial and is independent from Asadi and Backman's proof. A notable difference is that the attainability in their proof requires that the firing/reverse-firing sequences to be legal, while the legality restriction is not imposed in our proof. Lastly, we give an energy maximizing (resp. minimizing) characteristic for critical (resp. superstable) configurations (Theorem \ref{T:max2} and Corollary \ref{C:sma}). Unlike previous studies on CFG and potential theory~\cite{BS13, GK15} where energies are defined as norms, in this paper, the energies are defined as vectors and are compared using the containment order. This order is not unfamiliar in the CFG literature: its restriction on the set of accessible configurations (given an initial configuration) coincides with the accessibility order, which has been extensively investigated in connection with lattice structures~\cite{CPT13, LP01}. \smallskip The manuscript is organized as follows. Section \ref{} gives a brief summary of relevant definitions and fundamental results on $\CFG$ and critical configurations. Section \ref{} presents some properties and the maximizing characterization of critical configurations. Section \ref{} discusses the duality and the minimizing characterization of superstable configurations. \bigskip \section*{Acknowledgment}\label{} \addcontentsline{toc}{section}{Acknowledgements} The manuscript was conceptualized when the authors visited the Vietnam Institute for Advanced Study in Mathematics. We would like to thank Prof. Robert Cori and Prof. Spencer Backman for their very useful comments and suggestions on a short version of this manuscript \cite{Tra16}. The work is partially funded by the PSSG no.6 and the NAFOSTED grant under the project number 101.99-2016.16. \bibliographystyle{plain} \input{biblio} \addcontentsline{toc}{section}{References} \end{document} \section{Preliminaries}\label{} Unless stated otherwise, the vectors in this paper are row vectors. Let $a$ be a vector in $\mathbb Z^n$. We denote by $a_i$ the $i$-th component of $a$. We say that a vector is non-negative (resp. positive) if all of its components are non-negative (resp. positive). The zero vector is denoted by $\pmb 0$ and the all-one vector by $\pmb 1$. The {\em support} of $a$ is defined by $\mbox{supp}(a) = \{i\ \vert\ a_i \neq 0\}$. The {\em weight} of $a$ is $w(a) = \sum_{i = 1}^n a_i$. For a matrix $M$, let $M_i$ denote the $i$-th row vector of $M$ and $M_{ij}$ the entry at row $i$ and column $j$ of $M$. The {\em containment order} on $\mathbb Z^n$ is defined by: $a \succeq b \iff a_i \ge b_i$ for all $i=1,2,\cdots,n$. \medskip Let us now give some basic notions of chip-firing games on digraphs with a global sink. Let $G = (V, E)$ be a directed multigraph without loops, with $n + 1$ vertices, {\em i.e.} $V = \{1, 2, \cdots, n + 1\}$. The \emph{out-going degree} (resp. \emph{in-going degree}) of a vertex $i$ is denoted by $d^+_i$ (resp. $d^-_i$). The number of edges going from $i$ to $j$ is denoted by $e_{i, j}$. A {\em source component} (resp. {\em sink component}) of $G$ is a strongly connected component without in-going (resp. out-going) edge from other strongly connected components. A vertex $s$ of $V$ is a {\em sink} if $\{s\}$ is a sink component. It is a {\em global sink} if it is a sink and for every other vertex $i$, there exists a directed path from $i$ to $s$. Note that a global sink is unique if it exists. In the rest of the paper, we suppose that $G$ has a global sink at the vertex $n + 1$. \smallskip The graph $G$ can be represented by its {\em Laplacian matrix} $\widetilde \Delta \in \mathbb Z^{(n + 1) \times (n + 1)}$ whose entries are defined as follows: \begin{equation*} \widetilde{\Delta}_{ij} = \begin{cases} d^+_i \quad & \text { if } i = j\,,\\ -e_{i,j} \quad & \text{ if } i \ne j\,. \end{cases} \end{equation*} The {\em reduced Laplacian matrix} $\Delta$ is the matrix obtained from $\widetilde \Delta$ by removing the row and the column corresponding to the sink. The Laplacian and reduced Laplacian matrices play an important role in the study of combinatorial properties of graphs and of chip-firing games on graphs (defined hereafter) as well, see \cite{BL92, BLS91} and the references therein. We recall here two useful properties for the Laplacian and the reduced Laplacian of digraphs with a global sink. \begin{enumerate}[label=\roman*)] \item The kernel of $\widetilde{\Delta}$ is spanned by a non-negative vector $v$ such that $v_{n + 1} > 0$. Consequently, the rank of $\widetilde{\Delta}$ is $n$. \item The matrix $\Delta$ is non-singular. Moreover, all the entries of $\Delta^{-1}$ are non-negative. \end{enumerate} The first property is a particular case of \cite[Proposition 3.1]{BL92}. The non-singularity is a corollary of the first property. The non-negativity of $\Delta^{-1}$ is a property of non-singular M-matrix (see, for instance, \cite{plemmons77, PS04}). \medskip A {\em chip-firing game} ($\CFG$) on $G$, denoted by $\CFG(G)$, is a discrete dynamical system in which: \begin{itemize} \item[-] A {\em chip configuration} ({\em configuration} for short) is a vector in $\mathbb Z^n$ whose coordinates represent the numbers of chips at non-sink vertices \footnote{Typically, the configurations must be non-negative. In this paper, we allow the coordinates of a configuration to be negative.}. We will use bold letters such as $\pmb a$ to denote configuration vectors. \item[-] The transitions occur according to the {\em firing rule}, defined as follows. A vertex is {\em active} if it has at least as many chips as its out-going degree. An active vertex can {\em fire} by sending one chip along each of its out-going edge to its neighbors. That is, if $\pmb b$ is the configuration obtained from $\pmb a$ by firing an active vertex $i$, we write $\pmb a \xrightarrow{i} \pmb b$, then: \[ b_j = \begin{cases} a_i - d^+_i & \mbox{ if } j = i\,,\\ a_j + e_{i,j} & \mbox{ otherwise,} \end{cases} \] or equivalently: \[ \pmb b = \pmb a - {\Delta}_i\,. \] \end{itemize} \smallskip A (unconstrained) {\em firing sequence} is a sequence of non-sink vertices $(s_1, s_2, \cdots, s_k)$ in $V$. The firing sequence $(s_1, s_2, \cdots, s_k)$ is called {\em legal} from $\pmb a$ if there exists a sequence of configurations $\pmb a = \pmb a_0, \pmb a_1, \cdots, \pmb a_k$ such that $\pmb a_{r - 1} \xrightarrow{s_{r}} \pmb a_{r}$ for all $1 \le r \le k$ and $s_r$ is active in $\pmb a_{r - 1}$. The {\em firing script} of a legal firing sequence $(s_1, s_2, \cdots, s_k)$ is a vector $\sigma \in \mathbb N^n$, where $\sigma_i$ is the number of occurrences of the vertex $i$ in the sequence $(s_1,s_2,\cdots,s_k)$. The firing script provides a direct way to compute the result of a firing sequence: if $\pmb b$ is obtained from $\pmb a$ by firing sequentially $s_1, s_2, \cdots, s_k$ and $\sigma$ is the corresponding firing script, then \begin{equation} \pmb b = \pmb a - \sigma \Delta\,. \label{} \end{equation} When referring to vectors in $\mathbb N^n$ without mentioning a firing sequence, we will use the term {\em $n$-scripts}, or simply {\em scripts}. \medskip We say that two configurations $\pmb a$ and $\pmb b$ are {\em linearly equivalent}, denoted by $\pmb a \sim \pmb b$, if there exists a vector $\sigma \in \mathbb Z^n$ verifying (\ref{}). The linear equivalence is an equivalent relation on $\mathbb Z^{n}$. The quotient group $\mathbb Z^n/\langle\Delta_1,\dots,\Delta_n\rangle$ is called the {\em Sandpile group} of $G$ and is denoted by $\mathcal{SP}(G)$. When the context is clear, we refer to elements of $\mathcal{SP}(G)$ simply as equivalent classes. \smallskip Note that $\pmb a \sim \pmb b$ does not imply that $\pmb b$ can be obtained from $\pmb a$ by a legal firing sequence. In fact, even if $\pmb b = \pmb a - \sigma \Delta$ for some non-negative $\sigma$, it is not always the case that there exists a legal firing sequence leading from $\pmb a$ to $\pmb b$. \medskip A configuration is said to be {\em stable} if it does not have any active vertex. It is known that for each $\CFG$ with a global sink and for any non-negative configuration $\pmb a$, there exists a unique stable configuration, denoted by $\pmb a^\circ$, which can be obtained from $\pmb a$ by a legal firing sequence. A process of firing from $\pmb a$ to $\pmb a^\circ$ is called a {\em stabilization} of $\pmb a$. Note that there may be many legal firing sequences from $\pmb a$ to $\pmb a^\circ$, but they all have the same firing script. In fact, they are the longest legal firing sequences from $\pmb a$. It is straightforward that if $\pmb a$ and $\pmb b$ are non-negative configurations, then \[ (\pmb a+\pmb b)^\circ = (\pmb a+\pmb b^\circ)^\circ =(\pmb a^\circ+\pmb b^\circ)^\circ\,. \] Moreover, if $\pmb b$ is obtained from $\pmb a$ by a legal firing sequence with firing script $\sigma$, then $$(\pmb a-\sigma\Delta)^\circ =\pmb b^\circ.$$ We refer the interested reader to \cite{BLS91,HLMPPW08,LP01} for more details on these results. \medskip We end this section with some important properties of critical configurations, also known as recurrent configurations. There are several definitions of critical configurations in the literature. In \cite{HLMPPW08}, it is shown that all those definitions are equivalent for $\CFG$ on a digraph with a global sink. We recall here the definition that we consider most useful for our purposes in this paper. Let $\pmb c_{max} = (d_1^+-1, d_2^+-1,\cdots, d_n^+-1)$ be the maximum stable configuration by the containment order. \begin{definition}[Critical configuration]\mbox{} A non-negative configuration $\pmb a$ is critical if and only if for any configuration $\pmb b$, there exists a non-negative configuration $\pmb c$ such that $\pmb a = (\pmb b + \pmb c)^\circ$. \end{definition} \smallskip The following lemma presents some known properties of critical configurations. \begin{lemma}[\cite{HLMPPW08}] Consider a $\CFG$ on a graph $G$ with a global sink. \begin{enumerate}[label=\roman*)] \item Each equivalent class of $\mathcal{SP}(G)$ has exactly one critical configuration. \item If $\pmb a$ is critical, $\pmb b$ is stable, and $\pmb b \succeq \pmb a$, then $\pmb b$ is critical. \item A non-negative configuration $\pmb a$ is {\em critical} if it is stable and there exists a non-negative configuration $\pmb c$ such that $\pmb a = (\pmb c_{max} + \pmb c)^\circ$. \end{enumerate} \label{} \end{lemma} It is difficult to decide if a configuration is critical or not using the above definitions. However, there exist efficient algorithms to recognize critical configurations, for instance Dhar's \emph{burning algorithm} for undirected graphs~\cite{Dha90} and Speer's \emph{script algorithm} for connected digraphs~\cite{Spe93}. In the next section, we give a maximizing characterization for the critical configurations on their equivalent classes.

\section{Introduction} \label{} Understanding the origins of the patterns of neutrino mixing and of neutrino mass squared differences, revealed by the data obtained in the neutrino oscillation experiments (see, e.g., \cite{}), is one of the most challenging problems in neutrino physics. It is part of the more general fundamental problem in particle physics of understanding the origins of flavour, i.e., of the patterns of quark, charged lepton and neutrino masses, and of the quark and lepton mixing. We believe, and we are not alone in holding this view, that with the observed pattern of neutrino mixing Nature is ``sending'' us a Message. The Message is encoded in the values of the neutrino mixing angles, leptonic CP violation (CPV) phases in the Pontecorvo, Maki, Nakagawa and Sakata (PMNS) neutrino mixing matrix \cite{BPont57,MNS62,BPont67} and neutrino masses. We do not know at present what is the content of Nature's Message. However, on the basis of the current ideas about the possible origins of the observed pattern of neutrino mixing, the Nature's Message can have two completely different contents, each of which can be characterised by one word: ANARCHY or SYMMETRY. In the ANARCHY approach \cite{Anarchy} to understanding the pattern of neutrino mixing it is assumed that Nature ``threw dice'' when ``choosing'' the values of the neutrino masses, mixing angles and leptonic CPV phases. The main prediction of the ANARCHY explanation of the pattern of neutrino mixing is the absence of whatever correlations between the values of the neutrino masses, between the values of the neutrino mixing angles, and between the values of the neutrino mixing angles and the CPV phases, all of them being random quantities. As a consequence, no specific values of, e.g., neutrino mixing angles are predicted: the predictions of these (and other leptonic) measurable quantities are in the form of distributions. In contrast, one of the main characteristic features of the SYMMETRY approach to neutrino mixing is the prediction of the values of some of the mixing angles and/or of the existence of correlations between the values of at least some of the observables (angles, CPV phases) of the the neutrino mixing matrix. Within the SYMMETRY approach, the observed pattern of neutrino mixing, which differs drastically from the quark mixing pattern, can be naturally understood on the basis of specific class of symmetries - the class of non-Abelian discrete flavour symmetries (see, e.g., \cite{}). Thus, the specific form of the neutrino mixing can have its origin in the existence of new fundamental symmetry in the lepton sector. The most distinctive feature of the approach to neutrino mixing based on non-Abelian discrete flavour symmetries is the predictions of the values of some of the neutrino mixing angles and leptonic CPV phases, and/or of existence of correlations between the values of at least some the neutrino mixing angles and/or between the values of the neutrino mixing angles and the Dirac CPV phase in the PMNS matrix, etc. (see, e.g., \cite{, }). Combining the discrete symmetry approach with the idea of generalised CP invariance \cite{} ~--~ a generalisation of the standard CP invariance requirement~--~ allows to obtain predictions also for the Majorana CPV phases in the PMNS matrix in the case of massive Majorana neutrinos (see, e.g., \cite{} and references quoted therein). Most importantly, these predictions and predicted correlations, and thus the discrete symmetry approach itself, can be tested experimentally (see, e.g., \cite{} and \cite{}). In the present article we review aspects of the symmetry approach to neutrino mixing based on the class of non-Abelian discrete flavour symmetries, which is widely explored at present (see, e.g., \cite{} and references quoted therein \footnote{For early attempts see, e.g., \cite{Early}.} ). We will discuss also the typical phenomenological predictions of the approach and their possible tests in currently running and future planned neutrino oscillation experiments. Before discussing the discrete flavour symmetry approach to neutrino mixing we would like to review briefly the current status of our knowledge of neutrino masses, neutrino mixing and leptonic CPV phases, the remaining fundamental problems in neutrino physics and the future prospects in this field. \section{The Three-Neutrino Mixing} \label{} The experiments with solar, atmospheric, reactor and accelerator neutrinos have provided compelling evidences for the existence of neutrino oscillations \cite{BPont57,MNS62}~--~ transitions in flight between the different flavour neutrinos $\nu_e$, $\nu_\mu$, $\nu_\tau$ (antineutrinos $\bar{\nu}_e$, $\bar{\nu}_\mu$, $\bar{\nu}_\tau$)~--~ caused by nonzero neutrino masses and neutrino mixing (see, e.g., ref. \cite{} for review of the relevant data). The existence of flavour neutrino oscillations implies the presence of mixing in the weak charged lepton current: \begin{equation} \label{CC} {\cal L}_{\rm CC} = - ~\frac{g}{\sqrt{2}}\, \sum_{l=e,\mu,\tau} \overline{l_L}(x)\, \gamma_{\alpha} \nu_{l\mathrm{L}}(x)\, W^{\alpha \dagger}(x) + h.c.\,,~ \nu_{l \mathrm{L}}(x) = \sum^n_{j=1} U_{l j} \nu_{j \mathrm{L}}(x)\,, \end{equation} \noindent where $\nu_{lL}(x)$ are the flavour neutrino fields, $\nu_{j \mathrm{L}}(x)$ is the left-handed (LH) component of the field of the neutrino $\nu_j$ having a mass $m_j$, and $U$ is a unitary matrix - the PMNS neutrino mixing matrix \cite{BPont57,MNS62,BPont67}. All compelling neutrino oscillation data can be described assuming 3-neutrino mixing in vacuum, $n=3$. The number of massive neutrinos $n$ can, in general, be bigger than 3 if, e.g., there exist RH sterile neutrinos \cite{BPont67} and they mix with the LH flavour neutrinos. It follows from the current data that at least 3 of the neutrinos $\nu_j$, say $\nu_1$, $\nu_2$, $\nu_3$, must be light, i.e., must have masses smaller than roughly 1 eV, $m_{1,2,3} \ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$<$}\ 1$ eV, and must have different masses, $m_1\neq m_2 \neq m_3$ \footnote{At present there are several experimental inconclusive hints for existence of one or two light sterile neutrinos at the eV scale, which mix with the flavour neutrinos, implying the presence in the neutrino mixing of additional one or two neutrinos, $\nu_4$ or $\nu_{4,5}$, with masses $m_4~(m_{4,5})\sim 1~{\rm eV}$ (see, e.g., ref.~\cite{}). For a recent discussion of these hints and of the related implications see, e.g., ref.~\cite{}. }. In the case of 3 light neutrinos, the $3\times 3$ unitary neutrino mixing matrix $U$ can be parametrised, as is well known, by 3 angles and, depending on whether the massive neutrinos $\nu_j$ are Dirac or Majorana particles, by one Dirac, or one Dirac and two Majorana, CP violation (CPV) phases \cite{}: \begin{equation} U= VP\,,~~~ P = {\rm diag}(1, e^{i \frac{\alpha_{21}}{2}}, e^{i \frac{\alpha_{31}}{2}})\,, \label{VP} \end{equation} where $\alpha_{21,31}$ are the two Majorana CPV phases and in the ``standard'' parametrisation the matrix $V$ is given by: \begin{equation} \begin{array}{c} \label{} V = \left(\begin{array}{ccc} c_{12} c_{13} & s_{12} c_{13} & s_{13} e^{-i \delta} \\[0.2cm] -s_{12} c_{23} - c_{12} s_{23} s_{13} e^{i \delta} & c_{12} c_{23} - s_{12} s_{23} s_{13} e^{i \delta} & s_{23} c_{13} \\[0.2cm] s_{12} s_{23} - c_{12} c_{23} s_{13} e^{i \delta} & - c_{12} s_{23} - s_{12} c_{23} s_{13} e^{i \delta} & c_{23} c_{13} \\ \end{array} \right)\,. \end{array} \end{equation} \noindent In eq.~(\ref{}), $c_{ij} = \cos\theta_{ij}$, $s_{ij} = \sin\theta_{ij}$, the angles $\theta_{ij} \in [0,\pi/2]$, and $\delta \in [0,2\pi)$ is the Dirac CPV phase. Thus, in the case of massive Dirac neutrinos, the neutrino mixing matrix $U$ is similar, in what concerns the number of mixing angles and CPV phases, to the Cabibbo, Kobayashi, Maskawa (CKM) quark mixing matrix. The PMNS matrix $U$ contains two additional physical CPV phases if $\nu_j$ are Majorana particles due to the special properties of Majorana fermions (see, e.g., refs.~\cite{}). On the basis of the existing neutrino data it is impossible to determine whether the massive neutrinos are Dirac or Majorana fermions. The probabilities of neutrino oscillations are functions of the neutrino energy, $E$, the source-detector distance $L$, of the elements of $U$ and, for relativistic neutrinos used in all neutrino experiments performed so far, of the neutrino mass squared differences $\Delta m^2_{ij} \equiv (m^2_{i} - m^2_j)$, $i\neq j$ (see, e.g., ref.~\cite{BiPet87}). In the case of 3-neutrino mixing there are only two independent $\Delta m^2_{ij}$, say $\Delta m^2_{21}\neq 0$ and $\Delta m^2_{31} \neq 0$. The numbering of neutrinos $\nu_j$ is arbitrary. We will employ the widely used convention which allows to associate $\theta_{13}$ with the smallest mixing angle in the PMNS matrix, and $\theta_{12}$, $\Delta m^2_{21}> 0$, and $\theta_{23}$, $\Delta m^2_{31}$, with the parameters which drive the solar ($\nu_e$) and the dominant atmospheric $\nu_{\mu}$ and $\bar{\nu}_{\mu}$ oscillations, respectively. In this convention $m_1 < m_2$, $ 0 < \Delta m^2_{21} < |\Delta m^2_{31}|$, and, depending on ${\rm sgn}(\Delta m^2_{31})$, we have either $m_3 < m_1$ or $m_3 > m_2$. \begin{table} \centering \renewcommand{\arraystretch}{1.2} \begin{tabular}{lccc} \toprule Parameter & Best fit value & $2\sigma$ range & $3\sigma$ range \\ \midrule $\sin^2\theta_{12}/10^{-1}$ & $2.97$ & $2.65 - 3.34$ & $2.50 - 3.54$ \\ $\sin^2\theta_{13}/10^{-2}$~(NO) & $2.15$ & $1.99 - 2.31$ & $1.90 - 2.40$\\ $\sin^2\theta_{13}/10^{-2}$~(IO) & $2.16$ & $1.98 - 2.33$ & $1.90 - 2.42$\\ $\sin^2\theta_{23}/10^{-1}$~(NO) & $4.25$ & $3.95 - 4.70$ & $3.81 - 6.15$\\ $\sin^2\theta_{23}/10^{-1}$~(IO) & $5.89$ & $3.99 - 4.83 \oplus 5.33 - 6.21$ & $3.84 - 6.36$\\ $\delta/\pi$~(NO) & $1.38$ & $1.00 - 1.90$ & $0 - 0.17 \oplus 0.76 - 2$\\ $\delta/\pi$~(IO) & $1.31$ & $0.92 - 1.88$ & $0 - 0.15 \oplus 0.69 - 2$\\ \midrule $\Delta m_{21}^{2}/10^{-5}$~eV$^2$ & $7.37$ & $7.07 - 7.73$ & $6.93 - 7.96$\\ $\Delta m_{31}^{2}/10^{-3}$~eV$^2$~(NO) & $2.56$ & $2.49 - 2.64$ & $2.45 - 2.69$\\ $\Delta m_{23}^{2}/10^{-3}$~eV$^2$~(IO) & $2.54$ & $2.47 - 2.62$ & $2.42 - 2.66$\\ \bottomrule \end{tabular} \caption{The best fit values, $2\sigma$ and 3$\sigma$ ranges of the neutrino oscillation parameters obtained in the global analysis of the neutrino oscillation data performed in~\cite{}. (The Table is taken from ref. \cite{}.) } \label{} \end{table} The existing data, accumulated over many years of studies of neutrino oscillations, allow us to determine $\Delta m^2_{21}$, $\theta_{12}$, and $|\Delta m^2_{31(32)}|$, $\theta_{23}$ and $\theta_{13}$, with a relatively high precision \cite{}. Since 2013 there are also persistent hints that the Dirac CPV phase $\delta$ has a value close to $3\pi/2$ (see \cite{}). The best fit values (b.f.v.) and the 2$\sigma$ and 3$\sigma$ allowed ranges of $\Delta m^2_{21}$, $s^2_{12}$, $|\Delta m^2_{31(32)}|$, $s^2_{23}$, $s^2_{13}$ and $\delta$, found in the analysis of global neutrino oscillation data performed in \cite{} are given in Table \ref{}. Similar results were obtained in ref.~\cite{}. In both analyses \cite{} the authors find, in particular, that the best fit value of the Dirac CPV phases $\delta$ is close to $3\pi/2$: in \cite{}, for example, the authors find $\delta = 1.38\pi~(1.31\pi)$ for $\Delta m^{2}_{31(32)}>0$ ($\Delta m^{2}_{31(32)} < 0$). The absolute $\chi^2$ minimum takes place for $\Delta m^{2}_{31(32)}>0$. According to ref. \cite{}, the CP conserving values $\delta = 0$ or $2\pi$ are disfavored at $2.4\sigma$ ($3.2\sigma$) for $\Delta m^2_{31(32)}>0$ ($\Delta m^2_{31(32)}<0$); the CP conserving value $\delta = \pi$ in the case of $\Delta m^2_{31(32)}>0$ ($\Delta m^2_{31(32)}<0$) is statistically approximately $2.0\sigma$ ($2.5\sigma$) away from the best fit value $\delta \cong 1.38\pi~(1.31\pi)$. In what concerns the CP violating value $\delta = \pi/2$, it is strongly disfavored at $3.4\sigma$ ($3.9\sigma$) for $\Delta m^2_{31(32)}>0$ ($\Delta m^2_{31(32)}<0$)~ \footnote{The quoted confidence levels for $\delta = 0,\pi$ and $\pi/2$ are all with respect to the absolute $\chi^2$ minimum.}. At $3\sigma$, $\delta/\pi$ is found to lie in the case of $\Delta m^{2}_{31(32)}>0$ ($\Delta m^{2}_{31(32)} < 0$) in the following intervals \cite{}: $(0.00-0.17(0.15))\oplus (0.76(0.69)-2.00))$. The results on $\delta$ obtained in \cite{} differ somewhat from, but are compatible at $1\sigma$ confidence level (C.L.) with, those found in \cite{}. It follows also from the results quoted in Table \ref{} that $\Delta m^2_{21}/|\Delta m^2_{31(32)}| \cong 0.03$. We have $|\Delta m^2_{31}| = |\Delta m^2_{32} - \Delta m^2_{21}| \cong |\Delta m^2_{32}|$. The angle $\theta_{12}$ is definitely smaller than $\pi/4$: the value of $\theta_{12} = \pi/4$, i.e., maximal solar neutrino mixing, is ruled out at high confidence level (C.L.) by the data - one has $\cos2\theta_{12} \geq 0.29$ at $99.73\%$ C.L. The quoted results imply also that the value of $\theta_{23}$ can deviate by approximately $\pm 0.1$ from $\pi/4$, $\theta_{12} \cong \pi/5.4$ and that $\theta_{13} \cong \pi/20$. Thus, the pattern of neutrino mixing differs drastically from the pattern of quark mixing. It should be noted that in the more recent global analyses \cite{NuFITv32Jan2018,}, which used, in particular, updated results on $\sin^2\theta_{23}$ from the NO$\nu$A experiment, the best fit value of $\sin^2\theta_{23}$ for NO spectrum was found to be larger than 0.5~ \footnote{In what concerns the other two neutrino mixing angles $\theta_{12}$ and $\theta_{13}$, the results reported in \cite{} and in \cite{NuFITv32Jan2018,} differ insignificantly. }: \begin{equation} \sin^2\theta_{23} = 0.538~(0.554)~[46]\,,~~~ \sin^2\theta_{23} = 0.551~(0.557)~[47]\,,~~~{\rm NO~(IO)}\,. \label{} \end{equation} Apart from the hint that the Dirac phase $\delta \sim 3\pi/2$, no other experimental information on the Dirac and Majorana CPV phases in the neutrino mixing matrix is available at present. Thus, the status of CP symmetry in the lepton sector is essentially unknown. With $\theta_{13} \cong 0.15 \neq 0$, the Dirac phase $\delta$ can generate CP violating effects in neutrino oscillations \cite{}, i.e, a difference between the probabilities of the $\nu_l \rightarrow \nu_{l'}$ and $\bar{\nu}_l \rightarrow \bar{\nu}_{l'}$ oscillations, $l\neq l'=e,\mu,\tau$. The magnitude of CP violation in $\nu_l \rightarrow \nu_{l'}$ and $\bar{\nu}_l \rightarrow \bar{\nu}_{l'}$ oscillations in vacuum, $l\neq l'=e,\mu,\tau$, is determined by \cite{PKSP3nu88} the rephasing invariant $J_{CP}$, associated with the Dirac CPV phase in $U$: \begin{equation} J_{\rm CP} = {\rm Im}\, \left (U_{\mu 3}\,U^*_{e3}\,U_{e2}\,U^*_{\mu 2}\right )\,. \label{JCP02} \end{equation} It is analogous to the rephasing invariant associated with the Dirac CPV phase in the CKM quark mixing matrix \cite{CJ85}. In the standard parametrisation of the neutrino mixing matrix (\ref{}), $J_{\rm CP}$ has the form: \begin{equation} J_{\rm CP} \equiv {\rm Im}\,(U_{\mu 3}\,U^*_{e3}\,U_{e2}\,U^*_{\mu 2}) = \frac{1}{8}\,\cos\theta_{13} \sin 2\theta_{12}\,\sin 2\theta_{23}\,\sin 2\theta_{13}\,\sin \delta\,. \label{JCPstandparam} \end{equation} Thus, given the fact that $\sin 2\theta_{12}$, $\sin 2\theta_{23}$ and $\sin 2\theta_{13}$ have been determined experimentally with a relatively good precision, the size of CP violation effects in neutrino oscillations depends essentially only on the magnitude of the currently not well determined value of the Dirac phase $\delta$. The current data imply $0.026(0.027)|\sin\delta|\lesssim |J_{\rm CP}| \lesssim 0.035 |\sin\delta|$, where we have used the $3\sigma$ ranges of $\sin^2\theta_{12}$, $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$ given in Table \ref{}. For the current best fit values of $\sin^2\theta_{12}$, $\sin^2\theta_{23}$, $\sin^2\theta_{13}$ and $\delta$ we find in the case of $\Delta m^2_{31(2)} > 0$ ($\Delta m^2_{31(2)} < 0$): $J_{\rm CP} \cong 0.032\sin\delta \cong -\; 0.030$ ($J_{\rm CP} \cong 0.032\sin\delta \cong -\; 0.027$). Thus, if the indication that $\delta$ has a value close to $3\pi/2$ is confirmed by future more precise data, i) the $J_{\rm CP}$ factor in the lepton sector would be approximately by 3 orders of magnitude larger in absolute value than corresponding $J_{\rm CP}$ factor in the quark sector, and ii) the CP violation effects in neutrino oscillations would be relatively large and observable. If the neutrinos with definite masses $\nu_i$, $i=1,2,3$, are Majorana particles, the 3-neutrino mixing matrix contains two additional Majorana CPV phases \cite{}. However, the flavour neutrino oscillation probabilities $P(\nu_l \rightarrow \nu_{l'})$ and $P(\bar{\nu}_l \rightarrow \bar{\nu}_{l'})$, $l,l' =e,\mu,\tau$, do not depend on the Majorana phases\cite{}. The Majorana phases can play important role, e.g, in $|\Delta L| = 2$ processes like neutrinoless double beta ($(\beta\beta)_{0\nu}$-) decay $(A,Z) \rightarrow (A,Z+2) + e^- + e^-$, $L$ being the total lepton charge, in which the Majorana nature of massive neutrinos $\nu_i$ manifests itself (see, e.g, refs. \cite{BiPet87,STPNuNature2013,BPP1}). Our interest in the CPV phases present in the neutrino mixing matrix is stimulated also by the intriguing possibility that the Dirac phase and/or the Majorana phases in $U_{\rm PMNS}$ can provide the CP violation necessary for the generation of the observed baryon asymmetry of the Universe (BAU) \cite{} (for specific models in which this possibility is realised see, e.g., \cite{}; for a recent review see \cite{}). More specifically, if, e.g., all CP violation necessary for the generation of BAU is due to the Dirac phase $\delta$, which is possible within the ``flavoured'' leptogenesis scenario \cite{FlavLG,nardietal} of generation of baryon asymmetry, a necessary condition for reproducing the observed BAU in this scenario (with hierarchical heavy Majorana neutrinos) is \cite{} $|\sin\theta_{13}\,\sin\delta|\ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$>$}\ 0.09$. This condition is comfortably compatible with the measured value of $\sin\theta_{13}$ and with the best fit value of $\delta \sim 3\pi/2$. The sign of $\Delta m^2_{31(32)}$ cannot be directly and unambiguously determined from the existing data \footnote{In the recent analysis of the global neutrino oscillation data performed in \cite{} it was found that the case of $\Delta m^2_{31(32)} < 0$ is disfavored at $3.1\sigma$ with respect to the case of $\Delta m^2_{31(32)} > 0$. }. In the case of 3-neutrino mixing, the two possible signs of $\Delta m^2_{31(32)}$ correspond to two types of neutrino mass spectrum. In the convention of numbering of neutrinos $\nu_j$ employed by us the two spectra read:\\ {\it i) spectrum with normal ordering (NO)}: $m_1 < m_2 < m_3$, $\Delta m^2_{31(32)} >0$, $\Delta m^2_{21} > 0$, $m_{2(3)} = (m_1^2 + \Delta m^2_{21(31)})^{1\over{2}}$; \\~~ {\it ii) spectrum with inverted ordering (IO)}: $m_3 < m_1 < m_2$, $\Delta m^2_{32(31)}< 0$, $\Delta m^2_{21} > 0$, $m_{2} = (m_3^2 + \Delta m^2_{23})^{1\over{2}}$, $m_{1} = (m_3^2 + \Delta m^2_{23} - \Delta m^2_{21})^{1\over{2}}$.\\ Depending on the values of the lightest neutrino mass, ${\rm min}(m_j)$, the neutrino mass spectrum can also be:~\\ {\it a) Normal Hierarchical (NH)}: $m_1 \ll m_2 < m_3$, $m_2 \cong (\Delta m^2_{21})^ {1\over{2}}\cong 8.6\times 10^{-3}$ eV, $m_3 \cong (\Delta m^2_{31})^{1\over{2}} \cong 0.0506$ eV; or\\ {\it b) Inverted Hierarchical (IH)}: $m_3 \ll m_1 < m_2$, $m_{1} \cong (|\Delta m^2_{32}| - \Delta m^2_{21})^{1\over{2}}\cong 0.0497$ eV, $m_{2} \cong (|\Delta m^2_{32}|)^{1\over{2}}\cong 0.0504$ eV; or\\ {\it c) Quasi-Degenerate (QD)}: $m_1 \cong m_2 \cong m_3 \cong m_0$, $m_j^2 \gg |\Delta m^2_{31(32)}|$, $m_0 \ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$>$}\ 0.10$ eV.\\ All three types of spectrum are compatible with the constraints on the absolute scale of neutrino masses. Determining the type of neutrino mass spectrum is one of the main goals of the future experiments in the field of neutrino physics \footnote{For a brief discussion of experiments which can provide data on the type of neutrino mass spectrum see, e.g., ref. \cite{XQPVogel2015}; for some specific proposals see, e.g., ref.~\cite{NOIO1,NOIO11,NOIO2,NOIO3}.} (see, e.g., refs.~ \cite{}). Data on the absolute neutrino mass scale (or on ${\rm min}(m_j)$) can be obtained, e.g., from measurements of the spectrum of electrons near the end point in $\mbox{}^3 {\rm H}$ $\beta$-decay \ experiments \cite{Fermi34,Mainz,MoscowH3} and from cosmological and astrophysical observations. The most stringent upper bound on the $\bar{\nu}_e$ mass was reported by the Troitzk~\cite{MoscowH3b} experiment: \begin{equation*} m_{\bar{\nu}_e} < 2.05~\mathrm{eV}~~~\mbox{at}~95\%~\mathrm{C.L.} \label{H3beta} \end{equation*} \noindent Similar result was obtained in the Mainz experiment \cite{Mainz}~: $m_{\bar{\nu}_e} < 2.3~\rm{eV}$ at 95\% CL. We have $m_{\bar{\nu}_e} \cong m_{1,2,3}$ in the case of QD spectrum. The KATRIN experiment~\cite{MainzKATRIN}, which was commissioned on June 11, 2018, is designed to reach sensitivity of $m_{\bar{\nu}_e} \sim 0.20$~eV, i.e., to probe the region of the QD spectrum. Constraints on the sum of the neutrino masses can be obtained from cosmological and astrophysical data (see, e.g., ref.~\cite{summj}). Depending on the model complexity and the input data used one typically obtains \cite{summj}: $\sum_j m_j\ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$<$}\ (0.3 - 1.3)$ eV, 95\% C.L. Assuming the existence of three light massive neutrinos and the validity of the $\Lambda$ CDM (Cold Dark Matter) model, and using their data on the CMB temperature power spectrum anisotropies, polarisation, on gravitational lensing effects and the low $l$ CMB polarization spectrum data (the "low P" data), etc. the Planck Collaboration reported an updated upper limit on the sum of the neutrino masses \cite{}, which, depending on the data-set used, varies in the interval: $\sum_j m_j\, < \,(0.340 - 0.715)$ eV, 95\%~C.L. Adding data on Baryon Acoustic Oscillations (BAO) lowers the limit to \cite{}: \begin{equation} \sum_j m_j\, < \, 0.170~{\rm eV},~~~ 95\%~{\rm C.L.} \label{Planck1} \end{equation} In spite of the remarkable progress made in the last 19 years in establishing the existence of neutrino oscillations caused by non-zero neutrino masses and neutrino mixing and in measuring the 3-neutrino oscillation parameters, one has to admit that we are still completely ignorant about the fundamental aspects of neutrino mixing. We do not know whether the massive neutrinos are Dirac or Majorana particles, what is the neutrino mass ordering, what is the status of the CP symmetry in the lepton sector and what is the absolute neutrino mass scale (i.e., the lightest neutrino mass). Determining the nature~---~Dirac or Majorana~---~of massive neutrinos, the type of spectrum the neutrino masses obey, establishing the status of the CP symmetry in the lepton sector and determining the absolute neutrino mass scale are among the highest priority goals of the programme of future experimental research in neutrino physics (see, e.g., \cite{}), which extends beyond 2030. The principal goal of the theoretical studies in this field is the understanding at a fundamental level the mechanism giving rise to neutrino masses and mixing and to $L_l-$non-conservation. Are the observed patterns of $\nu$-mixing and of $\Delta m^2_{21,31}$ related to the existence of a new fundamental symmetry of particle interactions? Is there any relation between quark mixing and neutrino mixing? What is the physical origin of CPV phases in the neutrino mixing matrix $U$? Is there any relation (correlation) between the (values of) CPV phases and mixing angles in $U$? Progress in the theory of neutrino mixing might also lead, in particular, to a better understanding of the mechanism of generation of baryon asymmetry of the Universe. \vspace{-0.3cm} \section{Origins of the Pattern of Neutrino Mixing: the Discrete Symmetry Approach } \label{sec3} \subsection{The General Framework} \label{subsec31} The observed pattern of neutrino mixing in the reference 3-neutrino mixing scheme we are going to consider in what follows is characterised, as we have seen, by two large mixing angles $\theta_{12}$ and $\theta_{23}$, and one small mixing angle $\theta_{13}$: $\theta_{12}\cong 33^\circ$, $\theta_{23} \cong 45^\circ \pm 6^\circ$ and $\theta_{13} \cong 8.4^\circ$. These values can naturally be explained by extending the Standard Model (SM) with a flavour symmetry corresponding to a non-Abelian discrete (finite) group $G_f$. This symmetry is supposed to exist at some high-energy scale and to be broken at lower energies to residual symmetries of the charged lepton and neutrino sectors, described respectively by subgroups $G_e$ and $G_{\nu}$ of $G_f$. Flavour symmetry groups $G_f$ that have been used in this approach to neutrino mixing and lepton flavour include $A_4$ \cite{A4}, $S_4$ \cite{S4}, $T'$ \cite{Tprime}, $A_5$ \cite{A5}, $D_{n}$ (with $n=10,12$) \cite{D10,HGM}, $\Delta(27)$ \cite{Delta27}, the series $\Delta(6n^2)$ \cite{Delta6n2}, to name several \footnote{Some of the groups $T'$, $A_5$, etc. can be and have been used also for a unified description of the quark and lepton flavours, see, e.g., refs. \cite{} and references quoted therein. } (see, e.g., ref. \cite{} for definitions of these groups and discussion of their properties \footnote{ $A_4$ is the group of even permutations of 4 objects and the symmetry group of the regular tetrahedron. $S_4$ is the group of permutations of 4 objects and the symmetry group of the cube. $T'$ is the double covering group of $A_4$. $A_5$ is the icosahedron symmetry group of even permutations of five objects, etc. All these groups are subgroups of the group $SU(3)$. }). The numbers of elements, of generators and of irreducible representations of the groups $S_4$, $A_4$, $T'$, $A_5$, $D_{10}$ and $D_{12}$ are given in Table \ref{}. In what concerns the group $S_4$, it is well known that $S_4$ can be generated by two transformations, $S$ and $T$ (see, e.g.,\cite{}). However, in the context of non-Abelian discrete symmetry approach to neutrino mixing it often proves convenient to use the three generators $S$, $T$ and $U$ of $S_4$, indicated in Table 2, and these generators are widely used in the literature on the subject (see, e.g., the review article \cite{}). We will use the two generator formalism for the group $S_4$ in the analysis performed in sub-section \ref{}. The choice of the non-Abelian discrete groups $A_4$, $S_4$, $T'$, $A_5$, etc. is related, in particular, to the fact that they describe symmetries with respect to rotations on fixed large mixing angles and, correspondingly, lead to values of the neutrino mixing angles $\theta_{12}$ and $\theta_{23}$, which can differ from the measured values at most by sub-leading perturbative corrections, with $\theta_{13}$ typically (but not universally) predicted to be zero. \begin{table}[t!] \centering \renewcommand*{\arraystretch}{1.5} \begin{tabular}{|c|c|c|c|} \hline Group & Number of elements & Generators & Irreducible representations \\ \hline $S_4$ & $24$ & $S$, $T$ ($U$) & $\mathbf{1}$, $\mathbf{1^{\prime}}$, $\mathbf{2}$, $\mathbf{3}$, $\mathbf{3^{\prime}}$ \\ $A_4$ & $12$ & $S$, $T$ & $\mathbf{1}$, $\mathbf{1^{\prime}}$, $\mathbf{1^{\prime \prime}}$, $\mathbf{3}$ \\ $T^{\prime}$ & $24$ & $S$, $T$ ($R$) & $\mathbf{1}$, $\mathbf{1^{\prime}}$, $\mathbf{1^{\prime \prime}}$, $\mathbf{2}$, $\mathbf{2^{\prime}}$, $\mathbf{2^{\prime \prime}}$, $\mathbf{3}$ \\ $A_5$ & $60$ & $\tilde{S}$, $\tilde{T}$ & $\mathbf{1}$, $\mathbf{3}$, $\mathbf{3^{\prime}}$, $\mathbf{4}$, $\mathbf{5}$ \\ $D_{10}$ & $20$ & $A$, $B$ & $\mathbf{1}_1$, $\mathbf{1}_2$, $\mathbf{1}_3$, $\mathbf{1}_4$, $\mathbf{2}_1$, $\mathbf{2}_2$, $\mathbf{2}_3$, $\mathbf{2}_4$ \\ $D_{12}$ & $24$ & $\tilde{A}$, $\tilde{B}$ & $\mathbf{1}_1$, $\mathbf{1}_2$, $\mathbf{1}_3$, $\mathbf{1}_4$, $\mathbf{2}_1$, $\mathbf{2}_2$, $\mathbf{2}_3$, $\mathbf{2}_4$, $\mathbf{2}_5$ \\ \hline \end{tabular} \caption{Number of elements, generators and irreducible representations of some discrete groups.} \label{} \end{table} The requisite corrections can most naturally be provided by the unitary matrix $U_e$ which originates from the diagonalisation of the charged lepton mass term and enters into the expression of the PMNS neutrino mixing matrix (see, e.g., \cite{} and references quoted therein): \begin{equation} U_{\rm PMNS} = U_e^\dagger\,U_\nu\,. \label{UPMNSUeUnu} \end{equation} where $U_\nu$ is a unitary matrix coming from the diagonalisation of the neutrino mass term. More specifically, $U_e$ diagonalises the product $M_e M_e^\dagger$, where $M_e$ is the charged lepton mass matrix in the charged lepton mass term $\mathcal{L}_{\ell}(x)$ (written in the left-right convention): \begin{eqnarray} \label{Le} & \mathcal{L}_{\ell}(x) = -\, \overline{\tilde{l}_L(x)}\, (M_e)_{\tilde{l}\tilde{l}'}\, \tilde{l}'_R(x) +~{\rm h.c.}\,,\\[0.25cm] &U^\dagger_e\, M_e M_e^\dagger\, U_e = {\rm diag}(m^2_e,m^2_\mu,m^2_\tau)\,, \label{UeMe} \end{eqnarray} $\tilde{l}_L(x)$ and $\tilde{l}'_R(x)$, $\tilde{l},\tilde{l}' =\tilde{e},\tilde{\mu},\tilde{\tau}$, being respectively the $SU(2)$ doublet and singlet left-handed (LH) and right-handed (RH) components of the charged lepton fields in the basis in which the charged lepton mass term $\mathcal{L}_{\ell}(x)$ is not diagonal, while $m_e$, $m_\mu$ and $m_\tau$ are the masses of the charged leptons \footnote{The LH components of the fields of the electron, muon, and tauon, $l_L(x)$, $l=e,\mu,\tau$, are related to the fields $\tilde{l}_L(x)$ via the matrix $U_e$: $l_L(x) = (U^\dagger_e)_{l\tilde{l}}\tilde{l}_L(x)$. }. In certain classes of models, however, $U_e$ coincides with the unit $3\times3$ matrix and the requisite corrections are incorporated in a factor contained in the matrix $U_\nu$ (see, e.g., \cite{}). We shall assume in what follows that the weak-eigenstate neutrino fields (in the basis in which charged lepton mass term is not diagonal), $\nu_{\tilde{e}}(x)$, $\nu_{\tilde{\mu}}(x)$ and $\nu_{\tilde{\tau}}(x)$, possess a Majorana mass term, $\mathcal{L}_{M}^{\nu}(x)$, and thus the neutrinos with definite mass $\nu_1$, $\nu_2$ and $\nu_3$, are Majorana particles. In this case $U_\nu$ diagonalises the neutrino Majorana mass matrix $M_{\nu}$: \begin{eqnarray} \label{nuMajMass} & \mathcal{L}_{M}^{\nu}(x) = ~\frac{1}{2}~\nu^{\rm T}_{\tilde{l}' L}(x)~C^{-1}~M_{\nu \tilde{l}'\tilde{l}}~\, \nu_{\tilde{l} L}(x)~+~h.c.\,,~~~~ C^{-1}\,\gamma_{\alpha}\,C = -\,\gamma_{\alpha}^{\rm T}\,,\\[0.25cm] & U^{{\rm T}}_\nu\, M_{\nu}\,U_\nu = {\rm diag}(m_1,m_2,m_3)\,, \label{UnuMnu} \end{eqnarray} where $C$ is the charge conjugation matrix (see, e.g., \cite{BiPet87}). It should be noted, however, that the approach to neutrino mixing we are discussing can be employed also if $\nu_{1,2,3}$ are Dirac fermions (see, e.g., \cite{}), e.g., when the theory contains right-handed neutrino fields $\nu_{\tilde{l} R}(x)$ which form a Dirac mass term with the LH neutrino fields $\nu_{\tilde{l}'L}(x)$ $\tilde{l},\tilde{l}'=\tilde{e},\tilde{\mu},\tilde{\tau}$, and \footnote{The neutrino Dirac mass term in question originates \cite{} from an $SU(2)_L\times U(1)_{Y_w}$ invariant Yukawa-type term in the Lagrangian after the spontaneous breaking of the Standard Theory $SU(2)_L\times U(1)_{Y_w}$ symmetry.} the total lepton charge $L=L_e + L_\mu + L_\tau$ is conserved. In the approach under discussion it is standardly assumed that the LH neutrino fields, $\nu_{\tilde{l}L}(x)$, and the LH components of the charged lepton fields (in the basis in which charged lepton mass term is not diagonal) $\tilde{l}_L(x)$, which form an $SU(2)_L$ doublet in the Standard Theory, are assigned to the same r-dimensional irreducible unitary representation $\rho_r(g_f)$ of the Group $G_f$, $g_f$ being an element of $G_f$. Thus, under the action of $G_f$, $\nu_{\tilde{l}L}(x)$ and $\tilde{l}_L(x)$ transform as follows: \begin{eqnarray} \label{gfnu} & \nu_{\tilde{l}L}(x)\,\xrightarrow{G_f}\, (\rho_r(g_f))_{\tilde{l}\tilde{l}'}\, \nu_{\tilde{l}'L}(x)\,,~~~ \quad\, g_f \in G_f\,,\\[0.25cm] & \tilde{l}_L(x)\, \xrightarrow{G_f}\, (\rho_r(g_f))_{\tilde{l}\tilde{l}'} \tilde{l}'_L(x)\,,~~~ \tilde{l} = \tilde{e},\tilde{\mu},\tilde{\tau}\,. \label{gflL} \end{eqnarray} In the cases of $G_f = A_4$, $S_4$, $T'$ and $A_5$, which possess 3-dimensional irreducible representations, $\rho(g_f)$ is standardly taken to be a 3-dimensional irreducible unitary representation {\bf 3}, $\rho_r(g_f) = \rho_3(g_f)$. This is equivalent to the assumption of unification of the three lepton families at some high energy scale. We are going to consider this choice in what follows \footnote{In specific models the choice $\rho_r(g_f) = \rho_3(g_f)$ is usually accompanied by the assumption that $\tilde{e}_R(x)$, $\tilde{\mu}_R(x)$ and $\tilde{\tau}_R(x)$ transform as singlet irreducible representations of $G_f$ (see, e.g., \cite{}). }. At low energies the flavour symmetry $G_f$ has necessarily to be broken so that the electron, muon and tauon as well as the three neutrinos with definite mass $\nu_1$, $\nu_2$ and $\nu_3$, can get different masses. The breaking of $G_f$ is realised in specific models by scalar ``flavon'' fields, which are singlets with respect to the Standard Theory gauge group but transform under certain irreducible representations of $G_f$ and acquire non-zero vacuum expectation values (VEVs), thus breaking $G_f$ spontaneously. The breaking of the flavour symmetry $G_f$ can leave certain subgroups of $G_f$, $G_e$ and $G_\nu$, unbroken in the charged lepton and neutrino sectors. The unbroken symmetries $G_e\in G_f$ and $G_\nu \in G_f$ are {\it residual symmetries} of the charged lepton and neutrino mass matrices. The existence of a residual symmetry $G_e \in G_f$ in the charged lepton sector implies that $M_e M_e^\dagger$ is invariant with respect to the action of $G_e$ on the LH components of the charged lepton fields $\tilde{l}_L(x)$, $\tilde{l} = \tilde{e},\tilde{\mu},\tilde{\tau}$: \begin{equation} \rho_r(g_e)^{\dagger} M_e M_e^{\dagger} \rho_r(g_e) = M_e M_e^{\dagger} \,, \label{GeMe} \end{equation} where $g_e$ is an element of $G_e$ and $\rho_r(g_e)$ gives the action of $G_e$ on $\tilde{l}_L(x)$. Similarly, if $G_\nu$ is the residual symmetry of the neutrino Majorana mass matrix $M_{\nu}$ one has: \begin{equation} \rho_r(g_{\nu})^T M_{\nu} \rho_r(g_{\nu}) = M_{\nu} \,, \label{GnuMnu} \end{equation} where $g_\nu$ is an element of $G_\nu$ and $\rho_r(g_{\nu})$ determines the action of $G_\nu$ on $\nu_{\tilde{l}L}(x)$, $\tilde{l} = \tilde{e},\tilde{\mu},\tilde{\tau}$. From eq. (\ref{GnuMnu}) we get: \begin{equation} \rho_r(g_{\nu})^\dagger M^\dagger_{\nu}\,M_{\nu}\,\rho_r(g_{\nu}) = M^\dagger_{\nu}\, M_{\nu}\,. \label{GnuMnuMnu} \end{equation} It follows from eqs. (\ref{GeMe}) and (\ref{GnuMnuMnu}) that $M_e M_e^{\dagger}$ commutes with $\rho_r(g_e)$, while $M^\dagger_{\nu}M_{\nu}$ commutes with $\rho_r(g_{\nu})$. This implies that $M_e M_e^{\dagger}$ and $\rho_r(g_e)$ are diagonalised with one and the same matrix $U_e$, and that similarly, $M^\dagger_{\nu}M_{\nu}$ and $\rho_r(g_{\nu})$ are diagonalised by the same matrix $U^\circ_\nu$: \begin{eqnarray} \label{Uerhoge} U^\dagger_e\,\rho_r(g_e)\,U_e = \rho_r^{\diag}(g_e)\,,\\[0.30cm] \label{Uonurhognu} (U^\circ_\nu)^\dagger\,\rho_r(g_{\nu})\,U^\circ_\nu = \rho_r^{\diag}(g_\nu)\,. \end{eqnarray} Given $G_f$, $\rho_r(g_f)$, and non-trivial $G_e$, $\rho_r(g_e)$ is uniquely determined. As a consequence, the matrix $U_e$ diagonalising $\rho_r(g_e)$ (and $M_e M_e^{\dagger}$), which enters into the expression for the PMNS matrix $U$, is either completely determined or significantly constrained \footnote{Obviously, if $G_e$ is trivial consisting just of the unit element of $G_f$, i.e., if $G_f$ is completely broken in the charged lepton sector, $U_e$ would not be constrained.}. Similarly, for given $G_f$, $\rho_r(g_f)$, and non-trivial $G_\nu$, the matrix $U^\circ_\nu$ disgonalising $\rho_r(g_\nu)$ (and $M^\dagger_{\nu}\,M_{\nu}$) will either be completely determined or strongly constrained. One can show that the matrix $U_\nu$ diagonalising the neutrino Majorana mass matrix $M_{\nu}$ and the matrix $U^\circ_\nu$ diagonalising $M^\dagger_{\nu}\,M_{\nu}$ are related, in general, in the following way: \begin{equation} U_{\nu} = U^\circ_{\nu}\, P^\circ\,,~~~~ P^\circ={\rm diag}(1,e^{i\frac{\xi_{21}}{2}},e^{i\frac{\xi_{31}}{2}})\,. \label{UnuU0nu} \end{equation} The phases $\xi_{21}$ and $\xi_{31}$ contribute respectively to the Majorana phases $\alpha_{21}$ and $\alpha_{31}$ of the PMNS matrix (see eq. (\ref{VP})). Thus, within the discussed approach the PMNS neutrino mixing matrix $U = U^\dagger_e U_{\nu}$ is either completely determined or else has a constrained form. The form of $U$ one obtains depends on the choices of $G_f$, $\rho_r(g_f)$, $G_e$ and $G_{\nu}$. It should be clear from the preceding discussion that the residual symmetries $G_e$ and $G_\nu$, in particular, play a crucial role in obtaining a specific form of the PMNS matrix. If, in particular, $G_e \equiv G_\nu$, we would have $U_e = U^\circ_{\nu}$ and the PMNS matrix will be trivial, which is ruled out by the data. The largest possible exact symmetry of the charged lepton Dirac mass term $\mathcal{L}_{\ell}(x)$ (mass matrix $M_e$) is $U(1)\times U(1) \times U(1)$. The largest possible exact symmetry of the neutrino Majorana mass term $\mathcal{L}_{M}^{\nu}(x)$, with mass matrix $M_\nu$ having three non-zero non-degenerate eigenvalues, is $Z_2\times Z_2\times Z_2$. Making the standardly used simplifying assumption that $G_f$ is a subgroup of $SU(3)$, the largest possible symmetries of $\mathcal{L}_{\ell}(x)$ and $\mathcal{L}_{M}^{\nu}(x)$ reduce to $U(1)\times U(1)$ and $Z_2\times Z_2$ owing to the $SU(3)$ determinant condition. The residual symmetry group $G_e$ should be a subgroup of $U(1)\times U(1)$, while $G_\nu$ should be contained in $Z_2\times Z_2$ ($U(1)\times U(1)$) in the case of massive Majorana (Dirac) neutrinos. Thus, $G_e$ and $G_\nu$ should be Abelian groups. It follows from the preceding discussion that the possible discrete symmetries $G_e$ of the charged lepton mass term leaving $M_e M_e^{\dagger}$ invariant are: i) $G_e = Z_n$, with integer $n \geq 2$, or ii) $Z_m \times Z_k$, with integers $m,\,k \geq 2$. The maximal symmetry $G_{\nu}$ of the Majorana mass term of the LH flavour neutrino fields $\nu_{\tilde{l}L}(x)$ is the $Z_2 \times Z_2$ (sometimes referred to as the Klein four group) symmetry. $G_{\nu}$ can obviously be just $Z_2$. These two possible types of $G_{\nu}$ are associated with two approaches in constructing realistic models of lepton flavour: the {\it direct} approach with $G_{\nu} = Z_2 \times Z_2$ and the {\it semi-direct} approach with $G_{\nu} = Z_2$. Since the neutrino Majorana mass term (mass matrix $M_\nu$) possesses always a $Z_2 \times Z_2$ symmetry, the second $Z_2$ factor appears accidentally in models employing the semi-direct approach. The symmetry $G_f$ might be completely broken by the neutrino Majorana mass term $\mathcal{L}_{M}^{\nu}(x)$, i.e., the $Z_2 \times Z_2$ group of symmetry of $\mathcal{L}_{M}^{\nu}(x)$ might not necessarily be a subgroup of $G_f$. This corresponds to the so-called {\it indirect} approach in lepton flavour model building. The group $A_4$, for example, has three subgroups $Z_2$, four subgroups $Z_3$ and one subgroup $Z_2\times Z_2$, while $S_4$ has nine $Z_2$, four $Z_3$, three $Z_4$ and four $Z_2 \times Z_2$ subgroups. The bigger groups $T'$, $A_5$, etc. all have a certain number of $Z_2$, $Z_3$, $Z_2 \times Z_2$, etc. subgroups \footnote{For complete list of the subgroups of the groups $T'$, $A_5$, $\Delta(6n^2)$ and of the larger groups employed in the discrete flavour symmetry approach to neutrino mixing see, e.g., ref. \cite{}.}. As we have indicated in the Introduction, one of the main characteristics of the discussed approach to neutrino mixing based on discrete flavour symmetries is that it leads to certain specific predictions for the values of, and/or correlations between, the low-energy neutrino mixing parameters, which can be tested experimentally. These predictions depend on the chosen $G_f$, $\rho(g_f)$, $G_e$ and $G_{\nu}$. We give a few examples \cite{ \noindent {\bf I.} In a large class of models one gets $\sin^2\theta_{23} = 0.5$.\\ \noindent {\bf II.} In different class of models one finds that the values of $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$ are correlated: $\sin^2\theta_{23} = 0.5(1 \mp \sin^2\theta_{13} + O(\sin^4\theta_{13}))$.\\ \noindent {\bf III.} In certain models $\sin^2\theta_{23}$ is predicted to have specific values which differ significantly from those in cases {\bf I} and {\bf II} \cite{}: $\sin^2\theta_{23} = 0.455$; or 0.463; or 0.537; or 0.545, the uncertainties in these predictions being insignificant.\\ \noindent {\bf IV.} Certain class of models predict a correlation between the values of $\sin^2\theta_{12}$ and $\sin^2\theta_{13}$: $\sin^2\theta_{12} = 1/(3\cos^2\theta_{13}) = (1 + \sin^2\theta_{13} + O(\sin^4\theta_{13}))/3 \cong 0.340$, where we have used the b.f.v. of $\sin^2\theta_{13}$.\\ \noindent {\bf V.} In another class of models one still finds a correlation between the values of $\sin^2\theta_{12}$ and $\sin^2\theta_{13}$, which, however, differs from that in Case {\bf IV}: $\sin^2\theta_{12} = (1-3\sin^2\theta_{13})/(3\cos^2\theta_{13}) = (1 - 2\sin^2\theta_{13} + O(\sin^4\theta_{13}))/3 \cong 0.319$, where we have used again the b.f.v. of $\sin^2\theta_{13}$.\\ \noindent {\bf VI.} In large classes of models in which the elements of the PMNS matrix are predicted to be functions of just one real continuous free parameter (``one-parameter models''), the Dirac and the Majorana CPV phases have ``trivial'' CP conserving values 0 or $\pi$. In certain one-parameter schemes, however, the Dirac phase $\delta = \pi/2$ or $3\pi/2$.\\ \noindent {\bf VII.} In theories/models in which the elements of the PMNS matrix are functions of two (angle) or three (two angle and one phase) parameters, the Dirac phase $\delta$ satisfies a sum rule by which $\cos\delta$ is expressed in terms of the three neutrino mixing angles $\theta_{12}$, $\theta_{23}$, $\theta_{13}$ and one (or more) fixed (known) parameters $\theta^{\nu}$ which depend on the discrete symmetry $G_f$ employed and on the residual symmetries $G_e$ and $G_{\nu}$ \cite{}: \begin{equation} \cos\delta = \cos\delta(\theta_{12},\theta_{23},\theta_{13};\theta^{\nu}). \label{cosdSumR} \end{equation} In these cases the $J_{CP}$ factor which determines the magnitude of CP violation effects in neutrino oscillations, is also completely determined by the values of the three neutrino mixing angles and the symmetry parameter(s) $\theta_{\nu}$: \begin{equation} J_{CP} = J_{CP}(\theta_{12},\theta_{23},\theta_{13},\delta) = J_{CP}(\theta_{12},\theta_{23},\theta_{13};\theta^{\nu})\,. \label{JCPSumR} \end{equation} If in the model considered a correlation of the type corresponding to Case {\bf II} (case {\bf IV} or case {\bf V}) takes place, $\theta_{23}$ ($\theta_{12}$) in the sum rule for $\cos\delta$ and the expression for the $J_{CP}$ factor has to be expressed in terms of $\theta_{13}$ using the correlation. The predictions listed above, and therefore the respective models, can be and will be tested in the currently running (T2K \cite{T2K20152016} and NO$\nu$A \cite{NOvA2018}) and planned future (JUNO \cite{JUNO}, T2HK \cite{T2HK2015}, T2HKK \cite{} and DUNE \cite{DUNE2016}) experiments. As an illustration of the preceding discussion we will consider first the example of the tri-bimaximal mixing as an underlying symmetry form of the matrix $U_{\nu}$ ($U^\circ_{\nu}$). \subsection{Symmetry Forms of $U_{\nu}$: Tri-bimaximal Mixing} \label{subsec32} Consider the case of $G_f =S_4$, i.e., the group of permutations of four objects. $S_4$ is isomorphic to the group of rotational symmetries of the cube. It has 24 elements, two singlet, one doublet and two triplet irreducible representations. As was indicated earlier, we will assume that $\rho_r(g_f) = \rho_3(g_f)$, i.e., that $\nu_{\tilde{l}L}(x)$ and $\tilde{l}_L(x)$ transform under one of the two 3-dimensional irreducible unitary representations of $S_4$. In what follows, with the exception of sub-section \ref{}, we will work with the three generators of the group $S_4$, $S$, $T$ and $U$. These generators satisfy the following presentation rules (see, e.g., \cite{}): \begin{equation} S^2 = T^3 = (ST)^3 = U^2 = (TU)^2 = (SU)^2 = (STU)^4 = {\bf 1}\,, \label{S4STU} \end{equation} ${\bf 1}$ being the unit (identity) element of $S_4$, i.e., it is the $3\times3$ unit matrix in the case of the triplet representations of $S$, $T$ and $U$. In what follows we will use the basis \cite{} in which $S$, $T$ and $U$ have the following form in the two triplet representations \footnote{As can be shown, the results one obtains for the form of the PMNS matrix are independent of the chosen basis for the generators of the discrete symmetry group $G_f$.}: \begin{equation} S = \frac{1}{3} \begin{pmatrix} -1 & 2 & 2 \\ 2 & -1 & 2 \\ 2 & 2 & -1 \end{pmatrix}\,, \quad T = \begin{pmatrix} 1 & 0 & 0 \\ 0 & \omega^2 & 0 \\ 0 & 0 & \omega \end{pmatrix} \quad {\rm and} \quad U = \mp\, \begin{pmatrix} 1 & 0 & 0 \\ 0 & 0 & 1 \\ 0 & 1 & 0 \end{pmatrix}\,, \label{S4STU3drep} \end{equation} where $\omega = e^{2\pi i/3}$. For simplicity we use the same notation ($S$, $T$ and $U$) for the generators and their 3-dimensional representation matrices. Assume next that \cite{} (see also, e.g., \cite{}) \begin{equation} G_e = Z^T_3 = \{1,T,T^2\}\,,~~ G_\nu = Z^S_2\times Z^U_2 = \{1,S,U,SU\}\,, \label{GeZ3GnuZ2Z2} \end{equation} where $Z^T_3$ and $Z^S_2\times Z^U_2$ are two specific $Z_3$ and $Z_2\times Z_2$ subgroups of $S_4$~ \footnote{$S$ and $U$ are order two elements of $S_4$ (since $S^2 = U^2 = {\bf 1}$) and they commute. Correspondingly, $Z^S_2\times Z^U_2 = \{1,S,U,SU\}$ is a subgroup of $S_4$. Similarly, $T$ is order 3 element of $S_4$ and $Z^T_3 = \{1,T,T^2\}$ is a subgroup of $S_4$.}. In this case we have, in particular: $\rho_3(g_e) = 1,T,T^2$, $T$ being the diagonal matrix given in eq. (\ref{S4STU3drep}). As a consequence, $U_e$, which diagonalises $\rho_3(g_e)= T$, is just a diagonal phase matrix, whose phases are unphysical (they can be absorbed by the charged lepton fields in the weak charged lepton current of the weak interaction Lagrangian), while $M_e$ is a diagonal matrix with the masses of the electron, muon and tauon as diagonal elements. It follows from eq. (\ref{S4STU}) that $\rho_3(g_\nu) = S$ and $\rho_3(g'_\nu) = U$ commute. In the triplet representation of the generators of $S_4$ employed by us, eq. (\ref{S4STU3drep}), $S$ and $U$ are real symmetric matrices. Thus, they are diagonalised by a real orthogonal matrix. The matrix which diagonalises both $\rho_3(g_\nu) = S$ and $\rho_3(g'_\nu) = U$, with $S$ and $U$ given in eq. (\ref{S4STU3drep}), is the orthogonal tri-bimaximal (TBM) mixing matrix \cite{TBM}: \begin{equation} U^\circ_{\nu}= V_{\rm TBM} = \begin{pmatrix} \sqrt{\dfrac{2}{3}} & \dfrac {1}{\sqrt{3}} & 0 \vspace{0.2cm} \\ - \dfrac{1}{\sqrt{6}} & \dfrac{1}{\sqrt{3}} & - \dfrac{1}{\sqrt{2}} \vspace{0.2cm} \\ - \dfrac{1}{\sqrt{6}} & \dfrac{1}{\sqrt{3}} & \dfrac{1}{\sqrt{2}} \end{pmatrix} \;. \label{VTBM} \end{equation} Indeed, it is not difficult to check that \begin{eqnarray} \label{VTBMS} V^\dagger_{\rm TBM}\,S\,V_{\rm TBM} = {\rm diag}(-1,1,-1)\,,\\[0.30cm] \label{VTBMU} V^\dagger_{\rm TBM}\,U\,V_{\rm TBM} = \pm\, {\rm diag}(1,1,-1)\,.% \end{eqnarray} Thus, in the discussed case of $S_4$ symmetry and residual symmetries $G_e = Z^T_3$ and $G_\nu = Z^S_2\times Z^U_2$, the PMNS matrix has the TBM form \cite{}, $U = U^\circ_{\nu}\, P^\circ = V_{\rm TBM}\, P^\circ$. We can cast $V_{\rm TBM}$ in the form: \begin{equation} V_{\rm TBM} = R_{23} \left (\theta^\nu_{23} \right) R_{13}\left (\theta^\nu_{13} \right) R_{12}\left(\theta^\nu_{12}\right)\,,~~ \theta^\nu_{23} = -\,\pi/4,~\theta^\nu_{13} = 0\,, ~\theta^\nu_{12}=\sin^{-1}\frac{1}{\sqrt{3}}\,, \end{equation} where $R_{23}(\theta^\nu_{23})$, $R_{13}(\theta^\nu_{13})$ and $R_{12}(\theta^\nu_{12})$ are $3\times3$ orthogonal matrices describing rotations in the 2-3, 1-3 and 1-2 planes, respectively. We see that in the case of the TBM symmetry form we have $\sin^2\theta^\nu_{12} = 1/3$, $\sin^2\theta^\nu_{23} = 1/2$ and $\sin^2\theta^\nu_{13} = 0$. Without additional corrections leading to $\theta_{13} \cong 0.15 \neq 0$, the TBM symmetry from of the PMNS matrix is ruled out by the data. We will consider next two cases of realistic models based on the flavour symmetries $G_f = A_4$ and $G_f = S_4$, in which the corrections to the underlying symmetry form of the PMNS matrix are obtained by ``decreasing'' the residual symmetry $G_\nu$ from $Z_2\times Z_2$ symmetry to $Z_2$, \subsection{Neutrino Mixing from $A_4$ Symmetry} \label{subsec33} The group $A_4$ has two generators $S$ and $T$, which satisfy the presentation rules given in eq. (\ref{S4STU}). In the triplet representation of interest and in the Altarelli-Feruglio basis \cite{}, $S$ and $T$ have the form given in eq. (\ref{S4STU3drep}). Assume next that (see, e.g., \cite{}) \begin{equation} G_e = Z^T_3 = \{1,T,T^2\}\,,~~~G_\nu = Z^S_2 = \{1,S\}\,, \label{GeZ3GnuZ2} \end{equation} where $Z^T_3$ and $Z^S_2$ are two specific $Z_3$ and $Z_2$ subgroups of $A_4$. In this case we have, in particular: $\rho(g_e) = 1,T,T^2$, $T$ being the diagonal matrix given in eq. (\ref{S4STU3drep}). As a consequence, $U_e$, which diagonalises $\rho(g_e)= T$, as in the case discussed in the preceding subsection, is just a diagonal phase matrix, whose phases are unphysical, while $M_e$ is a diagonal matrix with the masses of the electron, muon and tauon as diagonal elements. The most general matrix which diagonalises $\rho(g_\nu) = S$, with $S$ given in eq. (\ref{S4STU3drep}) has the form: \begin{equation} U^\circ_{\nu} = V_{\rm TBM}\,U_{13}(\theta^\nu_{13},\alpha)\,, \label{UonuA4} \end{equation} where $V_{\rm TBM}$ is the tri-bimaximal (TBM) mixing matrix given in eq. (\ref{VTBM}), and \begin{equation} U_{13}(\theta^\nu_{13},\alpha) = \begin{pmatrix} \cos \theta^{\nu}_{13} & 0 & \sin \theta^{\nu}_{13}\,e^{i\alpha} \vspace{0.2cm} \\ 0 & 1 & 0 \\ -\,\sin \theta^{\nu}_{13}\,e^{-i\alpha} & 0 & \cos \theta^{\nu}_{13} \end{pmatrix} \;. \label{U13} \end{equation} The angle $\theta^{\nu}_{13}$ and the phase $\alpha$ in $U_{13}(\theta^\nu_{13},\alpha)$ are arbitrary free parameters. Indeed, it is not difficult to convince oneself that \begin{equation} S = U^\circ_{\nu}\,{\rm diag}(-1,1,-1)\,(U^\circ_{\nu})^\dagger = V_{\rm TBM}\,{\rm diag}(-1,1,-1)\,V^T_{\rm TBM}\,. \label{Sdiag} \end{equation} Thus, the matrix $U_{13}(\theta^\nu_{13},\alpha)$ appears in the matrix $U^\circ_{\nu}$ diagonalising $S$ as a consequence of the degeneracy of the first and third eigenvalues of $S$. We see that in the $A_4$ model considered, the underlying symmetry form of the PMNS matrix is the tri-bimaximal mixing, $V_{\rm TBM}$. The matrix $U_{13}(\theta^\nu_{13},\alpha)$ provides the necessary corrections to $V_{\rm TBM}$ that lead, in particular, to $\theta_{13} \neq 0$. Thus, the model considered contains two free parameters - the angle $\theta^\nu_{13}$ and the phase $\alpha$. Taking into account the results for the forms of $U_e$ and $U^\circ_{\nu}$ we have obtained and eq. (\ref{UnuU0nu}), we get the following expression for the PMNS matrix: \begin{equation} U_{\rm PMNS} = U^\circ_{\nu}\,P^\circ = V_{\rm TBM}\,U_{13}(\theta^\nu_{13},\alpha)\,P^\circ = \begin{pmatrix} \sqrt{\dfrac{2}{3}}c & \dfrac {1}{\sqrt{3}} & \sqrt{\dfrac{2}{3}}s\,e^{i\alpha} \vspace{0.2cm} \\ - \dfrac{c}{\sqrt{6}} +\dfrac{s}{\sqrt{2}}e^{-i\alpha} & \dfrac{1}{\sqrt{3}} & - \dfrac{c}{\sqrt{2}} - \dfrac{s}{\sqrt{6}}\,e^{i\alpha}\vspace{0.2cm} \\ - \dfrac{c}{\sqrt{6}} - \dfrac{s}{\sqrt{2}}e^{-i\alpha} & \dfrac{1}{\sqrt{3}} & \dfrac{c}{\sqrt{2}} - \dfrac{s}{\sqrt{6}}\,e^{i\alpha} \end{pmatrix}\,P^\circ\;, \label{A4PMNS} \end{equation} where $c \equiv \cos\theta^\nu_{13}$ and $s \equiv \sin\theta^\nu_{13}$. We will consider next the phenomenological predictions of the discussed $A_4$ model of neutrino mixing. Comparing, for example, the absolute values of the elements of the first rows of the PMNS matrix in eq. (\ref{A4PMNS}) and in the standard parametrisation, eqs. (\ref{VP}) and (\ref{}), we get: \begin{equation} \sin^2\theta_{13} = \dfrac{2}{3}\,s^2\,,~~ \sin^2\theta_{12} \cos^2\theta_{13} = \dfrac{1}{3}\,. \label{A4Resth12th12} \end{equation} Comparing the $U_{\mu 3}$ elements and using the first relation in the preceding equation we find: \begin{equation} \sin^2\theta_{23} = \dfrac{1}{c^2_{13}}\, \large |\dfrac{c}{\sqrt{2}} + \dfrac{s}{\sqrt{6}}\,e^{i\alpha}\, \large |^2 = \dfrac{1}{2} + \dfrac{s_{13}}{2}\, \dfrac{(2 - 3\,s^2_{13})^{\frac{1}{2}}}{(1 - s^2_{13})}\,\cos\alpha\,. \label{A4s2231} \end{equation} To leading order in $s_{13}$ we have: \begin{equation} \dfrac{1}{2} - \dfrac{s_{13}}{\sqrt{2}} \ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$<$}\ \sin^2\theta_{23} \ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$<$}\ \dfrac{1}{2} + \dfrac{s_{13}}{\sqrt{2}} \,,~~{\rm or}~~ 0.391\ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$<$}\ \sin^2\theta_{23} \ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$<$}\ 0.609\,, \label{A4s2232} \end{equation} where the numerical values correspond to the maximal allowed value of $\sin^2\theta_{13}$ at $3\sigma$ C.L. The interval of possible values of $\sin^2\theta_{23}$ in eq. (\ref{A4s2232}) lies within the $3\sigma$ ranges of experimentally allowed values of $\sin^2\theta_{23}$ for NO and IO spectra, quoted in Table \ref{}. Further, using the constraint $|U_{\mu 2}|^2 = 1/3$ (or $|U_{\tau 2}|^2 = 1/3$) following from the form of $U_{\rm PMNS}$ in eq. (\ref{A4PMNS}), we obtain the following sum rule for the Dirac phase $\delta$: \begin{equation} \cos\delta = \dfrac{\cos2\theta_{23}\,\cos2\theta_{13}} {\sin2\theta_{23}\,\sin\theta_{13}\,(2-3\sin^2\theta_{13})^{\frac{1}{2}}}\,, \label{cosdA4} \end{equation} where we have expressed $\cos\theta_{12}\sin\theta_{12}$ in terms of $\sin\theta_{13}$ using eq. (\ref{A4Resth12th12}). It follows from the preceding brief discussion that $\theta_{13}$ and $\theta_{23}$ of the standard parametrisation of the PMNS matrix are equivalent to the two independent parameters $\theta^\nu_{13}$ and $\alpha$ of the considered $A_4$ model, while the angle $\theta_{12}$ and the Dirac phase $\delta$ can be considered as functions of $\theta_{13}$ and $\theta_{23}$~ \footnote{Actually, any pair of the four parameters $\theta_{12}$, $\theta_{23}$, $\theta_{13}$ and $\delta$ can play the role of the two independent parameters of the model.}. The phase $\alpha$ and the Dirac phase $\delta$ are related via \begin{equation} \sin2\theta_{23}\, \sin\delta = \sin\alpha\,. \label{sindsinalphaA4} \end{equation} This relation follows from the equality between the expressions of the rephasing invariant $J_{\rm CP}$, eq. (\ref{JCPstandparam}), in the standard parametrisation of the PMNS matrix and in the parametrisation defined in eq. (\ref{A4PMNS}). As it is not difficult to show, the phase $\alpha$ contributes also to the Majorana phase $\alpha_{31}$ of the PMNS matrix, eqs. (\ref{VP}) and (\ref{}): \begin{equation} \dfrac{\alpha_{31}}{2} = \dfrac{\xi_{31}}{2} + \alpha_{2} + \alpha_3\,, \label{alpha31A4} \end{equation} where \begin{equation} \label{alpha2alpha3} \alpha_2 = {\rm arg} \big(-\,\dfrac{c}{\sqrt{2}} - \dfrac{s}{\sqrt{6}}\,e^{i\alpha}\big)\,,~~~ \alpha_3 = {\rm arg} \big(\dfrac{c}{\sqrt{2}} - \dfrac{s}{\sqrt{6}}\,e^{i\alpha}\big)\,, \end{equation} \begin{eqnarray} \label{alpha2} \sin\alpha_2 = -\,\dfrac{s}{\sqrt{6}}\,\dfrac{\sin\alpha}{s_{23}\,c_{13}} = -\,\tan\theta_{13}\,\cos\theta_{23}\,\sin\delta\,,\\[0.30cm] \label{alpha3} \sin\alpha_3 = -\,\dfrac{s}{\sqrt{6}}\,\dfrac{\sin\alpha}{c_{23}\,c_{13}} = -\,\tan\theta_{13}\,\sin\theta_{23}\,\sin\delta\,, \end{eqnarray} where we have used eq. (\ref{sindsinalphaA4}). In eqs. (\ref{alpha2}) and (\ref{alpha3}), $\sin\delta$ can be considered as a function of $\theta_{23}$ and $\theta_{13}$ (see eq. (\ref{cosdA4})). We also have: \begin{equation} \sin(\alpha - \alpha_2 - \alpha_3) = -\,\sin\delta\,. \label{sindsinalphab2b3A4} \end{equation} That the phases $\alpha_{2}$ and $\alpha_3$ contribute to the Majorana phase $\alpha_{31}$ can be seen by casting the parametrisation of $U_{\rm PMNS}$ in eq. (\ref{A4PMNS}) in the standard parametrisation form, eqs. (\ref{VP}) and (\ref{}). This can be done by multiplying the matrix in eq. (\ref{A4PMNS}) on the right by $P^*_{33}P_{33}$ with $P_{33} = {\rm diag}(1,1,e^{i(\alpha_2 + \alpha_3)})$, and absorbing $P_{33}$ in $P^\circ$. The phases $e^{-i\alpha_3}$ and $e^{-i\alpha_2}$, which after that appear respectively in the $U_{\mu 3}$ and $U_{\tau 3}$ elements of $U_{\rm PMNS}$ in eq. (\ref{A4PMNS}), are removed from these elements by phase redefinition of the $\mu^\mp$ and $\tau^\mp$ fields in the weak charged lepton current (\ref{CC}). As a consequence of these simple manipulations the phase factor $e^{i\alpha_3}$ ($e^{i\alpha_2}$) appears in the $U_{\mu 1}$ and $U_{\mu 2}$ ($U_{\tau 1}$ and $U_{\tau 2}$) elements of $U_{\rm PMNS}$, while the phase factor $e^{i\alpha}$ of the $U_{e 3}$ element (see eq. (\ref{A4PMNS})) changes to $e^{i(\alpha - \alpha_2 - \alpha_3)}$. The phases in $P_{33}P^\circ$ contribute to the Majorana phases $\alpha_{21}/2$ and $\alpha_{31}/2$. The phenomenology of neutrino mixing described by the PMNS matrix given in (\ref{A4PMNS}), apart from the relation (\ref{sindsinalphaA4}) and the contribution of the phase $\alpha$ to the Majorana phase $\alpha_{31}$, eqs. (\ref{alpha31A4}) - (\ref{alpha3}), as well as of the relation (\ref{sindsinalphab2b3A4}), was discussed in \cite{}. The prediction for $\sin^2\theta_{12}$ in eq. (\ref{A4Resth12th12}) and the sum rule for the Dirac phases $\delta$, eq. (\ref{cosdA4}), can also be obtained from the general results on neutrino mixing in the case of $A_4$ lepton flavour symmetry derived in \cite{}. Thus, the $A_4$ model considered predicts \footnote{The result for $\sin^2\theta_{12}$ and the sum rule for $\cos\delta$ can be obtained respectively from eq. (58) in subsection 4.1 and Table 3 (Case B1) in \cite{} by setting $\sin^2\theta^\circ_{12} = 1/3$ and $\sin^2\theta^\circ_{23} = 1/2$. } i) a correlation between the values of $\sin^2\theta_{12}$ and $\sin^2\theta_{13}$: $\sin^2\theta_{12} = 1/(3(1-\sin^2\theta_{13}))$, ii) an interval of possible values of $\sin^2\theta_{23}$, which depends on $\sin\theta_{13}$, and ii) a sum rule for the Dirac CPV phase $\delta$ by which $\cos\delta$ is expressed in terms of the two measured neutrino mixing angles $\theta_{13}$ and $\theta_{23}$. In this model the Majorana phases $\alpha_{21}$ and $\alpha_{31}$ remain undetermined due to the contribution respectively of the phases $\xi_{21}$ and $\xi_{31}$, which are not fixed. The correlation between $\sin^2\theta_{12}$ and $\sin^2\theta_{13}$ leads to the prediction $\sin^2\theta_{12} = 0.340$, where we have employed the best fit value of $\sin^2\theta_{13}$ in Table \ref{}. This value lies outside the $2\sigma$, but is inside the $3\sigma$, currently allowed intervals of values of $\sin^2\theta_{12}$. Using the best fit values of $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$ for the NO and IO neutrino mass spectrum, given in Table \ref{}, and the sum rule for $\cos\delta$, eq. (\ref{cosdA4}), we find: \begin{eqnarray} \label{cosdA4NO} & \cos\delta = 0.728\,(-\,0.865)\,,~ \delta = \pm 43.32^\circ\,( 180^\circ \pm 30.07^\circ)\,,~~{\rm NO~(IO)}\,. \label{cosdA4IO} \end{eqnarray} If instead we use the best fit values for NO (IO) spectrum of $\sin^2\theta_{23} = 0.538~(0.554)$ and $\sin^2\theta_{13} = 0.02206~(0.02227)$ reported in \cite{NuFITv32Jan2018} we get: \begin{eqnarray} \label{cosdA4NO2018} & \cos\delta = -\,0.353\,(-500)\,, ~\delta = 180^\circ \pm 69.4^\circ\, (180^\circ \pm 60.0^\circ)\,,~~{\rm NO~(IO)}\,. \label{cosdA4IO2018} \end{eqnarray} Thus, as a consequence primarily of the fact that $\cos\delta\propto \cos2\theta_{23}$, the predictions for $\cos\delta$, and correspondingly of $\delta$, depend strongly on the values of $\sin^2\theta_{23}$ and can differ significantly for the two neutrino mass orderings. The values of $\delta = 43.32^\circ$, $110.6^\circ$ and $120^\circ$ are strongly disfavored (if not ruled out) by the current data. It should be added that the difference between the predictions of $\cos\delta$ ($\delta$) for NO and IO neutrino mass spectra are due to the difference between the best fit values of $\sin^2\theta_{23}$ for the two spectra (see Table \ref{} and eq. (\ref{})). For $\sin^2\theta_{23} = 0.5$ we have for both spectra $\cos\delta=0$, or $\delta = \pi/2,3\pi/2$, with $\delta = \pi/2$ strongly disfavored by the current data. It follows from the preceding results that the high precision measurement of $\sin^2\theta_{12}$ combined with the data on $\sin^2\theta_{13}$ will allow to critically test the predicted correlation between $\sin^2\theta_{12}$ and $\sin^2\theta_{13}$ of the considered $A_4$ model. The high precision measurement of $\sin^2\theta_{23}$, the data on $\sin^2\theta_{13}$ and a sufficiently precise determination of $\delta$ will make it possible to test the sum rule predictions for $\delta$ of the model. With the indicated tests the $A_4$ model of neutrino mixing discussed in the present subsection will be either verified or ruled out. \subsection{Neutrino Mixing from $S_4$ Symmetry} \label{subsec34} We will consider next a second rather simple example of generation of neutrino mixing based on the $S_4$ symmetry. We recall that the three $S_4$ generators $S$, $T$ and $U$ satisfy the presentation rules given in eq. (\ref{S4STU}). In the triplet representation of interest and in the basis employed by us $S$, $T$ and $U$ are given in eq. (\ref{S4STU3drep}). In this case let us assume that (see, e.g., \cite{}) \begin{equation} G_e = Z^T_3 = \{1,T,T^2\}\,,~~~ G_\nu = Z^{SU}_2 = \{1,SU\}\,, \label{GeZ3GnuZ2} \end{equation} where $Z^{T}_3$, as we have discussed, is a $Z_3$ subgroup also of $S_4$ and $Z^{SU}_2$ is one of the $Z_2$ subgroups of $S_4$. As in the case of $A_4$ symmetry considered in the preceding subsection, $U_e$, which diagonalises $\rho(g_e)= T$, is effectively a unit $3\times3$ matrix and $M_e$ is a diagonal matrix containing the masses of the charged leptons. The matrix $U^\circ_\nu$, which diagonalises the element $\rho(g_{\nu}) = SU$ of $Z^{SU}_2$ (and $M^\dagger_{\nu}\,M_{\nu}$), with $S$ and $U$ given in eq. (\ref{S4STU3drep}), has the following general form: \begin{equation} U^\circ_{\nu} = V_{\rm TBM}\,U_{23}(\theta^\nu_{23},\beta)\,, \label{UonuA4} \end{equation} where $V_{\rm TBM}$ is the TBM mixing matrix and \begin{equation} U_{23}(\theta^\nu_{23},\beta) = \begin{pmatrix} 1 & 0 & 0 \vspace{0.2cm} \\ 0 & \cos \theta^{\nu}_{23} & \sin \theta^{\nu}_{23}\,e^{i\beta} \\ 0 & -\,\sin \theta^{\nu}_{23}\,e^{-i\beta} & \cos \theta^{\nu}_{23} \end{pmatrix} \;, \label{U23} \end{equation} the angle $\theta^{\nu}_{23}$ and the phase $\beta$ being arbitrary free parameters. The form of $U^\circ_\nu$ follows from the fact that the element $\rho(g_\nu) = SU$, as is easy to verify, is diagonalised by $V_{\rm TBM}$. However, in the resulting diagonal matrix the 2nd and the 3rd eigenvalues are degenerate and thus it is invariant with respect to a unitary transformation with $U_{23}(\theta^\nu_{23},\beta)$: \begin{equation} SU =\pm\, V_{\rm TBM}\,{\rm diag}(-1,1,1)\,V^T_{\rm TBM} = \pm\, V_{\rm TBM}\,U_{23}(\theta^\nu_{23},\beta)\,{\rm diag}(-1,1,1)\, (V_{\rm TBM}\,U_{23}(\theta^\nu_{23},\beta))^\dagger\,. \label{SUdiag} \end{equation} We see that also in the model with $S_4$ symmetry under discussion, the underlying symmetry form of the PMNS matrix is again the TBM one, $V_{\rm TBM}$. The matrix $U_{23}(\theta^\nu_{23},\beta)$ provides the necessary corrections to $V_{\rm TBM}$ leading, e.g., to $\theta_{13} \neq 0$. Similarly to the model based on the $A_4$ symmetry discussed in the previous subsection, the $S_4$ model we are discussing contains two free parameters - the angle $\theta^\nu_{23}$ and the phase $\beta$. However, as we show below, the testable phenomenological predictions of the model with $S_4$ symmetry differ significantly from the analogous predictions of the $A_4$ model. From eqs. (\ref{VTBM}), (\ref{U23}) and (\ref{UnuU0nu}) we get for the PMNS matrix: \begin{equation} U_{\rm PMNS} = U^\circ_{\nu}\,P^\circ = V_{\rm TBM}\,U_{23}(\theta^\nu_{23},\beta)\,P^\circ = \begin{pmatrix} \sqrt{\dfrac{2}{3}} & \dfrac{c^\nu_{23}}{\sqrt{3}} & \dfrac{s^\nu_{23}}{\sqrt{3}}\,e^{i\beta} \vspace{0.2cm} \\ -\,\dfrac{1}{\sqrt{6}} & \dfrac{c^\nu_{23}}{\sqrt{3}} + \dfrac{s^\nu_{23}}{\sqrt{2}}e^{-i\beta} & -\,\dfrac{c^\nu_{23}}{\sqrt{2}} + \dfrac{s^\nu_{23}}{\sqrt{3}}\,e^{i\beta} \vspace{0.2cm} \\ -\,\dfrac{1}{\sqrt{6}} & \dfrac{c^\nu_{23}}{\sqrt{3}} - \dfrac{s^\nu_{23}}{\sqrt{2}}e^{-i\beta} & \dfrac{c^\nu_{23}}{\sqrt{2}} + \dfrac{s^\nu_{23}}{\sqrt{3}}\,e^{i\beta} \end{pmatrix}\,P^\circ\;, \label{S4PMNS} \end{equation} where $c^\nu_{23} \equiv \cos\theta^\nu_{23}$ and $s^\nu_{23} \equiv \sin\theta^\nu_{23}$. Proceeding as in subsection \ref{subsec33} we find: \begin{equation} \sin^2\theta_{13} = \dfrac{1}{3}\sin^2\theta^\nu_{23}\,,~~ \sin^2\theta_{12} = \dfrac{1}{3}\cos^2\theta^\nu_{23} = \dfrac{1 - 3\sin^2\theta_{13}}{3(1-\sin^2\theta_{13})}\,. \label{S4s213s212} \end{equation} The neutrino mixing parameter $\sin^2\theta_{23}$ is determined by $\theta^\nu_{23}$ (or $\theta_{13}$) and $\beta$ and its value is not predicted: \begin{equation} \sin^2\theta_{23} = \dfrac{1}{c^2_{13}}\, \large |-\,\dfrac{c^\nu_{23}}{\sqrt{2}} + \dfrac{s^\nu_{23}}{\sqrt{3}}\,e^{i\beta}\, \large |^2 = \dfrac{1}{2} - \sqrt{2}\,s_{13}\, \dfrac{(1 - 3\,s^2_{13})^{\frac{1}{2}}}{(1 - s^2_{13})}\,\cos\beta\,. \label{S4s2231} \end{equation} To leading order in $s_{13}$ we have: \begin{equation} \dfrac{1}{2} - \sqrt{2}\, s_{13} \ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$<$}\ \sin^2\theta_{23} \ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$<$}\ \dfrac{1}{2} + \sqrt{2}\,s_{13} \,,~~{\rm or}~~ 0.293\ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$<$}\ \sin^2\theta_{23} \ \raisebox{-.4ex}{\rlap{$\sim$}} \raisebox{.4ex}{$<$}\ 0.707\,, \label{S4s2232} \end{equation} where the numerical values are obtained for the maximal value of $\sin^2\theta_{13}$ allowed at $3\sigma$ C.L. The interval of values of $\sin^2\theta_{23}$ in eq. (\ref{S4s2232}) is larger than the $3\sigma$ experimentally allowed NO and IO intervals of values of $\sin^2\theta_{23}$ (see Table \ref{}). The Dirac phase $\delta$ satisfies the following sum rule: \begin{equation} \cos\delta = \dfrac{ \dfrac{1}{6} - c^2_{23} + \dfrac{2}{3c^2_{13}}\,(c^2_{23} - s^2_{23}\,s^2_{13})} {2c_{23}\,s_{23}\,s_{13}\,c_{12}s_{12}} = \dfrac{(-\,1 + 5s^2_{13})\,\cos2\theta_{23}} {2\sqrt{2}\,\sin2\theta_{23}\,s_{13}\,(1 - 3s^2_{13})^{\frac{1}{2}}}\,, \label{cosdS4} \end{equation} where we expressed $c_{12}s_{12}$ in terms of $\theta_{13}$ using eq. (\ref{S4s213s212}), \begin{equation} c_{12}\,s_{12} = \dfrac{\sqrt{2}}{3c^2_{13}}\, (1-3s^2_{13})^{\frac{1}{2}}\,. \label{c12s12th13} \end{equation} We also have: \begin{equation} \sin2\theta_{23}\, \sin\delta = \sin\beta\,. \label{sindsinbetaS4} \end{equation} Similarly to the phases $\alpha$ of the $A_4$ model considered in the preceding subsection, the phase $\beta$ of the discussed $S_4$ model contributes to the Majorana phase $\alpha_{31}$ in the standard parametrisation of the PMNS matrix (see eqs. (\ref{VP}) and (\ref{})): \begin{equation} \dfrac{\alpha_{31}}{2} = \dfrac{\xi_{31}}{2} + \beta_{2} + \beta_3\,, \label{alpha31S4} \end{equation} where \begin{equation} \label{beta2beta3} \beta_2 = {\rm arg} \big(-\,\dfrac{c^\nu_{23}}{\sqrt{2}} + \dfrac{s^\nu_{23}}{\sqrt{3}}\,e^{i\beta}\big)\,,~~~ \beta_3 = {\rm arg} \big(\dfrac{c^\nu_{23}}{\sqrt{2}} + \dfrac{s^\nu_{23}}{\sqrt{3}}\,e^{i\beta}\big)\,, \end{equation} \begin{eqnarray} \label{beta2} \sin\beta_2 = \dfrac{s^\nu_{23}}{\sqrt{3}}\,\dfrac{\sin\beta}{s_{23}\,c_{13}} = \tan\theta_{13}\,\cos\theta_{23}\,\sin\delta\,,\\[0.30cm] \label{beta3} \sin\beta_3 = \dfrac{s^\nu_{23}}{\sqrt{3}}\,\dfrac{\sin\beta}{c_{23}\,c_{13}} = \tan\theta_{13}\,\sin\theta_{23}\,\sin\delta\,, \end{eqnarray} where we have used eqs. (\ref{S4PMNS}) and (\ref{sindsinbetaS4}) and $\sin\delta$ in eqs. (\ref{beta2}) and (\ref{beta3}) can be considered as a function of $\theta_{23}$ and $\theta_{13}$ (via eq. (\ref{cosdS4})). We also have: \begin{equation} \sin(\beta - \beta_2 - \beta_3) = -\,\sin\delta\,. \label{sindsinalphab2b3S4} \end{equation} The model with $U_{\rm PMNS}= V_{\rm TBM}\,U_{23}(\theta^\nu_{23},\beta)$ was discussed on general phenomenological grounds in \cite{}, where the predictions given in eqs. (\ref{S4s213s212}) and (\ref{cosdS4}) were obtained and the dependence of $\delta$ on $\sin^2\theta_{23}$ for a set of different values of $\theta_{13}$ was studied graphically. The correlation between $\sin^2\theta_{21}$ and $\sin^2\theta_{13}$ and the sum rule for $\cos\delta$ can also can be obtained from the general results for the group $S_4$ derived in \cite{ \footnote{In \cite{} a different basis for the $S_4$ generators $S$, $T$ and $U$ has been employed. The results of interest for, e.g., $\sin^2\theta_{12}$ in eqs. (\ref{S4s213s212}) and the sum rule for $\cos\delta$, eq. (\ref{cosdS4}), follow respectively from eq. (66) in subsection 4.2 and Table 3 (Case B2) in \cite{} by setting $\sin^2\theta^\circ_{12} = 1/6$ and $\sin^2\theta^\circ_{13} = 1/5$. }. Thus, as in the $A_4$ model, $\theta_{13}$ and $\theta_{23}$, or any pair of the four parameters $\theta_{12}$, $\theta_{23}$ $\theta_{13}$ and $\delta$, can be considered as the two independent parameters of the $S_4$ model. The model predicts a correlation between the values of $\sin^2\theta_{12}$ and $\sin^2\theta_{13}$, which for the best fit value of $\sin^2\theta_{13}$ implies $\sin^2\theta_{12} = 0.319$. This prediction lies in the current $1\sigma$ allowed interval of values of $\sin^2\theta_{12}$. Using eqs. (\ref{cosdS4}), (\ref{c12s12th13}) and the best fit values of $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$ from Table \ref{}, we also get the following predictions for $\cos\delta$ in the cases of NO and IO neutrino mass spectra: \begin{eqnarray} \label{cosdS4NO} &\cos\delta = -\,0.338\,(0.402)\,,~ \delta = \pm 109.73^\circ\,(\pm 66.27^\circ)\,,~~{\rm NO~(IO)}\,. \label{cosdS4IO} \end{eqnarray} Using instead the best fit values of $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$ for NO (IO) spectrum from \cite{NuFITv32Jan2018} we find rather different results due essentially to the difference in the best fit values of $\sin^2\theta_{23}$: \begin{eqnarray} \label{cosdS4NO2018} & \cos\delta = 0.167\,(0.237)\,,~ \delta = \pm 80.38^\circ\,(\pm 76.30^\circ)\,,~~{\rm NO~(IO)}\,. \label{cosdS4IO2018} \end{eqnarray} The values $\delta = 109.73^\circ$, $66.27^\circ$, $80.38^\circ$ and $76.30^\circ$ are strongly disfavored by the current data. As in the $A_4$ model, the difference between the predictions of $\cos\delta$ ($\delta$) for NO and IO neutrino mass spectra are a consequence of the difference between the best fit values of $\sin^2\theta_{23}$ for the two spectra (see Table \ref{} and eq. (\ref{})). For $\sin^2\theta_{23} = 0.5$ we have for both spectra $\cos\delta=0$, or $\delta = \pi/2,3\pi/2$, also in the $S_4$ model, with $\delta = \pi/2$ strongly disfavored by the current data. As we have seen, the $A_4$ and $S_4$ models considered lead to largely different predictions for $\sin^2\theta_{12}$ and, if $\theta_{23} \neq \pi/4$, for $\cos\delta$ ($\delta$) as well. These predictions can be used to discriminate experimentally between the two models. In both $A_4$ and $S_4$ models we have discussed the Majorana phases are not predicted. \subsection{Comment on the Symmetry Breaking} \label{subsec35} The discrete symmetry approach to neutrino mixing we have discussed so far allows to explain quantitatively the observed pattern of neutrino mixing. A complete self-consistent (renormalisable) theory based on this approach should include also a mechanism of neutrino mass generation as well as details of breaking of the flavour symmetry $G_f$ to the residual symmetries $G_e$ and $G_\nu$ in the charged lepton and neutrino sectors. As a rule, the non-Abelian flavour symmetry $G_f$ is broken spontaneously by a set of scalar fields, flavons, which are singlets with respect to the Standard Theory $SU(2)_L\times U(1)_{\rm Y_W}$ gauge symmetry but transform according certain irreducible representations of $G_f$, couple in a $G_f$-invariant manner to the LH lepton doublet fields and RH charged lepton $SU(2)_L$ singlet fields via Yukawa-type (typically non-renormalisable effective) interactions. These Yukawa-type effective interactions appear in the low-energy limit of a theory which is renormalisable at some high energy scale $\Lambda$ where the symmetry $G_f$ is exact (see, e.g., \cite{, }). The flavons develop non-zero vacuum expectation values in specific directions (``vacuum alignment''). In the case of the $A_4$ model considered in subsection \ref{subsec33}, for example, the $A_4$ symmetry breaking leading to $G_e = Z^T_3$ and $G_\nu = Z^S_2$ and generating Majorana mass term for the LH flavour neutrino fields can be achieved i) by assigning $\tilde{e}_R(x)$, $\tilde{\mu}_R(x)$ and $\tilde{\tau}_R(x)$ to the three different singlet representations of $A_4$ ${\bf 1}$, ${\bf 1''}$ and ${\bf 1'}$ (see Table 2), respectively, ii) by introducing two $A_4$ triplet and two $A_4$ singlet flavon scalar fields, which develop vacuum expectation values in specific directions, and iii) by using the rules of tensor products of irreducible representations for $A_4$ (for further details see, e.g., \cite{}). Discussing the flavon sectors of the models considered is beyond the scope of the present article. Examples of complete self-consistent (renormalisable) models, in which the breaking of the flavour symmetry $G_f$ to desired residual symmetries $G_e$ and $G_\nu$ with the help of sets of flavon fields developing non-zero vacuum expectation values in requisite directions and, thus, generating $G_e-$ invariant charged lepton mass term and $G_\nu-$ invariant neutrino Majorana mass term, include, e.g., the models in refs. \cite{, }; for a review see, e.g., ref. \cite{}. \subsection{Alternative Symmetry Forms of $U_{\nu}$: Bimaximal, Golden Ratio and Hexagonal Mixing } \label{subsec36} Thus, TBM can only be an underlying approximate symmetry form of the PMNS neutrino mixing matrix. Other widely discussed underlying (approximate) symmetry forms of the PMNS matrix include: i) bimaximal (BM) mixing \footnote{Bimaximal mixing can also be a consequence of the conservation of the lepton charge $L' = L_e - L_{\mu} - L_{\tau}$ (LC) \cite{SPPD82}, supplemented by $\mu - \tau$ symmetry.} \cite{BM}, ii) the golden ratio type A (GRA) mixing \cite{}, iii) the golden ratio type B (GRB) mixing \cite{GRBM}, and iv) hexagonal (HG) mixing \cite{HGM}. For all these forms, including the TBM one, the matrix $U^\circ_{\nu}$ has the form: $U^\circ_{\nu} = R_{23}(\theta^\nu_{23}) R_{13}(\theta^\nu_{13}) R_{12}(\theta^\nu_{12})$ with $\theta^\nu_{23} = -\,\pi/4$ and $\theta^\nu_{13} = 0$: \begin{equation} U^\circ_{\nu} = R_{23} \left ( \theta^\nu_{23}=-\,\pi/4 \right) R_{12}\left( \theta^\nu_{12}\right) = \begin{pmatrix} \cos \theta^{\nu}_{12} & \sin \theta^{\nu}_{12} & 0 \vspace{0.2cm} \\ - \dfrac{\sin \theta^{\nu}_{12}}{\sqrt{2}} & \dfrac{\cos \theta^{\nu}_{12}}{\sqrt{2}} & - \dfrac{1}{\sqrt{2}} \vspace{0.2cm} \\ - \dfrac{\sin \theta^{\nu}_{12}}{\sqrt{2}} & \dfrac{\cos \theta^{\nu}_{12}}{\sqrt{2}} & \dfrac{1}{\sqrt{2}} \end{pmatrix} \;. \label{Unu12} \end{equation} The value of the angle $\theta^{\nu}_{12}$, and thus of $\sin^2\theta^{\nu}_{12}$, depends on the symmetry form of $U^\circ_{\nu}$. For the TBM, BM, GRA, GRB and HG forms we have: i) $\sin^2\theta^{\nu}_{12} = 1/3$ (TBM), ii) $\sin^2\theta^{\nu}_{12} = 1/2$ (BM), iii) $\sin^2\theta^{\nu}_{12} = (2 + \tilde{r})^{-1} \cong 0.276$ (GRA), $\tilde{r}$ being the golden ratio, $\tilde{r} = (1 +\sqrt{5})/2$, iv) $\sin^2\theta^{\nu}_{12} = (3 - \tilde{r})/4 \cong 0.345$ (GRB), and v) $\sin^2\theta^{\nu}_{12} = 1/4$ (HG). As we have seen in subsections \ref{subsec32} and \ref{subsec34}, the TBM form of $U^\circ_{\nu}$ can originate from $G_f=S_4$ symmetry \cite{S4} (with residual symmetry $G_\nu = Z^S_2\times Z^U_2$). It can be obtained also from a $G_f = A_4$ symmetry \cite{A4} (with $G_\nu= Z^S_2$ and the presence of accidental $\mu-\tau$ (i.e., $Z_2$) symmetry, see, e.g., \cite{})) \footnote{ The TBM form of $U^\circ_{\nu}$ can also be derived from $G_f = T'$ - the double covering group of $A_4$ (see, e.g., \cite{}) - with $G_\nu = Z^{\rm S}_2$, provided the left-handed (LH) charged lepton and neutrino fields each transform as triplets of $T^{\prime}$ (see, e.g., \cite{} for details). Actually, as can be shown \cite{}, when working with 3-dimensional and 1-dimensional representations of $T^{\prime}$, there is no way to distinguish $T^{\prime}$ from $A_4$. }. The group $G_f = S_4$ can also be used to generate the BM from of $U^\circ _{\nu}$ (e.g., by choosing $G_{\nu} = Z_2$ combined with an accidental $\mu-\tau$ symmetry) \cite{S4,}. The GRA form of $U^\circ_{\nu}$ can be obtained from the group $A_5$ \cite{A5}, which is the group of even permutations of five objects and is isomorphic to the group of rotational symmetries of the icosahedron. In this case $\sin^2 \theta^{\nu}_{12} = 1/(\tilde{r}\sqrt{5}) \cong 0.276$. The GRB and HG forms of $U^\circ_{\nu}$ can be generated using the groups $G_f = D_{10}$ \cite{D10} and $G_f = D_{12}$ \cite{HGM}, respectively. The dihedral groups $D_{10}$ and $D_{12}$ are the groups of symmetries (rotations and reflections) of the regular decagon and dodecagon \footnote{ The groups $D_{10}$ and $D_{12}$, as it is indicated in Table 2, have 1-dimensional and 2-dimensional irreducible representations, but do not have 3-dimensional irreducible unitary representations. The problem of how the GRB and HG forms of $U^\circ_{\nu}$ can be generated using the groups $G_f = D_{10}$ and $G_f = D_{12}$, respectively, is discussed in refs. \cite{D10} and \cite{HGM} and we refer the interested reader to these articles.}. $D_{10}$ and $D_{12}$ lead respectively to $\theta^{\nu}_{12} = \pi/5$ (or $\sin^2 \theta^{\nu}_{12} = (3 - \tilde{r})/4 \cong 0.345$) and $\theta^{\nu}_{12} = \pi/6$ (or $\sin^2 \theta^{\nu}_{12} = 1/4$). The angles $\pi/5$ and $\pi/6$ are the external angles of the decagon and dodecagon. For all the five underlying symmetry forms of $U^\circ_{\nu}$ listed above we have i) $\theta^{\nu}_{13} = 0$, which should be corrected to the measured value of $\theta_{13} \cong 0.15$, and ii) $\sin^2\theta^{\nu}_{23} = 0.5$, which might also need to be corrected if it is firmly established that $\sin^2\theta_{23}$ deviates significantly from 0.5. In the case of the BM form $\sin^2\theta_{12} = 0.5$, which is ruled out by the existing data and should be corrected. Finally, the value of $\sin^2\theta^{\nu}_{12}$ for the HG form lies outside the current $2\sigma$ allowed range of $\sin^2\theta_{12}$ and might need also to be corrected. The requisite corrections to the discussed underlying symmetry forms of the PMNS matrix can be generated in each specific case by ``decreasing'' the residual symmetry $G_\nu$ which leads to a given symmetry form. In the case of $G_f = S_4$ ($G_f = A_4$), as we have seen, this can be achieved by ``decreasing'' $G_\nu$ from $Z_2\times Z_2$ ($Z_2$ + the ``accidental'' $\mu-\tau$ (i.e., $Z_2$)) symmetry to $Z_2$, leading to additional ``correcting'' matrix factor in $U^\circ_{\nu}$. As we have mentioned earlier, the corrections can also be provided by the matrix $U_e$. This approach was followed in \cite{} and corresponds to the case of $G_f$ i) either broken to $G_e = Z_2$, or ii) completely broken, by the charged lepton mass term. In this case the PMNS matrix has the following general form \cite{}: \begin{equation} U = U_e^{\dagger}\, U_{\nu} = (\tilde{U}_{e})^\dagger\, \Psi U^\circ_{\nu} \, P^\circ\,. \label{PMNS22} \end{equation} Here $\tilde{U}_e$ is a $3\times 3$ unitary matrix and $\Psi$ is a diagonal phase matrix. The matrix $\tilde{U}_e$ was chosen in \cite{} to have the following two forms: \begin{eqnarray} {\bf A0}:~ \tilde{U}_{e} = R^{-1}_{23}(\theta^e_{23})\,R^{-1}_{12}(\theta^e_{12})\,;~~ {\bf B0}:~\tilde{U}_{e} = R^{-1}_{12}(\theta^e_{12})\,. \label{Ue23122} \end{eqnarray} where $\theta^e_{12}$ and $\theta^e_{23}$ are free real angle parameters. These two forms appear in a large class of theoretical models of flavour and studies, in which the generation of charged lepton masses is an integral part (see, e.g., \cite{}). The phase matrix $\Psi$ in cases {\bf A0} and {\bf B0} is given by \cite{}: \begin{eqnarray} {\bf A0}:~\Psi = {\rm diag} \left(1,\text{e}^{-i \psi}, \text{e}^{-i \omega} \right)\,;~~ {\bf B0}:~\Psi = {\rm diag} \left(1,\text{e}^{-i \psi}, 1 \right)\,. \label{PsiAB2} \end{eqnarray} The phases $\omega$ and/or $\psi$ serve as a source for the Dirac CPV phase $\delta$ of the PMNS matrix and contribute to the Majorana CPV phases of the PMNS matrix $\alpha_{21}$ and $\alpha_{31}$ \cite{}. We recall that the diagonal phase matrix $P^\circ$ in eq. (\ref{PMNS22}) is given in eq. (\ref{UnuU0nu}): it contains two phases, $\xi_{21}$ and $\xi_{31}$, which also contribute to the Majorana phases $\alpha_{21}$ and $\alpha_{31}$, respectively. \vspace{-0.2cm} \subsection{Predictions for the Dirac CPV Phase} \label{subsec37} \subsubsection{The Cases of TBM, BM, GRA, GRB and HG Symmetry Forms Corrected by $U_e$} \label{subsub371} Consider the case of the five underlying symmetry forms of $U^\circ_\nu$ - TBM, BM, GRA, GRB and HG - corrected by the matrix $U_e$, with the PMNS matrix given in eq. (\ref{PMNS22}) and the matrices $\tilde{U}_{e}$ and $\Psi$ as given in eqs. (\ref{Ue23122}) and (\ref{PsiAB2}). In this setting the Dirac phase $\delta$ of the PMNS matrix was shown to satisfy the following sum rule \cite{}: \begin{equation} \cos\delta = \frac{\tan\theta_{23}}{\sin2\theta_{12}\sin\theta_{13}}\, \left [\cos2\theta^{\nu}_{12} + \left (\sin^2\theta_{12} - \cos^2\theta^{\nu}_{12} \right )\, \left (1 - \cot^2\theta_{23}\,\sin^2\theta_{13}\right )\right ]\,. \label{cosdthnu2} \end{equation} Within the approach employed this sum rule is exact \footnote{The renormalisation group corrections to the sum rule for $\cos\delta$, eq. (\ref{cosdthnu2}), in the cases of neutrino Majorana mass term generated by the Weinberg (dimension 5) operator added to i) the Standard Model, and ii) the minimal SUSY extension of the Standard Model, have been investigated in \cite{}. They were found in \cite{} to be negligible, e.g., when the Weinberg operator was added to the Standard Model. } and is valid for any value of the angle $\theta^{\nu}_{23}$ \cite{} (and not only for $\theta^{\nu}_{23} = -\,\pi/4$ of the five discussed symmetry forms of $U^\circ_\nu$). As we see, via the sum rule $\cos\delta$ is expressed in terms of the three neutrino mixing angles $\theta_{12}$, $\theta_{23}$, $\theta_{13}$ and one fixed (known) parameter $\theta^{\nu}_{12}$ which depends on the underlying symmetry form (TBM, BM, GRA, GRB, HG) of the PMNS matrix. The difference between the cases ${\bf A0}$ and ${\bf B0}$ of forms of $\tilde{U}_{e}$ in eq.~(\ref{Ue23122}) is, in particular, in the correlation between the values of $\sin^2 \theta_{23}$ and $\sin^2 \theta_{13}$ they lead to. In case ${\bf A0}$ of $\tilde{U}_{e}$, the values of $\sin^2 \theta_{23}$ and $\sin^2 \theta_{13}$ are not correlated and $\sin^2 \theta_{23}$ can differ significantly from 0.5 \cite{}. For the form ${\bf B0}$ of $\tilde{U}_{e}$ we have \cite{}: \begin{equation} \sin^2 \theta_{23} = \frac{1}{2}\, \frac{1 - 2\,\sin^2 \theta_{13}}{1 - \sin^2 \theta_{13}} \cong \frac{1}{2}\, (1 - \sin^2 \theta_{13})\,. \label{} \end{equation} Thus, in contrast to the case {\bf A0}, in case {\bf B0} the value of $\sin^2 \theta_{23}$ is correlated with the value of $\sin^2 \theta_{13}$ and as a consequence $\sin^2 \theta_{23}$ can deviate from 0.5 insignificantly - only by $0.5\sin^2 \theta_{13}$. Qualitatively, the result in eq. (\ref{cosdthnu2}) for $\delta$ can be understood as follows. In the parametrisation defined in eq. (\ref{PMNS22}) with $U^\circ_{\nu}$, $\tilde{U}_{e}$ and $\Psi$ given in (\ref{Unu12}) and, e.g., by forms ${\bf B0}$ in eqs. (\ref{Ue23122}) and (\ref{PsiAB2}), we have: \begin{align} &U_{\rm PMNS} = R_{12}(\theta^e_{12}) \, \Psi \, R_{23}(\theta^{\nu}_{23}) \, R_{12}(\theta^{\nu}_{12}) \, P^\circ \,. \label{} \end{align} The phase $\psi$ in the phase matrix $\Psi$ serves as a source for the Dirac phase $\delta$ (and gives a contribution to the Majorana phases $\alpha_{21,31}$ \cite{}). It follows from eq. (\ref{}) that in the case under discussion, the three angles $\theta_{12}$, $\theta_{23}$, $\theta_{13}$ and the Dirac phase $\delta$ of the standard parametrisation of $U_{\rm PMNS}$ are expressed in terms of the three parameters $\theta^e_{12}$, $\psi$ and $\theta^{\nu}_{12}$ ($\theta^{\nu}_{23} = -\,\pi/4$). This suggests that it will be possible to express one of the four parameters $\theta_{12}$, $\theta_{23}$, $\theta_{13}$ and $\delta$, namely $\delta$, in terms of the other three, hence eq. (\ref{cosdthnu2}). Although the case of $\tilde{U}_{e}$ having the form ${\bf A0}$ in eq. (\ref{Ue23122}) is somewhat more complicated, in what concerns $\cos\delta$ one arrives to the same conclusion and result \cite{}. Given the values of $\sin\theta_{12}$, $\sin\theta_{23}$, $\sin\theta_{13}$ and $\theta^{\nu}_{12}$, $\cos\delta$ is determined uniquely by the sum rule (\ref{cosdthnu2}). This allows us to determine also $|\sin\delta|$ uniquely. However, in absence of additional information, ${\rm sgn}(\sin\delta)$ remains undetermined, which leads to a two-fold ambiguity in the determination of the value of $\delta$ from the given value of $\cos\delta$. The fact that the value of the Dirac CPV phase $\delta$ is determined (up to an ambiguity of the sign of $\sin\delta$) by the values of the three mixing angles $\theta_{12}$, $\theta_{23}$ and $\theta_{13}$ of the PMNS matrix and the value of $\theta^{\nu}_{12}$ of the matrix $U^\circ_\nu$, eq. (\ref{Unu12}), is the most striking prediction of the models considered. This result implies that in the schemes under discussion, the rephasing invariant $J_{\text{CP}}$ associated with the Dirac phase $\delta$, eq. (\ref{JCPstandparam}), is also a function of the three angles $\theta_{12}$, $\theta_{23}$ and $\theta_{13}$ of the PMNS matrix and of $\theta^{\nu}_{12}$: \begin{equation} J_{\text{CP}} = J_{\text{CP}}(\theta_{12},\theta_{23},\theta_{13}, \delta(\theta_{12},\theta_{23},\theta_{13},\theta^{\nu}_{12})) = J_{\text{CP}}(\theta_{12},\theta_{23},\theta_{13},\theta^{\nu}_{12})\,. \label{JCPNO2} \end{equation} This allows to obtain predictions for the values of $J_{\text{CP}}$ for the different symmetry forms of $\tilde{U}_\nu$ (specified by the value of $\theta^{\nu}_{12}$) using the current data on $\theta_{12}$, $\theta_{23}$ and $\theta_{13}$. In \cite{}, by using the sum rule in eq.~(\ref{cosdthnu2}), predictions for $\cos\delta$, $\delta$ and the $J_{\text{CP}}$ factor were obtained in the TBM, BM, GRA, GRB and HG cases for the b.f.v. of $\sin^2\theta_{12}$, $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$. It was found that the predictions of $\cos\delta$ vary significantly with the symmetry form of $\tilde{U}_{\nu}$. For the b.f.v. of $\sin^2\theta_{12}=0.308$, $\sin^2\theta_{13}= 0.0234$ and $\sin^2\theta_{23}= 0.437$~found for NO spectrum in \cite{}, for instance, one gets\cite{} $\cos\delta = (-0.0906)$, $(-1.16$), $0.275$, $(-0.169)$ and $0.445$, for the TBM, BM (LC), GRA, GRB and HG forms, respectively. For the TBM, GRA, GRB and HG forms these values correspond to $\delta = \pm 95.2^{\circ}, \pm 74.0^{\circ}, \pm 99.7^{\circ}, \pm 63.6^{\circ}$. For the b.f.v. given in Table \ref{} and obtained in the global analysis \cite{} one finds in the cases of the TBM, BM (LC), GRA, GRB and HG forms the values given in Table~\ref{}. \begin{table}[t] \centering \caption{Predicted values of $\cos\delta$ and $\delta$ for the five symmetry forms, TBM, BM, GRA, GRB and HG, and $\tilde{U}_{e}$ given by the form ${\bf A0}$ in eq. (\ref{Ue23122}), obtained using eq. (\ref{cosdthnu2}) and the best fit values of $\sin^2\theta_{12}$, $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$ for NO and IO neutrino mass spectra from ref. \cite{}. (From refs. \cite{}.) } \begin{tabular}{lcccc} \hline Scheme & $\cos\delta$ (NO) & $\delta$ (NO) & $\cos \delta$ (IO) & $\delta$ (IO)\\ \hline TBM & $-0.16$ & $ \pm 99^\circ$ & $-0.27$ & $\pm 106^\circ$\\ BM (LC) & $-1.26$ & $\cos\delta-$unphysical & $-1.78$ &$\cos\delta-$unphysical\\ GRA & $\phantom{-}0.21$ & $\pm 78^\circ$ & $\phantom{-}0.24$ & $\pm 76^\circ $\\ GRB & $-0.24$ & $\pm 105^\circ$ & $-0.38$ &$\phantom{-}\pm 112^\circ$\\ HG & $\phantom{-}0.39$ & $\pm 67^\circ$ & $\phantom{-}0.48$ & $\pm 62^\circ$\\ \hline \end{tabular} \label{} \end{table} Due to the different NO and IO b.f.v. of $\sin^2\theta_{23}$, the predicted values of $\cos\delta$ and $\delta$ for IO spectrum differ (in certain cases significantly) from those for the NO spectrum. Three comments are in order. First, according to the results found in \cite{} and quoted in Table~\ref{}, the predicted values of $\delta$ lying in the first quadrant are strongly disfavored (if not ruled out) by the current data. Second, the unphysical value of $\cos\delta$ in the BM (LC) case is a reflection of the fact that the scheme under discussion with BM (LC) form of the matrix $U^\circ_{\nu}$ does not provide a good description of the current data on $\theta_{12}$, $\theta_{23}$ and $\theta_{13}$~ \cite{}. Physical values of $\cos \delta$ can be obtained in the case of the NO spectrum, e.g., for the b.f.v. of $\sin^2 \theta_{13}$ if the value of $\sin^2 \theta_{12}$ ($\sin^2 \theta_{23}$) is larger (smaller) than the current best fit value \footnote{For, e.g., $\sin^2 \theta_{12} = 0.34$ allowed at $2\sigma$ by the current data, we have $\cos \delta = -0.943$. Similarly, for $\sin^2 \theta_{12} = 0.32$, $\sin^2 \theta_{23}=0.41$ and $\sin\theta_{13}=0.158$ we have \cite{}: $\cos \delta = -0.978$. } \cite{}. However, with the b.f.v. of $\sin^2\theta_{23}$ quoted in eq. (\ref{}), the BM (LC) form is strongly disfavored for both NO and IO spectra. Third, the $A_4$ and $S_4$ models considered subsections \ref{subsec33} and \ref{subsec34} lead to largely different predictions for $\sin^2\theta_{12}$ and, if $\theta_{23} \neq \pi/4$, for $\cos\delta$ as well, which differ also from the predictions for $\cos\delta$ we have obtained in the cases of the five different symmetry forms - TBM, BM, GRA, GRB, HG - and the matrix $\tilde{U}_{e}$ given by the forms ${\bf A0}$ and ${\bf B0}$ in eq. (\ref{Ue23122}). These predictions can be used to discriminate experimentally between the different models. The results quoted above imply~\cite{} that a measurement of $\cos\delta$ can allow to distinguish between at least some of the different symmetry forms of $U^\circ_{\nu}$ provided $\theta_{12}$, $\theta_{13}$ and $\theta_{23}$ are known, and $\cos\delta$ is measured, with sufficiently high precision \footnote{Detailed results on the dependence of the predictions for $\cos\delta$ on $\sin^2\theta_{12}$, $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$ when the latter are varied in their respective $3\sigma$ experimentally allowed ranges can be found in \cite{}.}. Even determining the sign of $\cos\delta$ will be sufficient to eliminate some of the possible symmetry forms of $\tilde{U}_{\nu}$. These conclusions were confirmed by the statistical analyses performed in ref.~\cite{} where predictions of the sum rule (\ref{cosdthnu2}) for i) $\delta$, $\cos\delta$ and the rephasing invariant $J_{\rm CP}$ using the ``data'' (best fit values and $\chi^2-$distributions) on $\sin^2\theta_{12}$, $\sin^2\theta_{13}$, $\sin^2\theta_{23}$ and $\delta$ from \cite{}, and ii) for $\cos\delta$, using prospective uncertainties on $\sin^2\theta_{12}$, $\sin^2\theta_{13}$ and $\sin^2\theta_{23}$, were derived for the TBM, BM (LC), GRA, GRB and HG symmetry forms of the matrix $\tilde{U}_{\nu}$. Both analyses were performed for the case of NO neutrino mass spectrum. The results for the IO spectrum are similar. \begin{figure}[!t] \begin{center} \hspace{-0.9cm} \includegraphics[width=13cm]{Figchi2JCP.pdf} \end{center} \vspace{-1.2cm} \caption{ \label{} $N_{\sigma} \equiv \sqrt{\chi^2}$ as a function of $J_{\rm CP}$. The dashed lines represent the results of the global fit \cite{}, while the solid lines represent the results we obtain for the TBM, BM (LC), GRA (upper left, central, right panels), GRB and HG (lower left and right panels) neutrino mixing symmetry forms. The blue (red) lines are for NO (IO) neutrino mass spectrum. (From ref.~\protect\cite{}.) } \end{figure} \noindent The aim of the first analysis, the results of which for $J_{\rm CP}$ are shown in Fig. \ref{} and are summarised in Table \ref{}, was to derive the allowed ranges for $\delta$ and $J_{\rm CP}$, predicted on the basis of the current data on the neutrino mixing parameters for each of the symmetry forms of $U^\circ_{\nu}$ considered (see \cite{} for details of the analysis). It was found in \cite{}, in particular, that the CP-conserving value of $J_{\rm CP} = 0$ is excluded in the cases of the TBM, GRA, GRB and HG neutrino mixing symmetry forms, respectively, at approximately $5\sigma$, $4\sigma$, $4\sigma$ and $3\sigma$ C.L. with respect to the C.L. of the corresponding best fit values which all lie in the interval $J_{\rm CP} = (-0.034) - (-0.031)$ (see Table \ref{}). The best fit value for the BM (LC) form is much smaller and close to zero: $J_{\rm CP} = (-5\times 10^{-3})$. For the TBM, GRA, GRB and HG forms at $3\sigma$ we have $0.020 \leq |J_{\rm CP}| \leq 0.039$. Thus, for these four forms the CP violating effects in neutrino oscillations are predicted to be relatively large and observable in the T2HK and DUNE experiments \cite{DUNE2016,T2HK2015}. \begin{table}[t] \centering \caption{ Best fit values of $J_{\rm CP}$ and $\cos \delta$ and corresponding 3$\sigma$ ranges (found fixing $\chi^2-\chi^2_{\rm min} = 9$) for the five symmetry forms, TBM, BM, GRA, GRB and HG, and $\tilde{U}_{e}$ given by the form ${\bf A0}$ in eq. (\ref{Ue23122}) obtained using the data from \protect\cite{} for NO neutrino mass spectrum. (From ref. \protect\cite{}, where results for IO spectrum are also given.) } \begin{tabular}{lcccc} \toprule Scheme & $J_{\rm CP}/10^{-2}$ (b.f.v.) & $J_{\rm CP}/10^{-2}$ ($3\sigma$ range) & $\cos \delta$ (b.f.v.) & $\cos \delta$ ($3\sigma$ range)\\ \midrule TBM & $-3.4$ & $[-3.8,-2.8] \cup [3.1,3.6]$ & $-0.07$ & $[-0.47,\phantom{-}0.21]$\\ BM (LC) & $-0.5$ & $[-2.6,2.1]$ & $-0.99$ & $[-1.00,-0.72]$\\ GRA & $-3.3$ & $[-3.7,-2.7] \cup [3.0,3.5]$ & $\phantom{-}0.25$ & $[-0.08,\phantom{-}0.69]$\\ GRB & $-3.4$ & $[-3.9,-2.6] \cup [3.1,3.6]$ & $-0.15$ & $[-0.57,\phantom{-}0.13]$\\ HG & $-3.1$ & $[-3.5,-2.0] \cup [2.6,3.4]$ & $\phantom{-}0.47$ & $[\phantom{-}0.16,\phantom{-}0.80]$\\ \bottomrule \end{tabular} \label{} \end{table} \noindent These conclusions hold if one uses in the analysis the results on the neutrino mixing parameters and $\delta$, obtained in the more recent global analyses \cite{}. In Fig.~\ref{Fig:23} (left panel) we present the results of the statistical analysis of the predictions for $\cos\delta$, namely the likelihood function versus $\cos \delta$ within the Gaussian approximation (see \cite{} for details) performed using the b.f.v. of the mixing angles for NO neutrino mass spectrum given in ref.~\cite{} and the prospective rather small $1\sigma$ uncertainties i) of 0.7\% on $\sin^2 \theta_{12}$, planned to be reached in JUNO experiment \cite{JUNO}, ii) of 3\% on $\sin^2 \theta_{13}$, foreseen to be obtained in the Daya Bay experiment \cite{}, and iii) of 5\% on $\sin^2 \theta_{23}$, expected to be reached in the currently running and future planned long baseline neutrino oscillation experiments. In the proposed upgrading of the currently taking data T2K experiment \cite{T2K20152016}, for example, $\theta_{23}$ is estimated to be determined with a $1\sigma$ error of $1.7^\circ$, $0.5^\circ$ and $0.7^\circ$ if the best fit value of $\sin^2\theta_{23} = 0.50$, $0.43$ and $0.60$, respectively. This implies that for these three values of $\sin^2\theta_{23}$ the absolute (relative) $1\sigma$ error would be 0.0297 (5.94\%), 0.0086 (2\%) and 0.0120 (2\%). This error on $\sin^2\theta_{23}$ is expected to be further reduced in the future planned T2HK \cite{T2HK2015} and DUNE \cite{DUNE2016} experiments. \begin{figure}[!t] \begin{center} \hspace{-0.9cm} \subfigure {\includegraphics[width=8cm]{Fig2.pdf}} {\includegraphics[width=8cm]{Fig3.pdf}} \vspace{5mm} \end{center} \vspace{-1.0cm} \caption{ \label{Fig:23} The likelihood function versus $\cos\delta$ for NO neutrino mass spectrum after marginalising over $\sin^2\theta_{13}$ and $\sin^2\theta_{23}$, for the TBM, BM (LC), GRA, GRB and HG symmetry forms of the mixing matrix $U^\circ_{\nu}$. The figure is obtained by using the prospective $1\sigma$ uncertainties in the determination of $\sin^2\theta_{12}$, $\sin^2\theta_{13}$ and $\sin^2\theta_{23}$ within the Gaussian approximation. In the left (right) panel $\sin^2 \theta_{12}$ is set to its b.f.v. of \protect\cite{} $0.308$ (is set to $0.332$), the other mixing angles being fixed to their NO best fit values taken from \protect\cite{}. See text for further details. (From ref.~\protect\cite{}.) } \end{figure} As we have already remarked, the BM (LC) case is very sensitive to the b.f.v. of $\sin^2 \theta_{12}$ and $\sin^2 \theta_{23}$ and is disfavored at more than $2\sigma$ for the b.f.v. found in \cite{} for the NO spectrum. This case might turn out to be compatible with the data for larger (smaller) measured values of $\sin^2 \theta_{12}$ ($\sin^2 \theta_{23}$). This is illustrated in Fig. \ref{Fig:23} (right panel). The measurement of $\sin^2 \theta_{12}$, $\sin^2 \theta_{13}$ and $\sin^2 \theta_{23}$ with the quoted precision will open up the possibility to distinguish between the BM (LC), TBM/GRB, GRA and HG forms of $\tilde U_{\nu}$. Distinguishing between the TBM and GRB forms seems to require unrealistically high precision measurement of $\cos \delta$~ \footnote{Self-consistent models or theories of (lepton) flavour which lead to the GRB form of $U^\circ_{\nu}$ might still be possible to distinguish from those leading to the TBM form using the specific predictions of the two types of models for the neutrino mixing angles. The same observation applies to models which lead to the GRA and HG forms of $U^\circ_{\nu}$.}. Assuming that $|\cos \delta| < 0.93$, which means for 76\% of values of $\delta$, the error on $\delta$, $\Delta \delta$, for an error on $\cos \delta$, $\Delta(\cos \delta) = 0.10 \, (0.08)$, does not exceed $\Delta \delta \lesssim \Delta(\cos \delta) /\sqrt{1 - 0.93^2} = 16^{\circ} \, (12^{\circ})$. This accuracy is planned to be reached in the future neutrino experiments like T2HK, T2HKK (ESS$\nu$SB) \cite{T2HK2015,}. Therefore a measurement of $\cos \delta$ in the quoted range with $\Delta(\cos \delta) = 0.08$ will allow one to distinguish between the TBM/GRB, BM (LC) and GRA/HG forms at approximately $3\sigma$ C.L., if the precision achieved on $\sin^2 \theta_{12}$, $\sin^2 \theta_{13}$ and $\sin^2 \theta_{23}$ is the same as in Fig.~\ref{Fig:23}. A more detailed study of the possibility to distinguish between BM (LC), TBM, GRB, GRA and HG forms of $U^\circ_{\nu}$ using the prospective data from DUNE and T2HK experiments was performed in \cite{}. \begin{figure}[!t] \begin{center} \hspace{-0.9cm} \includegraphics[width=15cm]{chisq_sinth12nu.pdf} \end{center} \vspace{-0.5cm} \caption{ \label{} Sensitivities of the experiments DUNE, T2HK and their combined (prospective) data to the symmetry form parameter $\sin^2\theta^\nu_{12}$ allowing to distinguish between the TBM, GRA, GRB, and HG symmetry forms under the assumption that one of them is realised in Nature. In the top left and right panels the assumed true symmetry forms are respectively TBM ($\sin^2\theta^\nu_{12} = 1/3$) and GRA ($\sin^2\theta^\nu_{12} = 0.276$), while in the bottom left and right panel these forms are GRB ($\sin^2\theta^\nu_{12} = 0.345$) and HG ($\sin^2\theta^\nu_{12} = 0.25$). See text for further details. (From ref.~\protect\cite{}.) } \end{figure} \noindent Some of the results of this study are illustrated in Fig. \ref{}. As is shown in \cite{} and is indicated by Fig. \ref{}, the combined analysis of the data from the DUNE and T2HK experiments would allow to distinguish between TBM and HG (GRA) symmetry forms of the PMNS matrix at approximately $3\sigma$ ($2\sigma$) confidence level; and the same data would allow to distinguish between GRB and HG (GRA) forms at more than $3\sigma$ (at approximately $2\sigma$) confidence level. Using the data from the T2HK, T2HKK and DUNE experiments is expected to lead to a better discrimination between the different symmetry forms of $U_{\rm PMNS}$ owing to the better prospective sensitivity to $\delta$ of the combined data from the T2HK and T2HKK experiments. In what concerns the BM (LC) form, as we have already discussed, it is not compatible with the best fit values of the neutrino mixing angles (leading to $|\cos\delta| > 1$), but is viable if the $2\sigma$ ranges of the neutrino mixing angles are taken into account. If, e.g., one keeps $\sin^2\theta_{13}$ and $\sin^2\theta_{23}$ fixed at their best fit values for NO spectrum, one finds $\cos\delta = -\,1$, and thus a viable BM (LC) mixing form, for $\sin^2\theta_{12} = 0.334$, which is the upper limit of the allowed $2\sigma$ range of $\sin^2\theta_{12}$ (see Table \ref{}). For the indicated choice of values of $\sin^2\theta_{13}$, $\sin^2\theta_{23}$ and $\sin^2\theta_{12}$ the BM form, as was shown in \cite{}, can be distinguished from the other four symmetry forms -- TBM, GRB, GRA and HG -- at more than $5\sigma$ using only the data from the DUNE experiment. \vspace{-0.3cm} \subsubsection{Alternative Cases and the Power of Data} \label{subsubsect372} In \cite{} the analyses performed in \cite{} was extended by obtaining sum rules for $\cos\delta$ for $U_{\rm PMNS}$ having the general form given in eq. (\ref{PMNS22}) and the following forms of $\tilde{U}_e$ and $U^\circ_\nu$~ \footnote{In \cite{} a systematic analysis of the forms of $\tilde{U}_e$ and $U^\circ_\nu$, for which sum rules for $\cos\delta$ of the type of eq.~(\ref{cosdthnu2}) could be derived, but did not exist in the literature, was performed.}: \begin{itemize} \item[{\bf C0}.] $U^\circ_\nu = R_{23}(\theta^{\nu}_{23})R_{12}(\theta^{\nu}_{12})$ with $\theta^{\nu}_{23} = -\pi/4$ and $\theta^{\nu}_{12}$ as dictated by TBM, BM, GRA, GRB or HG mixing, and i) $\tilde{U}_e = R^{-1}_{13}(\theta^{e}_{13})$ ($\Psi = {\rm diag}(1,1,e^{-i \omega})$), ii) $\tilde{U}_e = R^{-1}_{23}(\theta^{e}_{23})R^{-1}_{13}(\theta^{e}_{13})$ ($\Psi = {\rm diag}(1,e^{-i \psi},e^{-i \omega})$), and iii) $\tilde{U}_e = R^{-1}_{13}(\theta^{e}_{13})R^{-1}_{12}(\theta^{e}_{12})$ ($\Psi={\rm diag}(1,e^{-i \psi},e^{-i \omega})$); \vspace{-0.2cm} \item[{\bf D0}.] $U^\circ_\nu = R_{23}(\theta^{\nu}_{23})R_{13}(\theta^{\nu}_{13}) R_{12}(\theta^{\nu}_{12})$ with $\theta^{\nu}_{23}$, $\theta^{\nu}_{13}$ and $\theta^{\nu}_{12}$ fixed by arguments associated with symmetries, and iv) $\tilde{U}_e = R^{-1}_{12}(\theta^{e}_{12})$ ($\Psi ={\rm diag}(1,e^{-i \psi}, 1)$), and v) $\tilde{U}_e = R^{-1}_{13}(\theta^{e}_{13})$ ($\Psi = {\rm diag}(1,1,e^{-i \omega})$). \end{itemize} \vspace{-0.3cm} The sum rules for $\cos\delta$ were derived first for $\theta^{\nu}_{23} = -\,\pi/4$ for the cases listed in point {\bf C0}, and for the specific values of (some of) the angles in $U^\circ_{\nu}$, characterising the cases listed in point {\bf D0}, as well as for arbitrary fixed values of all angles contained in $U^\circ_{\nu}$. In certain models with $\sin^2\theta^\nu_{13}\neq 0$, $\sin^2\theta_{23}$ is predicted to have specific values which differ significantly from those in case {\bf B0} \cite{}: $\sin^2\theta_{23} = 0.455$; or 0.463; or 0.537; or 0.545, the uncertainties in these predictions being insignificant. Predictions for correlations between neutrino mixing angle values and/or sum rules for $\cos\delta$, which can be tested experimentally, were further derived in \cite{} for a large number of models based on $G_f = S_4$, $A_4$, $T^\prime$ and $A_5$ and all symmetry breaking patterns, i.e., all possible combinations of residual symmetries, which could lead to the correlations and sum rules of interest:\\ ({\bf A}) $G_e = Z_2$ and $G_{\nu} = Z_n$, $n > 2$ or $Z_n \times Z_m$, $n,m \geq 2$;\\ ({\bf B}) $G_e = Z_n$, $n > 2$ or $G_e = Z_n \times Z_m$, $n,m \geq 2$ and $G_{\nu} = Z_2$;\\ ({\bf C}) $G_e = Z_2$ and $G_{\nu} = Z_2$;\\ ({\bf D}) $G_e$ is fully broken and $G_{\nu} = Z_n$, $n > 2$ or $Z_n \times Z_m$, $n,m \geq 2$;\\ ({\bf E}) $G_e = Z_n$, $n > 2$ or $Z_n \times Z_m$, $n,m \geq 2$ and $G_{\nu}$ is fully broken. The three LH neutrino fields and the three LH charged lepton fields were assumed in \cite{} to transform under the action of $G_f$ by a 3-dimensional irreducible representation of $G_f$. In this case, as we have already remarked, the results obtained for $A_4$ and $T^\prime$ coincide. For each pattern, sum rules, i.e., relations between the neutrino mixing angles and/or between the neutrino mixing angles and the Dirac CPV phase $\delta$, when present, were derived. We note that neutrino mixing sum rules can exist also in the case of pattern {\bf D} ({\bf E}) if due to additional assumptions (e.g., additional symmetries) the otherwise unconstrained unitary matrix $U_e$ ($U_\nu$) is constrained to have the specific form of a matrix of $U(2)$ transformation in a plane or of the product of two $U(2)$ transformations in two different planes \cite{}. Thus, the cases of patterns {\bf D} and {\bf E} leading to interesting phenomenological predictions are ``non-minimal'' from the point of view of the symmetries employed (see, e.g., \cite{FlavourG,, }), compared to patterns {\bf A}, {\bf B} and {\bf C} characterised by non-trivial residual symmetries present in both charged lepton and neutrino sectors, which originate from just one non-Abelian flavour symmetry. As was shown in \cite{}, in the case of pattern {\bf A}, $U^\circ_\nu$ is fixed by $G_\nu$. There are three different general sub-cases, A1, A2 and A3~, corresponding to $U_e$ determined up to a unitary rotation in the 1-2, 1-3 and 2-3 plane, respectively. In sub-cases A1 and A2 one obtains a correlation between $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$ and a sum rule for $\cos\delta$, while in sub-case A3, $\sin^2\theta_{13}$ and $\sin^2\theta_{12}$ are predicted and $\delta$ is not constrained: \begin{equation} \sin^2 \theta_{23} = 1 - \frac{\cos^2 \theta^{\circ}_{13} \cos^2 \theta^{\circ}_{23}} {1-\sin^2\theta_{13}}\,,~~~{\rm A1}\,, \label{} \end{equation} \begin{equation} \cos \delta = \frac{\cos^2 \theta_{13} (\sin^2 \theta^{\circ}_{23} - \cos^2 \theta_{12}) + \cos^2 \theta^{\circ}_{13} \cos^2 \theta^{\circ}_{23} (\cos^2 \theta_{12} - \sin^2 \theta_{12} \sin^2 \theta_{13})}{\sin 2 \theta_{12} \sin \theta_{13} |\cos \theta^{\circ}_{13} \cos \theta^{\circ}_{23}| (\cos^2 \theta_{13} - \cos^2 \theta^{\circ}_{13} \cos^2 \theta^{\circ}_{23})^{\frac{1}{2}}}\,, ~{\rm A1}\,, \label{} \end{equation} \begin{equation} \sin^2 \theta_{23} = \frac{\sin^2 \theta^{\circ}_{23}}{1 - \sin^2 \theta_{13}}\,,~~~{\rm A2}\,, \label{} \end{equation} \begin{equation} \cos \delta = -\frac{\cos^2 \theta_{13} (\cos^2 \theta^{\circ}_{12} \cos^2 \theta^{\circ}_{23} - \cos^2 \theta_{12}) + \sin^2 \theta^{\circ}_{23} (\cos^2 \theta_{12} - \sin^2 \theta_{12} \sin^2 \theta_{13})}{\sin 2 \theta_{12} \sin \theta_{13} |\sin \theta^{\circ}_{23}| (\cos^2 \theta_{13} - \sin^2 \theta^{\circ}_{23})^{\frac{1}{2}}}\,,~{\rm A2}\,, \label{} \end{equation} \begin{equation} \sin^2\theta_{13} = \sin^2\theta^\circ_{13}\,, ~\sin^2\theta_{12} = \sin^2\theta^\circ_{12}\,,~ \cos\delta - {\rm unconstrained}\,,~~{\rm A3}\,, \label{} \end{equation} where the angles $\theta^\circ_{13}$, $\theta^\circ_{23}$ and $\theta^{\circ}_{12}$ are fixed once the flavour symmetry group $G_f$ and the residual symmetry subgroups $G_e$ and $G_\nu$ are specified. In the case of pattern {\bf B}, of which there are also of three different sub-cases, B1, B2 and B3, corresponding to $U_e$ fixed by $G_e$ and $U^\circ_\nu$ determined up to a unitary rotation in the 1-3, 2-3 and 1-2 plane, respectively, there exist a correlation between $\sin^2\theta_{12}$ and $\sin^2\theta_{13}$ and a sum rule for $\cos\delta$ (sub-cases B1, B2), or $\sin^2\theta_{13}$ and $\sin^2\theta_{23}$ are predicted while $\delta$ remains unconstrained (sub-case B3): \begin{equation} \sin^2 \theta_{12} = \frac{\sin^2 \theta^{\circ}_{12}}{1 - \sin^2 \theta_{13}}\,,~~~{\rm B1}\,, \label{} \end{equation} \begin{equation} \cos \delta = -\frac{\cos^2 \theta_{13} (\cos^2 \theta^{\circ}_{12} \cos^2 \theta^{\circ}_{23} - \cos^2 \theta_{23}) + \sin^2 \theta^{\circ}_{12} (\cos^2 \theta_{23} - \sin^2 \theta_{13} \sin^2 \theta_{23})} {\sin 2 \theta_{23} \sin \theta_{13} |\sin \theta^{\circ}_{12}| (\cos^2 \theta_{13} - \sin^2 \theta^{\circ}_{12})^{\frac{1}{2}}}\,,~{\rm B1}\,, \label{} \end{equation} \begin{equation} \sin^2 \theta_{12} = 1 - \frac{\cos^2\theta^{\circ}_{12} \cos^2\theta^{\circ}_{13} }{1 - \sin^2\theta_{13}}\,,~~~{\rm B2}\,, \label{} \end{equation} \begin{equation} \cos \delta = \frac{\cos^2 \theta_{13} (\sin^2 \theta^{\circ}_{12} - \cos^2 \theta_{23}) + \cos^2 \theta^{\circ}_{12} \cos^2 \theta^{\circ}_{13} ( \cos^2 \theta_{23} - \sin^2 \theta_{13} \sin^2 \theta_{23} )} {\sin 2 \theta_{23} \sin \theta_{13} | \cos \theta^{\circ}_{12} \cos \theta^{\circ}_{13}| (\cos^2 \theta_{13} - \cos^2 \theta^{\circ}_{12} \cos^2 \theta^{\circ}_{13} )^{\frac{1}{2}}}\,,~{\rm B2}\,, \label{} \end{equation} \begin{equation} \sin^2\theta_{13} = \sin^2\theta^\circ_{13}\,, ~\sin^2\theta_{23} = \sin^2\theta^\circ_{23}\,,~ \cos\delta - {\rm unconstrained}\,,~~{\rm B3}\,, \label{} \end{equation} where, as in the case of pattern {\bf A}, $\theta^\circ_{12}$, $\theta^\circ_{23}$ and $\theta^\circ_{13}$ are fixed once the symmetries are specified. Finally, in the case of pattern {\bf C}, of which there are altogether nine sub-cases, corresponding to $U_e$ and $U^\circ_\nu$, each determined by $G_e$ and $G_\nu$ up to a unitary rotations in the $i$-$j$ and $k$-$l$ planes, respectively, $i$-$j$=1-2,1-3,2-3, $k$-$l$= 1-2,1-3,2-3, there is either a correlation between $\sin^2\theta_{13}$ and $\sin^2\theta_{12}$, or between $\sin^2\theta_{13}$ and $\sin^2\theta_{23}$, or else a sum rule for $\cos\delta$. We number them as in \cite{}, i.e., cases C1--C9. Four of them lead to sum rules for $\cos\delta$, which have the form: \begin{align} &\text{C1, $(ij,kl) = (12,13)$:} \quad \cos \delta = \dfrac{\sin^2 \theta^{\circ}_{23} - \cos^2 \theta_{12} \sin^2 \theta_{23} - \cos^2 \theta_{23} \sin^2 \theta_{12} \sin^2 \theta_{13}} {\sin \theta_{13} \sin 2 \theta_{23} \sin \theta_{12} \cos \theta_{12}}\,, \label{] &\text{C3, $(ij,kl) = (12,23)$:} \quad \cos \delta = \dfrac{\sin^2 \theta_{12} \sin^2 \theta_{23} - \sin^2 \theta^{\circ}_{13} + \cos^2 \theta_{12} \cos^2 \theta_{23} \sin^2 \theta_{13}} {\sin \theta_{13} \sin 2 \theta_{23} \sin \theta_{12} \cos \theta_{12}}\,, \\[0.2cm] &\text{C4, $(ij,kl) = (13,23)$:} \quad \cos \delta = \dfrac{\sin^2 \theta^{\circ}_{12} - \cos^2 \theta_{23} \sin^2 \theta_{12} - \cos^2 \theta_{12} \sin^2 \theta_{13} \sin^2 \theta_{23}} {\sin \theta_{13} \sin 2 \theta_{23} \sin \theta_{12} \cos \theta_{12}}\,, \\[0.2cm] &\text{C8, $(ij,kl) = (13,13)$:} \quad \cos \delta = \dfrac{\cos^2 \theta_{12} \cos^2 \theta_{23} - \cos^2 \theta^{\circ}_{23} + \sin^2 \theta_{12} \sin^2 \theta_{23} \sin^2 \theta_{13}} {\sin \theta_{13} \sin 2 \theta_{23} \sin \theta_{12} \cos \theta_{12}}\,. \label{} \end{align} The neutrino mixing angles in these cases should be treated as free parameters. Other two cases, C5 and C9, yield correlations between $\sin^2\theta_{12}$ and $\sin^2\theta_{13}$: \begin{align} &\text{C5, $(ij,kl) = (23,13)$:} \quad \sin^2 \theta_{12} = \frac{\sin^2 \theta^{\circ}_{12}} {1 - \sin^2 \theta_{13}}\,,~\cos\delta - {\rm unconstrained}\,, \\[0.2cm] &\text{C9, $(ij,kl) = (23,23)$:} \quad \sin^2 \theta_{12} = \frac{\sin^2 \theta^\circ_{12} - \sin^2 \theta_{13}} {1 - \sin^2 \theta_{13}}\,,~\cos\delta - {\rm unconstrained}\,. \label{} \end{align} In cases C2 and C7, instead, there are correlations between $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$: \begin{align} &\text{C2, $(ij,kl) = (13,12)$:} \quad \sin^2 \theta_{23} = \frac{\sin^2 \theta^{\circ}_{23}} {1 - \sin^2 \theta_{13}}\,,~\cos\delta - {\rm unconstrained}\,, \label{] &\text{C7, $(ij,kl) = (12,12)$:} \quad \sin^2 \theta_{23} = \frac{\sin^2 \theta^\circ_{23} - \sin^2 \theta_{13}} {1 - \sin^2 \theta_{13}}\,,~\cos\delta - {\rm unconstrained}\,. \label{} \end{align} Finally, in case C6, $(ij,kl) = (23,12)$, $\cos\delta$ is unconstrained and $\sin^2\theta_{13}$ is predicted to be equal to $\sin^2\theta^\circ_{13}$. In cases C2, C5, C6, C7 and C9, as is indicated above, $\cos\delta$ remains unconstrained. Given the fact that the group $A_4$ has eight Abelian subgroups (three $Z_2$, four $Z_3$ and one Klein group $K_4$ isomorphic to $Z_2\times Z_2$), the group $S_4$ possesses 20 Abelian subgroups (nine $Z_2$, four $Z_3$, three $Z_4$ and four $Z_2\times Z_2$ groups, see, e.g., \cite{}), and $A_5$ has 36 Abelian subgroups (fifteen $Z_2$, ten $Z_3$, five $Z_2\times Z_2$ and six $Z_5$, see, e.g., \cite{}) the total number of the different residual symmetry patterns {\bf A}, {\bf B} and {\bf C} to be analysed is extremely large. For the group $A_4$ ($T^\prime$) alone there are altogether 64 cases (up to permutations of rows and columns of the predicted neutrino mixing matrix). It is quite remarkable that of these extremely large number of cases only a very limited number turned out to be phenomenologically viable, i.e., to be compatible with the existing data on the neutrino mixing angles \cite{}. In the case of the group $G_f = A_4~(T^\prime)$, for example, only one case was found to be phenomenologically viable \cite{}, i.e., to be compatible with the experimentally determined values (including the $3\sigma$ uncertainties) of the three neutrino mixing parameters $\sin^2 \theta_{12}$, $\sin^2 \theta_{13}$ and $\sin^2 \theta_{23}$. Namely, this is case B1 with $(G_e,G_\nu) = (Z_3,Z_2)$, which yields $(\sin^2\theta^\circ_{12},\sin^2\theta^\circ_{23}) = (1/3,1/2)$ and corresponds to the TBM mixing matrix corrected from the right by the $U_{13}(\theta^\nu_{13},\alpha)$ transformation in the 1-3 plane (see sub-section 3.3). The case B1 is common also to the two other groups $S_4$ and $A_5$. For $G_f = S_4$, there are 6 more viable cases. The $A_5$ flavour symmetry leads to 7 additional phenomenologically viable cases. One arrives at this results for the number of phenomenologically viable cases in the following way. For the groups $S_4$ and $A_5$ there are respectively altogether 8 and 13 cases, which are acceptable {\it a priori}, i.e., which lead to $U_{\rm PMNS}$ without zero entries. They are summarised in Table~3 (for $S_4$) and Table~4 (for $A_5$) of Ref.~\cite{}. In Tables~3 and 4 in~\cite{} the specific values of $\sin^2\theta^\circ_{ij}$ in each case are also given. However, the case B1, as we have already noticed, is common to all the three flavour symmetry groups $A_4~(T^\prime)$, $S_4$ and $A_5$, while four cases, C1, C3, C4 and C8, are shared by $S_4$ and $A_5$. Thus, there are 16 cases in total, which lead to different predictions for $\sin^2\theta_{12}$ or $\sin^2\theta_{23}$ and/or $\cos\delta$. A statistical analysis of these predictions showed \cite{} that two cases, namely, C4 (for both $S_4$ and $A_5$) and B2A$_5$II (i.e., the second of the two B2 cases with $G_f = A_5$, characterised by different fixed values of $\theta^\circ_{12}$ and $\theta^\circ_{13}$) are globally disfavored at more than $3\sigma$ confidence level by the current data (including the uncertainties) on $\sin^2\theta_{12}$, $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$ \cite{NuFITv32Jan2018}. As a consequence, only 14 cases altogether turned out to be phenomenologically viable at present. Five of them, B1, B1$A_5$, B2$S_4$, B2$A_5$, C9$A_5$ \footnote{The notation X$G_f$ means case X, X=A1,A2,...,B1,...,C1,...,C9, corresponding to the group $G_f$, $G_f=A_4,S_4,A_5$. The group is not indicated in cases B1, C1, C3 and C8 (see below) because case B1 is common to the $A_4$, $S_4$ and $A_5$ groups, while each of the cases C1, C3 and C8 is shared by the $S_4$ and $A_5$ groups.}, lead to sharp predictions for $\sin^2\theta_{12}$, and four others, A1$A_5$, A2$A_5$, C2$S_4$, C7$S_4$, to similarly sharp predictions for $\sin^2\theta_{23}$. The six phenomenologically viable cases ${\bf A}$ and ${\bf B}$ lead also to predictions for $\cos\delta$, while five out of the eight viable cases ${\bf C}$, C1, C3, C3$A_5$ (which differs from C3 that is common to $S_4$ and $A_5$), C4$A_5$ (which differs from C4) and C8, also lead to predictions for $\cos\delta$. Statistical analysis of these 14 cases was performed in \cite{} using the best fit values of the three neutrino mixing parameters $\sin^2\theta_{ij}$ from \cite{NuFITv32Jan2018} and taking into account the prospective ($1\sigma$) uncertainties in the determination of the mixing angles, planned to be achieved in currently running (Daya Bay \cite{}) and the next generation (JUNO \cite{JUNO}, T2HK \cite{T2HK2015}, DUNE \cite{DUNE2016}) of neutrino oscillation experiments: 3\% on $\sin^2\theta_{13}$ \cite{}, 0.7\% on $\sin^2\theta_{12}$ \cite{JUNO} and 3\% on $\sin^2\theta_{23}$ \cite{DUNE2016,T2HK2015}. This analysis revealed that only six cases would be compatible with the indicated prospective data from the Daya Bay, JUNO, T2HK, DUNE neutrino oscillation experiments. In Fig.~\ref{}, we present the likelihood functions for $\sin^2\theta_{12}$ and $\sin^2\theta_{23}$, obtained for NO and IO spectra in all the cases compatible at $3\sigma$ with the current global data~\cite{NuFITv32Jan2018}. The corresponding likelihood profiles are very narrow because their widths are determined by the small experimental uncertainty on $\sin^2\theta_{13}$. In the upper (lower) panel, the dashed line corresponds to the likelihood for $\sin^2\theta_{12}$ ($\sin^2\theta_{23}$) extracted from the global analysis. \begin{figure}[t] \centering \includegraphics[width=14.0cm]{LikelihoodSS12.pdf} \includegraphics[width=14.0cm]{LikelihoodSS23.pdf} \caption{Upper [lower] panel: predictions for $\sin^2\theta_{12}$ [$\sin^2\theta_{23}$] obtained using the current global data on the neutrino mixing parameters. ``Future'' refers to the scenario with $\sin^2\theta_{12}^\mathrm{bf} = 0.307$ [$\sin^2\theta_{23}^\mathrm{bf} = 0.538~(0.554)$ for NO (IO)] (current best fit values) and the relative $1\sigma$ uncertainty of $0.7\%$ [$3\%$] expected from JUNO [DUNE and T2HK]. See text for further details. (From Ref.~\cite{}.)} \label{} \end{figure} \noindent The dotted line represents the prospective precision on $\sin^2\theta_{12}$ ($\sin^2\theta_{23}$) corresponding to $1\sigma$ uncertainty of $0.7\%$ ($3\%$), which is planned to be achieved by JUNO \cite{JUNO} (DUNE \cite{DUNE2016} and T2HK \cite{T2HK2015}). It is obtained under the assumption that the best fit value(s) of $\sin^2\theta_{12}$ ($\sin^2\theta_{23}$) will not change in the future. If it is indeed the case, then, as is clear from Fig.~\ref{}, all five models leading to the predictions for $\sin^2\theta_{12}$ will be ruled out by the JUNO measurement of this parameter. The results of statistical analysis of the predictions for $\cos\delta$ are summarised in Fig.~\ref{}. The dashed line stands for the likelihood extracted from the global analysis~\cite{NuFITv32Jan2018}. At present, all (almost all) values of $\cos\delta$ are allowed at $3\sigma$ for NO (IO) spectrum. We also show the dash-dotted and dotted lines which represent two benchmark cases. The first case, marked as ``Future 1'', corresponds to the current best fit NO (IO) value~\cite{NuFITv32Jan2018} $\delta^\mathrm{bf} = 234^\circ~(278^\circ)$ and the prospective $1\sigma$ uncertainty on $\delta$ of $10^\circ$. The second case, ``Future 2'', corresponds to the potential best fit value $\delta^\mathrm{bf} = 270^\circ$ (for both NO and IO cases) and the same $10^\circ$ error on $\delta$. The likelihoods in cases C peak at values of $|\cos\delta|\sim 0.5 - 1$. Looking at the dotted line, we see that if in the future the best fit value of $\delta$ shifted to $270^\circ$ and the next generation of long-baseline experiments managed to achieve the $1\sigma$ uncertainty on $\delta$ of $10^\circ$, all cases C viable at the moment would be disfavored at approximately $3\sigma$ C.L. only by the measurement of $\delta$. \begin{figure}[t] \centering \includegraphics[width=14.0cm]{LikelihoodCosDeltaAB.pdf} \includegraphics[width=14.0cm]{LikelihoodCosDeltaC.pdf} \caption{Predictions for $\cos\delta$ obtained using the current global data on the neutrino mixing parameters. ``Future 1'' refers to the scenario with $\delta^{\rm bf} = 234^\circ~(278^\circ)$ for NO (IO) spectrum (current best fit values) and the $1\sigma$ uncertainty on $\delta$ of $10^\circ$. ``Future 2'' corresponds to $\delta^{\rm bf} = 270^\circ$ and the $1\sigma$ uncertainty on $\delta$ of $10^\circ$. See text for further details. (From Ref.~\cite{}.) } \label{} \end{figure} The results of the studies \cite{} summarised in the present subsection lead to the important conclusion that although the number of cases of non-Abelian discrete symmetry groups and their subgroups that can be used for description of lepton mixing is extremely large, only a very limited number survive when confronted with the existing data on the three neutrino mixing angles. This limited number of presently phenomenologically viable cases will be further considerably reduced by the precision measurements of the three neutrino mixing angles and the Dirac phase $\delta$ in the currently running (Daya Bay) and future planned (JUNO, T2HK, T2HKK, DUNE) neutrino oscillation experiments. As was shown in \cite{} and we have briefly discussed, if the best fit values of $\sin^2\theta_{12}$, $\sin^2\theta_{23}$ and $\sin^2\theta_{13}$ as found in \cite{NuFITv32Jan2018} would not change significantly in the future, only six cases would be compatible with the prospective data from the Daya Bay, JUNO, T2HK and DUNE neutrino oscillation experiments. This number would be further reduced by a precision measurement of the Dirac phase $\delta$. \section{Flavour Symmetry Combined with Generalised CP Symmetry} \label{sec4} In all models discussed by us the Majorana phases $\alpha_{21}$ and $\alpha_{31}$ remain undetermined. The values of the Majorana CPV phases are instead constrained to lie in certain narrow intervals, or are predicted, in theories which in addition to a flavour symmetry possess at a certain high-energy scale a ``generalised CP'' (GCP) symmetry \cite{}. The GCP symmetry, as the term suggests, is a generalisation of the traditional (canonical) CP symmetry. The GCP symmetry should be implemented in a theory based on a discrete flavour symmetry in a way that is consistent with the flavour symmetry \cite{}. At low energies the GCP symmetry is broken, in general, to residual CP symmetries of the charged lepton and neutrino sectors. The GCP transformations are applied on the LH and RH components of the charged lepton fields $\tilde{l}_L(x)$ and $\tilde{l}_R(x)$ and on the LH neutrino fields $\nu_{\tilde{l}L}(x)$ - the fields in terms of which the general charged lepton and neutrino Majorana mass terms are formed, eqs. (\ref{Le}) and (\ref{nuMajMass}). The transformations of interest are defines as follows: \begin{align} \tilde{l}_L(x) &\xrightarrow{CP} i (X_L)_{\tilde{l}\tilde{l}'} \gamma_0 C\, \overline{\tilde{l}'_L(x')}^T\,,~ \label{} \\[0.2cm] \tilde{l}_R(x) &\xrightarrow{CP} i (X_R)_{\tilde{l}\tilde{l}'} \gamma_0 C\, \overline{\tilde{l}'_R(x')}^T\,,~ \label{} \\[0.2cm] \nu_{\tilde{l}L}(x) &\xrightarrow{CP} i (X_L)_{\tilde{l}\tilde{l}'} \gamma_0 C\, \overline{\nu_{ \tilde{l}' L}(x')}^T\,, \label{} \end{align} where $\tilde{l} = \tilde{e},\tilde{\mu},\tilde{\tau}$, $X_L$ and $X_R$ are $3\times 3$ unitary matrices and $x'= (t,-\mathbf{x})$. The transformations of $\tilde{l}_L(x)$ and $\nu_{\tilde{l}L}(x)$ should involve the same matrix $X_L$ in order to ensure the CP invariance of the CC weak interaction Lagrangian, expressed in terms of the SM $SU(2)_L$ lepton doublet fields $\tilde{l}_L(x)$ and $\nu_{\tilde{l}L}(x)$: \begin{equation} \mathcal{L_{\rm CC}} = -\,\frac{g}{\sqrt{2}}\, \sum_{\tilde{l} = \tilde{e},\tilde{\mu},\tilde{\tau}} \overline{\tilde{l}_L(x)}\, \gamma_\alpha\, \nu_{\tilde{l}L}(x)\, W^{\alpha\dagger}(x) + \mathrm{h.c.}\,. \label{} \end{equation} The GCP symmetry will hold then in the lepton sector if it is a symmetry of the charged lepton and neutrino mass terms, eqs. (\ref{Le}) and (\ref{nuMajMass}), i.e., if the charged lepton and neutrino Majorana mass matrices satisfy the following constraints: \begin{align} X_L^\dagger\, M_e\, X_R = M_e^*\,,~or\\[0.2cm] X_L^\dagger\, M_e\,M^\dagger_e X_L = (M_e\,M^\dagger_e)^*\,,\\[0.2cm] X_L^T\, M_\nu\, X_L = M_\nu^*\,. \label{} \end{align} In the presence of flavour symmetry the form of the GCP transformations is significantly constrained. Indeed, consider a GCP transformation on a generic field $\varphi(x)$ which is assigned to an r-dimensional irreducible unitary representation $\rho_r(g)$ of $G_f$: \begin{equation} \varphi(x) \xrightarrow{CP} X_r\, \varphi^*(x')\,, \label{} \end{equation} where $X_r$ is a unitary matrix. The action of the CP transformation on the spinor indices in the case of $\varphi$ being a spinor (shown explicitly in eqs.~\eqref{}) has been omitted here for simplicity. If both the flavour symmetry and the GCP symmetry hold, the theory under study should be invariant also under the following sequence of transformations: a GCP transformation, followed by a flavour symmetry transformation, which in turn is followed by an (inverse) GCP transformation, i.e., under \begin{equation} \varphi(x) \xrightarrow{CP} X_r\, \varphi^*(x') \xrightarrow{G_f} X_r\, \rho_r(g_f)^*\, \varphi^*(x') \xrightarrow{CP^{-1}} X_r\, \rho_r(g_f)^*\, X_r^{-1}\, \varphi(x)\,. \end{equation} In order for the theory to be invariant under this sequence of transformations the resulting transformation should be a flavour symmetry transformation of $\varphi(x)$ corresponding to an element $g'_f$ of $G_f$, which can differ from $g_f$: \begin{equation} X_r\, \rho_r(g_f)^*\, X_r^{-1} = \rho_r(g'_f)\,, \qquad g_f\,,\, g'_f \in G_f\,. \label{} \end{equation} This equation represents a {\it consistency condition} which has to be satisfied in order for the implementation of the GCP symmetry in the theory to be compatible with the presence of a flavour symmetry \cite{}. For a discrete flavour symmetry group $G_f$, the consistency condition (\ref{}) will hold if it is satisfied by the group's generators. Let us denote by $H_\textrm{CP} = \{X_L\}$ the full set of GCP transformations $X_L$ acting on $\nu_{\tilde{l}L}(x)$ and $\tilde{l}_L(x)$, which are compatible with a given flavour symmetry group $G_f$, i.e., which satisfy the consistency condition (\ref{}) in which $X_r$ is replaced by $X_L$ and $\rho_r(g_f)$ is the irreducible unitary representation of $G_f$ to which $\nu_{\tilde{l}L}(x)$ and $\tilde{l}_L(x)$ are assigned. We denote further by $H^\nu_\textrm{CP} = \{X_\nu\}$ and $H^\ell_\textrm{CP} = \{X_\ell\}$ the sets of GCP transformations $X_\nu$ and $X_\ell$ which are compatible respectively with the residual flavour symmetries $G_{\nu}$ and $G_{\ell}$ of the neutrino and charged lepton sectors, i.e., of the neutrino Majorana and charged lepton mass terms. $X_\nu$ and $X_\ell$ satisfy consistency conditions in which $\rho_r(g_f)$ ($\rho_r(g'_f)$) is replaced respectively by $\rho_r(g_\nu)$ ( $\rho_r(g'_\nu)$) and $\rho_r(g_\ell)$ ($\rho_r(g'_\ell)$), where $g_\nu,g'_\nu \in G_{\nu}$ and $g_\ell,g'_\ell \in G_{\ell}$. The sets $H_\textrm{CP}$, $H^\nu_\textrm{CP}$ and $H^\ell_\textrm{CP}$ are not groups by themselves. They are sets of GCP transformations, which always involve conjugation of the fields they act upon; $H^\nu_\textrm{CP}$ and $H^\ell_\textrm{CP}$ are subsets of $H_\textrm{CP}$. These sets become groups only if they are extended by (at least) an identity element which does not conjugate fields (see, for instance, Appendix B of ref. \cite{}). When one writes a semi-direct product of $G_f$ and $H_{\rm CP}$, $G_f \rtimes H_{\rm CP}$, and the semi-direct \footnote{In the case of $G_\nu$ or $G_e$ being a $Z_2$ symmetry, the corresponding product becomes direct.} products of $G_\nu$ and $H_{\rm CP}^\nu$, $G_\nu \rtimes H_{\rm CP}^\nu$, and of $G_\ell$ and $H_{\rm CP}^\ell$, $G_\ell \rtimes H_{\rm CP}^\ell$, it is always implicitly assumed that $H_{\rm CP}$, $H_{\rm CP}^\nu$ and $H_{\rm CP}^\ell$ are appropriate groups, which are obtained from a single generating GCP transformation, as explained in Appendix B of ref. \cite{}. In this case $H^\nu_\textrm{CP}$ and $H^\ell_\textrm{CP}$ are subgroups of $H_\textrm{CP}$. \subsection{Implications for the Majorana Phases} \label{} As we have discussed in section \ref{subsec31}, the unitary matrix $U_{\nu}$ which diagonalises the neutrino Majorana mass matrix $M_\nu$ and enters into the expression for the PMNS matrix, eq. (\ref{UPMNSUeUnu}), is related to the unitary matrix $U^\circ_{\nu}$ diagonalising $M^\dagger_{\nu}\,M_{\nu}$ and $\rho_r(g_\nu)$ in the following way: $U_\nu \equiv U^\circ_\nu P^\circ$, where $P^\circ={\rm diag}(1,e^{i\frac{\xi_{21}}{2}},e^{i\frac{\xi_{31}}{2}})$ (see eq. (\ref{UnuU0nu})). The phases $\xi_{21}$ and $\xi_{31}$ contribute respectively to the Majorana phases $\alpha_{21}$ and $\alpha_{31}$ of the PMNS matrix, eq. (\ref{VP}). These phases remain undetermined by the flavour symmetries under discussion. We will consider next the implications of a residual GCP symmetry $H_{\rm CP}^\nu \subset H_{\rm CP}$, which is preserved in the neutrino sector, for the determination of $\xi_{21}$ and $\xi_{31}$, and thus of the Majorana phases $\alpha_{21}$ and $\alpha_{31}$. In the case of a residual GCP symmetry $H_{\rm CP}^\nu$, the neutrino Majorana mass matrix satisfies the condition given in eq. (\ref{}), $X_L \in H_{\rm CP}^\nu$ being the GCP transformation defined in eq. (\ref{}). Using eq.~(\ref{UnuMnu}) we find: \begin{equation} (X_L^{\rm d})^T\, M_\nu^{\rm d}\, X_L^{\rm d} = M_\nu^{\rm d}\,,~~\text{with}~~ M_\nu^{\rm d} = {\rm diag}(m_1,m_2,m_3)\,,~~ X_L^{\rm d} = U_\nu^\dagger \, X_L \, U_\nu^*\,. \label{} \end{equation} For $m_1 \neq m_2 \neq m_3$ and \footnote{It follows from the neutrino oscillation data that $m_1 \neq m_2 \neq m_3$, and that at least two of the three neutrino masses, $m_{2,3}$ ($m_{1,2}$) in the case of the NO (IO) spectrum, are non-zero. However, even if $m_1 = 0$ ($m_3 = 0$) at tree level and the zero value is not protected by a symmetry, $m_1$ ($m_3$) will get a non-zero contribution at least at two loop level \cite{CNLSPST83} and in the framework of a self-consistent (renormalisable) theory of neutrino mass generation this higher order contribution will be finite. } ${\rm min}(m_j)\neq 0$, $j=1,2,3$, as it is not difficult to show (see, e.g., \cite{}), the unitary matrix $X_L^{\rm d}$ can have only the following form: \begin{equation} X_L^{\rm d} = \diag(\pm1, \pm1, \pm1)\,, \end{equation} where the signs of the three non-zero entries in $X_L^{\rm d}$ are not correlated. Further, using that $U_\nu = U^\circ_\nu P^\circ$, we obtain from eq. (\ref{}) \cite{}: \begin{equation} (U^\circ_\nu)^\dagger \, X_L \, (U^\circ_\nu)^* = P^\circ \, X_L^{\rm d} \, P^\circ = {\rm diag} \left (\pm 1,\pm e^{i \xi_{21}}, \pm e^{i \xi_{31}}\right )\,. \label{} \end{equation} Thus, we come to the conclusion that the phases $\xi_{21}$ and $\xi_{31}$ will be determined provided i) the matrix $U^\circ_\nu$, which diagonalises $M^\dagger_{\nu}\,M_{\nu}$ and $\rho_r(g_\nu)$ (see eq. ({\ref{Uonurhognu}) and the related discussion) is fixed by the residual flavour symmetry $G_\nu$, and ii) the GCP transformations $X_L \in H_{\rm CP}^\nu$, which are consistent with $G_\nu$, are identified. \subsection{Concrete Examples of Symmetries} \label{} Now we turn to concrete examples \cite{}. For $G_f = A_4$ we choose to work in the Altarelli-Feruglio basis \cite{}, in which the $S$ and $T$ generators have the form given in eq. (\ref{S4STU3drep}). Preserving the $S$ generator, i.e., choosing $G_{\nu} = Z^S_2 = \{1,S\}$, leads to $U^\circ_\nu = V_{\rm TBM}$, provided there is an additional accidental $\mu$\,--\,$\tau$ symmetry \cite{}. The twelve GCP transformations consistent with the $A_4$ flavour symmetry for the triplet representation in the chosen basis have been found in \cite{}, solving the consistency condition \begin{equation} X_L \, \rho^*(g) \, X^{-1}_L = \rho (g^\prime)\,, \quad g, g^\prime \in A_4\,. \label{} \end{equation} These transformations can be summarised in a compact way as follows: \begin{equation} X_L = \rho (g)\,, \quad g \in A_4\,, \end{equation} i.e., the GCP transformations consistent with the $A_4$ flavour symmetry are of the same form as the flavour symmetry group transformations \cite{}. They are given in Table~1 in \cite{} together with the elements $\hat S$ and $\hat T$ to which the generators $S$ and $T$ of $A_4$ are mapped by the consistency condition in eq.~(\ref{}). Further, since in our case the residual flavour symmetry $G_\nu = Z^S_2 \times Z_2$, where the $Z^S_2$ factor corresponds to the preserved $S$ generator, only those $X$ are acceptable, for which $\hat S = S$. From Table~1 in \cite{} it follows that there are four such GCP transformations, namely, $\rho(E)$, $\rho(S)$, $\rho(T^2ST)$ and $\rho(TST^2)$, where $E$ is the identity element of the group. The last two transformations are not symmetric in the chosen basis, and, as shown in \cite{}, lead to partially degenerate neutrino mass spectrum with two equal masses (see also \cite{}), which is ruled out by the existing neutrino oscillation data. Thus, we are left with two allowed generalised CP transformations, $\rho(E)$ and $\rho(S)$, for which we have: \begin{align} & V_{\rm TBM}^{\dagger} \, \rho(E) \, V^*_{\rm TBM} = \rho(E) = \diag(1, 1, 1)\,, \\ & V_{\rm TBM}^{\dagger} \, \rho(S) \, V^*_{\rm TBM} = \diag(-1, 1, -1)\,. \end{align} Finally, according to eq.~(\ref{}), this implies that the phases $\xi_{21}$ and $\xi_{31}$ can be either $0$ or $\pi$. The same conclusion holds for a $T^{\prime}$ flavour symmetry, because restricting ourselves to the triplet representation for the LH charged lepton and neutrino fields, there is no way to distinguish $T^{\prime}$ from $A_4$ \cite{}. In the case of $G_f = S_4$ considered in \cite{}, the authors chose to work with the two generators $S$ and $T$ of $S_4$ in the basis given in \cite{}. In this case the generators $S$ and $T$ satisfy the following presentation rules: \begin{equation} S^2 = T^4 = (ST)^3 = (TS)^3 = {\bf 1}\,. \label{S4ST} \end{equation} Although the presentation rules for $S$ and $T$ given above differ from the presentation rules when the third generator $U$ for $S_4$ is employed, eq. (\ref{S4STU}), we will keep the notation $S$ and $T$ for the two generators satisfying the presentation rules (\ref{S4ST}) in the discussion which follows. In the basis chosen in \cite{} and used in \cite{}, $S$ and $T$ have the following form in the two triplet representations of interest: \begin{equation} S = \pm\, \begin{pmatrix} 0 &-\,\frac{1}{\sqrt{2}} &-\,\frac{1}{\sqrt{2}} \\ -\,\frac{1}{\sqrt{2}} & \frac{1}{2} & -\,\frac{1}{2} \\ -\,\frac{1}{\sqrt{2}} & -\,\frac{1}{2} & \frac{1}{2} \end{pmatrix}\,, \quad T = \pm\,\begin{pmatrix} -\,1 & 0 & 0 \\ 0 & -\,i & 0 \\ 0 & 0 & i \end{pmatrix}\,, \label{S4ST3drep} \end{equation} The residual symmetry $G_{\nu} = Z^S_2 \times Z_2$, where the $Z^S_2$ factor corresponds to preserved $S$ generator in the chosen basis and the second one arises accidentally (corresponding to a $\mu$\,--\,$\tau$ symmetry), leads to the bimaximal mixing, $U^\circ_\nu = V_{\rm BM}$ \cite{}. As in the previous example, the GCP transformations consistent with the $S_4$ flavour symmetry are of the same form as the flavour symmetry group transformations \cite{}. Solving the consistency condition in eq.~(\ref{}), in \cite{} ten symmetric GCP transformations consistent with the $S_4$ flavour symmetry for the triplet representation in the chosen basis were found. They are summarised in Table~\ref{} together with the elements $\hat T$ and $\hat S$ to which the consistency condition maps the group generators $T$ and $S$. \begin{table}[h] \centering \begin{tabular}{lcc} \toprule $g$, $X = \rho(g)$ & $T \rightarrow \hat T$ & $S \rightarrow \hat S$ \\ \midrule $(ST^2)^2$ & $T$ & $S$ \\ $T^3$ & $T^3$ & $T^3ST$ \\ $E$ & $T^3$ & $S$ \\ $T$ & $T^3$ & $TST^3$ \\ $T^2ST^2$ & $STS$ & $S$ \\ $ST^2S$ & $T$ & $T^2ST^2$ \\ $S$ & $TST$ & $S$ \\ $T^2$ & $T^3$ & $T^2ST^2$ \\ $STS$ & $ST^2$ & $ST^2ST$ \\ $TST$ & $T^2S$ & $TST^2S$ \\ \bottomrule \end{tabular} \caption{The ten symmetric generalised CP transformations $X = \rho(g)$ consistent with the $S_4$ flavour symmetry for the triplet representation $\rho$ in the chosen basis \cite{}, determined by the consistency condition in eq.~(\ref{}). The mapping $(T,S) \rightarrow (\hat T, \hat S)$ is realised via the consistency condition applied to the group generators $T$ and $S$, i.e., $X \rho^*(T) X^{-1} = \rho(\hat T)$ and $X \rho^*(S) X^{-1} = \rho(\hat S)$. $E$ denotes the identity element of $S_4$. (From \protect\cite{}.) \label{}} \end{table} From Table \ref{} we see that there are four symmetric GCP transformations consistent with the preserved $S$ generator, namely, $\rho(E)$, $\rho(S)$, $\rho(T^2ST^2)$ and $\rho(ST^2ST^2)$. Substituting them and $U^\circ_\nu = V_{\rm BM}$ in eq.~(\ref{}), one finds \cite{}: \begin{align} & V_{\rm BM}^{\dagger} \, \rho(E) \, V^*_{\rm BM} = \rho(E) = \diag(1,1,1)\,, \\ & V_{\rm BM}^{\dagger} \, \rho(S) \, V^*_{\rm BM} = \diag(1,-1,1)\,, \\ & V_{\rm BM}^{\dagger} \, \rho(T^2ST^2) \, V^*_{\rm BM} = \diag(-1,1,1)\,, \\ & V_{\rm BM}^{\dagger} \, \rho(ST^2ST^2) \, V^*_{\rm BM} = \diag(-1,-1,1)\,. \end{align} Therefore also in this case the phases $\xi_{21}$ and $\xi_{31}$ are fixed by the residual GCP symmetry to be either $0$ or $\pi$. As was shown in \cite{}, these results for the phases $\xi_{21}$ and $\xi_{31}$ hold also for $G_f = A_5$ and $G_\nu = Z_2\times Z_2$, generated by $\tilde{S}$ and $\tilde{T}^3\tilde{S}\tilde{T}^2\tilde{S}\tilde{T}^3$ and leading to the GRA mixing, $U^\circ_\nu = U_{\rm GRA}$ (see section \ref{subsec36}), when the flavour symmetry is combined with the GCP symmetry. If the matrix $U_e$ originating from the charged lepton sector is non-trivial, as like in the cases {\bf A} and {\bf B} defined by equations (\ref{PMNS22}), (\ref{Ue23122}) and (\ref{PsiAB2}), the Majorana phases $\alpha_{21}$ and $\alpha_{31}$ of the PMNS matrix receive also contributions from the phases associated with $U_e$ in the PMNS matrix \cite{}. In the specific examples of the forms {\bf A} and {\bf B} of $U_e$, the phases $\omega$ and/or $\psi$ of the matrix $\Psi$ in eqs. (\ref{PMNS22}) and (\ref{PsiAB2}), as we have already remarked, serve as a source for the Dirac CPV phase $\delta$ of the PMNS matrix and contribute to the Majorana phases $\alpha_{21}$ and $\alpha_{31}$ \cite{}. In these cases the Majorana phases $\alpha_{21}$ and $\alpha_{31}$ are determined by the sums respectively of the phases $\xi_{21}$ and $\xi_{31}$ and of the indicated additional contributions due to the phases in the matrix $\Psi$. As a consequence, $\alpha_{21}$ and $\alpha_{31}$ have non-trivial values which differ from $0$ or $\pi$ even when $\xi_{21}$ and $\xi_{31}$ are fixed by the employed residual GCP symmetry to be either $0$ or $\pi$ \cite{}. As we have indicated earlier, the Majorana phases play important role, e.g, in $|\Delta L| = 2$ processes like neutrinoless double beta ($(\beta\beta)_{0\nu}$-) decay $(A,Z) \rightarrow (A,Z+2) + e^- + e^-$, $L$ being the total lepton charge, in which the Majorana nature of massive neutrinos $\nu_i$ manifests itself (see, e.g, ref. \cite{}). Determining the values of the Majorana phases allows to make predictions for the basic $(\beta\beta)_{0\nu}-$ decay parameter -- the effective neutrino Majorana mass (see, e.g, refs. \cite{STPNuNature2013,}). \subsection{Examples of Models} \label{} In the scenarios involving a GCP symmetry, which were most widely explored so far (see, e.g., \cite{}), a non-Abelian flavour symmetry $G_f$ consistently combined with a GCP symmetry $H_{\rm CP}$ is broken to residual Abelian symmetries $G_e = Z_n$, $n > 2$, or $Z_m \times Z_k$, $m,\,k \geq 2$, and $G_\nu = Z_2 \times H^\nu_{\rm CP}$ of the charged lepton and neutrino mass terms, respectively~% \footnote{We note that in refs.~\cite{} the residual symmetry $G_e$ of the charged lepton mass term is augmented with a remnant CP symmetry $H^\ell_{\rm CP}$ as well. }. The factor $H^\nu_{\rm CP}$ in $G_\nu$ stands for a remnant GCP symmetry of the neutrino mass term. In such a set-up, $G_e$ fixes completely the form of the unitary matrix $U_e$ which diagonalises the product $M_e M_e^\dagger$ and enters into the expression of the PMNS matrix. At the same time, $G_\nu$ fixes the unitary matrix $U_\nu$, diagonalising the neutrino Majorana mass matrix $M_\nu$ up to a single free real parameter~---~% a rotation angle $\theta^\nu$. Given the fact that the PMNS neutrino mixing matrix $U_{\rm PMNS}$ is given by the product $U_{\rm PMNS} = U_e^\dagger\,U_\nu$, all three neutrino mixing angles are expressed in terms of this rotation angle. In this class of models one obtains specific correlations between the values of the three neutrino mixing angles, while the leptonic CPV phases are typically predicted to be exactly $0$ or $\pi$, or else $\pi/2$ or $3\pi/2$. For example, in the set-up considered in \cite{} and based on $G_f \rtimes H_{\rm CP} = S_4 \rtimes H_{\rm CP}$ broken to $G_e = Z_3^T$ and $G_{\nu} = Z_2^S \times H^\nu_{\rm CP}$ with $H^\nu_{\rm CP} = \{U,SU\}$~ \footnote{We recall that $S$, $T$ and $U$ are the generators of $S_4$ in the basis for its 3-dimensional representation specified in eq. (\ref{S4STU3drep}).}, the authors find: \begin{align} \sin^2\theta_{13} = \frac{2}{3}\sin^2\theta^\nu\,,\quad &\sin^2\theta_{12} = \frac{1}{2 + \cos2\theta^\nu} = \frac{1}{3\left(1 - \sin^2\theta_{13}\right)}\,,\quad \sin^2\theta_{23} = \frac{1}{2}\,, \label{Fer2012}\\[2mm] &|\sin\delta| = 1\,,\quad \sin\alpha_{21} = \sin\alpha_{31} = 0\,. \end{align} It follows, in particular, from the results on the neutrino oscillation parameters~---~best fit values, $2\sigma$ and $3\sigma$ allowed ranges~---~% obtained in the global fit of neutrino oscillation data \cite{} and summarised in Table~\ref{}, as well as in the more recent analyses \cite{NuFITv32Jan2018,}, that the prediction quoted in eq.~(\ref{Fer2012}) for $\sin^2\theta_{12}$ lies outside of its currently allowed $2\sigma$ range \footnote{We have used the best fit value of $\sin^2\theta_{13}$ to obtain the prediction of $\sin^2\theta_{12} = 0.341$ leading to the quoted conclusion. Using the $2\sigma$ allowed range for $\sin^2\theta_{13}$ leads to a minimal value of $\sin^2\theta_{12} = 0.340$, which is still larger than the maximal allowed value of $\sin^2\theta_{12}$ at $2\sigma$ C.L., but inside its $3\sigma$ allowed range.}. In what concerns the prediction $\sin^2\theta_{23} = 1/2$, according to \cite{NuFITv32Jan2018,}, it lies within the $1\sigma$ ($2\sigma$) allowed range of $\sin^2\theta_{23}$ for NO (IO) spectrum. Other examples of one (real angle) parameter models based on the flavour symmetry groups $A_4$, $S_4$ and $A_5$ combined with GCP symmetry can be found, e.g., in refs. \cite{}. Most of them share some of the properties of the model discussed above: the correlation between $\sin^2\theta_{12}$ and $\sin^2\theta_{13}$, the predictions that $|\sin\delta| = 1$, $\sin\alpha_{21} = \sin\alpha_{31} = 0$. Some of the models predict $\sin\delta = 0$, which is disfavored by the results of the global neutrino data analyses (see Table \ref{}). In certain set-ups the Majorana phases $\alpha_{21}$ and $\alpha_{31}$ take non-trivial values while $|\sin\delta| = 1$. In a class of models based on the groups $G_f = \Delta(3n^2)$ and $G_f = \Delta(6n^2)$ of flavour symmetry combined with GCP symmetry the neutrino mixing angles and the CPV phases are functions of one real angle and one or two discrete phase parameters, which depend on the parameter $n$ which characterises the size of the group $G_f$ (see, e.g., \cite{}). Due to the presence of the additional discrete valued phases, the CPV phase $\delta$ can have non-trivial and non-maximal values, i.e., one can have $|\sin\delta| \neq 1,0$. In this class of models, as a rule, there exist correlations between the values of $\sin^2\theta_{12}$, $\sin^2\theta_{23}$ and of $\sin^2\theta_{13}$ in the form of, e.g., the following relations \cite{}: $3\cos^2\theta_{13}\sin^2\theta_{12} = 1$ and $\sin^2\theta_{23} = 0.5 \pm 0.5\tan\theta_{13}\sqrt{2-\tan^2\theta_{13}}$. The first correlation $3\cos^2\theta_{13}\sin^2\theta_{12} = 1$ is typical for the models under discussion, while the second one or similar occur in most of them. For the best fit value of $\sin^2\theta_{13}$ given in Table ~\ref{}, the quoted relations lead to the predictions: $\sin^2\theta_{12} = 0.340$ and $\sin^2\theta_{23} = 0.396$ or 0.604. As we have already noticed, the value $\sin^2\theta_{12} = 0.340$ is outside the $2\sigma$, but within the $3\sigma$ allowed ranges of $\sin^2\theta_{12}$ found in \cite{}. As it follows from the results reported, e.g., in \cite{}, both predicted values of $\sin^2\theta_{23}$ lie outside the $3\sigma$ allowed range of $\sin^2\theta_{23}$. Theoretical models based on the approach to neutrino mixing that combines discrete symmetries and GCP invariance, in which the neutrino mixing angles and the leptonic CPV phases are functions of two or three parameters have also been considered in the literature (see, e.g., \cite{}). In these models the residual symmetry $G_e$ of the charged lepton mass term is typically assumed to be a $Z_2$ symmetry or to be fully broken. In spite of the larger number of parameters in terms of which the neutrino mixing angles and the leptonic CPV phases are expressed, the values of the CPV phases are still predicted to be correlated with the values of the three neutrino mixing angles. A set-up with $G_e = Z_2 \times H^e_{\rm CP}$ and $G_\nu = Z_2 \times H^\nu_{\rm CP}$ has been considered in \cite{}. The resulting PMNS matrix in such a scheme depends on two free real parameters~---~two angles $\theta^\nu$ and $\theta^e$. The authors have obtained several phenomenologically viable neutrino mixing patterns from $G_f = S_4$ combined with $H_{\rm CP}$, broken to all possible residual symmetries of the type indicated above. Models allowing for three free parameters (two real angles and one phase) have been investigated in \cite{}. In, e.g., \cite{}, the author has considered $G_f = A_5$ combined with $H_{\rm CP}$, which are broken to $G_e = Z_2$ and $G_\nu = Z_2 \times H^\nu_{\rm CP}$. In this case, the matrix $U_e$ depends on an angle $\theta^e$ and a phase $\delta^e$, while the matrix $U_\nu$ depends on an angle $\theta^\nu$. In these two scenarios the leptonic CPV phases possess non-trivial values. \vspace{-0.3cm} \section{ Outlook} \label{} \vspace{-0.3cm} The results obtained in refs.~\cite{} and in many other studies (quoted in the present and the cited articles) show that a sufficiently precise measurement of the Dirac phase $\delta$ of the PMNS neutrino mixing matrix in the current and future neutrino oscillation experiments, combined with planned improvements of the precision on the neutrino mixing angles, can provide unique information about the possible discrete symmetry origin of the observed pattern of neutrino mixing and, correspondingly, about the existence of new fundamental symmetry in the lepton sector. Thus, these experiments will not simply provide a high precision data on the neutrino mixing and Dirac CPV parameters, but will probe at fundamental level the origin of the observed form of neutrino mixing. These future data will show, in particular, whether Nature followed the discrete symmetry approach for fixing the values of the three neutrino mixing angles and of the Dirac and Majorana CP violation phases of the PMNS neutrino mixing matrix. We are looking forward to these data and to the future exciting developments in neutrino physics. \vspace{0.3cm} {\bf Acknowledgements.} I would like to thank my former Ph.D. students I. Girardi and A.V. Titov for the fruitful and enjoyable collaboration and the numerous illuminating discussions of the problems considered in the present article. This work was supported in part by the INFN program on Theoretical Astroparticle Physics (TASP), by the European Union Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grants 674896 and 690575, and by the World Premier International Research Center Initiative (WPI Initiative, MEXT), Japan.

\section{Introduction} Sequences (also known as right-infinite words; we use the terms interchangeably) that arise from the iteration of a morphism appear in many different places in the mathematical and theoretical computer science literature. To name just a few examples, they play an important role in the theory of words avoiding patterns \cite{}, in the theory of Sturmian sequences \cite{Berstel&}, in L-systems used in computer graphics \cite{Prusinkiewicz&}, and in the theory of algebraic series in positive characteristic \cite{}. Different kinds of morphisms give rise to different kinds of sequences, with different properties. In this paper we consider some of these variations and properties and classify sequences according to whether they do, or do not, satisfy these properties. The goal is to create a relatively complete taxonomy of the different kinds of behavior exhibited my morphic sequences, as well as to illustrate the many different techniques that can be used to show that a sequence exhibits, or does not exhibit, a certain property. Let us fix our notation. Let $\Sigma^*$ denote the set of all finite words over the alphabet $\Sigma$, including the empty word $\epsilon$, and let $\Sigma^\omega$ denote the set of all right-infinite words over $\Sigma$. We write $\Sigma^\infty = \Sigma^* \ \cup \ \Sigma^\omega$. Let $h:\Sigma^* \rightarrow \Sigma^*$ be a morphism, that is, a map obeying $h(xy) = h(x) h(y)$ for all words $x, y \in \Sigma^*$. From the definition, it suffices to define $h$ on each element of $\Sigma$. A morphism $h:\Sigma \rightarrow \Delta$ is said to be {\it non-erasing} if $h(a) \not= \epsilon$ for all $a \in \Sigma$. A letter $a$ is said to be {\it growing\/} for $h$ if $\lim_{n \rightarrow \infty} |h^n (a)| = \infty$; otherwise it is {\it non-growing}. We can define $h$ on $\Sigma^\omega$ in the obvious way: $h(a_1 a_2 \cdots) = h(a_1) h(a_2) \cdots$. We can iterate $h$, writing $h^2$ for the composition $h \circ h$, $h^3$ for $h \circ h \circ h$, etc. If there exists a letter $a \in \Sigma$ and $x \in \Sigma^*$ such that \begin{enumerate} \item $h(a) = ax$; and \item $h^i(x) \not= \epsilon$ for all $i \geq 0$, \end{enumerate} then $h$ is said to be {\it prolongable on $a$}. In this case, iterating $h$ on $a$ produces a sequence of words of increasing length, $$a, h(a), h^2(a), \ldots$$ where each word is a proper prefix of the word that follows. In the limit, this tends to the infinite word $$h^\omega(a) := a \, x \, h(x) \, h^2(x) \, h^3(x) \cdots \in \Sigma^\omega , $$ which is a fixed point of $h$; that is, $$ h(h^\omega(a)) = h^\omega(a).$$ We say that the infinite word $h^\omega(a)$ is {\it generated} by $h$. If $w$ is a finite word, then $|w|$ denotes the length of $w$. If $w$ is a finite or infinite word, then $w[n]$ denotes the $n$'th symbol of $w$, and $w[i..j]$ represents the factor of $\bf w$ beginning at position $i$ and ending at position $j$. If $x = uvw$, then we say that $u$ is a {\it prefix\/} of $x$, that $w$ is a {\it suffix\/} of $x$, and $v$ is a {\it factor\/} of $x$. Let $\Sigma_k = \{ 0,1,\ldots, k-1 \}$. If $w \in \Sigma_k^*$, we let $[w]_k$ denote the integer represented by $w$ in base $k$. We start by discussing different kinds of morphisms. \subsection{Pure morphic words} If $\bf w$ is an infinite word over $\Sigma$ and there exists a morphism $h$, prolongable on $a$, such that ${\bf w} = h^\omega (a)$, then $\bf w$ is said to be {\it pure morphic}. \begin{example} One of the most famous pure morphic words is the {\it Fibonacci word} $${\bf f} = 01001010 \cdots,$$ which is generated by the morphism $\varphi$ defined by $$0 \rightarrow 01; \quad\quad\quad 1 \rightarrow 0.$$ See, e.g., \cite{}. \label{fib} \end{example} \subsection{Pure uniform morphic words} A morphism $h$ is said to be {\it $k$-uniform} if $|h(a)| = k$ for all $a \in \Sigma$. It is said to be {\it uniform} if it is uniform for some $k \geq 2$. If an infinite word is generated by a uniform prolongable morphism, then it is said to be {\it pure uniform morphic}. \begin{example} One of the most famous pure uniform morphic words is the {\it Thue-Morse word} $$ {\bf t} = 01101001 \cdots,$$ which is generated by the morphism $\mu$ defined by $$0 \rightarrow 01; \quad\quad\quad 1 \rightarrow 10.$$ See, for example, \cite{}. \end{example} \subsection{Morphic words} We can also apply a coding (a $1$-uniform morphism from $\Sigma$ to a possibly different alphabet $\Delta$) to a morphic word. This has the effect of renaming the symbols. One way to think about this is to give letters in $\Sigma$ subscripts, and then the effect of $\tau$ is to erase the subscripts. We use the shorthand $a_1 a_2 \cdots a_n \rightarrow b_1 b_2 \cdots b_n$ to represent the coding $a_i \rightarrow b_i$ for $i = 1, 2, \ldots n$. If an infinite word is expressible as the coding of a pure morphic word, it is said to be ${\it morphic}$. \begin{example} The Fibonacci numbers are defined by $F_0 = 0$, $F_1 = 1$, and $F_n = F_{n-1} + F_{n-2}$. The {\it characteristic word} $\chi_F = (\chi (n))_{n \geq 0}$ of the Fibonacci numbers is defined by \begin{equation} \chi (n) = \begin{cases} 1, & \text{ if $n = F_i$ for some $i \geq 0$;}\\ 0, & \text{ otherwise.} \end{cases} \end{equation} Thus $$\chi = 11110100100001 \cdots .$$ This word is morphic, as it is generated by the morphism sending \begin{eqnarray*} c_0 & \rightarrow & c_0 e_0 \\ e_0 & \rightarrow & c_1 \\ c_1 & \rightarrow & c_1 e_1 \\ e_1 & \rightarrow & c_2 \\ c_2 & \rightarrow & c_2 e_1 \end{eqnarray*} followed by the coding $c_0 e_0 c_1 e_1 c_2 \rightarrow 11100$. See, for example, \cite[Example 7, p.~168]{}. \end{example} \subsection{Uniform morphic words} If an infinite word is expressible as the coding of a pure uniform morphic word, it is said to be {\it uniform morphic}. From Cobham's theorem \cite{}, we know that an infinite word is uniform morphic if and only if it is automatic. (A $k$-automatic word is one where the $n$'th term is computed by a finite automaton reading the base-$k$ expansion of $n$ as input and producing an output associated with the last state reached; a word is said to be automatic if it is $k$-automatic for some integer $k \geq 2$.) \begin{example} Consider the uniform morphic word $\bf s$ generated by the morphism \begin{eqnarray*} a & \rightarrow & ab \\ b & \rightarrow & ac \\ c & \rightarrow & db \\ d & \rightarrow & dc . \end{eqnarray*} Apply the coding sending $abcd \rightarrow 0011$ to $\bf s$; the result is $$ {\bf r} = 0001001000011101 \cdots,$$ the {\it Golay-Rudin-Shapiro} sequence. (See, for example, \cite{, }.) It is not hard to see that ${\bf r}[n]$ equals the parity of the number of (possibly overlapping) occurrences of $11$ in the binary expansion of $n$. \label{rudin} \end{example} \subsection{Pure primitive morphic words} A morphism $h:\Sigma^* \rightarrow \Sigma^*$ is called {\it primitive} if there exists an integer $n \geq 1$ such that for all $a \in \Sigma$, the word $h^n(a)$ contains at least one occurrence of each symbol in $\Sigma$. For example, the Fibonacci morphism $\varphi$ is primitive, since $\varphi^2 (0) = 010$ and $\varphi^2 (1) = 01$. An infinite word $\bf w$ is {\it pure primitive morphic} if it is the fixed point of some primitive morphism. \begin{example} Consider the morphism $\gamma$ defined by \begin{align*} \gamma(0) &= 03 \\ \gamma(1) &= 43 \\ \gamma(3) &= 1 \\ \gamma(4) &= 01. \end{align*} Then \begin{align*} \gamma^6(0) &= 03143011034343031011011 \\ \gamma^6(1) &= 03143011031011011031011011 \\ \gamma^6(3) &= 03143034343034343 \\ \gamma^6(4) &= 0314301103434303143034343034343 , \end{align*} so $\gamma$ is primitive. The infinite word $0314301103434303101101103\cdots$ generated by $\gamma$ is thus pure primitive morphic. It was studied in \cite{Cassaigne&Currie&Schaeffer&} and has the interesting property that it avoids additive cubes, that is, three consecutive blocks of the same length and same sum. \end{example} \subsection{Primitive morphic words} An infinite word is {\it primitive morphic} if it is the image, under a coding, of some primitive morphism. \begin{example} An interesting example of a primitive morphic word appears in \cite{Du&Mousavi&Rowland&Schaeffer&}. Consider the morphism $g$ and coding $\rho$ defined as follows: \begin{align*} g(a) &= abcab \quad & \rho(a) = 0 \\ g(b) &= cda \quad & \rho(b) = 0 \\ g(c) &= cdacd \quad & \rho(c) = 1 \\ g(d) &= abc \quad & \rho(d) = 1 \end{align*} and form the infinite word ${\bf R} = \rho(g^\omega(a))$. By considering $g^2$ we see that $g$ is primitive and so $\bf R$ is primitive morphic. The word $\bf R$, sometimes called the Rote-Fibonacci word, is an aperiodic word that avoids the pattern $x x x^R$, where $x^R$ denotes the reversal of $x$. \end{example} \subsection{Uniform primitive morphic words} Finally, an infinite word is {\it pure uniform primitive morphic} if it is the fixed point of a primitive morphism that is also uniform, and {\it uniform primitive morphic} if it is the image, under a coding of such a word. \subsection{Recurrence and uniform recurrence} An infinite word $\bf w$ is said to be {\it recurrent} if every finite factor $x$ appearing in $\bf w$ appears infinitely often in $\bf w$. If, in addition, for each factor $x$ there is a constant $c(x)$ such that every two consecutive occurrences of $x$ in $\bf w$ are separated by no more than $c(x)$ symbols, then we say that $\bf w$ is {\it uniformly recurrent}. Synonyms for uniformly recurrent in the literature include ``almost periodic'' \cite{} and ``minimal'' \cite{}. A basic result connecting uniform recurrence with other properties is the following theorem of Cobham \cite[Theorem 5, p.\ 178]{}: \begin{theorem} \leavevmode \begin{itemize} \item[(a)] If $\bf w$ is a primitive morphic word, then it is uniformly recurrent. \item[(b)] If $\bf w$ is uniformly recurrent and uniform morphic, then it is uniform primitive morphic. \end{itemize} \label{cobham} \end{theorem} We now mention some other useful results: \begin{theorem} Let $\bf u$ be a fixed point of a morphism $\sigma$ ``without useless letters" (i.e., the alphabet on which $\sigma$ is defined is exactly the set of letters occurring in $\bf u$). Furthermore suppose that for each letter $a$ in $\bf u$ the length of the iterates $|\sigma^k(a)|$ tends to infinity. If the sequence $\bf u$ is uniformly recurrent, then $\sigma$ is primitive. \end{theorem} \begin{proof} By the hypothesis on the letters being all useful, it follows that for every letter $a$ that $\sigma^k(a)$ is a factor of $\bf u$ for all $k \geq 0$. Since $u$ is uniformly recurrent, for every letter $b$ there exists a length $\ell_b$ such that every factor of $u$ of length $\geq \ell_b$ contains at least one $b$. Taking $\ell = \max_{b \in \Sigma} \ell_b$, every factor of $\bf u$ of length $\geq \ell$ contains at least one copy of each letter. On the other hand, for each $a$ we know that $|\sigma^k(a)|$ is unbounded. Hence there exists $k_a$ such that for all $k \geq k_a$ one has $|\sigma^{k_a}(a)| \geq \ell$. Taking $K = \max k_a$, we see that for each $k \geq K$ and for each letter $a$, the word $\sigma^k(a)$ contains at least one copy of each letter. In other words, $\sigma$ is primitive. \end{proof} \begin{corollary} If a sequence is pure uniform morphic and uniformly recurrent, then the sequence is pure uniform primitive morphic. \end{corollary} \begin{proof} Clean out the useless letters. \end{proof} \begin{corollary} If a sequence is pure uniform morphic and primitive morphic, then the sequence is pure uniform primitive morphic. \label{allou} \end{corollary} \begin{proof} The sequence is uniformly recurrent. Apply the previous corollary. \end{proof} We recall a result from Queff{\'e}lec \cite[Prop.~5.5, p.~130]{}: \begin{theorem} Let $h:\Sigma^* \rightarrow \Sigma^*$ be a morphism, prolongable on $a$, and suppose all letters of $\Sigma$ are growing. Then $h^\omega(a)$ is uniformly recurrent if and only if $h$ is primitive. \label{queffelec} \end{theorem} Finally, we also mention a recent important result of Durand \cite[Thm.~3, p.~124]{}: \begin{theorem} Let $\bf u$ be a morphic sequence that is uniformly recurrent. Then $\bf u$ is primitive morphic. \label{durand} \end{theorem} \begin{remark} More precisely, Durand proves in \cite{} that a sequence is uniformly recurrent and morphic if and only if it is ``primitive substitutive'', but his theorem also implies the statement above (F. Durand, private communication, June 2017). \end{remark} \section{The classification} From the preceding section, a word can be classified in ten different ways: \begin{itemize} \item[P1:] pure morphic \item[P2:] morphic \item[P3:] pure uniform morphic \item[P4:] uniform morphic \item[P5:] pure primitive morphic \item[P6:] primitive morphic \item[P7:] pure uniform primitive morphic \item[P8:] uniform primitive morphic \item[P9:] uniformly recurrent \item[P10:] recurrent \end{itemize} However, these ten properties are clearly not independent. We have the following trivial implications: \begin{itemize} \item P1 $\implies$ P2 \item P3 $\implies$ P1, P2, P4 \item P4 $\implies$ P2 \item P5 $\implies$ P1, P2, P6 \item P6 $\implies$ P2 \item P7 $\implies$ P1, P2, P3, P4, P5, P6, P8 \item P8 $\implies$ P2, P4, P6 \item P9 $\implies$ P10 \end{itemize} Theorem~\ref{cobham} (a) tells us that P6 $\implies$ P9, and Theorem~\ref{cobham} (b) tells us (P6 and P4) $\implies$ P8. Corollary~\ref{allou} tells us that (P3 and P9) $\implies$ P7. Theorem~\ref{durand} tells us that (P2 and P9) $\implies$ P6. All these restrictions lower the total number of possibilities from 1024 to 20: \begin{itemize} \item[(a)] Neither morphic nor recurrent. \item[(b)] Recurrent, but neither morphic nor uniformly recurrent. \item[(c)] Uniformly recurrent, but not morphic. \item[(d)] Morphic; but neither pure morphic, uniform morphic, primitive morphic, nor recurrent. \item[(e)] Morphic and recurrent; but neither pure morphic, uniform morphic, primitive morphic, nor uniformly recurrent. \item[(f)] Primitive morphic; but neither pure morphic nor uniform morphic. \item[(g)] Uniform morphic; but neither pure morphic, primitive morphic, nor recurrent. \item[(h)] Uniform morphic and recurrent; but neither pure morphic nor primitive morphic. \item[(i)] Uniform primitive morphic; but not pure morphic. \item[(j)] Pure morphic; but neither uniform morphic, primitive morphic, nor recurrent. \item[(k)] Pure morphic and recurrent; but neither uniform morphic, primitive morphic, nor uniformly recurrent. \item[(l)] Pure morphic and primitive morphic; but neither uniform morphic nor pure primitive morphic. \item[(m)] Pure primitive morphic; but not uniform morphic. \item[(n)] Pure morphic and uniform morphic, but neither pure uniform morphic, primitive morphic, nor recurrent. \item[(o)] Pure morphic and uniform morphic and recurrent, but neither pure uniform morphic nor primitive morphic. \item[(p)] Pure morphic and uniform primitive morphic; but neither pure uniform morphic nor pure primitive morphic. \item[(q)] Pure primitive morphic and uniform primitive morphic; but not pure uniform morphic. \item[(r)] Pure uniform morphic; but neither primitive morphic nor recurrent. \item[(s)] Pure uniform morphic and recurrent; but not primitive morphic. \item[(t)] Pure uniform primitive morphic. \end{itemize} In this note we give examples of all 20 possibilities. Of course, examples of some of these cases are very well-known; it is our point to collect these examples in one place and to illustrate each of the 20 classes. Inexplicably, the authors of \cite{Allouche&} failed to do this explicitly. Before we get to the examples, we recall several more useful results. The first few involve frequency of letters. Define the frequency of a symbol $a$ in an infinite word $\bf w$ to be the quantity $\lim_{n \rightarrow \infty} {{{{\bf w}[0..n-1]}_a} \over{n}}$, if it exists. \begin{theorem} Suppose the frequency $\alpha$ of the letter $a$ in the word $\bf w$ exists. Then \begin{itemize} \item[(a)] If $\bf w$ is morphic then $\alpha$ is algebraic. \item[(b)] If $\bf w$ is uniform morphic then $\alpha$ is rational. \end{itemize} \label{freq} \end{theorem} For a proof, see \cite[Theorem 8.4.5, p.~268]{Allouche&}. \begin{proposition} Let $\bf w$ be a morphic sequence and let $a$ be a letter occurring infinitely often in $\bf w$. Then the number of occurrences of $a$ in a prefix of length $n$ of $\bf w$ is $\Omega(\log n)$. \label{positions} \end{proposition} \begin{proof} This follows from the matrix representation of the morphism. See, for example, \cite[Cor.~8.2.4, p.~249]{Allouche&}. \end{proof} The last results involve subword complexity (or factor complexity), the number of distinct factors of length $n$. We recall the following theorem of Pansiot \cite{}: \begin{theorem} If $\bf w$ is a pure morphic word, then the subword complexity of $\bf w$ is in $\Theta(1)$, $\Theta(n)$, $\Theta (n\log\log n)$, $\Theta(n \log n)$, or $\Theta(n^2)$. \label{subword} \end{theorem} We also recall the following theorem: \begin{theorem} If $\bf w$ is a uniform morphic word or a primitive morphic word that is not ultimately periodic, then the subword complexity of $\bf w$ is $\Theta(n)$. \label{uniprim} \end{theorem} \begin{proof} For uniform morphic words, see \cite[Thm.~2, p.~171]{}. For primitive morphic words, see \cite{}. \end{proof} \section{The examples} We now turn to providing examples of all of the 20 possibilities listed in the previous section. \begin{example} {\bf (a) A word that is neither morphic nor recurrent.} Consider the binary word $01100010000000000000000010 \cdots$ that is the characteristic sequence of the factorials $1, 2, 6, 24, \ldots$. This word has $n = O((\log N)/(\log\log N))$ $1$'s in a prefix of length $N = n! + 1$, and so by Proposition~\ref{positions} it cannot be morphic. It is evidently not recurrent because the factor $11$ appears only once. \end{example} \begin{example} {\bf (b) A word that is recurrent, but neither morphic nor uniformly recurrent.} Consider the binary word $$ {\bf b} = 11011100101110111 \cdots$$ formed by the concatenation of the binary expansions of $1, 2, 3, \ldots$ in order. This word clearly has $2^n$ distinct factors of length $n$, and hence by Theorem~\ref{subword} cannot be morphic. Nor is $\bf b$ uniformly recurrent, because it contains arbitrarily long blocks of $0$'s. However, it is recurrent. \end{example} \begin{example} {\bf (c) A word that is uniformly recurrent, but not morphic.} Consider the Sturmian characteristic word ${\bf s}_{\alpha} = s_0 s_1 s_2 \cdots$ defined by $s_n = \lfloor (n+1) \alpha \rfloor - \lfloor n\alpha \rfloor$ for $n \geq 0$. It is well-known that all such words are uniformly recurrent, see, e.g., \cite[Proposition 3.17, p.~186]{}. However, for ${\bf s}_{\alpha}$ to be morphic, the number $\alpha$ must be a quadratic irrational \cite[Prop.~2.11]{Berthe&Holton&}. So take $\alpha = \pi$, for example. \end{example} \begin{example} {\bf (d) A word that is morphic; but neither pure morphic, uniform morphic, primitive morphic, nor recurrent.} Take the Fibonacci word mentioned above in Example~\ref{fib}, and change the first two symbols to $2$, giving the word $$ {\bf f}' = 22001010 \cdots .$$ Now ${\bf f}'$ cannot be pure morphic, because if it were, then we would be able to write it as $h^\omega(2)$ where $h(2)$ begins with $22$. Then iterating $h$ would produce infinitely many $2$'s, a contradiction. This word cannot be uniform morphic by Theorem~\ref{freq}, because the frequency of the symbol $0$ is the same as that in the Fibonacci word, namely $(\sqrt{5} -1)/2$, which would contradict Theorem~\ref{freq}. Finally, ${\bf f}'$ cannot be primitive morphic by Theorem~\ref{cobham} because the symbol 2 only occurs twice, and so ${\bf f}'$ is not recurrent. However, the word ${\bf f}'$ is morphic. It is generated by the morphism $a \rightarrow ab$, $b \rightarrow c$, $c \rightarrow cd$, $d \rightarrow c$, followed by the coding $abcd \rightarrow 2201$. \end{example} \begin{example} {\bf (e) A word that is morphic and recurrent; but neither pure morphic, uniform morphic, primitive morphic, nor uniformly recurrent.} Consider the fixed point $\bf x$ of the morphism $a\rightarrow ababb$, $b \rightarrow bc$, $c \rightarrow c$ coded by $\tau$ sending $a\rightarrow 0$ and $b,c\rightarrow 1$. An easy induction gives $$ {\bf x} = 0101111010111111111010111101 \cdots = \prod_{n \geq 1} 0 1^{a(n)} $$ where $a(n) = (\nu_2 (n)+1)^2$ and $\nu_2 (n)$ is the exponent of the highest power of $2$ dividing $n$. Then $0 1^n 0 $ occurs in $\bf x$ if and only if $n$ is a perfect square. Following the construction in \cite{}, it can be shown that $\bf x$ has subword complexity $\Theta(n \sqrt{n})$, so by Theorem~\ref{uniprim} $\bf x$ is not uniform morphic or primitive morphic. By Theorem~\ref{subword} it is not pure morphic. It is recurrent but not uniformly recurrent. \end{example} \begin{example} {\bf (f) A word that is primitive morphic, but neither pure morphic nor uniform morphic.} From a theorem of Yasutomi \cite{}, we know that the Sturmian word ${\bf s}_{\alpha, \rho} = s_0 s_1 s_2 \cdots$ defined by $s_n = \lfloor (n+1) \alpha + \rho \rfloor - \lfloor n\alpha + \rho \rfloor$ is pure morphic if and only if $\alpha$ is a quadratic irrational, $\rho \in {\mathbb{Q}}(\alpha)$, and either $\alpha' > 1$, $1 - \alpha' \leq \rho' \leq \alpha'$ or $\alpha' < 0$, $\alpha' \leq \rho' \leq 1- \alpha'$ where $\alpha'$ is the (algebraic) conjugate of $\alpha$. Now consider the case where $\alpha = (3-\sqrt{5})/2$. Then $\alpha' = (3+\sqrt{5})/2 > 1$. Take $\rho = 2\alpha = 3-\sqrt{5}$. Then $\rho' = 3+\sqrt{5} > \alpha'$, so ${\bf s}_{\alpha, \rho}$ is not pure morphic. However, this word is just the shift of the Fibonacci word $\bf f$, and hence is morphic. In fact, it is easy to see that ${\bf s}_{\alpha, \rho} = \tau(h^\omega(a))$, where $h: a\rightarrow ac, c \rightarrow b, b \rightarrow ac$ and $\tau(abc) = 100$. Then $h$ is primitive, as $h^3$ applied to each letter contains every letter. So ${\bf s}_{\alpha, \rho}$ is primitive morphic. It cannot be uniform morphic by Theorem~\ref{freq}. \end{example} \begin{example} {\bf (g) A word that is uniform morphic; but neither pure morphic, primitive morphic, nor recurrent.} As is well known, the morphism $g$ defined by $2 \rightarrow 210, 1 \rightarrow 20, 0 \rightarrow 1$ generates a squarefree word $g^\omega (2) = 210201\cdots$; see \cite{}. Now consider the morphism $h$ defined by \begin{eqnarray*} a & \rightarrow & ab \\ b & \rightarrow & ca \\ c & \rightarrow & cd \\ d & \rightarrow & ac \end{eqnarray*} and the coding $\tau$ defined by $\tau(abcd) = 2101$. Then it is known that $\tau(h^\omega(a)) = g^\omega(2)$; see \cite{}. Hence $g^\omega(2)$ is $2$-automatic and hence uniform morphic. Now take the word $g^\omega(2)$ and change the first $1$ to $2$, obtaining the word ${\bf w} = 220201210120\cdots$. The resulting word $\bf w$ is $2$-automatic since $g^\omega(2)$ is. (In fact, it is generated by iterating the morphism \begin{align*} 0 & \rightarrow 01 & \quad & 1 \rightarrow 23 \\ 2 & \rightarrow 24 & \quad & 3 \rightarrow 35 \\ 4 & \rightarrow 32 & \quad & 5 \rightarrow 23 \end{align*} followed by the coding $012345 \rightarrow 220211$.) But $\bf w$ is not pure morphic. Suppose ${\bf w} = \xi^\omega (2)$. Then $\xi$ maps $2$ to a word beginning with $22$, which means that $22$ occurs infinitely often in $\bf w$, a contradiction. Finally, $\bf w$ is neither primitive morphic nor recurrent because $22$ only occurs once. \label{example15} \end{example} \begin{example} {\bf (h) A word that is uniform morphic and recurrent; but neither pure morphic nor primitive morphic.} The idea is to construct a word that contains a nonzero but finite number of $k$-th powers. Such a word cannot be pure uniform morphic, and it can only be pure morphic if the $k$-th powers are made up of non-growing letters. Consider ${\bf v}$, the fixed point of the morphism $a \rightarrow abba$, $b\rightarrow bccb$, $c \rightarrow cbbc$. Apply the coding $abc \rightarrow 001$ to get $$ {\bf x} = 0000011001100000011010011001011001101001100101100000 \cdots .$$ Clearly $\bf x$ is uniform morphic and recurrent. However, $000$ occurs with unbounded gaps between occurrences, since arbitrarily long factors of Thue-Morse occur, so $\bf x$ is not uniformly recurrent. If $\bf x$ were pure morphic, it would be fixed by some morphism $f$ prolongable on $0$, then it would start with $f(0)^5$, since $f(0)$ starts with $000001$. But this is not possible, because there is no factor $000001u000001u0000$. To see this, note that $000001$ occurs only at positions that are a multiple of 4, and once $000001u$ is synchronized modulo 4, it can be factored into $\{0000, 0110, 1001 \}$ and decoded. An occurrence of $000001u000001u0000$ then corresponds to an overlap in $\bf v$, which is impossible since $\bf v$ is overlap-free (applying the coding $abc \rightarrow 010$ gives the Thue-Morse sequence). \bigskip An alternative construction creates a word with unbounded powers, but not of the right kind: Consider the fixed point of $a \rightarrow aba$, $b \rightarrow ccc$, $c \rightarrow ccc$ coded by $a,b \rightarrow 0$, $c \rightarrow 1$, generating the word ${\bf y} = 000111000111111111000111000 \cdots$. As in the previous example, $\bf y$ is uniform morphic, recurrent, and not uniformly recurrent. If it were pure morphic, fixed by some $f$, then it would start with $f(0)^3$, since $f(0)$ starts with $000$. But $000u000u000$ does not occur: if $k$ is the size of the largest block of ones in $u$ (it is a power of 3), then we have two occurrences of $01^k0$ without a larger block of ones between them, which is always the case. In other words: our word is the coding of $010201030102 \cdots$ (the ``ruler sequence'' $(\nu_2 (n))_{n \geq 1}$) under $i \rightarrow 0001^{3^{i+1}}$. An occurrence of $000u000u000$ could be decoded into a square in $010201030102 \cdots$, which is square-free. This ends our second construction for this case. \end{example} \begin{example} {\bf (i) A word that is uniform primitive morphic; but not pure morphic.} Here our example is $\bf r$, the Rudin-Shapiro sequence. The morphism and coding given in Example~\ref{rudin} show that $\bf r$ is uniform morphic. Assume it is pure morphic, generated by a morphism $g$. Since $\bf r$ starts with $000$, it must be that $g({\bf r}) = {\bf r}$ contains arbitrarily large cubes (namely, $g^n (0) g^n(0) g^n(0)$ for all $n \geq 1$). But from a well-known result \cite{Allouche&} (also see \cite{Kao&Rampersad&Shallit&}), the only cubes in $\bf r$ are $000$ and $111$, a contradiction. \end{example} \begin{example} {\bf (j) A word that is pure morphic; but neither uniform morphic, primitive morphic, nor recurrent.} Take the Fibonacci sequence $\bf f$ discussed above in Example~\ref{fib} and change the first symbol from $0$ to $2$. The resulting sequence ${\bf u} = 21001010\cdots$ is pure morphic, since it is generated by the morphism that sends $2 \rightarrow 21$, $1 \rightarrow 0$, $0 \rightarrow 01$. However, it is not automatic by Theorem~\ref{freq} since $0$ occurs with frequency $(\sqrt{5}-1)/2$, which is irrational. It is neither recurrent nor primitive morphic since $2$ occurs only once in ${\bf u}$. \end{example} \begin{example} {\bf (k) A word that is pure morphic and recurrent; but neither uniform morphic, primitive morphic, nor uniformly recurrent.} Consider the morphism $h$ defined by $h(0) = 010$ and $h(1) = 11$. Then $h^\omega(0)$ is evidently pure morphic. It is evidently recurrent because any block that appears must appear in $h^n (0)$ for some $n$, and then that block appears twice in $h^n(010) = h^{n+1} (0)$. It is not uniformly recurrent because there are arbitrarily long blocks of $1$'s. So it is also not primitive morphic. Suppose it is $k$-automatic for some $k$. Using the ``logical'' approach to automatic sequences \cite{Bruyere&Hansel&Michaux&}, the sequence ${\bf u} = (u_n)_{n \geq 0}$ defined by $$ u_n = \begin{cases} 1, & \text{if the position of the first occurrence of a block of $n$ consecutive $1$'s in $h^\omega(0)$} \\ & \text{is not the position of the first occurrence of a block of $n+1$ consecutive $1$'s}; \\ 0, & \text{otherwise;} \end{cases} $$ is also $k$-automatic. But this sequence (it is easy to see) is the characteristic sequence of powers of $2$. So we can assume $k = 2$. Again, using the ``logical" approach, the function $f(n)$ computing the starting position of the first occurrence of a block of $n$ consecutive $1$'s in the word is ``$k$-synchronized" \cite{Goc&Schaeffer&}, and hence by a theorem about $k$-synchronized sequences, we have $f(n) = O(n)$. But it is not hard to see that in fact $f(n) = g( \lceil \log_2 n \rceil )$, where $g(n) = (n+2)\cdot 2^{n-1} + 1$. So $f(n) \not\in O(n)$, a contradiction. \end{example} \begin{example} {\bf (l) A word that is pure morphic and primitive morphic; but neither uniform morphic nor pure primitive morphic.} The {\it Chacon morphism} is defined by the map $c: 0 \rightarrow 0012, 1 \rightarrow 12, 2 \rightarrow 012$ (see \cite{} and \cite[p.\ 133, \S 5.5.1]{}). Iterating $c$ on $0$ gives the infinite word ${\bf C} := c^\omega(0) = 0012001212012\cdots$. The morphism $c$ is primitive, as $c^2$ applied to each letter contains every letter. Now consider the coding $\ \rightarrow 010$ Applying $\tau$ to $\bf C$ gives the word ${\bf D} = \tau({\bf C}) = 0010001010010 \cdots$, which satisfies ${\bf D} = \delta^\omega(0)$, where $\delta:0 \rightarrow 0010, 1 \rightarrow 1$. It follows that $\bf D$ is pure morphic and primitive morphic. We first show that $\bf D$ is not pure primitive morphic. Let $h$ be {\it any\/} morphism such that $h({\bf D}) = {\bf D}$. If $h(0)=0$, we will show that $h(1)=1$. Suppose $h(1)=u$. If $u = \epsilon$, then $\bf D$ would be periodic, which it is not. Otherwise, if $u$ is nonempty and not equal to $1$, then $u$ begins with $1000$ (since ${\bf D}$ begins $001000\cdots$) and must end with $1$ (since $u$ can be followed by $000$). Now if $u$ ends in $001$, then since $u$ can be followed by $00u$ and hence by $001$ we get a contradiction, since $001001$ is not a factor of $\bf D$. If $u$ ends in $101$, then since $u$ can be followed by $0u$ and hence by $01$ we get a contradiction, since $10101$ is not a factor of $\bf D$. So $u=1$. Now let $h$ be a primitive morphism with $h({\bf D}) = {\bf D}$. If $uu$ is a prefix of $\bf D$ then either $u=0$ or $u=t^n(0)$. This is easy to see by induction on $|u|$. If $|u|>1$, then $u$ begins with $0010$, and since every occurrence of $0010$ in $\bf D$ comes from $t(0)$, it follows that $u =t(u')$ for some prefix $u'$ of $\bf D$. So $\bf D$ begins with $t(u'u')$ and hence begins with $u'u'$. Now since $\bf D$ begins with $h(0)h(0)$ it follows that $h(0)=t^n(0)$ for some $n>0$ (if $h(0)=0$, then $h$ would not be primitive). So $h(0)$ begins with $0010$ and so $h(1) =t(u)$ for some $u$. Note that $u$ must contain $0$ and $1$. We can suppose that $|h(1)|$ is minimal among all primitive morphisms $h$ fixing $\bf D$. So ${\bf D} =t^n(0)t^n(0)t(u)t^n(0)t^n(0)t^n(0)t(u)\cdots$ so ${\bf D}=t^{n-1}(0)t^{n-1}(0)ut^{n-1}(0)t^{n-1}(0)t^{n-1}(0)u \cdots$. By minimality of $|h(1)|$ it follows that the morphism $h': 0 \rightarrow t^{n-1}(0)$ and $1 \rightarrow u$ is not primitive. Since $u$ contains both $0$ and $1$ the only way this can fail to be primitive is if $n=1$, i.e., $h(0)=0$. But now the argument above completes the proof. Finally, we prove that for all $k \geq 2$, the word $\bf D$ is not $k$-automatic. By the argument in \cite[Thm.\ 3.1]{Allouche&Allouche&}, it suffices to prove that ${\bf D} = d(0) \, d(1) \, d(2) \, \cdots $ is not $3$-automatic. Suppose it were. Then, by a well-known result (e.g., \cite[Cor.\ 5.3.3]{Allouche&}) the sequence $(d(x_n))_{n \geq 0}$ is ultimately periodic, where $x_n = [(20)^n]_3$. We will show it is not. Note that $x_n = (3^{2n+1}-3)/4$. An easy induction shows that $|\delta^n (0)| = (3^{n+1}-1)/2$ and hence $|\delta^{2n-1}(0) \delta^{2n-2}(0) \cdots \delta^2(0) \delta(0) 0| = \sum_{0 \leq i \leq 2n-1} |\delta^i(0)| = \sum_{0 \leq i \leq 2n-1} (3^{i+1}-1)/2 = (3^{2n+1}-3)/4 - n = x_n - n$. Now another easy induction gives $$\delta^{2n} (0) = \delta^{2n-1}(0) \delta^{2n-2}(0) \cdots \delta(0)\, 0010 \, 1 \, \delta(0)\, 1 \, \delta^2(0)\, 1 \, \cdots 1 \, \delta^{2n-1} (0).$$ It follows that $d(x_n)$ is the $n$'th symbol of the infinite word ${\bf w} = 0101\delta(0) 1 \delta^2(0) 1 \delta^3(0) \cdots$. However, $\bf w$ is not ultimately periodic. If it were, then its subword complexity $\rho$ would satisfy $\rho(n) \leq n$ for some $n$; however $\bf w$ contains every prefix of $\bf D$ as a factor, and the subword complexity of $\bf D$ is well-known to be $2n+1$, a contradiction \cite{}. \end{example} \begin{example} {\bf (m) A word that is pure primitive morphic; but not uniform morphic.} The Fibonacci word $\bf f$ discussed above in Example~\ref{fib} is generated by the morphism $0 \rightarrow 01;\ 1 \rightarrow 0$ and hence is pure primitive morphic. But it is not automatic, as already mentioned above in Example~\ref{fib}, and hence not uniform morphic. \end{example} \begin{example} {\bf (n) A word that is pure morphic and uniform morphic, but neither pure uniform morphic, nor primitive morphic, nor recurrent.} Consider the morphism $h$ defined by \begin{eqnarray*} 3 &\rightarrow& 32 \\ 2 &\rightarrow& 102012 \\ 1 &\rightarrow& 1012 \\ 0 &\rightarrow& 02 \end{eqnarray*} An easy induction shows that $h^{n+1} (3) = 3 g^{2n} (2) g^{2n-2}(2) \cdots g^2 (2) 2$, where $g$ is the morphism defined in Example~\ref{example15}. Letting $n \rightarrow \infty$, we see that $h^\omega(3) = 3 g^\omega(2)$. Define ${\bf w} = h^\omega(3)$; then we claim $\bf w$ has the desired properties. It is clearly pure morphic, and it is $2$-automatic because $g^\omega(2)$ is (as remarked above in Example~\ref{example15}), and automatic sequences are closed under shift (see, e.g., \cite[Theorem 6.8.4]{Allouche&}). In fact, $\bf w$ is the image under the coding $\rho$ of the fixed point, starting with $a$, of $\delta$, where $\delta(a) = ab$, $\delta(b) = cd$, $\delta(c) = bd$, $\delta(d) = eb$, and $\delta(e) = db$, and $\rho(abcde) = 32101$. However, $\bf w$ is not primitive morphic because if it were, it would be uniformly recurrent. But $3$ only appears once, so it is not even recurrent, a contradiction. Finally, $\bf w$ is not pure uniform morphic. Suppose it is generated by a $k$-uniform morphism $f$. If $k$ is multiplicatively independent of $2$, then $\bf w$ is both $2$-automatic and $k$-automatic, and so by Cobham's theorem \cite{} it is ultimately periodic, a contradiction. Therefore $k$ is multiplicatively dependent on $2$, and hence $k = 2^n$ for some $n \geq 1$. But now ${\bf w}[2] = 1$ and ${\bf w}[6] = 1$. If $\bf w$ were the fixed point of the $2^n$-uniform morphism $f$ we would have the image of ${\bf w}[2]$ under $f$, which is ${\bf w}[2\cdot 2^n..3 \cdot 2^n-1]$, equal to the image of ${\bf w}[6]$ under $f$, which is ${\bf w}[6 \cdot 2^n..7\cdot 2^n - 1]$. However, from our description above we have ${\bf w} = \rho(\delta^\omega(a))$. Since $\delta^\omega(a)$ begins with $abcdbdeb$, it follows that ${\bf w}[2\cdot 2^n..3 \cdot 2^n-1] = \rho(\delta^n(c))$ and ${\bf w}[6 \cdot 2^n..7\cdot 2^n - 1] = \rho(\delta^n(e))$. However $\rho(\delta^n(c))$ begins with $20$ if $n$ is odd and $10$ if $n$ is even, whereas $\rho(\delta^n(e))$ begins with $02$ if $n$ is odd and $12$ if $n$ is even, a contradiction. \label{ex20} \end{example} \begin{example} {\bf (o) A word that is pure morphic and uniform morphic and recurrent, but neither pure uniform morphic nor primitive morphic.} Consider the fixed point of $f$: $a \rightarrow abcda$, $b \rightarrow bcdee$, and $c,d,e \rightarrow eeeee$, followed by the coding by $g:abcde \rightarrow 01123$. The resulting word $\bf q$ starts $$ 01120112333333333333011201123^{62}0 \cdots.$$ It is the fixed point of $h$ defined by $0 \rightarrow 01120$, $1 \rightarrow 1$, $2 \rightarrow 2333333333333$, $3 \rightarrow 33333$. (observe that $hgf = gf^2$). If $\bf q$ were fixed by a non-trivial uniform morphism $j$, then by Cobham's theorem \cite{} the length of $j$ would be $5^k$ for some $k>0$. Then we would have $jg = gf^k$. But $(gf^k)(b)$ starts with $1$, while $(gf^k)(c)$ starts with $3$, a contradiction. The word $\bf q$ is not uniformly recurrent, because it has arbitrarily long blocks of $3$'s. \end{example} \begin{example} {\bf (p) A word that is pure morphic and uniform primitive morphic; but neither pure uniform morphic nor pure primitive morphic.} Let ${\bf u} = acbcbcacbcacacbc \cdots$ be the image of the Thue-Morse word $\bf t$ under the morphism $0 \rightarrow ac$ and $1\rightarrow bc$. Then $\bf u$ is pure morphic, because it is generated by the morphism $a \rightarrow acb$, $b \rightarrow bca$, $c \rightarrow c$. Now $\bf u$ is uniform primitive morphic because it is the image, under the coding $0123 \rightarrow acbc$ of the fixed point of the word generated by the morphism $\eta$ defined by $0 \rightarrow 01$, $1 \rightarrow 23$, $2 \rightarrow 23$, $3 \rightarrow 01$. Note that $\eta^2$ applied to each letter contains every letter. However, $\bf u$ is neither pure uniform morphic, nor pure primitive morphic. To see this, assume ${\bf u} = f^\omega (a)$ for some morphism $f$. If $f$ is primitive or $k$-uniform, with $k \geq 2$, then $f(c)$ is neither $c$ nor $\epsilon$. Since $\bf u$ contains no occurrence of the factor $cc$, it must be that $f(c)$ contains an occurrence of $a$ or $b$. Now $\bf u$ can be factored over $\{f(acbc), f(bcac) \}$, which are words of the same length that have this $a$ or $b$ occurring at the same position. This implies that there is an arithmetic sequence of indices on which $\bf u$ is constantly $a$ or constantly $b$, so the Thue-Morse word $\bf t$ has the same property, which is not true: any sequence extracted from Thue-Morse by indexing from an arithmetic progression contains both $a$'s and $b$'s \cite{}. \end{example} \begin{example} {\bf (q) A word that is pure primitive morphic and uniform primitive morphic; but not pure uniform morphic.} Let $\bf T$ be the word generated by the morphism $g$ mentioned in Example~\ref{example15}: $g$ maps $2 \rightarrow 210; \ 1 \rightarrow 20;\ 0 \rightarrow 1$. Then, as in Example~\ref{example15}, the word $\bf T$ is $2$-automatic, and the underlying morphism is primitive, so it is uniform primitive morphic. Suppose $\bf T$ is pure uniform morphic. Then it is generated by iterating a $s$-uniform morphism for some $s \geq 2$. If $s$ is not a power of $2$, then $\bf T$ is both $2$-automatic and $s$-automatic where $2$ and $s$ are multiplicatively independent. Hence by Cobham's theorem \cite{}, $T$ is ultimately periodic. But in fact $T$ is a well-known squarefree word arising from the Thue-Morse sequence \cite{}. So $\bf T$ must be generated by iterating a morphism $h$ that is $2^k$-uniform for some $k \geq 1$. In this case, Berstel has shown that this is impossible \cite{}, because then $\bf T$ and $g({\bf T})$ differ at the position $5 \cdot 2^k$. \end{example} \begin{example} {\bf (r) A word that is pure uniform morphic; but neither primitive morphic nor recurrent.} The word generated by the morphism $a \rightarrow ab;\ b \rightarrow bc; \ c \rightarrow cc$ iterated on $a$. This is clearly pure uniform morphic. However, since $a$ only appears once, it is not recurrent and thus cannot be primitive morphic. \end{example} \begin{example} {\bf (s) A word that is pure uniform morphic and recurrent; but not primitive morphic.} Here we can take the word that is the fixed point of the morphism $0 \rightarrow 010$ and $1 \rightarrow 111$, as in \cite{}. This is evidently pure uniform morphic and recurrent, but as there are arbitrarily long blocks of $1$'s, it cannot be uniformly recurrent, and hence it is not primitive morphic. \end{example} \begin{example} {\bf (t) A word that is pure uniform primitive morphic.} The Thue-Morse word $\bf t$. \end{example} \section{Final remarks} None of the 20 examples we provided are ultimately periodic. One might ask whether every morphic word can be generated by a coding applied to a non-uniform morphism. The answer is yes: it suffices to prove this for uniform morphic words, which is done in \cite{Allouche&}. We thank Dirk Nowotka for his helpful comments. \newcommand{\noopsort}[1]{} \newcommand{\singleletter}[1]{#1}

\section{Introduction and main result} For fixed integers $M_1,M_2$, the divisor function with congruence conditions is defined by \begin{equation*} d(n;\ell_1,M_1,\ell_2,M_2):=\#\big\{(n_1,n_2)\in\mathbb{N}^2:n_1n_2=n,\,n_i\equiv\ell_i(\bmod M_i),\,i=1,2\big\}, \end{equation*} where $1\leqslant\ell_1\leqslant M_1$ and $1\leqslant\ell_2\leqslant M_2$. Its generating function is \begin{equation*} \sum_{n=1}^\infty\frac{d(n;\ell_1,M_1,\ell_2,M_2)}{n^s}=\zeta(s,\lambda_1)\zeta(s,\lambda_2)(M_1M_2)^{-s}, \end{equation*} where $\Re s>1,\lambda_i=\frac{\ell_i}{M_i},i=1,2$, and $\zeta(s,\lambda)$ denotes the Hurwitz zeta--function, which is defined by \begin{equation*} \zeta(s,\alpha):=\sum_{n=0}^{\infty}\frac{1}{(n+\alpha)^s},\qquad s=\sigma+it,\quad \sigma>1, \end{equation*} for $0<\alpha\leqslant1$. The ``divisor problem with congruence conditions'' is to study the error term \begin{align*} \Delta(x;\ell_1,M_1,\ell_2,M_2):= & \sum_{n\leqslant x}d(n;\ell_1,M_1,\ell_2,M_2) \nonumber \\ & -\sum_{s_0=0,1}Res_{s=s_0}\bigg(\frac{\zeta(s,\lambda_1)\zeta(s,\lambda_2)}{s}\bigg(\frac{x}{M_1M_2}\bigg)^s\bigg). \end{align*} It is conjectured that, for every $\varepsilon>0$, there holds \begin{equation*} \Delta(M_1M_2x;\ell_1,M_1,\ell_2,M_2)\ll x^{1/4+\varepsilon} \end{equation*} uniformly in $1\leqslant\ell_1\leqslant M_1$ and $1\leqslant\ell_2\leqslant M_2$. The best result up to date is \begin{equation*} \Delta(M_1M_2x;\ell_1,M_1,\ell_2,M_2)\ll x^{131/416}(\log x)^{26947/8320}, \end{equation*} uniformly again in $1\leqslant\ell_1\leqslant M_1$ and $1\leqslant\ell_2\leqslant M_2$, which follows from Richert's work \cite{Richert} and Huxley's estimates \cite{Huxley}. In the other direction, Nowak proved in \cite{Nowak-1989} that \begin{equation*} \Delta(Mx;\ell,M,1,1)\ll \Omega_*\Big((x\log x)^{1/4}(\log\log\log x)^{-1/4}\Big) \end{equation*} with $*=-$ if $\lambda_1^*<\frac{\ell}{M}<\lambda_2^*$, and $*=+$ else, where $\lambda_1^*=0.03728310\cdots,\lambda_2^*=0.67181895\cdots$ are the zeros of \begin{equation*} -\log(2\sin (\pi\lambda))+\pi\lambda-\frac{\pi}{2}=0 \end{equation*} in the unit interval. This result holds uniformly in $1\leqslant\ell \leqslant M\leqslant x$, as long as $\frac{\ell}{M}$ is bounded away from $\lambda_1^*,\lambda_2^*$. Another result follows from Theorem 2 of Nowak \cite{Nowak-1990}, i.e. \begin{equation*} \Delta(x;\ell,M,1,1)=\Omega_\pm\Big((x\log x)^{1/4}(\log\log\log x)^{-1/2}\Big) \end{equation*} as $x\to\infty$, where $\ell,M$ are fixed integers satisfying \begin{equation*} 0<\frac{\ell}{M}<\frac{1}{6} \qquad \textrm{or} \qquad \frac{1}{2}<\frac{\ell}{M}<\frac{5}{6}. \end{equation*} In 1995, M\"{u}ller and Nowak \cite{Muller-Nowak} studied the mean value of $\Delta(M_1M_2x;\ell_1,M_1,\ell_2,M_2)$. They proved that, for a large real parameter $T$, there hold \begin{equation*} \int_1^T\Delta(M_1M_2x;\ell_1,M_1,\ell_2,M_2)\mathrm{d}x\ll T^{3/4} \end{equation*} and \begin{equation*} \int_1^T\Delta^2(M_1M_2x;\ell_1,M_1,\ell_2,M_2)\mathrm{d}x=\mathfrak{S}_2\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2}\bigg)T^{3/2}+o(T^{3/2}) \end{equation*} (even with a good error term), uniformly in $1\leqslant\ell_1\leqslant M_1,1\leqslant\ell_2\leqslant M_2$. For the third power moment, they proved that, for any $\varepsilon>0$, there holds \begin{equation*} \int_1^T\Delta^3(M_1M_2x;\ell_1,M_1,\ell_2,M_2)\mathrm{d}x=\mathfrak{S}_3\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2}\bigg)T^{7/4}+o(T^{47/28+\varepsilon}) \end{equation*} uniformly in $1\leqslant\ell_1\leqslant M_1,1\leqslant\ell_2\leqslant M_2$. In 2011, Liu \cite{Liu-Kui} consider the higher--power moments of $\Delta(M_1M_2x;\ell_1,M_1,\ell_2,M_2)$. In his paper, Liu refer to a unified approach, which is originated from Zhai \cite{Zhai-2}, to show the asymptotic formula \begin{equation}\label{unified-Delta-asymptotic} \int_1^T\Delta^k(M_1M_2x;\ell_1,M_1,\ell_2,M_2)\mathrm{d}x=\mathcal{C}_k T^{1+k/4}+O(T^{1+k/4-\vartheta_k+\varepsilon}). \end{equation} holds for $3\leqslant k\leqslant9$, where $\mathcal{C}_k$ and $0<\vartheta_k<1$ are explicit constants. For the case $k=4$, Liu \cite{Liu-Kui} also gives a separate conclusion of the value $\vartheta_4=3/28$, which is better than that in the unified asymptotic formula (\ref{unified-Delta-asymptotic}). The aim of this paper is to improve the value of $\vartheta_4=3/28$, which is achieved by Liu \cite{Liu-Kui}. The main result is the following theorem. \begin{theorem}\label{theorem-1} We have \begin{equation*} \int_1^T\Delta^4(M_1M_2x;\ell_1,M_1,\ell_2,M_2)\mathrm{d}x=\frac{1}{32\pi^4}C_4\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2}\bigg) T^2+O(T^{2-\vartheta_4+\varepsilon}), \end{equation*} with $\vartheta_4=1/8$, where $C_4(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2})$ is defined by (\ref{C_4(l_1/M_1,l_2/M_2)-def}). \end{theorem} We will use the method, which is developed by Zhai \cite{Zhai-3}, to establish Theorem \ref{theorem-1}, but with Kong's Lemma (See Lemma \ref{kong-lamma}) instead of Lemma 2 in Zhai \cite{Zhai-3} by more careful analysis. This method can be applied to a series of fourth--power moment estimates , and obtained better estimates. \section{Notation} Throughout this paper, $\|x\|$ denotes the distance from $x$ to the nearest integer, i.e. $\|x\|=\min_{n\in\mathbb{Z}}|x-n|$. $[x]$ denotes the integer part of $x$; $\psi(t)=t-[t]-\frac{1}{2}$; $e(t)=e^{2\pi it}$. $n\sim N$ means $N<n\leqslant2N$; $n\asymp N$ means $C_1N\leqslant n\leqslant C_2N$ with positive constants $C_1,C_2$ satisfying $C_1<C_2$. $f(x)=\Omega_+(g(x))$ means that there exists a suitable constant $C>0$ such that $f(x)>Cg(x)$ holds for a sequence $x=x_n$ such that $\lim_{n\to\infty}x_n=\infty$; $f(x)=\Omega_-(g(x))$ means that there exists a suitable constant $C>0$ such that $f(x)<-Cg(x)$ holds for a sequence $x=x_n$ such that $\lim_{n\to\infty}x_n=\infty$; $f(x)=\Omega(g(x))$ means that $|f(x)|=\Omega_\pm(g(x))$. $\varepsilon$ always denotes an arbitrary small positive constant, which may not be the same at different occurrences. $d(n)$ denotes the Dirichlet divisor function, and we shall use the well--know estimate $d(n)\ll n^\varepsilon$. Suppose $f:\mathbb{N}\to\mathbb{R}$ is any arithmetic function satisfying $f(n)\ll n^\varepsilon$. Define \begin{equation*} \sideset{}{'}\sum_{\alpha\leqslant n\leqslant\beta}f(n)=\left\{ \begin{array}{ll} \displaystyle\sum_{\alpha<n<\beta}f(n), & \textrm{if $\alpha\not\in\mathbb{Z},\beta\not\in\mathbb{Z}$}, \\ \displaystyle\frac{1}{2}f(\alpha)+\displaystyle\sum_{\alpha<n<\beta}f(n), & \textrm{if $\alpha\in\mathbb{Z},\beta\not\in\mathbb{Z}$}, \\ \displaystyle\displaystyle\sum_{\alpha<n<\beta}f(n)+\frac{1}{2}f(\beta), & \textrm{if $\alpha\not\in\mathbb{Z},\beta\in\mathbb{Z}$}, \\ \displaystyle\frac{1}{2}f(\alpha)+\displaystyle\sum_{\alpha<n<\beta}f(n)+\frac{1}{2}f(\beta), & \textrm{if $\alpha\in\mathbb{Z},\beta\in\mathbb{Z}$}, \end{array} \right. \end{equation*} and, for $k\geqslant2$, define \begin{equation}\label{s_{k;v}-def} s_{k;v}:=s_{k;v}(f)=\sum_{\substack{n_1,\cdots,n_v,n_{v+1},\cdots,n_k\in\mathbb{N}^*\\ \sqrt{n_1}+\cdots+\sqrt{n_v}=\sqrt{n_{v+1}}+\cdots+\sqrt{n_k}}} \frac{f(n_1)f(n_2)\cdots f(n_k)}{(n_1n_2\cdots n_k)^{3/4}}, \quad 1\leqslant v<k. \end{equation} We shall use $s_{k;v}$ or $s_{k;v}(f)$ to denote both of the series (\ref{s_{k;v}-def}) and its value. Suppose $y>1$ is large parameter, and we define \begin{equation*} s_{k;v}(y):=s_{k;v}(f;y)=\sum_{\substack{n_1,\cdots,n_v,n_{v+1},\cdots,n_k\leqslant y\\ \sqrt{n_1}+\cdots+\sqrt{n_v}=\sqrt{n_{v+1}}+\cdots+\sqrt{n_k}}} \frac{f(n_1)f(n_2)\cdots f(n_k)}{(n_1n_2\cdots n_k)^{3/4}}, \quad 1\leqslant v<k. \end{equation*} \section{Preliminary Lemmas} \begin{lemma}\label{Heath-Brown-psi(u)} Let $H\geqslant2$ be any real number. Then we have \begin{equation*} \psi(u)=-\sum_{1\leqslant|h|\leqslant H}\frac{e(hu)}{2\pi ih}+O\bigg(\min\bigg(1,\frac{1}{H\|u\|}\bigg)\bigg). \end{equation*} \end{lemma} \begin{proof} See pp. 245 of Heath--Brown \cite{Heath-Brown-1}. \end{proof} \begin{lemma}\label{Min-lemma} Suppose $A_1,\cdots,A_5$ are absolute positive constants. Let $f(x)$ and $g(x)$ are algebraic functions on $[a,b]$, which satisfy \begin{equation*} \frac{A_1}{R}\leqslant|f''(x)|\leqslant\frac{A_2}{R}, \quad |f'''(x)|\leqslant\frac{A_3}{RU},\quad U\geqslant1, \end{equation*} \begin{equation*} |g(x)|\leqslant A_4G, \quad |g'(x)|\leqslant\frac{A_5G}{U_1},\quad U_1\geqslant1. \end{equation*} Suppose $[\alpha,\beta]$ is the image of $[a,b]$ under the mapping $y=f'(x)$, then \begin{align*} \sum_{a<n\leqslant b}g(n)e\big(f(n)\big) =& e^{\frac{\pi i}{4}}\sum_{\alpha\leqslant \nu\leqslant \beta}b_\nu\frac{g(n_\nu)}{\sqrt{f''(n_\nu)}}e\big(f(n_\nu)-\nu n_\nu\big) \\ & +O\big(G\log(\beta-\alpha+2)+G(b-a+R)(U^{-1}+U_1^{-1})\big) \\ & +O\bigg(G\min\bigg(\sqrt{R},\,\,\max\bigg(\frac{1}{\langle\alpha\rangle},\,\frac{1}{\langle\beta\rangle}\bigg)\bigg)\bigg), \end{align*} where $n\nu$ is the solution of $f'(n)=\nu$, \begin{align*} \langle t\rangle= &\left\{ \begin{array}{cl} \displaystyle\|t\|, & \textrm{if $t\not\in\mathbb{Z}$}, \\ \displaystyle\beta-\alpha, & \textrm{if $t\in\mathbb{Z}$}, \end{array} \right. \\ b_\nu=& \left\{ \begin{array}{cl} \displaystyle 1, & \textrm{if $\alpha<\nu<\beta$ or $\alpha\not\in\mathbb{Z},\beta\not\in\mathbb{Z}$}, \\ \displaystyle\frac{1}{2}, & \textrm{if $u=\alpha\in\mathbb{Z}$ or $u=\beta\in\mathbb{Z}$ }, \end{array} \right. \\ \sqrt{f''}=& \left\{ \begin{array}{cl} \sqrt{f''}, & \textrm{if $f''>0$}, \\ i\sqrt{|f''|}, & \textrm{if $f''<0$ }. \end{array} \right. \end{align*} \end{lemma} \begin{proof} See Theorem 2.2 in Chapter 2 in Min \cite{Min-si-he}, also see Theorem 1 in Chapter III in Karatsuba and Voronin \cite{Karatsuba-Voronin}. \end{proof} \begin{lemma}\label{Tsang-lemma-1} If $g(x)$ and $h(x)$ are continuous real--valued functions of $x$ and $g(x)$ is monotonic, then \begin{equation*} \int_a^bg(x)h(x)\mathrm{d}x\ll \bigg(\max_{a\leqslant x\leqslant b}|g(x)|\bigg)\bigg (\max_{a\leqslant u<v\leqslant b}\bigg|\int_u^vh(x)\mathrm{d}x\bigg|\bigg). \end{equation*} \end{lemma} \begin{proof} See Lemma 1 of Tsang \cite{Tsang}. \end{proof} \begin{lemma}\label{yijie} Suppose $A,B\in\mathbb{R},\,A\not=0$. Then for any $\alpha\in\mathbb{R}$, there holds \begin{equation*} \int_T^{2T} t^\alpha \cos\big(A\sqrt{t}+B\big)\mathrm{d}t\ll T^{1/2+\alpha}|A|^{-1}. \end{equation*} \end{lemma} \begin{proof} It follows from Lemma \ref{Tsang-lemma-1} easily. \end{proof} \begin{lemma}\label{s_k;v(f)=s_k;v(f;y)+error} Let f(n) be an arithmetic function. Then we have \begin{equation*} |s_{k;v}(f)-s_{k;v}(f;y)|\ll y^{-1/2+\varepsilon}, \qquad 1\leqslant v<k. \end{equation*} \end{lemma} \begin{proof} See Lemma 3.1 of Zhai \cite{Zhai-2}. \end{proof} \begin{lemma}\label{kong-lamma} If $n,m,k,\ell\in\mathbb{N}$ such that $\sqrt{n}+\sqrt{m}\pm\sqrt{k}-\sqrt{\ell}\not=0$, then there hold \begin{equation*} |\sqrt{n}+\sqrt{m}\pm\sqrt{k}-\sqrt{\ell}|\gg(nmk\ell)^{-1/2}\max(n,m,k,\ell)^{-3/2}, \end{equation*} respectively. \end{lemma} \begin{proof} See Kong~\cite{Kong}, Lemma~3.2.1. \end{proof} \begin{lemma}\label{Zhai-lemma-5} Suppose $1\leqslant N\leqslant M,\,1\leqslant K\leqslant L,\,N\leqslant K,\,M\asymp L,\,0<\Delta\ll L^{1/2}$. Let $\mathscr{A}_1(N,M,K,L;\Delta)$ denote the number of solutions of the following inequality \begin{equation*} 0<|\sqrt{n}+\sqrt{m}-\sqrt{k}-\sqrt{\ell}|<\Delta \end{equation*} with $n\sim N,\,m\sim M,\,k\sim K,\,\ell\sim L$. Then we have \begin{equation*} \mathscr{A}_1(N,M,K,L;\Delta)\ll \Delta L^{1/2}NMK+NKL^{1/2+\varepsilon}. \end{equation*} Especially, if $\Delta L^{1/2}\gg1$, then \begin{equation*} \mathscr{A}_1(N,M,K,L;\Delta)\ll \Delta L^{1/2}NMK. \end{equation*} \end{lemma} \begin{proof} See Zhai~\cite{Zhai-3}, Lemma 5. \end{proof} \begin{lemma}\label{Zhai-lemma-6} Suppose $1\leqslant N\leqslant M\leqslant K\asymp L,\,0<\Delta\ll L^{1/2}$, Let $\mathscr{A}_2(N,M,K,L;\Delta)$ denote the number of solutions of the inequality \begin{equation*} 0<|\sqrt{n}+\sqrt{m}+\sqrt{k}-\sqrt{\ell}|<\Delta \end{equation*} with $n\sim N,\,m\sim M,\,k\sim K,\,\ell\sim L$. Then we have \begin{equation*} \mathscr{A}_2(N,M,K,L;\Delta)\ll \Delta L^{1/2}NMK+NML^{1/2+\varepsilon}. \end{equation*} Especially, if $\Delta L^{1/2}\gg1$, then \begin{equation*} \mathscr{A}_1(N,M,K,L;\Delta)\ll \Delta L^{1/2}NMK. \end{equation*} \end{lemma} \begin{proof} See Zhai \cite{Zhai-3}, Lemma 6. \end{proof} \begin{lemma}\label{Zhai-lemma-3} Suppose $N_j\geqslant2\,(j=1,2,3,4),\,\Delta>0$. Let $\mathscr{A}_\pm(N_1,N_2,N_3,N_4;\Delta)$ denote the number of solutions of the following inequality \begin{equation*} 0<|\sqrt{n_1}+\sqrt{n_2}\pm\sqrt{n_3}-\sqrt{n_4}|<\Delta \end{equation*} with $n_j\sim N_j\,(j=1,2,3,4),\,n_j\in\mathbb{N}^*$. Then we have \begin{equation*} \mathscr{A}_\pm(N_1,N_2,N_3,N_4;\Delta)\ll\prod_{j=1}^4\big(\Delta^{1/4}N_j^{7/8}+N_j^{1/2}\big)N_j^{\varepsilon}. \end{equation*} \end{lemma} \begin{proof} See Zhai~\cite{Zhai-3}, Lemma 3. \end{proof} \section{Analogue of Vorono\"{i}'s Formula} For the Dirichlet divisor problem, there exists the following truncated Vorono\"{i}'s formula, i.e. \begin{equation*} \Delta(x)=\frac{x^{1/4}}{\sqrt{2}\pi}\sum_{n\leqslant N}\frac{d(n)}{n^{3/4}}\cos\big(4\pi\sqrt{nx}-\pi/4\big)+O\big(x^{1/2+\varepsilon}N^{-1/2}\big), \end{equation*} for $1\leqslant N\ll x$. This formula plays a significant role in the study of the higher--power moments of $\Delta(x)$. But for $\Delta(M_1M_2x;\ell_1,M_1,\ell_2,M_2)$, there is not such a convenient formula at hand. However, we can use the finite expression of $\psi(x)$ and van der Corput's B--process to establish an analogue of Vorono\"{i}'s formula. Suppose $T/2\leqslant x\leqslant T$. According to Dirichlet's hyperbolic summation method, we write \begin{equation*} \sum_{n\leqslant M_1M_2x}d(n;\ell_1,M_1,\ell_2,M_2)=\sum_{\substack{n_1n_2\leqslant M_1M_2x\\ n_1\equiv \ell_1(\!\bmod M_1)\\n_2\equiv\ell_2(\!\bmod M_2)}}1 =\Sigma_1+\Sigma_2-\Sigma_3, \end{equation*} where \begin{equation*} \Sigma_1=\sum_{\substack{n_1\leqslant\sqrt{M_1M_2x}\\ n_1\equiv \ell_1(\!\bmod M_1)}} \sum_{\substack{n_2\leqslant\frac{M_1M_2x}{n_1}\\n_2\equiv \ell_2(\!\bmod M_2) }}1, \qquad \Sigma_2=\sum_{\substack{n_2\leqslant\sqrt{M_1M_2x}\\ n_2\equiv \ell_2(\!\bmod M_2)}} \sum_{\substack{n_1\leqslant\frac{M_1M_2x}{n_2}\\n_1\equiv \ell_1(\!\bmod M_1) }}1, \end{equation*} \begin{equation*} \Sigma_3=\sum_{\substack{n_1\leqslant\sqrt{M_1M_2x}\\ n_1\equiv \ell_1(\!\bmod M_1)}} 1 \sum_{\substack{n_2\leqslant\sqrt{M_1M_2x}\\ n_2\equiv \ell_2(\!\bmod M_2)}} 1. \end{equation*} Computing the above three sums directly, it is easy to derive that \begin{equation}\label{Delta=F_12+F_21} \Delta(M_1M_2x;\ell_1,M_1,\ell_2,M_2)=F_{12}(x)+F_{21}(x)+O(1), \end{equation} where \begin{equation*} F_{12}(x):=-\sum_{\substack{n_1\leqslant\sqrt{M_1M_2x}\\ n_1\equiv\ell_1(\!\bmod M_1)}}\psi\bigg(\frac{M_1x}{n_1}-\frac{\ell_2}{M_2}\bigg), \end{equation*} \begin{equation*} F_{21}(x):=-\sum_{\substack{n_1\leqslant\sqrt{M_1M_2x}\\ n_2\equiv\ell_2(\!\bmod M_2)}}\psi\bigg(\frac{M_2x}{n_2}-\frac{\ell_1}{M_1}\bigg). \end{equation*} First, we consider $F_{12}(x)$. Suppose that $H\geqslant2$ is a parameter which is to be determined later. By Lemma \ref{Heath-Brown-psi(u)}, we get \begin{equation}\label{F_12=R_12+G_12} F_{12}(x)=R_{12}(x;H)+G_{12}(x;H), \end{equation} where \begin{equation*} R_{12}(x;H):=\frac{1}{2\pi i}\sum_{1\leqslant |h|\leqslant H}\frac{1}{h}\sum_{\substack{n_1\leqslant\sqrt{M_1M_2x}\\ n_1\equiv\ell_1(\!\bmod M_1)}} e\bigg(\frac{hM_1x}{n_1}-\frac{h\ell_2}{M_2}\bigg), \end{equation*} \begin{equation}\label{G_12(x;H)-def} G_{12}(x;H):=O\left(\sum_{\substack{n_1\leqslant\sqrt{M_1M_2x}\\ n_1\equiv\ell_1(\!\bmod M_1)}} \min\Bigg(1,\frac{1}{H\big\|\frac{M_1x}{n_1}-\frac{\ell_2}{M_2}\big\|}\Bigg)\right). \end{equation} By a splitting argument, we have \begin{equation*} R_{12}(x;H)=\frac{1}{2\pi i}\sum_{1\leqslant |h|\leqslant H}\frac{1}{h}\sum_{j=0}^J \sum_{\substack{\sqrt{\frac{M_1M_2x}{2^{j+1}}}<n_1\leqslant\sqrt{\frac{M_1M_2x}{2^j}} \\ n_1\equiv\ell_1(\!\bmod M_1)}} e\bigg(\frac{hM_1x}{n_1}-\frac{h\ell_2}{M_2}\bigg)+O(\mathscr{L}^2), \end{equation*} where \begin{equation*} J=\bigg[\frac{\mathscr{L}-2\log\mathscr{L}}{\log2}\bigg]. \end{equation*} It is easy to see that \begin{equation}\label{R_12-Sigma_12} R_{12}(x;H)=-\frac{\Sigma_{12}}{2\pi i}+\frac{\overline{\Sigma_{12}}}{2\pi i}+O(\mathscr{L}^2), \end{equation} where \begin{equation}\label{Sigma_{12}} \Sigma_{12}=\sum_{1\leqslant h\leqslant H}\frac{1}{h}\sum_{j=0}^J\mathcal{S}(x,h,j), \end{equation} \begin{equation}\label{S(x,h,j)-def} \mathcal{S}(x,h,j):= \sum_{\substack{\sqrt{\frac{M_1M_2x}{2^{j+1}}}<n_1\leqslant\sqrt{\frac{M_1M_2x}{2^j}} \\ n_1\equiv\ell_1(\!\bmod M_1)}} e\bigg(-\frac{hM_1x}{n_1}+\frac{h\ell_2}{M_2}\bigg). \end{equation} Applying Lemma \ref{Min-lemma} to (\ref{S(x,h,j)-def}), we obtain \begin{align}\label{S(x,h,j)-transfer} \mathcal{S}(x,h,j) = & \sum_{\substack{k\\ \sqrt{\frac{M_1M_2x}{2^{j+1}}}<kM_1+\ell_1\leqslant\sqrt{\frac{M_1M_2x}{2^j}} }} e\bigg(-\frac{hM_1x}{kM_1+\ell_1}+\frac{h\ell_2}{M_2}\bigg) \nonumber \\ = & \frac{x^{1/4}}{\sqrt{2}}\sideset{}{'}\sum_{\frac{2^jhM_1}{M_2}\leqslant r\leqslant \frac{2^{j+1}hM_1}{M_2}}\frac{h^{1/4}}{r^{3/4}} e\bigg(-2\sqrt{hrx}+\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}-\frac{1}{8}\bigg)+O(\mathscr{L}). \end{align} Putting (\ref{S(x,h,j)-transfer}) into (\ref{Sigma_{12}}), we get \begin{align}\label{Sigma_(12)-transfer} \Sigma_{12} = & \frac{x^{1/4}}{\sqrt{2}}\sum_{1\leqslant h\leqslant H}\sum_{j=0}^J \sideset{}{'}\sum_{\frac{2^jhM_1}{M_2}\leqslant r\leqslant \frac{2^{j+1}hM_1}{M_2}}\frac{1}{(hr)^{3/4}} e\bigg(-2\sqrt{hrx}+\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}-\frac{1}{8}\bigg)+O(\mathscr{L}^3) \nonumber \\ = & \frac{x^{1/4}}{\sqrt{2}}\sum_{1\leqslant h\leqslant H} \sideset{}{'}\sum_{\frac{hM_1}{M_2}\leqslant r\leqslant\frac{2^{J+1}hM_1}{M_2}}\frac{1}{(hr)^{3/4}} e\bigg(-2\sqrt{hrx}+\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}-\frac{1}{8}\bigg)+O(\mathscr{L}^3). \end{align} Define \begin{equation*} \tau_{12}(n,x;H):=\sideset{}{'}\sum_{\substack{n=hr\\ 1\leqslant h\leqslant H\\ \frac{hM_1}{M_2}\leqslant r\leqslant\frac{2^{J+1}hM_1}{M_2}}} \cos\bigg(4\pi\sqrt{nx}-2\pi\bigg(\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}+\frac{1}{8}\bigg)\bigg) \end{equation*} and \begin{equation*} R_{12}'(x;H):=\frac{x^{1/4}}{\sqrt{2}\pi}\sum_{n\leqslant \frac{2^{J+1}H^2M_1}{M_2}}\frac{\tau_{12}(n,x;H)}{n^{3/4}}. \end{equation*} Inserting (\ref{Sigma_(12)-transfer}) into (\ref{R_12-Sigma_12}), we obtain \begin{align}\label{R_12=R12'+L^3} & R_{12}(x;H) \nonumber \\ = & \frac{x^{1/4}}{2\sqrt{2}\pi i}\sum_{1\leqslant h\leqslant H}\sideset{}{'}\sum_{\frac{hM_1}{M_2}\leqslant r\leqslant\frac{2^{J+1}hM_1}{M_2}} \frac{1}{(hr)^{3/4}} \nonumber \\ & \times \bigg(e\bigg(2\sqrt{hrx}-\frac{h\ell_2}{M_2}-\frac{r\ell_1}{M_1}+\frac{1}{8}\bigg)- e\bigg(-2\sqrt{hrx}+\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}-\frac{1}{8}\bigg)\bigg) +O(\mathscr{L}^3) \nonumber \\ = & \frac{x^{1/4}}{\sqrt{2}\pi }\sum_{1\leqslant h\leqslant H}\sideset{}{'}\sum_{\frac{hM_1}{M_2}\leqslant r\leqslant\frac{2^{J+1}hM_1}{M_2}} \frac{1}{(hr)^{3/4}} \sin\bigg(4\pi\sqrt{hrx}-2\pi\bigg(\frac{h\ell_2}{M_1}+\frac{r\ell_1}{M_1}-\frac{1}{8}\bigg)\bigg)+O(\mathscr{L}^3) \nonumber \\ = & \frac{x^{1/4}}{\sqrt{2}\pi }\sum_{1\leqslant h\leqslant H}\sideset{}{'}\sum_{\frac{hM_1}{M_2}\leqslant r\leqslant\frac{2^{J+1}hM_1}{M_2}} \frac{1}{(hr)^{3/4}} \cos\bigg(4\pi\sqrt{hrx}-2\pi\bigg(\frac{h\ell_2}{M_1}+\frac{r\ell_1}{M_1}+\frac{1}{8}\bigg)\bigg)+O(\mathscr{L}^3) \nonumber \\ = & R'_{12}(x;H)+O(\mathscr{L}^3). \end{align} Define \begin{equation*} \tau_{12}(n,x):=\sideset{}{'}\sum_{\substack{n=hr\\ \frac{hM_1}{M_2}\leqslant r}} \cos\bigg(4\pi\sqrt{nx}-2\pi\bigg(\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}+\frac{1}{8}\bigg)\bigg). \end{equation*} It is easy to check that if $n\leqslant\min(H^2,T)\mathscr{L}^{-3}$, then $h\leqslant H$ and $r\leqslant\frac{2^{J+1}hM_1}{M_2}$. Therefore, we have \begin{equation*} \tau_{12}(n,x;H)=\tau_{12}(n,x),\qquad n\leqslant \min(H^2,T)\mathscr{L}^{-3}. \end{equation*} Suppose $T^\varepsilon<y\leqslant\min(H^2,T)\mathscr{L}^{-3}$ is a parameter which is to be determined later. Define \begin{equation}\label{R_(12)(x;y)-def} R_{12}(x;y):=\frac{x^{1/4}}{\sqrt{2}\pi}\sum_{n\leqslant y}\frac{\tau_{12}(n,x)}{n^{3/4}} \end{equation} and \begin{equation}\label{R_(12)(x;y,H)-def} R^*_{12}(x;y,H):=\frac{x^{1/4}}{\sqrt{2}\pi}\sum_{y<n\leqslant \frac{2^{J+1}H^2M_1}{M_2}}\frac{\tau_{12}(n,x;H)}{n^{3/4}}. \end{equation} Then from (\ref{F_12=R_12+G_12}), (\ref{R_12=R12'+L^3}), (\ref{R_(12)(x;y)-def}) and (\ref{R_(12)(x;y,H)-def}), we get \begin{equation}\label{F_12-fenjie} F_{12}(x)=R_{12}(x;y)+R_{12}^*(x;y,H)+G_{12}(x;H)+O(\mathscr{L}^3). \end{equation} By symmetry, we also have \begin{equation}\label{F_21-fenjie} F_{21}(x)=R_{21}(x;y)+R_{21}^*(x;y,H)+G_{21}(x;H)+O(\mathscr{L}^3), \end{equation} where \begin{equation*} R_{21}(x;y)=\frac{x^{1/4}}{\sqrt{2}\pi}\sum_{n\leqslant y}\frac{\tau_{21}(n,x)}{n^{3/4}}, \end{equation*} \begin{equation*} \tau_{21}(n,x)=\sideset{}{'}\sum_{\substack{n=hr\\ \frac{hM_2}{M_1}\leqslant r}} \cos\bigg(4\pi\sqrt{nx}-2\pi\bigg(\frac{h\ell_1}{M_1}+\frac{r\ell_2}{M_2}+\frac{1}{8}\bigg)\bigg), \end{equation*} \begin{equation*} R^*_{21}(x;y,H)=\frac{x^{1/4}}{\sqrt{2}\pi}\sum_{y<n\leqslant \frac{2^{J+1}H^2M_2}{M_1}}\frac{\tau_{21}(n,x;H)}{n^{3/4}}, \end{equation*} \begin{equation*} \tau_{21}(n,x;H)=\sideset{}{'}\sum_{\substack{n=hr\\ 1\leqslant h\leqslant H\\ \frac{hM_2}{M_1}\leqslant r\leqslant\frac{2^{J+1}hM_2}{M_1}}} \cos\bigg(4\pi\sqrt{nx}-2\pi\bigg(\frac{h\ell_1}{M_1}+\frac{r\ell_2}{M_2}+\frac{1}{8}\bigg)\bigg), \end{equation*} \begin{equation*} G_{21}(x;H)=O\left(\sum_{\substack{n_2\leqslant\sqrt{M_1M_2x}\\ n_2\equiv\ell_2(\!\bmod M_2)}} \min\Bigg(1,\frac{1}{H\big\|\frac{M_2x}{n_2}-\frac{\ell_1}{M_1}\big\|}\Bigg)\right). \end{equation*} Define \begin{equation*} \tau(n,x):=\sum_{n=hr}\cos\bigg(4\pi\sqrt{nx}-2\pi\bigg(\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}+\frac{1}{8}\bigg)\bigg) \end{equation*} and \begin{equation*} R_0(x;y):=\frac{x^{1/4}}{\sqrt{2}\pi}\sum_{n\leqslant y}\frac{\tau(n,x)}{n^{3/4}}. \end{equation*} From the definition of $\tau_{12}(n,x)$ and $\tau_{21}(n,x)$, we derive that \begin{align*} \tau_{12}(n,x)+\tau_{21}(n,x) = & \sideset{}{'}\sum_{\substack{n=hr\\ \frac{hM_1}{M_2}\leqslant r}} \cos\bigg(4\pi\sqrt{nx}-2\pi\bigg(\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}+\frac{1}{8}\bigg)\bigg) \nonumber \\ & +\sideset{}{'}\sum_{\substack{n=hr\\ \frac{hM_2}{M_1}\leqslant r}} \cos\bigg(4\pi\sqrt{nx}-2\pi\bigg(\frac{h\ell_1}{M_1}+\frac{r\ell_2}{M_2}+\frac{1}{8}\bigg)\bigg) \nonumber \\ = & \sideset{}{'}\sum_{\substack{n=hr\\ \frac{hM_1}{M_2}\leqslant r}} \cos\bigg(4\pi\sqrt{nx}-2\pi\bigg(\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}+\frac{1}{8}\bigg)\bigg) \nonumber \\ \end{align*} \begin{align*} & +\sideset{}{'}\sum_{\substack{n=hr\\ r\leqslant\frac{hM_1}{M_2}}} \cos\bigg(4\pi\sqrt{nx}-2\pi\bigg(\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}+\frac{1}{8}\bigg)\bigg) \nonumber \\ = & \sum_{n=hr}\cos\bigg(4\pi\sqrt{nx}-2\pi\bigg(\frac{h\ell_2}{M_2}+\frac{r\ell_1}{M_1}+\frac{1}{8}\bigg)\bigg) =\tau(n,x), \end{align*} which implies \begin{equation}\label{R_0=R_12+R_21} R_0(x;y)=R_{12}(x;y)+R_{21}(x;y). \end{equation} Combining (\ref{Delta=F_12+F_21}), (\ref{F_12-fenjie})--(\ref{R_0=R_12+R_21}), we obtain \begin{align}\label{Delta=R_0+R12+R_21+G_12+G_21} \Delta(M_1M_2x;\ell_1,M_1,\ell_2,M_2)= & R_0(x;y)+R_{12}^*(x;y,H)+R_{21}^*(x;y,H) \nonumber \\ & +G_{12}(x;H)+G_{21}(x;H)+O(\mathscr{L}^3). \end{align} For simplicity, we denote $\mathscr{R}_0=R_0(x;y),\,\mathscr{R}_1=R_{12}^*(x;y,H)$, $\mathscr{R}_2=R_{21}^*(x;y,H)$, $\mathscr{R}=\mathscr{R}_0+\mathscr{R}_1+\mathscr{R}_2$ and $\mathscr{G}=G_{12}(x;H)+G_{21}(x;H)$. Then (\ref{Delta=R_0+R12+R_21+G_12+G_21}) can be written as \begin {equation*} \Delta(M_1M_2x;\ell_1,M_1,\ell_2,M_2)= \mathscr{R}+\mathscr{G}+O(\mathscr{L}^3). \end{equation*} \section{Proof of Theorem \ref{theorem-1}} In this section, we shall prove Theorem \ref{theorem-1}. Take $H=T^8$ and $y=T^{3/4}$. By the elementary formula \begin{equation}\label{elementary-formula } (a+b)^4=a^4+O(|a|^3|b|+b^4), \end{equation} we have \begin {equation}\label{Delta(M_1M_2x)-fenjie} \Delta^4(M_1M_2x;\ell_1,M_1,\ell_2,M_2)= \mathscr{R}^4+O\big(|\mathscr{R}|^3|\mathscr{G}|+|\mathscr{R}|^3\mathscr{L}^3+\mathscr{G}^4+\mathscr{L}^{12}\big). \end{equation} By a splitting argument, it is sufficient to prove the result in the interval $[T/2,T]$. We will divide the process of the proof of Theorem \ref{theorem-1} into two parts. \begin{proposition}\label{proposition-1} For $T\geqslant10,\,y=T^{3/4}$, we have \begin{equation*} \int_{\frac{T}{2}}^T\mathscr{R}^4\mathrm{d}x=\frac{1}{32\pi^4}C_4\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2}\bigg)\int_{\frac{T}{2}}^Tx\mathrm{d}x +O(T^{2-1/8+\varepsilon}). \end{equation*} \end{proposition} \begin{proof} From (\ref{elementary-formula }), we get \begin{equation*} \mathscr{R}^4=\big(\mathscr{R}_0+(\mathscr{R}_1+\mathscr{R}_2)\big)^4=\mathscr{R}_0^4+O\big(|\mathscr{R}_0|^3|\mathscr{R}_1 +\mathscr{R}_2|+|\mathscr{R}_1+\mathscr{R}_2|^4\big) \end{equation*} Therefore, we have \begin{equation}\label{R_4-mean-explicit} \int_{\frac{T}{2}}^T\mathscr{R}^4\mathrm{d}x =\int_{\frac{T}{2}}^T\mathscr{R}_0^4\mathrm{d}x +O\bigg(\int_{\frac{T}{2}}^T|\mathscr{R}_0|^3|\mathscr{R}_1+\mathscr{R}_2|\mathrm{d}x+\int_{\frac{T}{2}}^T|\mathscr{R}_1+\mathscr{R}_2|^4\mathrm{d}x\bigg). \end{equation} According to the elementary formula \begin{equation*} \cos{a_1}\cos{a_2}\cdots\cos{a_k}=\frac{1}{2^{k-1}}\sum_{(i_1,i_2\cdots,i_{k-1})\in\{0,1\}^{k-1}}\cos\big(a_1+(-1)^{i_1}a_2+\cdots+(-1)^{i_{k-1}}a_k\big), \end{equation*} we have \begin{align*} \mathscr{R}_0^4 = &\frac{x}{4\pi^4}\sum_{n_1,n_2,n_3,n_4\leqslant y}\frac{\tau(n_1,x)\tau(n_2,x)\tau(n_3,x)\tau(n_4,x)}{(n_1n_2n_3n_4)^{3/4}} \nonumber \\ = & \frac{x}{4\pi^4}\sum_{n_1,n_2,n_3,n_4\leqslant y}\frac{1}{(n_1n_2n_3n_4)^{3/4}} \nonumber \\ & \qquad\times\sum_{\substack{n_j=h_jr_j\\j=1,2,3,4}}\prod_{j=1}^4 \cos\bigg(4\pi\sqrt{n_jx}-2\pi\bigg(\frac{h_j\ell_2}{M_2}+\frac{r_j\ell_1}{M_1}+\frac{1}{8}\bigg)\bigg) \nonumber \\ = & \frac{x}{32\pi^4}\sum_{(i_1,i_2,i_3)\in\{0,1\}^3}\sum_{n_1,n_2,n_3,n_4\leqslant y}\frac{1}{(n_1n_2n_3n_4)^{3/4}} \nonumber \\ & \qquad\times\sum_{\substack{n_j=h_jr_j\\j=1,2,3,4}} \cos\big(4\pi\boldsymbol{\alpha}(\mathbf{n},\mathbf{i})\sqrt{x}-2\pi\boldsymbol{\beta}(\mathbf{h},\mathbf{r},\mathbf{i})\big), \end{align*} where $\mathbf{n}=(n_1,n_2,n_3,n_4)\in\mathbb{N}^4,\,\mathbf{i}=(i_1,i_2,i_3)\in\{0,1\}^3$ and \begin{align*} \boldsymbol{\alpha}(\mathbf{n},\mathbf{i}) := & \sqrt{n_1}+(-1)^{i_1}\sqrt{n_2}+(-1)^{i_2}\sqrt{n_3}+(-1)^{i_3}\sqrt{n_4}, \\ \boldsymbol{\beta}(\mathbf{h},\mathbf{r},\mathbf{i}) := & \bigg(\frac{h_1\ell_2}{M_2}+\frac{r_1\ell_1}{M_1}+\frac{1}{8}\bigg)+(-1)^{i_1}\bigg(\frac{h_2\ell_2}{M_2}+\frac{r_2\ell_1}{M_1}+\frac{1}{8}\bigg) \\ & +(-1)^{i_2}\bigg(\frac{h_3\ell_2}{M_2}+\frac{r_3\ell_1}{M_1}+\frac{1}{8}\bigg) +(-1)^{i_3}\bigg(\frac{h_4\ell_2}{M_2}+\frac{r_4\ell_1}{M_1}+\frac{1}{8}\bigg). \end{align*} Therefore, we can write \begin{equation}\label{R_0=S_1(x)+S_2(x)} \mathscr{R}_0^4=S_1(x)+S_2(x), \end{equation} where \begin{align*} S_1(x)= & \frac{x}{32\pi^4}\sum_{(i_1,i_2,i_3)\in\{0,1\}^3}\sum_{\substack{n_1,n_2,n_3,n_4\leqslant y\\ \boldsymbol{\alpha}(\mathbf{n},\mathbf{i})=0 }} \frac{1}{(n_1n_2n_3n_4)^{3/4}} \sum_{\substack{n_j=h_jr_j\\j=1,2,3,4}}\cos\big(2\pi\boldsymbol{\beta}(\mathbf{h},\mathbf{r},\mathbf{i})\big), \\ S_2(x)=& \frac{x}{32\pi^4}\sum_{(i_1,i_2,i_3)\in\{0,1\}^3}\sum_{\substack{n_1,n_2,n_3,n_4\leqslant y\\ \boldsymbol{\alpha}(\mathbf{n},\mathbf{i})\not=0 }} \frac{1}{(n_1n_2n_3n_4)^{3/4}} \\ & \qquad\qquad\qquad\qquad\qquad\qquad\times \sum_{\substack{n_j=h_jr_j\\j=1,2,3,4}} \cos\big(4\pi\boldsymbol{\alpha}(\mathbf{n},\mathbf{i})\sqrt{x}-2\pi\boldsymbol{\beta}(\mathbf{h},\mathbf{r},\mathbf{i})\big). \end{align*} For $\mathbf{n}=(n_1,n_2,n_3,n_4)\in\mathbb{N}^4,\,\mathbf{i}=(i_1,i_2,i_3)\in\{0,1\}^3$, define \begin{equation*} \tau(\mathbf{n},\mathbf{i}):=\sum_{\substack{\mathbf{n}\in\mathbb{N}^4\\ \boldsymbol{\alpha}(\mathbf{n},\mathbf{i})=0}} \frac{1}{(n_1n_2n_3n_4)^{3/4}} \sum_{\substack{n_j=h_jr_j\\j=1,2,3,4}}\cos\big(2\pi\boldsymbol{\beta}(\mathbf{h},\mathbf{r},\mathbf{i})\big) \end{equation*} and \begin{equation*} \tau(\mathbf{n},\mathbf{i};y):=\sum_{\substack{n_1,n_2,n_3,n_4\leqslant y\\ \boldsymbol{\alpha}(\mathbf{n},\mathbf{i})=0}} \frac{1}{(n_1n_2n_3n_4)^{3/4}} \sum_{\substack{n_j=h_jr_j\\j=1,2,3,4}}\cos\big(2\pi\boldsymbol{\beta}(\mathbf{h},\mathbf{r},\mathbf{i})\big). \end{equation*} The condition $\boldsymbol{\alpha}(\mathbf{n},\mathbf{i})=0$ implies $\mathbf{i}\not=\mathbf{0}$, i.e. $(i_1,i_2,i_3)\not=(0,0,0)$. It is easy to see that for $\mathbf{i}=(i_1,i_2,i_3)\in\{0,1\}^3$ and $\mathbf{i}'=(i_1',i_2',i_3')\in\{0,1\}^3$ satisfying $i_1+i_2+i_3=i_1'+i_2'+i_3'=v$, there must hold \begin{equation*} \tau(\mathbf{n},\mathbf{i})= \tau(\mathbf{n},\mathbf{i}')=:s_{4;v},\qquad \tau(\mathbf{n},\mathbf{i};y)= \tau(\mathbf{n},\mathbf{i}';y)=:s_{4;v}(y), \end{equation*} say. Also, we write \begin{equation}\label{C_4(l_1/M_1,l_2/M_2)-def} C_4\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2}\bigg)=\sum_{v=1}^3{3\choose v}s_{4;v}. \end{equation} Thus, from Lemma \ref{s_k;v(f)=s_k;v(f;y)+error} we deduce that \begin{align*} B_4\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2};y\bigg):=\sum_{\mathbf{i}\in\{0,1\}^3}\tau(\mathbf{n},\mathbf{i};y) =\sum_{v=1}^3{3\choose v}s_{4;v}(y)=C_4\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2}\bigg)+O(y^{-1/2+\varepsilon}). \end{align*} Hence, we obtain \begin{align}\label{S_1(x)-asymptotic} \int_{\frac{T}{2}}^T S_1(x)\mathrm{d}x = & \frac{1}{32\pi^4}B_4\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2};y\bigg)\int_{\frac{T}{2}}^T x\mathrm{d}x \nonumber \\ = & \frac{1}{32\pi^4} C_4\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2}\bigg) \int_{\frac{T}{2}}^T x\mathrm{d}x + O(T^2y^{-1/2+\varepsilon}) \nonumber \\ = & \frac{1}{32\pi^4} C_4\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2}\bigg) \int_{\frac{T}{2}}^T x\mathrm{d}x + O(T^{2-1/8+\varepsilon}). \end{align} Now, we proceed to consider the contribution of $S_2(x)$. By Lemma \ref{yijie}, we get \begin{equation}\label{S_2(x)-upper-all} \int_{\frac{T}{2}}^TS_2(x)\mathrm{d}x\ll\sum_{\mathbf{i}\in\{0,1\}^3} \sum_{\substack{n_1,n_2,n_3,n_4\leqslant y\\ \boldsymbol{\alpha}(\mathbf{n},\mathbf{i})\not=0}} \frac{d(n_1)d(n_2)d(n_3)d(n_4)}{(n_1n_2n_3n_4)^{3/4}}\min\bigg(T^2,\frac{T^{3/2}}{|\boldsymbol{\alpha}(\mathbf{n},\mathbf{i})|}\bigg). \end{equation} For $\mathbf{i}=(i_1,i_2,i_3)\in\{0,1\}^3$, if $(i_1,i_2,i_3)=(0,0,0)$, then the contribution is \begin{align}\label{S_2(x)-all-positive} \ll & \,\, T^{3/2}\sum_{n_1,n_2,n_3,n_4\leqslant y}\frac{d(n_1)d(n_2)d(n_3)d(n_4)}{(n_1n_2n_3n_4)^{3/4}(\sqrt{n_1}+\sqrt{n_2}+\sqrt{n_3}+\sqrt{n_4})} \nonumber \\ \ll & \,\, T^{3/2+\varepsilon}\sum_{n_1\leqslant n_2\leqslant n_3\leqslant n_4\leqslant y}\frac{1}{(n_1n_2n_3n_4)^{3/4} n_4^{1/2}} \nonumber \\ \ll & \,\, T^{3/2+\varepsilon}\sum_{n_1\leqslant n_2\leqslant n_3\leqslant n_4\leqslant y}\frac{1}{(n_1n_2n_3)^{3/4}n_4^{5/4}} \nonumber \\ \ll & \,\, T^{3/2+\varepsilon}y^{1/2}\ll T^{2-1/8+\varepsilon}. \end{align} If $(i_1,i_2,i_3)\not=(0,0,0)$, we write \begin{equation}\label{not-all>0-Sigma1+Sigma2} \sum_{\mathbf{0}\not=\mathbf{i}\in\{0,1\}^3} \sum_{\substack{n_1,n_2,n_3,n_4\leqslant y\\ \boldsymbol{\alpha}(\mathbf{n},\mathbf{i})\not=0}} \frac{d(n_1)d(n_2)d(n_3)d(n_4)}{(n_1n_2n_3n_4)^{3/4}}\min\bigg(T^2,\frac{T^{3/2}}{|\boldsymbol{\alpha}(\mathbf{n},\mathbf{i})|}\bigg) =\Sigma_{(1)}+\Sigma_{(2)}, \end{equation} where \begin{equation*} \Sigma_{(1)}= \sum_{\substack{\mathbf{i}\in\{0,1\}^3\\ i_1+i_2+i_3=2}} \sum_{\substack{n_1,n_2,n_3,n_4\leqslant y\\ \boldsymbol{\alpha}(\mathbf{n},\mathbf{i})\not=0}} \frac{d(n_1)d(n_2)d(n_3)d(n_4)}{(n_1n_2n_3n_4)^{3/4}}\min\bigg(T^2,\frac{T^{3/2}}{|\boldsymbol{\alpha}(\mathbf{n},\mathbf{i})|}\bigg), \end{equation*} \begin{equation*} \Sigma_{(2)}= \sum_{\substack{\mathbf{i}\in\{0,1\}^3\\ i_1+i_2+i_3=1}} \sum_{\substack{n_1,n_2,n_3,n_4\leqslant y\\ \boldsymbol{\alpha}(\mathbf{n},\mathbf{i})\not=0}} \frac{d(n_1)d(n_2)d(n_3)d(n_4)}{(n_1n_2n_3n_4)^{3/4}}\min\bigg(T^2,\frac{T^{3/2}}{|\boldsymbol{\alpha}(\mathbf{n},\mathbf{i})|}\bigg). \end{equation*} Let \begin{equation*} g=g(n_1,n_2,n_3,n_4):=\frac{d(n_1)d(n_2)d(n_3)d(n_4)}{(n_1n_2n_3n_4)^{3/4}}. \end{equation*} For $\Sigma_{(1)}$, by a splitting argument, there exist $1\leqslant N_1,N_2,N_3,N_4\leqslant y$ such that \begin{equation}\label{Sigma_(1)-fenjie} \Sigma_{(1)}\ll T^\varepsilon \mathcal{G}(N_1,N_2,N_3,N_4) \end{equation} where \begin{align*} \mathcal{G}(N_1,N_2,N_3,N_4) = & \sum_{\substack{\sqrt{n_1}+\sqrt{n_2}\not=\sqrt{n_3}+\sqrt{n_4}\\ n_i\sim N_i,1\leqslant i\leqslant 4\\1\leqslant N_1\leqslant N_2\leqslant y\\ 1\leqslant N_3\leqslant N_4\leqslant y}}g \cdot \min\bigg(T^2,\frac{T^{3/2}}{|\sqrt{n_1}+\sqrt{n_2}-\sqrt{n_3}-\sqrt{n_4}|}\bigg) \end{align*} If $N_2\geqslant300N_4$, then $|\sqrt{n_1}+\sqrt{n_2}-\sqrt{n_3}-\sqrt{n_4}|\gg N_2^{1/2}$, so the trivial estimate yields \begin{equation*} \mathcal{G}(N_1,N_2,N_3,N_4)\ll \frac{T^{3/2+\varepsilon}N_1N_2N_3N_4}{(N_1N_2N_3N_4)^{3/4}N_2^{1/2}}\ll T^{3/2+\varepsilon}y^{1/2}\ll T^{2-1/8+\varepsilon}. \end{equation*} If $N_4\geqslant300N_2$, we can get the same estimate. So later we always suppose that $N_2\asymp N_4$. Let $\eta=\sqrt{n_1}+\sqrt{n_2}-\sqrt{n_3}-\sqrt{n_4}$. Write \begin{equation}\label{G-fenjie} \mathcal{G}(N_1,N_2,N_3,N_4)=\mathcal{G}_1+\mathcal{G}_2+\mathcal{G}_3, \end{equation} where \begin{eqnarray*} \mathcal{G}_1 & := & T^{2}\sum_{\substack{0<|\eta|\leqslant T^{-1/2}\\n_i\sim N_i,1\leqslant i\leqslant4 \\1\leqslant N_1\leqslant N_2\leqslant y\\ 1\leqslant N_3\leqslant N_4\leqslant y}}g, \\ \mathcal{G}_2 & := & T^{3/2}\sum_{\substack{ T^{-1/2}<|\eta|\leqslant 1 \\n_i\sim N_i,1\leqslant i\leqslant4 \\1\leqslant N_1\leqslant N_2\leqslant y\\ 1\leqslant N_3\leqslant N_4\leqslant y }}g|\eta|^{-1}, \\ \end{eqnarray*} \begin{eqnarray*} \mathcal{G}_3 & := & T^{3/2}\sum_{\substack{|\eta|>1 \\n_i\sim N_i,1\leqslant i\leqslant4 \\1\leqslant N_1\leqslant N_2\leqslant y\\ 1\leqslant N_3\leqslant N_4\leqslant y}}g|\eta|^{-1}. \end{eqnarray*} We estimate $\mathcal{G}_1$ first. By Lemma \ref{Zhai-lemma-5}, we get \begin{eqnarray}\label{G_1-111} \mathcal{G}_1 & \ll & \frac{T^{2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}}\mathscr{A}_1\big(N_1,N_2,N_3,N_4;T^{-1/2}\big) \nonumber \\ & \ll & \frac{T^{2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}}\big(T^{-1/2}N_4^{1/2}N_1N_2N_3+N_1N_3N_4^{1/2}\big) \nonumber \\ & \ll & T^{3/2+\varepsilon}(N_1N_3)^{1/4}+T^{2+\varepsilon}(N_1N_3)^{1/4}N_4^{-1} \nonumber \\ & \ll & T^{3/2+\varepsilon}y^{1/2}+T^{2+\varepsilon}(N_1N_3)^{1/4}N_4^{-1} \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{2+\varepsilon}(N_1N_3)^{1/4}N_4^{-1}. \end{eqnarray} On the other hand, by Lemma \ref{Zhai-lemma-3}, without loss of generality, we assume that $N_1\leqslant N_3\leqslant N_4$ and obtain \begin{eqnarray}\label{G-1-222} \mathcal{G}_1 & \ll & \frac{T^{2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}}\mathscr{A}_{-}(N_1,N_2,N_3,N_4;T^{-1/2}) \nonumber \\ & \ll & \frac{T^{2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}} \big(T^{-1/8}N_1^{7/8}+N_1^{1/2}\big) \big(T^{-1/8}N_3^{7/8}+N_3^{1/2}\big)\big(T^{-1/4}N_4^{7/4}+N_4\big) \nonumber \\ & \ll & T^{2+\varepsilon}(N_1N_3)^{-1/4}N_4^{-1/2}\big(T^{-1/8}N_1^{3/8}+1\big)\big(T^{-1/8}N_3^{3/8}+1\big)\big(T^{-1/4}N_4^{3/4}+1\big) \nonumber \\ & \ll & T^{2+\varepsilon}(N_1N_3)^{-1/4}N_4^{-1/2} \big(T^{-1/4}(N_1N_3)^{3/8}+T^{-1/8}N_3^{3/8}+1\big) \big(T^{-1/4}N_4^{3/4}+1\big) \nonumber \\ & \ll & T^{7/4+\varepsilon}(N_1N_3)^{1/8}N_4^{-1/2} \nonumber \\ & & +T^{2+\varepsilon}(N_1N_3)^{-1/4}N_4^{-1/2} \big(T^{-1/8}N_3^{3/8}+1\big) \big(T^{-1/4}N_4^{3/4}+1\big) \nonumber \\ & \ll & T^{7/4+\varepsilon}N_4^{-1/4}+T^{2+\varepsilon}(N_1N_3)^{-1/4}N_4^{-1/2}\big(T^{-3/8}N_4^{9/8}+1\big) \nonumber \\ & \ll & T^{7/4+\varepsilon}+T^{2+\varepsilon}(N_1K_3)^{-1/4}N_4^{-1/2}\big(T^{-3/8}N_4^{9/8}+1\big). \end{eqnarray} From (\ref{G_1-111}) and (\ref{G-1-222}), we get \begin{equation*} \mathcal{G}_1 \ll T^{2-1/8+\varepsilon}+T^{2+\varepsilon}\cdot \min\Bigg(\frac{(N_1N_3)^{1/4}}{N_4},\frac{T^{-3/8}N_4^{9/8}+1}{(N_1N_3)^{1/4}N_4^{1/2}}\Bigg). \end{equation*} \textbf{Case 1} If $N_4\gg T^{1/3}$, then $T^{-3/8}N_4^{9/8}\gg1,$ and thus \begin{eqnarray}\label{G_1-case-1} \mathcal{G}_1 & \ll & T^{2-1/8+\varepsilon}+T^{2+\varepsilon}\cdot \min\Bigg(\frac{(N_1N_3)^{1/4}}{N_4},\frac{T^{-3/8}N_4^{9/8}}{(N_1N_3)^{1/4}N_4^{1/2}}\bigg) \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{2+\varepsilon}\bigg(\frac{(N_1N_3)^{1/4}}{N_4}\bigg)^{1/2} \bigg(\frac{T^{-3/8}N_4^{9/8}}{(N_1N_3)^{1/4}N_4^{1/2}}\bigg)^{1/2} \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{2-3/16+\varepsilon}N_4^{-3/16}\ll T^{2-1/8+\varepsilon}. \end{eqnarray} \textbf{Case 2} If $N_4\ll T^{1/3}$, then $T^{-3/8}N_4^{9/8}\ll1$. By noting Lemma \ref{kong-lamma} and \begin{equation*} N_2\asymp N_4\asymp\max(N_1,N_2,N_3,N_4), \end{equation*} we have \begin{equation*} T^{-1/2}\gg|\eta|\gg(n_1n_2n_3n_4)^{-1/2}\max(n_1,n_2,n_3,n_4)^{-3/2}\asymp (N_1N_3)^{-1/2}N_4^{-5/2}. \end{equation*} Hence, we obtain \begin{eqnarray}\label{G_1-case-2} \mathcal{G}_1 & \ll & T^{2-1/8+\varepsilon}+T^{2+\varepsilon}\min\Bigg(\frac{(N_1N_3)^{1/4}}{N_4},\frac{1}{(N_1N_3)^{1/4}N_4^{1/2}}\Bigg) \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{2+\varepsilon}\Bigg(\frac{(N_1N_3)^{1/4}}{N_4}\Bigg)^{1/4}\Bigg(\frac{1}{(N_1N_3)^{1/4}N_4^{1/2}}\Bigg)^{3/4} \nonumber \\ & = & T^{2-1/8+\varepsilon}+T^{2+\varepsilon}(N_1N_3)^{-1/8}N_4^{-5/8} \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{2+\varepsilon}(T^{-1/2})^{1/4}\ll T^{2-1/8+\varepsilon}. \end{eqnarray} Combining (\ref{G_1-case-1}) and (\ref{G_1-case-2}), we get \begin{equation}\label{G-1-estimate} \mathcal{G}_1\ll T^{2-1/8+\varepsilon}. \end{equation} Now, we estimate $\mathcal{G}_2$. By a splitting argument, we get that there exists some $\delta$ satisfying $T^{-1/2}\ll \delta\ll 1$ such that \begin{equation*} \mathcal{G}_2\ll\frac{T^{3/2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}\delta}\times\sum_{\substack{\delta<|\eta|\leqslant2\delta\\ \eta\not=0}}1. \end{equation*} By Lemma \ref{Zhai-lemma-5}, we get \begin{eqnarray}\label{G-2-111} \mathcal{G}_2 & \ll & \frac{T^{3/2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}\delta}\mathscr{A}_1(N_1,N_2,N_3,N_4;2\delta) \nonumber \\ & \ll & \frac{T^{3/2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}\delta} (\delta N_4^{1/2}N_1N_2N_3+N_1N_3N_4^{1/2}) \nonumber \\ & = & T^{3/2+\varepsilon}(N_1N_3)^{1/4}+T^{3/2+\varepsilon}\delta^{-1}(N_1N_3)^{1/4}N_4^{-1} \nonumber \\ & \ll & T^{3/2+\varepsilon}y^{1/2}+T^{3/2+\varepsilon}\delta^{-1}(N_1N_3)^{1/4}N_4^{-1} \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{3/2+\varepsilon}\delta^{-1}(N_1N_3)^{1/4}N_4^{-1} . \end{eqnarray} On the other hand, by Lemma \ref{Zhai-lemma-3}, without loss of generality, we assume that $N_1\leqslant N_3\leqslant N_4$ and obtain \begin{eqnarray*} \mathcal{G}_2 & \ll & \frac{T^{3/2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}\delta}\times\mathscr{A}_{-}(N_1,N_2,N_3,N_4;2\delta) \nonumber \\ & \ll & \frac{T^{3/2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}\delta} \big(\delta^{1/4}N_1^{7/8}+N_1^{1/2}\big) \big(\delta^{1/4}N_3^{7/8}+N_3^{1/2}\big)\big(\delta^{1/2}N_4^{7/4}+N_4\big) \nonumber \\ & \ll & T^{3/2+\varepsilon}(N_1N_3)^{-1/4}N_4^{-1/2}\delta^{-1}\big(\delta^{1/4}N_1^{3/8}+1\big) \big(\delta^{1/4}N_3^{3/8}+1\big)\big(\delta^{1/2}N_4^{3/4}+1\big) \end{eqnarray*} \begin{eqnarray}\label{G-2-222} & \ll & T^{3/2+\varepsilon}(N_1N_3)^{-1/4}N_4^{-1/2}\delta^{-1} \big(\delta^{1/2}(N_1N_3)^{3/8}+\delta^{1/4}N_3^{3/8}+1\big) \big(\delta^{1/2}N_4^{3/4}+1\big) \nonumber \\ & \ll & T^{3/2+\varepsilon}(N_1N_3)^{1/8}N_4^{-1/2}\delta^{-1/2} \nonumber \\ & & +T^{3/2+\varepsilon}(N_1N_3)^{-1/4}N_4^{-1/2}\delta^{-1}\big(\delta^{1/4}N_3^{3/8}+1\big) \big(\delta^{1/2}N_4^{3/4}+1\big) \nonumber \\ & \ll & T^{7/4+\varepsilon}N_4^{-1/4}+T^{3/2+\varepsilon}(N_1N_3)^{-1/4}N_4^{-1/2}\delta^{-1}\big(\delta^{3/4}N_4^{9/8}+1\big). \end{eqnarray} From (\ref{G-2-111}) and (\ref{G-2-222}), we get \begin{equation*} \mathcal{G}_2 \ll T^{2-1/8+\varepsilon}+T^{3/2+\varepsilon}\delta^{-1} \cdot\min\Bigg(\frac{(N_1N_3)^{1/4}}{N_4},\frac{\delta^{3/4}N_4^{9/8}+1}{(N_1N_3)^{1/4}N_4^{1/2}}\Bigg). \end{equation*} \textbf{Case 1} If $\delta\gg N_4^{-3/2}$, then $\delta^{3/4}N_4^{9/8}\gg1$, we get (recall $\delta\gg T^{-1/2}$) \begin{eqnarray}\label{G_2-case-1} \mathcal{G}_2 & \ll & T^{2-1/8+\varepsilon}+T^{3/2+\varepsilon}\delta^{-1} \cdot\min\Bigg(\frac{(N_1N_3)^{1/4}}{N_4},\frac{\delta^{3/4}N_4^{9/8}}{(N_1N_3)^{1/4}N_4^{1/2}}\Bigg) \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+ T^{3/2+\varepsilon}\delta^{-1} \Bigg(\frac{(N_1N_3)^{1/4}}{N_4}\Bigg)^{1/2} \Bigg(\frac{\delta^{3/4}N_4^{9/8}}{(N_1N_3)^{1/4}N_4^{1/2}}\Bigg)^{1/2} \nonumber \\ & \ll & T^{2-1/8+\varepsilon} +T^{3/2+\varepsilon}\delta^{-5/8}N_4^{-3/16} \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{3/2+\varepsilon}T^{5/16}N_4^{-3/16}\ll T^{2-1/8+\varepsilon}. \end{eqnarray} \textbf{Case 2} If $\delta\ll N_4^{-3/2}$, then $\delta^{3/4}N_4^{9/8}\ll1$. By Lemma \ref{kong-lamma}, we have \begin{equation*} \delta\gg|\eta|\gg(n_1n_2n_3n_4)^{-1/2}\max(n_1,n_2,n_3,n_4)^{-3/2}\asymp (N_1N_3)^{-1/2}N_4^{-5/2}. \end{equation*} Therefore, we obtain (recall $\delta\gg T^{-1/2}$) \begin{eqnarray} \label{G_2-case-2} \mathcal{G}_2 & \ll & T^{2-1/8+\varepsilon}+T^{3/2+\varepsilon}\delta^{-1}\cdot \min\Bigg(\frac{(N_1N_3)^{1/4}}{N_4},\frac{1}{(N_1N_3)^{1/4}N_4^{1/2}}\Bigg) \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{3/2+\varepsilon}\delta^{-1}\Bigg(\frac{(N_1N_3)^{1/4}}{N_4}\Bigg)^{1/4} \Bigg(\frac{1}{(N_1N_3)^{1/4}N_4^{1/2}}\Bigg)^{3/4} \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{3/2+\varepsilon}\delta^{-1}(N_1N_3)^{-1/8}N_4^{-5/8} \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{3/2+\varepsilon}\delta^{-1}\delta^{1/4} \nonumber \\ & \ll & T^{2-1/8+\varepsilon}+T^{3/2+\varepsilon}T^{3/8}\ll T^{2-1/8+\varepsilon}. \end{eqnarray} Combining (\ref{G_2-case-1}) and (\ref{G_2-case-2}), we get \begin{equation}\label{G-2-estimate} \mathcal{G}_2\ll T^{2-1/8+\varepsilon}. \end{equation} For $\mathcal{G}_3$, by a splitting argument and Lemma \ref{Zhai-lemma-5} again, we get \begin{eqnarray*} \mathcal{G}_3 & \ll & \frac{T^{3/2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}\delta}\times \sum_{\substack{\delta<|\eta|\leqslant2\delta\\ \delta\gg 1}}1 \end{eqnarray*} \begin{eqnarray} \label{G-3-estimate} & \ll & \frac{T^{3/2+\varepsilon}}{(N_1N_2N_3N_4)^{3/4}\delta}\cdot \delta N_4^{1/2}N_1N_2N_3\ll T^{3/2+\varepsilon}(N_1N_3)^{1/4} \nonumber \\ & \ll & T^{3/2+\varepsilon}y^{1/2}\ll T^{2-1/8+\varepsilon}. \end{eqnarray} Comnining (\ref{Sigma_(1)-fenjie}), (\ref{G-fenjie}), (\ref{G-1-estimate}), (\ref{G-2-estimate}) and (\ref{G-3-estimate}), we deduce that \begin{equation}\label{Sigma_{(1)}-upper} \Sigma_{(1)}\ll T^{2-1/8+\varepsilon}. \end{equation} In the same way, by using Lemma \ref{Zhai-lemma-6} instead of Lemma \ref{Zhai-lemma-5}, we can follow the above arguments step by step and obtain \begin{equation}\label{Sigma_{(2)}-upper} \Sigma_{(2)}\ll T^{2-1/8+\varepsilon}. \end{equation} It follows from (\ref{S_2(x)-upper-all}), (\ref{S_2(x)-all-positive}), (\ref{not-all>0-Sigma1+Sigma2}), (\ref{Sigma_{(1)}-upper}) and (\ref{Sigma_{(2)}-upper}) that \begin{equation}\label{S_2(x)-upper-last} \int_\frac{T}{2}^TS_2(x)\mathrm{d}x\ll T^{2-1/8+\varepsilon}. \end{equation} According to (\ref{R_0=S_1(x)+S_2(x)}), (\ref{S_1(x)-asymptotic}) and (\ref{S_2(x)-upper-last}), we conclude that \begin{equation}\label{R_0-fourth-power-mean} \int_{\frac{T}{2}}^T\mathscr{R}_0^4\mathrm{d}x=\frac{1}{32\pi^4}C_4\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2}\bigg) \int_{\frac{T}{2}}^Tx\mathrm{d}x+O(T^{2-1/8+\varepsilon}). \end{equation} From (4.11) of Liu \cite{Liu-Kui}, we know that \begin{equation}\label{R_1+R_2-mean-square} \int_{\frac{T}{2}}^T|\mathscr{R}_1+\mathscr{R}_2|^2\mathrm{d}x\ll T^{1/2}y^{-1/2}\mathscr{L}^3. \end{equation} Also, by (4.14) of Liu \cite{Liu-Kui}, we have \begin{equation}\label{R_1+R_2-mean-8th-power} \int_{\frac{T}{2}}^T|\mathscr{R}_1+\mathscr{R}_2|^8\mathrm{d}x\ll T^{3+\varepsilon}. \end{equation} It follows from (\ref{R_1+R_2-mean-square}) and (\ref{R_1+R_2-mean-8th-power}) and Cauchy's inequality that \begin{align}\label{R_1+R_2-mean-4th-power} \int_{\frac{T}{2}}^T|\mathscr{R}_1+\mathscr{R}_2|^4\mathrm{d}x = & \int_{\frac{T}{2}}^T|\mathscr{R}_1+\mathscr{R}_2|^{4/3}|\mathscr{R}_1+\mathscr{R}_2|^{8/3}\mathrm{d}x \nonumber \\ \ll & \bigg(\int_{\frac{T}{2}}^T|\mathscr{R}_1+\mathscr{R}_2|^2\mathrm{d}x\bigg)^{2/3} \bigg(\int_{\frac{T}{2}}^T|\mathscr{R}_1+\mathscr{R}_2|^8\mathrm{d}x\bigg)^{1/3} \nonumber \\ \ll & (T^{1/2}y^{-1/2}\mathscr{L}^3)^{2/3}(T^{3+\varepsilon})^{1/3} \nonumber \\ \ll & T^{4/3+\varepsilon}y^{-\frac{1}{3}}\ll T^{13/12+\varepsilon}. \end{align} By (\ref{R_0-fourth-power-mean}), (\ref{R_1+R_2-mean-4th-power}) and H\"{o}lder's inequality, we deduce that \begin{align}\label{R_^3|R_1+R_2|-mean-value} \int_{\frac{T}{2}}^T|\mathscr{R}_0|^3|\mathscr{R}_1+\mathscr{R}_2|\mathrm{d}x \ll & \bigg(\int_{\frac{T}{2}}^T\mathscr{R}_0^4\mathrm{d}x\bigg)^{3/4}\bigg(\int_{\frac{T}{2}}^T|\mathscr{R}_1+\mathscr{R}_2|^4\mathrm{d}x\bigg)^{1/4} \nonumber \\ \ll & (T^2)^{3/4}(T^{13/12+\varepsilon})^{1/4}\ll T^{85/48+\varepsilon}\ll T^{2-1/8+\varepsilon}. \end{align} Combining (\ref{R_4-mean-explicit}), (\ref{R_0-fourth-power-mean}), (\ref{R_1+R_2-mean-4th-power}) and (\ref{R_^3|R_1+R_2|-mean-value}), we obtain \begin{equation}\label{R^4-asymp-conclusion} \int_{\frac{T}{2}}^T\mathscr{R}^4\mathrm{d}x=\frac{1}{32\pi^4}C_4\bigg(\frac{\ell_1}{M_1},\frac{\ell_2}{M_2}\bigg) \int_{\frac{T}{2}}^Tx\mathrm{d}x+O(T^{2-1/8+\varepsilon}). \end{equation} This completes the proof of Proposition \ref{proposition-1}. \end{proof} \begin{proposition}\label{proposition-2} For $H\gg T^\varepsilon$, we have \begin{equation*} \int_{\frac{T}{2}}^T\mathscr{G}^4\mathrm{d}x\ll T^{3/2}H^{-1/4}, \end{equation*} \begin{equation*} \int_{\frac{T}{2}}^T|\mathscr{R}|^3|\mathscr{G}|\mathrm{d}x\ll T^{2-1/8+\varepsilon}, \end{equation*} \begin{equation*} \int_{\frac{T}{2}}^T|\mathscr{R}|^3\mathscr{L}^3\mathrm{d}x\ll T^{2-1/8+\varepsilon}. \end{equation*} \end{proposition} \begin{proof} Recall that \begin{equation*} \mathscr{G}=G_{12}(x;H)+G_{21}(x;H). \end{equation*} From (\ref{G_12(x;H)-def}), we deduce that \begin{align}\label{G_12(x;H)-fourth-power} \int_{\frac{T}{2}}^T\big(G_{12}(x;H)\big)^4\mathrm{d}x \ll & \sum_{\substack{n_1\leqslant\sqrt{M_1M_2T}\\ n_1\equiv \ell_1(\!\bmod M_1)}} \int_{\frac{T}{2}}^T\min\Bigg(1,\frac{1}{H\big\|\frac{M_1x}{n_1}-\frac{\ell_2}{M_2}\big\|^4}\Bigg)\mathrm{d}x \nonumber \\ \ll & \sum_{\substack{n_1\leqslant\sqrt{M_1M_2T}\\ n_1\equiv \ell_1(\!\bmod M_1)}}\frac{n_1}{M_1} \int_{\frac{M_1T}{2n_1}-\frac{\ell_2}{M_2}}^{\frac{M_1T}{n_1}-\frac{\ell_2}{M_2}}\min\bigg(1,\frac{1}{H\|u\|^4}\bigg)\mathrm{d}u \nonumber \\ \ll & \,\,T\sum_{\substack{n_1\leqslant\sqrt{M_1M_2T}\\ n_1\equiv \ell_1(\!\bmod M_1)}} \int_0^1\min\bigg(1,\frac{1}{H\|U\|^4}\bigg)\mathrm{d}u \nonumber \\ \ll & \,\,T\sum_{\substack{n_1\leqslant\sqrt{M_1M_2T}\\ n_1\equiv \ell_1(\!\bmod M_1)}} \bigg\{\int_0^{\frac{1}{2}}+\int_{\frac{1}{2}}^1\bigg\}\min\bigg(1,\frac{1}{H\|U\|^4}\bigg)\mathrm{d}u \nonumber \\ \ll & \,\, T\sum_{\substack{n_1\leqslant\sqrt{M_1M_2T}\\ n_1\equiv \ell_1(\!\bmod M_1)}} \int_0^{\frac{1}{2}}\min\bigg(1,\frac{1}{Hu^4}\bigg)\mathrm{d}u \nonumber \\ \ll & \,\, T\sum_{\substack{n_1\leqslant\sqrt{M_1M_2T}\\ n_1\equiv \ell_1(\!\bmod M_1)}} \bigg(\int_0^{H^{-\frac{1}{4}}}1\mathrm{d}u+\int_{H^{-\frac{1}{4}}}^{\frac{1}{2}}\frac{1}{Hu^4}\mathrm{d}u\bigg) \nonumber \\ \ll & \,\, T^{3/2}H^{-1/4}. \end{align} Similarly, we also have \begin{equation}\label{G_21(x;H)-fourth-power} \int_{\frac{T}{2}}^T\big(G_{21}(x;H)\big)^4\mathrm{d}x \ll T^{3/2}H^{-1/4}. \end{equation} Therefore, from (\ref{G_12(x;H)-fourth-power}) and (\ref{G_21(x;H)-fourth-power}), we derive that \begin{equation}\label{G-fourth-power} \int_{\frac{T}{2}}^T\mathscr{G}^4\mathrm{d}x \ll T^{3/2}H^{-1/4}. \end{equation} It follows from (\ref{R^4-asymp-conclusion}), (\ref{G-fourth-power}) and Holder's inequality that \begin{align} \int_{\frac{T}{2}}^T|\mathscr{R}|^3|\mathscr{G}|\mathrm{d}x \ll & \,\,\bigg( \int_{\frac{T}{2}}^T|\mathscr{R}|^4\mathrm{d}x\bigg)^{3/4}\bigg(\int_{\frac{T}{2}}^T|\mathscr{G}|^4\mathrm{d}x\bigg)^{1/4} \nonumber \\ \ll & \,\,T^{15/8}H^{-1/16}\ll T^{2-1/8+\varepsilon}, \end{align} and \begin{equation} \int_{\frac{T}{2}}^T|\mathscr{R}|^3\mathscr{L}^3\mathrm{d}x \ll \bigg( \int_{\frac{T}{2}}^T\mathscr{R}^4\mathrm{d}x\bigg)^{3/4}\bigg(\int_{\frac{T}{2}}^T \mathscr{L}^{12}\mathrm{d}x\bigg)^{1/4} \ll T^{7/4}\mathscr{L}^3\ll T^{2-1/8+\varepsilon}. \end{equation} This completes the proof of Proposition \ref{proposition-2}. \end{proof} From (\ref{Delta(M_1M_2x)-fenjie}), Proposition \ref{proposition-1} and Proposition \ref{proposition-2}, we get the fourth--power moment of $\Delta(M_1M_2x;\ell_1,M_1,\ell_2,M_2)$. \section*{Acknowledgement} The authors would like to express the most sincere gratitude to Professor Wenguang Zhai for his valuable advice and constant encouragement.

\section{\label{Intro} Introduction} The low-temperature $\alpha$-phase of Sn has attracted considerable attention recently as a unique elemental three-dimensional topologically non-trivial material \cite{,}. Being a zero-gap semiconductor if unstrained, it can enter a Dirac semimetal or strong topological insulator (TI) phases under strain \cite{}. Interestingly, the non-trivial topology and, hence, the topological surface state (TSS) in $\alpha$-Sn\ exist independently of the strain owing to a robust band inversion between conduction and second valence bands, similar to HgTe \cite{Bernevig2006} or some ternary Heusler compounds \cite{Chadov2010}. While the electronic structure of $\alpha$-Sn\ has been reported in several experimental studies \cite{}, a detailed analysis of the TSS is still missing. Despite the degeneracy of the TSS with the projected bulk bands, $\alpha$-Sn\ has proven its potential for spintronics applications \cite{}. A weak hybridization of the TSS with the energetically coexisting bulk bands \cite{} calls for the analysis of the TSS quasiparticle lifetime, which has not been reported so far. Additional spintronic functionalities can be expected from introducing ferromagnetic impurities, which are suggested to break time-reversal symmetry, thereby opening a gap in the TSS. This is of interest for devices, such as transistors, as well as for achieving the quantum anomalous Hall state \cite{}. However, their experimental realization has proven challenging and has led to yet another controversy in the literature, i.e., whether adding ferromagnetic impurities in a TI can indeed open a band gap in the TSS \cite{}. Such effects were mainly studied in the Bi$_2$X$_3$\ family of compounds, and related ternary systems, and it has been argued that, e.g., the high intrinsic doping in Bi$_2$X$_3$\ prevents the proposed surface-state-mediated ferromagnetic alignment of the impurities \cite{}. Hence, the tunability of the Dirac point might be an essential requirement to obtain and study a band gap opening via magnetic impurities in the TSS. In this paper, we present different aspects of the electronic structure of $\alpha$-Sn\ films using spin- and angle-resolved photoemission (ARPES). We reveal that although the TSS is degenerate with the bulk band structure, the former maintains an isotropic {\bf k}$_\parallel$-space dispersion with circular constant energy contours (CECs) unaffected by different surface reconstructions of the film. We compensate the intrinsic $p$-type doping of the as-grown film by surface deposition of potassium, thereby shifting the Kramers degeneracy point of the surface state below the Fermi level. Additional control over the precise position of $E_{\rm F}$\ is gained through the variation of the thickness of the Te buffer layer between substrate and film. Despite this tunability, we present data that unambiguously demonstrates the absence of a resolvable gap in the TSS in $\alpha$-Sn\ after Fe adatom deposition [up to 0.25 monolayers (ML)], which is in line with previous findings in Bi$_2$Se$_3$\ and Bi$_2$Te$_3$\ \cite{}. Furthermore, we have measured the spin-vector alignment of electrons in the TSS by means of spin-resolved ARPES. Our results show that the spin vector lies fully in the sample plane along all directions in $k$-space. Further, our results on the circular dichroism in the angular distribution (CDAD) of photoelectrons confirm that the method does not provide a reliable measure of spin polarization \cite{}. Finally, we analyze the decay of photoholes introduced in the photoemission process, and in this way, get insight into the many-body interactions in the system. The extracted quasiparticle lifetimes are found to be comparable to other TIs where the TSSs appear within a bulk band gap. We argue that the main decay of photoholes is caused by intraband scattering and that scattering into bulk states is suppressed by virtue of different orbital symmetries and consequently small hybridization of bulk and surface states. \section{\label{Methods} Methods} $\alpha$-Sn\ thin film samples were grown by molecular beam epitaxy (MBE) on InSb(001) substrates. The 8-effusion cell MBE system is directly attached to the high-resolution ARPES system at beamline I05 at the Diamond Light Source (Didcot, UK), allowing for in-vacuum transfers \cite{}. The substrates were cleaned by several sputter and anneal cycles until a clear c(8$\times$2)-reconstruction was observed by low-energy-electron diffraction (LEED). Afterwards, a Te buffer layer was deposited, inducing a (1$\times$1) surface reconstruction. The amount of Te was monitored by x-ray photoemission spectroscopy (XPS) and LEED. Next, Sn was grown layer-by-layer, as evidenced by clear RHEED oscillations (not shown here). After depositing for $\sim$10-12 RHEED oscillations (monatomic Sn layers) the thickness of the film is sufficient to neglect the influence of the InSb/Te interface. ARPES measurements have been carried out with $s$ and $p$ linearly polarized light, as well as circularly polarized light, at varying photon energies at beamline I05. The endstation is equipped with a Scienta R4000 hemispherical electron analyzer that provides an ultimate energy and angular resolution of $\sim$5\,meV and 0.1$^{\circ}$, respectively. Spin-resolved photoemission spectra were measured at the SIS beamline at the Swiss Light Source using the COPHEE spectrometer with two 40\,kV classical Mott detectors \cite{Hoesch2002spin} at a photon energy of 19\,eV. Energy and angular resolutions were $\sim$60\,meV ($\sim$20\,meV in spin-integrated spectra) and 1.5$^{\circ}$ (0.5$^{\circ}$ in spin-integrated spectra), respectively. \begin{figure} [ht] \includegraphics[width=\linewidth]{Fig1_Remastered} \caption{(a) Illustration of the surface-truncated diamond crystal structure (top) and the reconstructed surface (bottom). Smaller size of atoms denotes atoms situated further from the (110) plane along [1$\bar{1}$0] direction. (b) Evolution of the LEED pattern from a c(8$\times$2) surface reconstruction for \textit{in situ} prepared InSb (top), to a nearly (1$\times$1) for the Te buffer on InSb (center), to a (2$\times$1) double domain pattern for Sn grown on top of the buffer layer (bottom). (c) Angle-dependent XPS (Al K$\alpha$) data.} \label{Fig1} \end{figure} \section{\label{Crystal}Atomic Surface Structure} Since $\alpha$-Sn\ crystallizes in the diamond structure, we want to briefly review the structural basics of the (001) surface. If the bulk is truncated at the (001) surface, the topmost atoms miss their partners to saturate bonds. Being of $sp^3$-character, these dangling bonds are strongly directional and lie diagonally in the (110) plane. Local density approximation calculations show that the unreconstructed surface is energetically not favorable due to a very high surface energy of $1.530\, \mathrm{eV} / ($1$\times$1$)\ \mathrm{cell}$ \cite{}. According to Lu \textit{et al.}\ \cite{}, a stabilization is achieved through the formation of asymmetric dimers, which leads to an energy gain of $0.618\, \mathrm{eV} / ($1$\times$1$)\ \mathrm{cell}$ [Fig.~\ref{Fig1}(a)]. This dimer formation is well known also for the Si(001) and Ge(001) surfaces \cite{}, and it is the building block for many reconstructions observed in experiments. For $\alpha$-Sn(001), the formation of reconstructions has been investigated by Yuen and coworkers, who found a dependence of the reconstruction type on the film thickness \cite{}. In the regime of interest, i.e., below 200\,\AA, the formation of a (2$\times$1) reconstruction is favored. It shows up in a double-domain fashion since the dangling bonds of the truncated bulk are oriented in two orthogonal directions on neighboring surface terraces separated by a mono-atomic step. As can be seen in Fig.~\ref{Fig1}(b), the LEED images show an evolution from a c(8$\times$2)-reconstructed surface (top) for pristine InSb(001) after surface preparation to a (2$\times$1) reconstruction for the Sn-covered surface (bottom). We found that adding Te prior, during, or after the Sn deposition improves the crystal quality as it apparently acts as a surfactant. Depending on the amount of Te added, the dangling bonds of Sn become saturated and the formation of dimers is avoided, resulting in an unreconstructed (1$\times$1) surface structure \cite{}. A similar effect is seen on the InSb surface after growth of a Te buffer layer [Fig.~\ref{Fig1}(b), center]. Figure~\ref{Fig1}(c) shows angle-dependent core-level spectra of a Sn film grown on InSb(001) where a Te buffer layer was added before Sn deposition. The $3d$ states of all elements that contribute to the signal are situated next to each other in binding energy such that the qualitative dependence of the intensities on the emission angle can be easily followed. The limited escape depth of photoelectrons leads to an enhanced surface sensitivity for off-normal emission where the escape depth is proportional to the cosine of the emission angle measured relative to the surface normal. Naively, one would expect Te, Sb, and In to decrease in intensity for 60$^\circ$ off normal emission (light gray line) as Sn was deposited on top of the three. As can be seen, this is not the case: with increasing surface sensitivity the intensity of Sb and In reduces, as expected. However, also the Sn intensity is decreasing, while the Te intensity appears unaltered, or slightly higher, for 60$^\circ$ off normal emission as compared to normal emission (0$^\circ$). This gives clear indication that Te segregates on top of the growing film, instead of buffering the interface between InSb and Sn. Such behavior is typical of surfactants which alter the surface free energies \cite{}. We quantify the results from the angle-dependent XPS by calculating the relative spectral-weight change according to \begin{eqnarray} \frac{A_{\rm 60}-A_{\rm NE}}{A_{\rm 60}+A_{\rm NE}}, \end{eqnarray} where $A_{\rm 60}$ and $A_{\rm NE}$ are the areas below each $3d_{5/2}$-peak for the measurement at 60$^\circ$ off normal emission and at normal emission, respectively. From each peak, a Shirley background has been removed prior to the analysis. The calculation gives an increase of $\sim$16\% for the Te state and a reduction of $\sim$19\% for the Sn state. Remarkably, the Sb state is reduced by $\sim$40\%, while the In state is reduced by less than $\sim$29\%. This indicates that diffusion of In into the film may persist despite the Te buffer layer. Interdiffusion of In can be further enhanced by formation of metallic In islands that accompanies InSb substrate surface preparation via sputtering and annealing \cite{}. Furthermore, Te seems to act as a $n$-type dopant (as it does in group-IV semiconductors like Si and Ge \cite{Sze1981physics}), or, it is at least able to compensate intrinsic $p$-type doping that is likely to be caused by the interdiffusion of In atoms from the substrate into the film \cite{}. \begin{figure} [ht] \includegraphics[width=\linewidth]{Fig2_Remastered} \caption{(a) Sketch of the bulk band structure of $\alpha$-Sn\ on InSb(001) along the in-plane $\Gamma$-K direction and the $\Gamma$-Z direction perpendicular to the (001) surface. (b) Bulk and surface Brillouin zone of strained $\alpha$-Sn\ with symmetry points labeled.} \label{BZ} \end{figure} \section{\label{Electronic}Electronic Structure} \begin{figure*} [!ht] \includegraphics[width=\linewidth]{Fig3_Remastered} \caption{ Experimental electronic structure data. (a) Band maps of an as-grown $\alpha$-Sn\ film measured with $h \nu$=18\,eV at $T$=8\,K along the $\overline{\Gamma}$-$\overline{\rm X}$\ direction (left) and $\overline{\Gamma}$-$\overline{\rm M}$\ direction (right). (b) Photon energy scan between $h \nu$=18\,eV and $h \nu$=32\,eV taken along the the $\overline{\Gamma}$-$\overline{\rm M}$\ direction at normal emission at $E_{\rm F}$\ (left) and $E_{\rm B}$=250\,meV (right). Spectra have been normalized to have equal intensity for each $k_{\rm z}$. The features that show up between $k_{\rm z}$=2.5\,\AA$^{-1}$\ and $k_{\rm z}$=2.7\,\AA$^{-1}$\ correspond to core levels that appear due to higher order excitations and can thus be ignored. (c) Stack of experimental CECs with black dotted lines following the TSS dispersion. (d) MDCs extracted at $E_{\rm F}$\ for the $\overline{\Gamma}$-$\overline{\rm X}$\ and $\overline{\Gamma}$-$\overline{\rm M}$\ directions. (e, f) Sketches of the experimental CECs for different binding energies show the overlap of circular TSS (red) and elliptic bulk bands (black), as well as surface state SS (green).} \label{ARPES_standard} \end{figure*} Recently it was shown that the compressive strain in $\alpha$-Sn\ films induced by the InSb substrate opens only a local band gap at the $\Gamma$-point of the bulk Brillouin zone [BBZ, Fig.~\ref{BZ}(b)], which closes along the direction perpendicular to the (001) surface, i.e., along the $\Gamma$-Z direction [Fig.~\ref{BZ}(a)] \cite{Xu2017,}. This gives rise to two three-dimensional Dirac points along the line Z-$\Gamma$-Z and defines $\alpha$-Sn\ as a topological Dirac semimetal. We note that this gap closing does not alter the bulk topological properties as the strain induced band crossing occurs between the two \GEP bands that form the topmost valence and conduction band in unstrained $\alpha$-Sn. As both share the same parity eigenvalues, the local gap opening induced by strain does not change the $\mathbb{Z}_2$-invariant. Note, that the size of the gap in the sketch of Fig.~\ref{BZ}(a) is exaggerated, and due to its small real size ($\sim$ 30\,meV) and photoemission {\bf k}$_\perp$\ broadening \cite{Strocov2003} we have no experimental resolution to clearly show the 3D Dirac points in ARPES. We approach the topological transition by reviewing the electronic structure of InSb in comparison with $\alpha$-Sn. Based on nonlocal pseudopotential calculations, Chelikowsky and Cohen were able to show that in InSb the \GSM band lies above $E_{\rm F}$\ for all time reversal symmetric momenta \cite{}. In contrast, the \GSM band is pushed below $E_{\rm F}$\ in $\alpha$-Sn \cite{Groves1963} due to mainly scalar relativistic effects that affect the s-electrons \cite{Cardona2010,Zhu2012}, thus giving rise to a change of the $\mathbb{Z}_2$-invariant \cite{}. There is yet another band inversion in $\alpha$-Sn\ as the \GSP\ band is pushed below the \GSM\ band [Fig.~\ref{BZ}(b)] \cite{}. While this inversion has no effect on the $\mathbb{Z}_2$-invariant, the presence of a second TSS in $\alpha$-Sn\ was recently discovered, that indeed connects the \GSM and \GSP bands in agreement with this inversion \cite{}. In Fig.~\ref{ARPES_standard}(a) we present energy dispersive maps of $\alpha$-Sn\ acquired at $T$=8\,K with an excitation energy of $h \nu$=18\,eV translating into a surface perpendicular momentum $k_{\rm z}$ = $2.33\times(2\pi/c$), assuming an inner potential of $V_0$=5.8\,eV (Ref.\ [\onlinecite{}]). The left-hand panel shows the dispersion along the $\overline{\Gamma}$-$\overline{\rm X}$\ direction (note that due to the dual domain nature of the surface reconstruction in our samples the $\overline{\rm X}$\ and $\overline{\rm Y}$\ points of the surface Brillouin zone (SBZ) can not be distinguished). At 0\,eV binding energy we observe a single sharp feature that crosses $E_{\rm F}$, corresponding to the TSS reported recently \cite{}. In addition, we observe a weak background intensity caused by the projected bulk \GEP band that also crosses $E_{\rm F}$, revealing a metallic behavior. The apparent $p$-type doping is likely caused by diffusion of In from the substrate into the film \cite{}, whereas the amount of Te in the buffer layer was not enough to compensate for it. From the momentum distribution curves (MDCs) taken at $E_{\rm F}$\ we extract the Fermi wavevector $k_{\rm F}$\ for the $\overline{\Gamma}$-$\overline{\rm X}$\ direction to $\sim$0.035\,\AA$^{-1}$ [Fig.~\ref{ARPES_standard}(d)]. When following the TSS dispersion to higher binding energy, a reduction in intensity, accompanied by a broadening, is observed. This broadening is ascribed to the presence of the bulk \GSM band, and a hybridization with it. The hybridization is even more pronounced for the $\overline{\Gamma}$-$\overline{\rm M}$\ direction as shown on the right-hand side of Fig.~\ref{ARPES_standard}(a). At $E_{\rm F}$\ we get a similar picture as along the $\overline{\Gamma}$-$\overline{\rm X}$\ direction with the only intense feature being the TSS. For higher binding energies, a second feature is observed, which we will refer to as the conventional surface state (SS), and which gains intensity and sharpness with increasing binding energy. In Fig.~\ref{ARPES_standard}(b) we show a photon energy scan between $h \nu$=18\,eV and $h \nu$=32\,eV, i.e., along the $\Gamma$-Z direction of the BBZ, while the surface is oriented with the $\overline{\Gamma}$-$\overline{\rm M}$\ direction along the angle-dispersive axis of the analyzer. At the Fermi level [Fig.~\ref{ARPES_standard}(b), left] we observe a single intense feature that corresponds to the TSS, in agreement with the results of Fig.~\ref{ARPES_standard}(a). This feature appears as a straight line along the {\bf k}$_\perp$\ direction, which is clear proof of its two-dimensional character, i.e., its surface localization. At 250\,meV binding energy [Fig.~\ref{ARPES_standard}(b), right], in addition to the TSS, one can observe the second sharp feature that originates from the SS band [see Fig.~\ref{ARPES_standard}(a)]. Clearly, this additional feature shows no dispersion in $k_z$ and dresses the TSS with a constant separation in {\bf k}$_\parallel$. We can therefore assign a two-dimensional character to this state as well. To elucidate the topography of the states in more detail, we present a set of experimental CECs in Fig.~\ref{ARPES_standard}(c). Sketches of CECs are shown in Fig.~\ref{ARPES_standard}(e),(f). The energy-dependence of the intense TSS in the stack of CECs in Fig.~\ref{ARPES_standard}(c) is illustrated by the black guide lines, which suggest a Dirac-point energy of $E_{\rm D}$ $\approx$150\,meV above $E_{\rm F}$. The shape of the Fermi contour can be resolved as a cloverleaf-like feature, resulting from two orthogonal elliptic contours that cross each other symmetrical about $\Gamma$. They originate from the bulk states and appear to be degenerate with the central TSS. Since the intensities of the elliptic bulk features add up at the crossing points along the four $\overline{\Gamma}$-$\overline{\rm M}$\ directions, the overall shape of the TSS appears to be quadratic. These crossing points are seen better at higher binding energies ($>$200\,meV) where the TSS shows a weaker intensity. The quadratic shape would fulfill a strong nesting condition, with the possibility of an emerging spin-density wave \cite{}, however, the $k_{\rm F}$\ values for TSS extracted from MDCs shown in Fig.~\ref{ARPES_standard}(d) are 0.035\,\AA$^{-1}$ and 0.037\,\AA$^{-1}$\ for the $\overline{\Gamma}$-$\overline{\rm X}$\ and $\overline{\Gamma}$-$\overline{\rm M}$\ direction, respectively. The latter is far off the expected value of $\sqrt{2} k_F= 0.049$\,\AA$^{-1}$\ along the $\overline{\Gamma}$-$\overline{\rm M}$\ direction for a quadratic shape. Therefore, for the available binding-energy range, we establish a circular shape of the TSS-related Fermi contour, while it only appears to be quadratic due to overlap with bulk elliptic features. \section{\label{DopStud}Surface doping study} Previously, we have studied the effect of Te co-deposition during the growth of Sn on InSb(001) \cite{}. It was shown that an enhanced flux of Te leads to a reduction of the intrinsic $p$-type doping, however, it has not been clarified whether Te prevents the outdiffusion of In atoms into the film, thus reducing the $p$-type doping, or whether Te itself acts as an electron donor. As already described above, samples in this study were grown with a Te buffer on InSb(001) instead, without an additional Te supply during the Sn deposition. \begin{figure*} [!htb] \includegraphics[width=\linewidth]{Fig4_Remastered} \caption{Study of K and Te adatom deposition. (a) ARPES band maps for the pristine sample (left), an intermediate K coverage (center, 3\,min. K deposition), and the final K coverage (right, 15\,min. K deposition) along the $\overline{\Gamma}$-$\overline{\rm X}$\ direction with $h \nu$=18\,eV at 8\,K. (b) ARPES band maps for the pristine sample (left), an intermediate Te coverage (center), and high Te coverage (right) along the $\overline{\Gamma}$-$\overline{\rm M}$\ direction at room temperature. Plots have been normalized to equal MDC intensity. (c) While the as-grown sample shows a (2$\times$1) double-domain diffraction pattern in LEED (left), we observe a (1$\times$1) pattern at the end of K deposition study (right). (d) By comparing the CECs at the same energy relative to $E_{\rm D}$\ for the pristine and highly Te-doped sample, we find no obvious difference (left-hand panel). The right-hand panel compares MDCs for the two orthogonal directions $ k_x $ and $ k_y $ for pristine ($E_{\rm B}$=200\,meV, black) and high Te-doped ($E_{\rm B}$=250\,meV, red) samples. } \label{Doping} \end{figure*} \textbf{K deposition.} To be able to show that the TSS discussed with respect to Fig.~\ref{ARPES_standard} is indeed the TSS reported in previous studies \cite{}, we have conducted a series of K depositions on the surface of $\alpha$-Sn\ at $T$=8\,K, monitoring the changes in the band structure after each step. Figure \ref{Doping}(a) shows band maps for the pristine sample (left), an intermediate K coverage (center, 3\,min K deposition), and the final K coverage (right, 15\,min K deposition). To determine the TSS peak positions, MDCs were fitted with Lorentzian peaks in a small energy window. Subsequently, by fitting a straight line to the peak positions, one can extrapolate $E_{\rm D}$\ to quantify the shift of $E_{\rm F}$. For the pristine sample $E_{\rm D}$\ is found $\sim$150\,meV above $E_{\rm F}$. One can clearly observe a shift of the Fermi level towards the Dirac point for the intermediate K dose. The band shift appears to be rigid since the group velocities determined from the linear approximation agree within the errors of our analysis. The Dirac point is now found $\sim$50\,meV above $E_{\rm F}$. For the highest K dose, shown in the right-hand panel, the intensity from the Sn states is suppressed due to the highly disordered surface. However, the extrapolation of the Dirac point matches well with the \textit{observed} Dirac point that shows an enhanced intensity. The crossing point is found $\sim$65\,meV below $E_{\rm F}$, therefore we can quantify the total shift of $E_{\rm F}$\ induced by K to $\sim$215\,meV. In addition to this shift, we observe a transition from a (2$\times$1) reconstruction of the pristine sample to an unreconstructed (1$\times$1) surface after the final K deposition in LEED [Fig.~\ref{Doping}(c)]. \textbf{Te deposition.} The same transition was observed in LEED for Te surface deposition that has been conducted at room temperature (RT) in order to reduce possible cross-contamination due to otherwise extreme temperature changes between deposition and measurement steps. The overall spectral quality becomes reduced at RT due to some k-space broadening of the states. On the other hand, due to the larger width of the Fermi-Dirac edge (of order $\sim$\ 4$ kT $) we are now able to probe also states slightly above the Fermi level. This spectral information can be enhanced through a normalization of the individual MDC intensities, as shown in Fig.~\ref{Doping}(b). For this sample, the initial $E_{\rm D}$\ is resolved at $\sim$15\,meV above $E_{\rm F}$, marked by the black line in the left panel. The reduction of the intrinsic $p$-type doping is ascribed to a higher initial Te amount in the buffer layer. Two subsequent surface depositions of Te after the growth are shown in the middle and right panel of Fig.~\ref{Doping}(b). Again, we observe a clear shift of $E_{\rm F}$, and the total shift amounts to $\sim$50\,meV with $E_{\rm D}$=35\,meV for the highest Te content. We attribute this shift to an electron transfer from Te to Sn, providing evidence that the $n$-type doping reported in Ref.\ [\onlinecite{}] is not only caused by a suppressed $p$-type doping from In diffusion. Interestingly, the overall quality of the measurement appears slightly enhanced for the Te-doped sample. Especially the trivial surface state SS in the $\overline{\Gamma}$-$\overline{\rm M}$\ direction [Fig.~\ref{ARPES_standard}(a)] is more pronounced. This is in stark contrast to the effect of K doping, where the overall quality was diminished. On the one hand, one may argue that Te forms an ordered overlayer that may smooth out the surface roughness of as-grown $\alpha$-Sn, in agreement with the surfactant character of Te. Potassium, on the other hand, may be unordered or form clusters, and an increased surface roughness could explain the low signal-to-noise ratio. Figure~\ref{Doping}(d) shows CECs extracted at 200\,meV (top) binding energy in the pristine sample, and at 250\,meV (bottom), for the highly Te doped sample as marked by the white dashed lines in Fig.~\ref{Doping}(b). Hence, the energetic distance to $E_{\rm D}$\ is the same and allows for a comparison. As can be seen, the CECs are, despite some intensity differences, identical. In the right panel of Fig.~\ref{Doping}(d) we show MDCs extracted at $k_x$=0 (bottom) and $k_y$=0 (top) from the CECs to the left. Clearly, the peak maxima that stem from the TSS agree for both directions and doping levels. This underpins the argument that the surface reconstruction does not affect the TSS --- even though it is indeed strongly localized in the topmost layers \cite{}. This is in line with the findings by Ohtsubo \textit{et al.}\ who showed that $\alpha$-Sn\ films covered with a (2$\times$1) reconstructed Bi layer exhibit no (2$\times$1) surface periodicity in the electronic structure \cite{}. \textbf{Fe deposition.} Both Te and K overlayers have no magnetic moment and, therefore, the TRS of the system is preserved. As theory predicts, such perturbations to a TI are neither able to change the metallic behavior of the surface nor to destroy the TSS \cite{}, as long as the TSS bridges the gap between two different bulk bands. In contrast, the presence of a magnetic field in a topologically non-trivial system has been proposed to break the TRS and introduce a band gap in the TSS at $E_{\rm D}$. For two-dimensional surface states in 3D TIs many studies address this symmetry-breaking by deposition of ferromagnetic impurities or interfacing the TIs with a ferromagnetic overlayer \cite{}. Figure~\ref{Doping_Fe} presents the results of Fe deposition on the surface of an $\alpha$-Sn\ film. The left panel shows the dispersion along $\overline{\Gamma}$-$\overline{\rm X}$\ of an as-grown pristine $\alpha$-Sn\ sample, measured with photons of $h \nu$\,=\,18\,eV at $T$\,=\,8\,K. The TSS is clearly observed on top of the bulk background, as described in the previous section. The thickness of the Te buffer between substrate and $\alpha$-Sn\ film was adjusted such that the Dirac point is slightly, but clearly, below the Fermi level. In the right panel of Fig.~\ref{Doping_Fe}, we show the ARPES data on the same sample and measurement setup after deposition of 0.25\, monolayer (ML) Fe and subsequent annealing at $T$\,=\,373\,K. The bulk bands are not visible anymore due to the increased background signal from Fe \textit{d}-states. In contrast, the TSS appears very pronounced on top of this diffuse background signal. As outlined by the dashed line between the two panels of the figure, we again observe a clear shift of the Fermi level of the $n$-type sample. We quantify this shift by the position of the Dirac point prior and after Fe deposition to be $\sim$15\,meV. At $E_{\rm F}$, we clearly observe the two branches of the TSS at $\pm k$ that was not resolved in the pristine sample. Below $E_{\rm D}$, the TSS dispersion of the pristine and the Fe-covered sample agree well with each other, apart from the small energy shift. Most importantly, we find no sign of a band gap opening within our experimental resolution. \begin{figure} [ht] \includegraphics[width=\linewidth]{Fig5_Remastered} \caption{Study of Fe deposition. Pristine sample (left) and after deposition of $\sim$0.25 ML Fe (right).} \label{Doping_Fe} \end{figure} Interestingly, for a $E_{\rm D}$\ located close to the $E_{\rm F}$\, as in case of our pristine $\alpha$-Sn\ film, theory predicts that the TSS may mediate an alignment of the magnetic moments of the adatoms via a Ruderman-Kittel-Kasuya-Yosida (RKKY) interaction \cite{}. Typically, the RKKY interaction oscillates between the ferromagnetic and antiferromagnetic coupling as a function of 1/$k_{\rm F}$, where $k_{\rm F}$\ refers to the electronic state that mediates the coupling. Here, since $E_{\rm D}$\ was adjusted close to $E_{\rm F}$\ ($k_{\rm F}$$\sim$0.005\AA$^{-1}$), the period of the RKKY interaction oscillation $2\pi/$$k_{\rm F}$\ increases to 100\, nm, which is much above the average distance between Fe adatoms. This means that the coupling between Fe adatoms should be ferromagnetic. However, we neither resolve an energy gap at $E_{\rm D}$\ nor do we observe a change in the band dispersion in terms of band mass or group velocity. Apart from the trivial absence of ferromagnetic order, the gap size could be still too small to be resolved with our given experimental resolution. Another reason for the gapless TSS dispersion could be intragap states induced by magnetic dopants \cite{sessi2016dual} that hide the clear gap opening in ARPES. In line with previous experimental findings in Bi$_2$Se$_3$\ \cite{, } and Bi$_2$Te$_3$\ \cite{}, as well as theoretical considerations \cite{}, we conclude that a simple 0.25\,ML Fe deposition onto the surface of $\alpha$-Sn\ is not able to open a resolvable band gap in the TSS. \section{\label{SARPES}Spin-momentum locking} Figure~\ref{Spin-ARPES} presents an in-depth study of the electron spin polarization of a sample that was optimized in terms of intrinsic doping, i.e., it has the Dirac point just slightly below the Fermi level as shown in Fig.~\ref{Spin-ARPES}(a). We have performed Mott-polarimetry of MDCs at $E_{\rm B}$$\sim$280\,meV as marked by the yellow dashed line in Fig.~\ref{Spin-ARPES}(a) and at different azimuthal rotations as marked by the green dashed lines on sketches on the right side of Fig.~\ref{Spin-ARPES}(b). Provided a 4-fold rotational symmetry, the selected cuts are representative for the whole surface state. \begin{figure*} [!] \includegraphics[width=\linewidth]{Fig6_Remastered} \caption{Spin-resolved MDCs for varying azimuthal angles $\alpha$. (a) Spin-integrated ARPES map of an $\alpha$-Sn\ sample. The yellow dashed line shows the $E_{\rm B}$\ of spin-resolved MDCs in panels (b-d). (b-d) All three components (S$_x$, S$_y$, S$_z$) of the spin vector measured for three azimuthal rotations $\alpha$ = 0$^{\circ}$, 23$^{\circ}$ and 45$^{\circ}$, respectively. The circles represent CECs of TSS, while the red (blue) color corresponds to a positive (negative) sign of the S$_x$ component of the spin vector. } \label{Spin-ARPES} \end{figure*} In Ref.\ [\onlinecite{}] it was reported that a spin-vector was perpendicular to the momentum, as expected for an ideal Dirac cone. In Fig.\ \ref{Spin-ARPES}(b-d) we present all three vector components of the spin polarization (S$_x$, S$_y$, S$_z$) measured in experimental geometries shown on the right side of each panel respectively. Note that the label $k_y$ refers to the experimental geometry which does not necessarily coincide with the sample high-symmetry directions. Since MDCs in Figs.\ \ref{Spin-ARPES}(b-d) were acquired by varying the tilt angle, we always move along the $k_y$-direction in $k$-space. In the same manner, the reference frame of the spin-polarization agrees always with the reference frame of the sample. For all azimuthal rotations we get a similar result for all three components of the polarization. The polarization is most pronounced in the $x$-component and the reversal of the sign of polarization is clearly visible. $S_x$ is found to be positive for $-k_y$, and negative for $+k_y$, thus resulting in the opposite helicity as compared to TSS at $E_{\rm F}$\ in Bi$_2$Se$_3$ \cite{Landolt2014spin}. The $z$-component shows vanishing polarization in agreement with the theoretical expectation that even rotational crystal symmetries confine the spin to the surface plane. As a refinement of our previous result \cite{}, we observe a finite polarization in the $y$-component as well. The effect is small, but significantly above the experimental error, and can not be simply explained with a small misalignment of the sample, i.e., an offset in $k_x$, since for that case the sign of the polarization should not reverse across $k_y$=0. Therefore, our measurements qualitatively give a picture as shown in the sketches on the right side of Figs.\ \ref{Spin-ARPES}(b)-(d) where the arrows visualize the orientation of the spin (angular-momentum) vector. We see that the spin-momentum locking is not perpendicular, but is pointing inside of the constant energy contour with a finite component antiparallel to the momentum of the spin. Resolving whether this is a pure spin effect or influenced by the coupling to the orbital momentum calls for further investigations, ideally assisted by one-step-photoemission theory. We emphasize that despite the reduced $k$-resolution as compared to conventional ARPES, the spin-resolved MDCs are capable to distinguish states below the $k$-resolution as long as there is a difference in the polarization \cite{Dil2009spin}. In agreement with the results of the conventional ARPES [see Fig.~\ref{Doping}(d)], we do not observe a $\sqrt{2}$-factor between the MDC peak maxima at azimuth $\alpha$=45$^{\circ}$ [$\overline{\Gamma}$-$\overline{\rm M}$\, Fig.~\ref{Spin-ARPES}(d)] and $\alpha$=0$^{\circ}$ [$\overline{\Gamma}$-$\overline{\rm X}$\, Fig.~\ref{Spin-ARPES}(b)], giving further evidence that the constant energy shape of the TSS is close to circular. To conclude the discussion of the spin-polarization, we show in Fig.~\ref{Circ_Dich} conventional ARPES measurements acquired with circularly polarized light. CDAD of photoelectrons has recently been suggested as a measure of spin in strong topological insulators \cite{}. In the study by Ohtsubo \textit{et al.}\ \cite{}, and also in the study by Liu \textit{et al.}\ for HgTe \cite{}, this method has been applied and appears to confirm the measurements using Mott-polarimetry. As can be seen in our results presented in Fig.~\ref{Circ_Dich}, already a small difference in the excitation energy as compared to the results by Ohtsubo \textit{et al.}, i.e., 18\,eV (here) vs. 19\,eV [\onlinecite{}], leads to an almost perfect cancellation of the CDAD effect in $\alpha$-Sn. We note that this is yet another confirmation that CDAD in spin-orbit coupled systems is no reliable stand-alone tool to measure the spin polarization of electronic states \cite{}. \begin{figure} [!h] \includegraphics[width=\linewidth]{Fig7_Remastered} \caption{Analysis of the circular dichroism in $\alpha$-Sn\ at $h \nu$=18\,eV. Panels (a) and (b) show the intensity distribution measured with circular right- and left-polarized light, respectively. Panel (c) shows the normalized difference between (a) and (b). (d) Direct comparison of MDCs extracted at 30\,meV binding energy from (a) (red curve) and (b) (blue curve).} \label{Circ_Dich} \end{figure} \section{\label{ImSigma}Many-body interactions} Having confirmed the spin-momentum locking in the previous section, we now turn to a topic related to the spin-polarization of the surface state. On the surface of a three-dimensional TI, 180$^{\circ}$ electron backscattering is forbidden as in case of the edge states in a two-dimensional TI. In contrast to the latter, scattering under angles smaller than 180$^{\circ}$ is only suppressed in three dimensions with a scatter probability that is dominated by the overlap of initial and final state spin states in the scattering process. According to theory \cite{}, the probability of a hole introduced at a certain {\bf k}$_\parallel$\ being filled with an electron of distinct {\bf k}$_\parallel$\ continuously decreases between a scattering angle of 0$^{\circ}$ and $\pm$180$^{\circ}$ on a circular constant-energy contour. This holds for the case of a pure Rashba interaction, where the spin is locked perpendicular to the momentum. For the case presented here, the spin-dependent scattering probability should not be altered by the deviation from the perpendicular locking, as we observe only a moderate modification. A much more significant impact is expected from the peculiarity that the TSS in $\alpha$-Sn\ is not situated in a projected bulk band gap, but exists completely degenerate with the projected \GEP state. One would naively expect strong hybridization effects as it is typical of such surface resonances. In this case, a hole created in the TSS by the photoemission process is readily filled with electrons from the bulk band. Such \textit{interband} scattering would not violate any conservation laws since the bulk states are not expected to be spin-polarized. This would lead to a strong broadening of the spectral function of the TSS, amounting to very short lifetimes of the excited state. Surprisingly, and as evident from the already discussed data, this is not the case in $\alpha$-Sn. The TSS appears pronounced and sharp against the bulk background of the \GEP band. To quantify the subjective impression from the false-color plots, we analyze the peak width from MDCs through a fit of Lorentzian curves convolved with a fixed Gaussian width to account for the experimental resolution (0.005\,\AA$^{-1}$). The results are shown in Fig.~\ref{Many_Body}. The analysis was applied to the data from a sample with increased Te amount in the buffer layer, which resulted in a Dirac point $\sim$35\,meV below the Fermi level. The bulk background was modeled with two pairs of peaks for the binding energy range above $\sim$90\,meV and with one pair of peaks below. Another pair of peaks has been fitted to the TSS in the whole binding energy range. Within each pair the peaks have been restricted to be symmetric about zero momentum and to have the same Lorentzian width at $\pm k$. These constraints gave meaningful results over almost the complete binding energy range. Only close to $E_{\rm D}$, where the routine was unable to fit two peaks to the TSS, we manually fixed the peak positions to get conclusive results. An example of MDC fitting is shown at the bottom-left inset in Fig.~\ref{Many_Body}. \begin{figure} [!hb] \includegraphics[width=\linewidth]{Fig8_Remastered_v2} \caption{Analysis of MDCs of an $\alpha$-Sn\ sample. The straight white line is a linear fit to the peak positions (gray diamonds) of the TSS. The bottom-left inset shows an MDC at $E_{\rm B}$=0.31\,eV fitted with 2 pairs of background peaks (green) and 1 pair of TSS peaks (red). The bottom-right inset shows a to-scale comparison of the linear fit to the experimental peak position (black line) with our $GW$-calculations. The size of the symbol is a measure of the surface localization, while the color represents the spin polarization (adapted from Ref.~[\onlinecite{}]). The imaginary part of the electron self-energy (Im$\Sigma$) (white circles), extracted from the data, is shown in the top panel. } \label{Many_Body} \end{figure} In Fig.~\ref{Many_Body} the gray filled diamonds represent the peak positions of the TSS from the described routine and the black bars give the full width at half maximum (FWHM). For clarity reasons we show only an average over two neighboring points of the fit results. The raw data has been preprocessed such that the energy step agrees with the experimental resolution of $\sim$10\,meV. In addition, we applied a linear regression to the peak positions as represented by the straight white line in the figure. As can be seen, the line fits very well to the fitted peak positions, confirming the linear dispersion of the TSS. We extract a group velocity of (4.5 $\pm$ 0.5)\,eV\AA, i.e., (6.8 $\pm$ 0.8) $\times$ 10$^5$\,m/s, which compares well with the Fermi velocity reported by Ohtsubo \textit{et al.}\ of 7.3 $\times$ 10$^5$\,m/s \cite{}. In the inset (bottom-right of Fig.~\ref{Many_Body}) we compare the group velocity from our analysis (black lines) with the one from theoretical \textit{GW}-calculations reported in Ref.\ [\onlinecite{}], which are in excellent agreement. To provide some insight into the decay mechanisms of created photoholes, we calculate the imaginary part of the electron self-energy (Im$\Sigma$) by multiplying the half width at half maximum (HWHM) of the Lorentzian peaks with the experimental group velocity \cite{}. The result is shown in the top panel of Fig.~\ref{Many_Body}. At $E_{\rm F}$, we observe an offset of 80\,meV that is explained by the presence of impurities on the surface, which lead to an energy-independent inelastic scattering. We then observe an increase of Im$\Sigma$ \xspace to $\sim$120\,meV from $E_{\rm F}$\ to $\sim$50\,meV binding energy. A possible decay of photoholes in this range is through electron-hole pair creation or electron-phonon interaction. However, a characteristic quasi-particle kink is not clearly resolvable in the dispersion due to the overlap of the two broadened peaks from the TSS branches at $\pm k$. Between binding energies of 50\,meV and 100\,meV, Im$\Sigma$ \xspace stays constant before a slight reduction to Im$\Sigma$ \xspace$\approx$100\,meV is observed. We note that this reduction appears in the vicinity of $E_{\rm D}$, and a detailed analysis of this behavior would be speculative due to the intricacies in the modeling. In this range we have reduced the number of peaks to model the bulk background from four to two. However, a similar decrease was observed for graphene on SiC and explained by the limited number of decays with enough momentum transfer for the creation of electron-hole pairs \cite{}. At $\sim$150\,meV binding energy we identify a very small increase that would again fit to a decay by electron-hole pair creation. So far, we have not taken into account the presence of the projected bulk states for the decay of photoholes. As mentioned above, one could expect a huge impact on the decay from these states. However, if we compare the average lifetimes of photoholes created in the TSS of $\alpha$-Sn\ ($\sim$3.3\,fs at Im$\Sigma$ \xspace$\approx$100\,meV) to those observed in other systems, where the surface state exists in a projected bulk band gap, like Bi$_2$Se$_3$\ ($\sim$11-30\,fs) \cite{} or Bi$_2$Te$_3$\ ($\sim$8-16\,fs) \cite{}, the difference is not so significant. Even in two-dimensional graphene on SiC, the lifetime of $\sim$5\,fs \cite{} is comparable to $\alpha$-Sn. In a recent study, we were able to show that the TSS we investigated is derived mainly from $p_z$-orbitals, while the \GEP band, the projection of which is degenerate with the TSS, has mainly $p_x+p_y$ character \cite{}. One may, therefore, argue about `orbital protection' against interband scattering for the TSS in $\alpha$-Sn. Moreover, this difference in the orbital character avoids hybridization and leads to a strong surface localization of the TSS despite its surface resonance character. In a real-space picture, one may argue that the density of states (DOS) at the surface originates mainly from the TSS since the bulk bands show a strong decay into the vacuum. Hence, the TSS DOS alone is left as a phase space for scattering events, which might explain the surprisingly high lifetimes of the photoholes. \section{\label{Conclude}Conclusions} In conclusion, we have analyzed the electronic structure of the TSS in strained $\alpha$-Sn\ grown on InSb(001) by spin- and angle-resolved photoemission. We found that the TSS has an almost ideal Dirac cone shape with a circular Fermi contour. With the help of photon energy scans, we identified a second two-dimensional feature that appears only away from the $\overline{\Gamma}$-$\overline{\rm X}$\ direction in $k$-space. We have shown that Te, as well as K, act as electron donors if deposited on the surface of $\alpha$-Sn\, and they allow to shift the Dirac point by at least $\sim$50\,meV and $\sim$215\,meV, respectively. Most importantly, we have demonstrated that deposition of 0.25\,ML of ferromagnetic Fe induces a slight $n$-type doping, however, no resolvable gap in ARPES. Moreover, we analyzed the spin-momentum locking of TSS electrons with spin-resolved photoemission and showed that the spin is captured within the surface plane. It has a finite component that is antiparallel to the momentum. Although the TSS is clearly spin-polarized, we observed that the circular dichroism in the angular distribution vanishes. Finally, we have shown that the quasi-particle lifetimes of photoholes created in the TSS of $\alpha$-Sn\ compare very well with those of other TIs --- despite the peculiarity of being fully degenerate with the bulk states. We argue that the long lifetimes are due to an `orbital protection' (low hybridization) against interband scattering between the TSS and background of bulk bands. \section{\label{Aknowledge}Acknowledgments} We thank A. Fleszar for support with calculations and discussions on the manuscript. This work was supported by the Deutsche Forschungsgemeinschaft (DFG) under Grant SCHA1510/5, the SPP 1666 Priority Program "Topological Insulators", the DFG Collaborative Research Center SFB 1170 "ToCoTronics" in W\"urzburg. Diamond Light Source (Didcot, UK) is gratefully acknowledged for beamtime under proposals SI10289, SI10244, SI12892, and SI15285. LCM and LBD acknowledge financial support from EPSRC (UK), and LBD also from the Science and Technology Facilities Council (UK). \bibliographystyle{apsrev4-1.bst}

\section{Introduction} Over the past two decades, significant effort has been invested in the development of nano- and micro-particles that can propel themselves through an aqueous environment \MP{\cite{ebbens10,gompper15,bechinger16}}. These synthetic swimmers have myriad potential applications in, e.g., cell sorting and manipulation \cite{baraban12b,solovev12}, micromanufacturing \cite{goodrich17}, and the assembly of dynamic and programmable materials \cite{stenhammar16}. For these and other applications, a longstanding challenge is to endow the synthetic swimmers with a semblance of ``intelligence'', i.e., the ability to autonomously sense the local environment (e.g., ambient flow fields \cite{palacci15,uspal15b}, or their position with respect to confining surfaces \cite{das15,simmchen16}) and to respond according to their design, \MP{i.e., to exhibit some form of ``taxis.'' For example, recent studies have reported, experimentally and theoretically, on the gravitaxis of asymmetrically shaped colloids \cite{hagen14}, as well as on phototaxis of light-activated colloids in a spatially modulated light intensity landscape \cite{lozano16,geiseler17}.} Among artificial swimmers, catalytically active particles constitute a large and important class \cite{paxton06,moran17,aubret17}. These particles typically move by self-phoresis, i.e., by consuming molecular ``fuel'' -- available in the surrounding solution -- they generate gradients of one or more thermodynamic variables (e.g., chemical composition or electrical potential) in the solution. These gradients, in conjunction with the interaction of the particle surface with the molecules of the solution, drive flow of the solution past the particle, powering directed motion \cite{golestanian05,golestanian07}. For example, bimetallic Au/Pt rods consume hydrogen peroxide and release charged molecules into the solution, and hence move by self-electrophoresis \cite{paxton04}. Janus spheres with an inert core and a platinum cap, when exposed to hydrogen peroxide, move by self-diffusiophoresis \cite{howse07} or, as proposed recently, by a more complex mechanism involving self- electrophoretic contributions \cite{brown14,ebbens14,brown17}. In addition to spheres and rods, catalytically active particles with more complex shapes have been fabricated, e.g., dimers of two connected particles \cite{valadares10,palacci15}, ``matchsticks,'' \cite{morgan14} and microstars \cite{simmchen17}. Self-phoretic particles exhibit complex dynamics when moving in the vicinity of a confining boundary. As with other swimmers, e.g., those moving due to mechanical surface deformations, self-phoretic particles create flow fields that extend into the surrounding solution, are reflected from nearby boundaries, and couple back to the particles \cite{spagnolie12}. These hydrodynamic interactions can lead to bound states of motion in which the swimming particle always remains in the vicinity of the boundary \cite{ishimoto13}. Notably, an additional level of complexity arises for self-phoretic particles, because the thermodynamic gradients that drive particle motion will also be modified by confining boundaries. For instance, if the particle creates gradients of chemical composition, a solid boundary will present a no-flux boundary condition for the number density fields of the various molecular species diffusing in the solution. In a previous study, we have shown that a no-flux boundary condition, in combination with hydrodynamic interactions, can lead to novel bound states, such as motionless ``hovering'' of a catalytic Janus sphere above a planar solid boundary (or ``wall'') \cite{uspal15}. \WEU{All of these phenomena are enriched even further if the boundary itself is ``responsive,'' i.e., if thermodynamic gradients along the boundary translate into tangential surfaces stresses driving flow of the solution.} Indeed, in analogy with phoresis, ``chemi-osmotic'' surface flows can occur if gradients in chemical composition occur along the surface of a solid boundary \cite{anderson89,derjaguin:47}. For a catalytic particle, chemi-osmotic surface flows, \textit{locally} induced by the particle, will drive flow in the bulk solution, coupling back to the particle. Recently, the propulsion of catalytic ``surfers'' near a solid boundary was attributed to such an effect \cite{palacci13}. Significantly, chemi-osmotic flows depend on the material identity of the solid boundary. This naturally leads to the question of whether, by patterning the chemical composition of the responsive wall, one can guide the motion of a catalytically active particle. Recently, we investigated the interplay of chemi-osmosis and self-diffusiophoresis for a catalytically active Janus sphere above a patterned wall \cite{uspal16,popescu17}. By analytical and numerical calculations we demonstrated that a sphere that swims \textit{away} from its catalytic cap can ``dock'' at a chemical step between two materials. In contrast, a sphere that swims \textit{towards} its cap can align with and follow a chemical stripe. We identified the physical mechanisms driving these behaviors. One key finding is that a chemically patterned wall can create fluid \textit{vorticity} in the bulk solution, and that this wall-induced vorticity completely accounts for rotation of the spherical particle. \begin{figure}[t] \includegraphics[scale=0.8]{./Fig1.pdf} \caption{\label{} (a, b) Schematic views of a Janus sphere and a Janus spherocylinder, respectively, at a distance $h$ from a solid planar wall. Each particle produces solute molecules (green spheres) over the area of its catalytic cap (black). The effective interaction between the solute and the particle surface drives a self-diffusiophoretic flow (purple arrows) at the surface of the particle. Similarly, the solute drives chemi-osmotic surface flow at the wall (blue arrows). The orientation vector $\mathbf{\hat{d}}$ is defined to point from the catalytic pole of the particle to the inert pole. For repulsive interactions between the solute molecules and the wall (which corresponds to the direction of the flow shown here), the surface flow converges towards the region of maximum density of solute. This surface flow entrains flow in the bulk solution (not shown). Near the point of convergence, due to fluid incompressibility the bulk flow must lift off the surface into the volume of the solution, coupling back to the particle. (c) Top-down schematic view of a spherocylinder near a chemically patterned surface. In this case, the pattern is a chemical step, as defined by Eq. \ref{}. The orientation vector $\mathbf{\hat{d}}$ and the $\hat{x}$ axis form the angle $\phi$.} \end{figure} The orientational dynamics of a solid particle in a spatially varying ambient flow field is a classical problem in microhydrodynamics \cite{kim1991}. A natural starting point is to perform a Taylor expansion of the flow velocity with respect to the particle centroid (i.e., its geometrical center). At first order in such an expansion, the vorticity can be mathematically identified with the antisymmetric component of the first derivative of the flow velocity. The symmetric component of this tensor, the so-called \textit{rate of strain} tensor, does not contribute to the rotation of a sphere. However, it does contribute to rotation of a non-spherical or rod-like particle, such as a spheroid or spherocylinder. Consequently, the orientational dynamics of a rod-like particle in a spatially varying flow field is typically much more complex than for a sphere. For instance, a sphere in a linear shear flow will simply rotate around the axis of vorticity at a rate proportional to the vorticity, while any other solid body of revolution will execute a complex three-dimensional periodic motion known as a Jeffery orbit \cite{jeffery22,bretherton62}. These previous findings naturally raise the question of how a rod-like catalytically active particle will behave near a chemically patterned wall. The present study addresses this question. Via detailed numerical calculations, we show that a rod-like particle, as compared with a sphere, exhibits significantly richer dynamics near a patterned wall. We find a new ``edge-following'' steady state, in which a rod-like particle moves along the edge of a chemical step at a steady distance from the step and with a steady orientation. Moreover, within a certain range of system parameters, the edge-following state co-exists with a docking state, i.e., the dynamics is bistable. We develop approximate analytical expressions that isolate the contributions of fluid vorticity and the rate of strain to particle rotation. We show that the rich phenomenology of rod-like particles is indeed a consequence of the competition between fluid vorticity and rate of strain. \section{Formulation of the model} We consider a catalytic Janus particle in the vicinity of a hard planar wall. The particle is a solid body of revolution with axially symmetric surface chemistry. Two examples of particles of this kind -- a sphere and a spherocylinder -- are shown schematically in Fig. \ref{}. We assume a stationary reference frame in which the wall is located at $z = 0$ and the instantaneous position of the particle centroid is $\mathbf{x}_{p}(t) = (x_p(t), y_p(t), z_p(t))$; for brevity of notations, in the following we shall indicate the dependence on time only if it is explicitly required. We denote by $\mathbf{x}$ a point in the liquid surrounding the particle. Over half of its surface (i.e., the catalytic ``cap''), the particle consumes ``fuel'' molecules, taken to be present in abundance in the solution, and emits product molecules (``solute''). We additionally assume that the reaction accounting for the conversion of the fuel molecules into solute molecules is such that the system operates within the so-called regime of reaction limited kinetics (see, e.g., Ref. \cite{oshanin17} and references therein). Under these assumptions, the chemical activity of the particle can be modeled simply as a release of solute molecules at a constant rate $\kappa$ (areal number density per time) over half of its surface (the catalytic ``cap''), which corresponds to a so-called ``constant-flux'' model \cite{golestanian05}. The orientation of the particle is defined by a unit vector $\mathbf{\hat{d}}$, which is aligned with the axis of symmetry of the particle, and points from the catalytic cap to the inert ``face'' of the particle. The solute diffuses in the surrounding liquid with a diffusion constant $D$. We define the P\'{e}clet number $Pe \equiv U_{0} R_{1}/D$, where $U_0$ and $R_1$ are the characteristic velocity and the characteristic length scale of the particle, respectively. If the P\'{e}clet number is small, the solute number density field $c(\mathbf{x})$ is quasi-static, and can be taken to obey the Laplace equation $\nabla^{2} c = 0$. The wall and the inert face are taken to be impenetrable to solute, so that the number density field is subject to the boundary conditions $- D [\mathbf{\hat{n}} \cdot \nabla c] = \kappa$ on the catalytic cap, and $- D [\mathbf{\hat{n}} \cdot \nabla c] = 0$ on the inert face and on the wall, where the normal $\mathbf{\hat{n}}$ is defined to point into the liquid. Additionally, the number density field is required to attain a fixed value $c^{\infty}$ far away from the particle, i.e., $c(|\mathbf{x} - \mathbf{x}_{p}| \rightarrow \infty) = c^{\infty}$. The suspending solution, consisting of solvent, fuel, and solute molecules, is an incompressible Newtonian fluid with dynamic viscosity $\eta$ and mass density $\rho$. We use the classical theory of charge neutral diffusiophoresis in order to model how concentration gradients drive particle motion through the liquid \cite{anderson89}. Within this framework, the interaction between the solute molecules and a confining surface (i.e., the wall or one of the particle faces) drives a surface flow which is described by an effective slip velocity $\mathbf{v}_{s}(\mathbf{x}_{s}) = -b(\mathbf{x}_{s}) \nabla_{||} c(\mathbf{x})|_{\mathbf{x}=\mathbf{x}_{s}}$, where $\nabla_{||} \equiv (\mathcal{I} - \mathbf{\hat{n}} \mathbf{\hat{n}}) \cdot \nabla$, $\mathcal{I}$ is the identity tensor, and $\mathbf{x}_{s}$ denotes a position on that surface. The so-called surface mobility $b\mathbf(\mathbf{x}_{s})$ is a material-dependent parameter which encodes the molecular interaction between the solute and that surface. We consider small Reynolds numbers $Re \equiv \rho U_0 R_{1} / \eta$, so that the hydrodynamic flow $\mathbf{u}(\mathbf{x})$ and the pressure $P(\mathbf{x})$ of the solution obey the Stokes equation $-\nabla P + \eta \nabla^{2} \mathbf{u} = 0$ and the incompressibility condition $\nabla \cdot \mathbf{u} = 0$. The fluid velocity obeys the boundary conditions $\mathbf{u}\lvert_{wall}=\mathbf{v}_{s}(\mathbf{x}_{s})$ on the wall, and $\mathbf{u}\lvert_{part}=\mathbf{U}+\bm{\Omega}\times(\mathbf{x}_{s}-\mathbf{x}_{p})+ \mathbf{v}_{s}(\mathbf{x}_{s})$ on the particle surface; note that the dependence on $\mathbf{x}_{s}$ accounts also for the, in general, distinct surface mobilities. Additionally, the fluid is quiescent far away from the particle, so that $\mathbf{u}(| \mathbf{x} - \mathbf{x}_{p}| \rightarrow \infty) = 0$. Here, $\mathbf{U}$ and $\bm{\Omega}$ are unknown translational and angular velocities, respectively, of the particle. In order to close the system of equations for $\mathbf{U}$ and $\bm{\Omega}$, we invoke the force balance, $\int \, \bm{\sigma} \cdot \mathbf{\hat{n}} \, dS = -\mathbf{F}^{ext}$, and the torque balance, $\int \, (\mathbf{x} - \mathbf{x}_{p}) \times \bm{\sigma} \cdot \mathbf{\hat{n}} \, dS = -\bm{\tau}^{ext}$, for the particle. The fluid stress tensor is $\bm{\sigma} = -P \mathcal{I} + \eta \left[ \nabla \mathbf{u} + \nabla^{T} \mathbf{u} \right]$. Here, $\mathbf{F}^{ext}$ and $\bm{\tau}^{ext}$ are the net external force and net external torque, respectively; these can represent (for instance) particle buoyancy, bottom-heaviness, and any external forces and torques of constraint. The above two integrals are taken over the surface of the particle. The above considerations completely specify how to determine the instantaneous particle velocities $\mathbf{U}$ and $\bm{\Omega}$ as functions of the instantaneous position $\mathbf{x}_{p}$ and the orientation $\mathbf{\hat{d}}$ of the particle. This scheme includes various contributions to the particle velocities, such as from slip on the surface of the particle (self-diffusiophoresis); slip on the surface of the wall (chemi-osmosis), which drives hydrodynamic flow in the volume of the solution; and external forces and torques. At this point, we make some conceptual distinctions which will prove to be helpful later. We note that the Stokes equations are linear. Therefore, the full problem can be split into three subproblems which isolate the \MP{individual contributions of wall slip (\textbf{ws}), self-diffusiophoresis (\textbf{sd}), and external forces and torques (\textbf{ext}).} In subproblem \textbf{(sd)}, we take the surface velocity on the wall to be zero, i.e., $\mathbf{v}_{s} (\mathbf{x}_{s})\lvert_{wall} \,= 0$, as well as $\mathbf{F}^{ext} = 0$ and $\bm{\tau}^{ext} = 0$, but otherwise the subproblem is posed as given above. This subproblem can be solved for $\mathbf{U}^{sd}$ and $\bm{\Omega}^{sd}$. Similarly, in subproblem \textbf{(ws)}, we take $\mathbf{v}_{s}(\mathbf{x}_{s})\lvert_{part} \,= 0$, $\mathbf{F}^{ext} = 0$ and $\bm{\tau}^{ext} = 0$. We obtain $\mathbf{U}^{ws}$ and $\bm{\Omega}^{ws}$ by solving subproblem \textbf{(ws)}. Finally, if there are external forces and torques, we consider a subproblem \textbf{(ext)}, in which $\mathbf{v}_{s}(\mathbf{x}_{s})\lvert_{wall} \,= 0$ and $\mathbf{v}_{s}(\mathbf{x}_{s})\lvert_{part} \,= 0$, and which can be solved for $\mathbf{U}^{ext}$ and $\bm{\Omega}^{ext}$. We superpose the individual solutions to obtain the total particle velocities $\mathbf{U} = \mathbf{U}^{sd} + \mathbf{U}^{ws} + \mathbf{U}^{ext}$ and $\bm{\Omega} = \bm{\Omega}^{sd} + \bm{\Omega}^{ws} + \bm{\Omega}^{ext}$. For a sphere, the self-diffusiophoretic contributions $\mathbf{U}^{sd}$ and $\bm{\Omega}^{sd}$ had been considered in Ref. \cite{uspal15}. The very same framework developed in that study can be applied to spherocylinders. Additionally, the contributions $\mathbf{U}^{ext}$ and $\bm{\Omega}^{ext}$ of the external forces and torques can be solved with standard methods (e.g., see Ref. \cite{simmchen16}.) In the remainder of the present study, we focus on obtaining the contributions from chemi-osmotic wall slip. At this stage we make some simplifying assumptions. We take the particle height $z_p(t)$ to be a fixed quantity $z_p(t) = h$, and assume that $\mathbf{\hat{d}}$ is always oriented within the plane of the wall, i.e., $\mathbf{\hat{d}} \cdot \hat{z} = 0$, where $\hat{z}$ is the unit vector in $z$-direction. In this case the orientation is completely described by the angle $\phi$, as shown in Fig. \ref{. For a spherocylindrical particle, these assumptions can be viewed as realistic. Catalytic particles are typically heavy, and they will sediment to the bottom surface of a bounding container. Concerning the restriction on $\mathbf{\hat{d}}$, if $h$ is small (compared to the pole-to-pole length of the particle), ``rocking'' of the particle orientation $\mathbf{\hat{d}}$ out of the $xy$ plane is restricted to small angles by the presence of the hard wall; this is apparent from inspecting Fig. \ref{. For spheres, the restriction on $\mathbf{\hat{d}}$ could be imposed by an external magnetic torque. \MP{Furthermore, we note that for heavy colloids the surface-to-surface distance is expected to be small (when compared to the radius of the particle): it is set mainly by the balance between the weight of the particle and the DLVO repulsion. If the vertical component of the osmotic flow is not very large, which is expected to be the case for our system, it is a reasonable assumption that the change in height due to the particle being dragged by the vertical component of the osmotic flow at the location of the particle is small compared to $h$, i.e., $h/R$ remains practically unchanged.} (For a further discussion of this assumption in relation to spheres, the reader is invited to peruse Ref. \cite{uspal16}.) We note that based on these assumptions in Ref. \cite{uspal15} it was shown via symmetry considerations that $\mathbf{U}^{sd}$ and $\bm{\Omega}^{sd}$ are functions of $h$ and $\mathbf{\hat{d}} \cdot \hat{z}$ only, and that ${\Omega}^{sd}_{z} = 0$. Since we take $h$ to be constant and $\mathbf{\hat{d}} \cdot \hat{z} = 0$, one finds $\mathbf{U}^{sd} = U^{sd} \, \mathbf{\hat{d}}$ and ${\Omega}^{sd}_{z} = 0$. In the present study, we treat $U^{sd}$ as an input parameter; however, for a given particle geometry and surface chemistry, it can be calculated via the methods outlined in Ref. \cite{uspal15}. The position and orientation of the particle evolve according to the equations \begin{equation} \label{} \begin{split} \dot{x}_{p} &= U^{sd} \cos(\phi) + U_{x}^{ws}(x_p, y_p, \phi), \\ \dot{y}_{p} &= U^{sd} \sin(\phi) + U_{y}^{ws}(x_p, y_p, \phi),~\mathrm{and}\\ \dot{\phi} &= \Omega^{ws}_{z}(x_p, y_p, \phi), \end{split} \end{equation} where the over-dot denotes the time derivative of the corresponding observable. These equations can be integrated to obtain complete particle trajectories \footnote{Note that we have assumed that external forces and torques do not contribute to the in-plane motion of the particle. In some cases, this assumption represents an additional approximation. For instance, for a sphere subject to a magnetic torque of constraint, there is a contribution to the in-plane velocity of the particle which is due to wall-induced translational-rotational coupling. As discussed at length in Ref. \cite{popescu17}, this contribution is negligible for the particle-wall separations considered here.}. We consider two types of surface patterns: a ``chemical step'' and a ``chemical stripe.'' For the chemical step, the surface mobility $b(\mathbf{x}_{s})$ is \begin{equation} \label{} b(\mathbf{x}_{s}) = \begin{cases} b_{w}^{l},~ x_s < 0,\\ b_{w}^{r},~ x_s \geq 0,\hspace*{1cm} \end{cases} \end{equation} where $b_{w}^{l}$ and $b_{w}^{r}$ hold on the \textit{l}eft and \textit{r}ight hand sides of the \textit{w}all, respectively. The chemical stripe is defined by \begin{equation} \label{} b(\mathbf{x}_{s}) = \begin{cases} b_{w},~ x_s < -W,\\ b_{w}^{c},~ -W \leq x_s \leq W, \\ b_{w},~ x_s > W, \hspace*{1cm} \end{cases} \end{equation} i.e., the \textit{c}entral stripe has $b=b_{w}^{c}$, while $b=b_{w}$ on the rest of the wall. Both the chemical step and the chemical stripe have translational symmetry in the $\hat{y}$-direction. Therefore, the dependence of $U_{x}^{ws}$, $U_{y}^{ws}$, and $\Omega_{z}^{ws}$ on ${y}_p$ drops out of Eq. (\ref{}), and the dynamics is completely described by two degrees of freedom, ${x}_{p}$ and $\phi$. We express lengths in units of $R_{1}$, where $2 R_{1}$ is the largest dimension of the particle. The solute density field has a characteristic number density $\kappa R_{1}/D$. Combining this with the surface mobility gives a characteristic velocity $U_0 = |b_0 \kappa/D|$, where $b_0$ is a characteristic value of the surface mobility. For the chemical step, we choose $b_0 = b_w^{l}$, and for the stripe, we choose $b_0 = b_w^{c}$. \WEU{We express $U^{sd}$ in units of $U_0$ so that, for a given particle geometry and height, the dimensionless parameter $U^{sd}/U_{0}$ characterizes the relative strengths of the self-diffusiophoretic and the chemi-osmotic contributions to particle motility.} For reasons of space, in this study we restrict our considerations to $U^{sd} > 0$, i.e., to a particle which tends to move away from its catalytic cap via self-diffusiophoresis. This ``inert-forward'' direction of motion is more typically seen in experiments than the ``catalyst-forward'' direction of motion. (Nevertheless, a catalyst-forward particle can also exhibit an interesting behavior near a chemically patterned surface, as we demonstrated for spheres in Ref. \cite{uspal16}.) An additional dimensionless parameter characterizes the surface: $b_{w}^{r}/b_{w}^{l}$ for the chemical step and $b_{w}/b_{w}^{c}$ for the stripe, respectively.\newline \section{Numerical results and discussion} In order to numerically solve the problem posed in the previous section, we resort to the so-called Boundary Element Method (BEM) \cite{uspal15}. The effect of a non-spherical shape is revealed by considering two particle geometries: a Janus sphere and a spherocylinder, each of them half covered by catalyst. These geometries are shown schematically in Fig. \ref{ and (b). The sphere has a radius $R$, and the height $h$ of the sphere is fixed as $h/R = 1.1$. The major axis of the spherocylinder has a tip-to-tip length $2c$, and the radius of the cylindrical part of the particle is chosen to be $a = 0.4 c$. The height of the spherocylinder is fixed as $h = 0.6 c$. \begin{figure}[!htb] \includegraphics[scale=0.45]{./Fig2.pdf} \caption{\label{} (a-b) Trajectories of a self-diffusiophoretic sphere, half-covered by catalyst, at height $h/R = 1.1$. For (a) and (b), the self-propulsion velocity is $U^{sd}/U_{0} = 0.1$ and $U^{sd}/U_{0} = 0.4$, respectively. Star symbols match the trajectories to the initial conditions shown in the phase portraits in (c) and (d), respectively. (c-d) Phase portraits for a half-covered sphere at a height $h/R = 1.1$ for two values of the self-diffusiophoretic velocity $U^{sd}$. The white circle indicates an attractor, whereas the white triangle indicates a saddle point. The color field encodes the component $U_{x}$ of the particle velocity in $\hat{x}$ direction.} \end{figure} The motion of a sphere near a chemical step had previously been studied in Ref. \cite{uspal16}. We briefly revisit those findings here, choosing $b_{w}^{r}/b_{w}^{l} = 4$ and $b_{w}^{l} < 0$. Therefore, both regions of the substrate have a repulsive effective interaction with the solute molecules, but the region to the right of the step ($x_p > 0$) has a stronger repulsive interaction. At low self-diffusiophoretic velocity $U^{sd}$, and depending on the initial orientation $\phi$ and distance $x_{p}$ from the step, the sphere can spontaneously ``dock'' at the step. In the docking state, the particle is trapped and motionless near the step (i.e., $|x_p/R| \ll 1$), with the orientation vector $\mathbf{\hat{d}}$ pointing towards the region where the substrate is more repulsive to the solute. A particle trajectory with $U^{sd}/U_0 = 0.1$, which is attracted to the docking state, is illustrated in Fig. \ref{. The initial condition of this trajectory is matched by the magenta star symbol to Fig. \ref{, which is a phase portrait showing the time evolution of any initial state $(x_{p}, \phi)$. From the phase portrait, it is clear that if the particle approaches the edge from the left ($x_p < 0$), the particle will dock for nearly any angle of incidence $0 \leq \phi < \pi/2$. If the particle approaches from the right ($x_p > 0$), the angle of incidence must be oblique ($\phi \gtrsim \pi/2$) for the particle to dock; otherwise, the particle will pass over the step. At higher speeds $U^{sd}$, the particle always passes over the step. Figure \ref{ shows the trajectory of a particle with $U^{sd}/U_0 = 0.4$; the corresponding phase portrait is shown in Fig. \ref{. Interestingly, if the particle passes from the left $(x_p < 0$, $\phi < \pi/2$), there is a narrow range of exit angles $\phi \ll 1$ for a broad range of angles of approach $0 \leq \phi < \pi/2$. In other words, in this case the chemical step operates as an ``angular focusing'' device. If the particle passes from the right ($x_p > 0$, $\phi > \pi/2$), the deviation of the exit angle from the initial angle is modest. \begin{figure}[!htb] \includegraphics[scale=0.5]{./Fig3.pdf} \caption{\label{} Trajectories of a half-covered spherocylinder with height $h/c = 0.6$ near a chemical step. For (a), the self-propulsion velocity is $U^{sd}/U_{0} = 0.12$. For (b)-(c), the speed is $U^{sd}/U_{0} = 0.15$, and for (d) it is $U^{sd}/U_{0} = 0.35$. The star symbols match the trajectories to the initial conditions in Fig. \ref{}.} \end{figure} Now we consider the spherocylindrical particle near the same chemical step (Fig. \ref{}). We find that the phenomenology is considerably enriched by the elongated shape. At low speed $U^{sd}$, there is a stable ``edge-following'' state in which the particle remains at a fixed distance $|x_p/c| \ll 1$ and angle $\phi \lesssim \pi/2$. Accordingly, the particle orientation $\mathbf{\hat{d}}$ aligns almost parallel to the step edge, but tilted slightly towards the more repulsive region of the substrate, and the particle moves steadily along the edge. A phase portrait for $U^{sd}/U_0 = 0.12$ is shown in Fig. \ref{; the white circle represents the edge-following attractor. The cyan star in the phase portrait gives the initial condition for the trajectory in Fig. \ref{. At a slightly higher speed $U^{sd}/U_0 = 0.15$, the edge-following state (Fig. \ref{) co-exists with a stable docking state (Fig. \ref{), i.e., the dynamics is bistable. For the speed $U^{sd}/U_0 = 0.15$ edge-following and docking trajectories are illustrated in Figs. \ref{ and (c); the complete phase portrait corresponding to $U^{sd}/U_0 = 0.15$ is shown in Fig. \ref{. The bistability occurs within a finite range of speed values $U^{sd}$. As $U^{sd}$ is increased within this range, the attractors and the saddle points of the phase diagram move in the phase space $(\phi, x_{p}/c)$. Within this process, at the upper threshold of this range, the saddle point and the edge-following state collide and annihilate, leaving only the docking state. The docking state persists for a certain range of high speeds $U^{sd}$. In Fig. \ref{, we show the phase portrait for $U^{sd}/U_0 = 0.25$. As $U^{sd}$ is increased even further, the docking state attractor also collides with a saddle point and annihilates. Therefore, for very high speed $U^{sd}$, the spherocylinder passes over the step for any inital angle with $\mathbf{\hat{d}}$ pointing towards the step. A corresponding phase portrait for $U^{sd}/U_{0} = 0.35$ is shown in Fig. \ref{. As in the case of the sphere, a spherocylinder passing from the left has a narrow band of exit angles, while a spherocylinder passing from the right shows only modest deviations of the exit angles $\phi$ from the incoming ones. Such a trajectory, which passes from the left, is shown in Fig. \ref{. \begin{figure}[!htb] \includegraphics[scale=0.5]{./Fig4.pdf} \caption{\label{ portraits of a half-covered spherocylinder near a chemical step for various self-diffusiophoretic velocities $U^{sd}$. The particle is at a height $h/c = 0.6$. The star symbols indicate the initial conditions corresponding to the trajectories shown in Fig. \ref{}. White circles indicate attractors, whereas white triangles indicate saddle points. The color field encodes the component $U_{x}$ of the particle velocity in $\hat{x}$ direction.} \end{figure} We note that we have obtained a very similar phenomenology for a prolate spheroid, i.e., edge-following at low speed $U^{sd}$, bistable edge-following and docking at intermediate speeds $U^{sd}$, etc. This agreement indicates that the enrichment of the phenomenology as compared to that for a sphere is a generic consequence of non-spherical shapes, rather than being specific for spherocylinders. We omit these calculations in order to avoid an overly repetitive presentation. Additionally, we note that for both spheres and spherocylinders the phenomenology we have discussed above is observed for a range of other values of the surface mobility contrast $(b_{w}^{r}/b_{w}^{l} - 1)$ which are of $\mathcal{O}(1)$. \begin{figure} \includegraphics[scale=0.9]{./Fig5.pdf} \caption{\label{} (a) Phase portrait for a half-covered spherocylinder near a chemical stripe. The particle is at height $h/c = 0.6$ and has a self-diffusiophoretic velocity $U^{sd}/U_0 = 0.1$. Star symbols indicate initial conditions corresponding to the trajectories shown in panels (b) and (c). White circles indicate attractors, whereas white triangles indicate saddle points. The color field in (a) encodes the component $U_x$ of the particle velocity in $\hat{x}$ direction. Panels (b) and (c) show a docking and a stripe-following trajectory, respectively, of a spherocylinder near a chemical stripe. } \end{figure} We briefly discuss the behavior of a rod-like particle near a chemical stripe. As one might expect, if the stripe width $2W$ is large, the dynamics near the stripe resembles the one near two isolated chemical steps. However, as the width is decreased and becomes comparable to the particle size, the effects of the two edges start to interfere with each other. We consider a stripe with $2W = 3 c$, $b_{w}/b_{w}^{c} = 3$, and $b_{w}^{c} < 0$. Both the sides and the center of the stripe have a repulsive effective interaction with the solute, but this interaction is stronger for the sides than for the center. In Fig. \ref{, we show the phase portrait for a spherocylinder near this stripe with $U^{sd}/U_{0} = 0.1$. The dynamics is multistable. There are two docking states, associated with the two edges. But instead of two edge-following states, we obtain a ``stripe-following'' attractor in which the particle is positioned at the center of the stripe ($x_p = 0$) and is perfectly aligned with it ($\phi = \pi/2$). Docking and stripe- following trajectories are shown in Fig. \ref{ and (c). For comparison, in previous studies we had obtained a stripe-following state for Janus spheres, but only for $U^{sd} < 0$, i.e., for a particle which tends to move towards its catalytic face. \WEU{We also briefly consider the effect of changing the particle aspect ratio $e = a/c$. In order to address this issue, we consider spherocylinders with various aspect ratios located near the same chemical step as in Figs. \ref{}. We fix the ratio $h/a$, i.e., the ratio of the height of the particle center to the radius of the spherical end-caps of the particle, to $h/a = 1.5$. By decreasing $e = a/c$ from $e = 1$, the shape of the particle is smoothly deformed from spherical ($e = 1$) to quasi-spherical ($e \lesssim 1$) to rod-like ($e < 1$), while maintaining a constant thickness of the gap between the particle and the wall (as measured by the smallest dimension of the particle.) For each aspect ratio, we study particle behavior, in terms of the presence or absence of steady states, as we increase the self-propulsion velocity $U^{sd}$ from $U^{sd} = 0$. Our results are shown in Fig. \ref{}. We find that even quasi-spherical particles ($e = 0.95$) have edge-following steady states, but these are restricted to a very small band near $U^{sd} = 0$. Decreasing the aspect ratio significantly increases the range of $U^{sd}$ over which edge-following states can be obtained. We also observe a region of bistability for $e \leq 0.9$, further extending the range of propulsion speeds for edge-following. For each aspect ratio considered, there are no steady states localized near the step for very large $U^{sd}$. However, decreasing the aspect ratio increases the threshold value of $U^{sd}$ for the disappearance of steady states.} \begin{figure} \includegraphics[scale=0.45]{./Fig6.pdf} \caption{\label{} \WEU{Phase map showing the presence or absence of steady states for spherocylinders of various aspect ratios $e = a/c$ and self-propulsion speeds $U^{sd}/U_{0}$. The ratio of the particle height to the radius of the spherical end-caps is fixed to $h/a = 1.5$. Blue squares indicate that there is a monostable edge-following steady state at the corresponding combination of parameters $e$ and $U^{sd}$. Green circles indicate that there is a monostable docking state, and purple diamonds indicate that there are bistable edge-following and docking states. Red crosses indicate that there are no steady states localized near the step.}} \end{figure} \MP{The results in Figs. \ref{} - \ref{} summarize the various phenomenology emerging from self-diffusiophoresis near chemically patterned walls. They clearly show that the shape of the particle is a relevant parameter for the dynamics.} Based on these observations, we seek to understand which new physical aspects are introduced by the non-spherical shape, as well as the mechanism by which they enrich the particle dynamics. We particularly focus on the rotation of the particle angle $\phi$. For a sphere, we note that a chemical step always drives rotation of the catalytic face towards the region of the substrate which is less repulsive to the solute. This fact can be clearly seen in the phase portraits of Fig. \ref{}, because all trajectories in phase space always exhibit $\dot{\phi} \leq 0$, i.e., they always move from right to left in the diagrams. This observation suggests that there is a single dominant physical effect which drives the rotation of a sphere. For the spherocylinders, instead, the existence of edge-following attractors with $\phi$ between $\phi = 0$ and $\phi = \pi$ suggests that there might be multiple competing contributions to the angular velocity of the particle, and that these contributions balance for a particle in the edge-following state. \WEU{In order to understand} the interplay of the physical effects introduced by a non-spherical shape, in the following section we take an approximate analytical approach. We seek to identify and isolate the various competing contributions to translational and rotational velocities of a non-spherical particle. \section{Analytical results and discussion} \subsection{Point-particle approach for the fluid velocity} In order to obtain analytical expressions for $\mathbf{U}^{ws}$ and $\bm{\Omega}^{ws}$, we neglect the finite size of the particle under consideration, and treat it as a point-like object located at $\mathbf{x}_{p}$. Within this approximation, the fluid domain can be regarded as an infinite half-space $\lbrace \mathbf{x}_{f} \in \mathbb{R}^3 | z_f > 0 \rbrace$ bounded by a planar wall; the hydrodynamic boundary condition on the surface of the particle is neglected. There is a slip velocity $\mathbf{v}_{s}(\mathbf{x}_{s})$ on the wall, where $\mathbf{x}_{s}$ denotes a position on the wall. The wall slip drives flow in the volume of the solution, and there are no external forces or torques applied to the system. We seek an expression for the fluid velocity $\mathbf{u}(\mathbf{x}_{f})$ at an \textit{arbitrary} point $\mathbf{x}_{f}$ in the fluid, i.e., $\mathbf{x}_{f}$ is not necessarily identical to $\mathbf{x}_{p}$. We use the Lorentz reciprocal theorem, which relates the fluid stresses $(\bm{\sigma}, \bm{\sigma}')$ and velocity fields $(\mathbf{u}, \mathbf{u}')$ of two solutions of the Stokes equations which share the same geometry \cite{lorentz}: \begin{equation} \label{} \int \mathbf{u} \cdot \bm{\sigma}' \cdot \mathbf{\hat{n}} \, dS = \int \mathbf{u}' \cdot \bm{\sigma} \cdot \mathbf{\hat{n}} \, dS \,, \end{equation} where the integrals are performed over the wall with the surface normal $\mathbf{n} = \hat{\mathbf{z}}$ pointing into the fluid. We take the ``unprimed'' problem to be the one under consideration: the velocity on the wall is $\mathbf{u}(\mathbf{x}_{s}) = \mathbf{v}_{s}(\mathbf{x}_{s})$, and there are no external forces or torques. We consider three ``primed'' or auxiliary subproblems. For each of them, we take the wall to be a no-slip surface (i.e., $\mathbf{u}'(\mathbf{x}_{s}) = 0$). Additionally, there is an external force exerted on the fluid at point $\mathbf{x}_{f}$. The force has a magnitude $F_{0}$ and is oriented in $j$ direction, where $j \in \{x, y, z\}$ designates the auxiliary subproblem. The force creates a stress $F_0 \, \bm{\sigma}'^{(j)}(\mathbf{x}_{s}; \, \mathbf{x}_{f})$ in the fluid. Reformulating Eq. (\ref{}) leads to \cite{uspal16} \begin{equation} \label{} {u}_{j}(\mathbf{x}_{f}) = - \int \mathbf{v}_{s} (\mathbf{x}_{s}) \cdot \bm{\sigma}'^{(j)}(\mathbf{x}_{s}; \, \mathbf{x}_{f}) \cdot \mathbf{\hat{n}} \, dS \,. \end{equation} For each $j$ the integration kernel $\bm{\sigma}'^{(j)}(\mathbf{x}_{s}; \, \mathbf{x}_{f}) \cdot \mathbf{\hat{n}}$ is \cite{uspal16}: \begin{equation} \begin{split} \label{} {\sigma}'^{(x)}_{iz}(\mathbf{x}_{s}; \, \mathbf{x}_{f}) &= - \frac{3 z_{f}}{2 \pi} \frac{r_i r_x}{r^{5}} \,,\\ {\sigma}'^{(y)}_{iz}(\mathbf{x}_{s}; \, \mathbf{x}_{f}) &= - \frac{3 z_{f}}{2 \pi} \frac{r_i r_y}{r^{5}} \,,~\mathrm{and}\\ {\sigma}'^{(z)}_{iz}(\mathbf{x}_{s}; \, \mathbf{x}_{f}) &= \frac{3 z_{f}^{2}}{2 \pi} \frac{r_i}{r^{5}} \,, \end{split} \end{equation} where $\mathbf{r} \equiv \mathbf{x}_{s} - \mathbf{x}_{f}$ and $r \equiv |\mathbf{r}|$. We note that although textbook statements of the reciprocal theorem are often restricted to Eq. (\ref{}), an expression equivalent to Eq. (\ref{}) was already given in the original paper by Lorentz \cite{lorentz}. Now we consider how the fluid velocity in Eq. (\ref{}) will affect the motion of the particle. We apply Fax\'{e}n's Laws for the velocity of a point-like tracer particle in an ambient flow field \cite{spagnolie12}: \begin{equation} \label{} \mathbf{U}^{ws} = \mathbf{u}(\mathbf{x}_{p}) \end{equation} and \begin{equation} \label{} \bm{\Omega}^{ws} = \frac{1}{2} \bm{\omega}(\mathbf{x}_{p}) + \Gamma \, \mathbf{\hat{d}} \times (\mathbf{E}(\mathbf{x}_{p}) \cdot \mathbf{\hat{d}}). \end{equation} The {vorticity} vector evaluated at the particle position is \begin{equation} \bm{\omega}(\mathbf{x}_{p}) = \nabla \times \mathbf{u}(\mathbf{x}_{f}) \big|_{\mathbf{x}_{f} = \mathbf{x}_{p}}, \end{equation} where the derivative is taken with respect to $\mathbf{x}_{f}$. The {strain rate tensor} evaluated at the particle position is \begin{equation} \mathbf{E}(\mathbf{x}_{p}) = \frac{1}{2} \left( \nabla \mathbf{u}(\mathbf{x}_{f}) + \nabla^{T} \mathbf{u}(\mathbf{x}_{f}) \right) \big|_{\mathbf{x}_{f} = \mathbf{x}_{p}}. \end{equation} In Eq. (\ref{}), the parameter $\Gamma$, with $0 \leq \Gamma \leq 1$, characterizes the elongation of the particle. It is zero for a sphere and close to unity for a long, slender object. For a prolate spheroid, there is an exact expression for $\Gamma$: \begin{equation} \label{} \Gamma_{prolate} = \frac{1 - e^2}{1 + e^2}\,, \end{equation} where the spheroid aspect ratio $e = a/c$ is the ratio of the semi-minor axis length $a$ and the semi-major axis length $c$. For spherocylinders, there is no exact expression for $\Gamma$, but theoretical \cite{bretherton62} and experimental \cite{trevelyan51,snook14} studies have established that Eq. (\ref{}) holds as an excellent approximation if $e$ is replaced by an empirically determined effective aspect ratio $e_{eff} \approx 1.25 e$: \begin{equation} \label{} \Gamma_{spherocyl} \approx \frac{1 - e_{eff}^2}{1 + e_{eff}^2}. \end{equation} In order to facilitate the application of Fax\'{e}n's Laws, by using Eq. (\ref{}) we have obtained \begin{equation} \label{} u_j(\mathbf{x}_{p}) = - \int \mathbf{v}_{s} (\mathbf{x}_{s}) \cdot \bm{\sigma}'^{(j)}(\mathbf{x}_{s}; \, \mathbf{x}_{s}) \cdot \mathbf{\hat{n}} \, dS \,, \end{equation} \begin{equation} \label{} \omega_i(\mathbf{x}_{p}) = - \int \left[ \epsilon_{ijk} \frac{\partial}{\partial x_{f,j}} (\mathbf{v}_{s}(\mathbf{x}_{s}) \cdot \bm{\sigma}'^{(k)}(\mathbf{x}_{s}; \mathbf{x}_{f}) \cdot \mathbf{\hat{n}} ) \right]_{\mathbf{x}_{f} = \mathbf{x}_{p}} \, dS \,, \end{equation} and \begin{equation} \label{} {E}_{ij}(\mathbf{x}_{p}) = - \frac{1}{2} \int \left[ \frac{\partial}{\partial {x}_{f,i}} (\mathbf{v}_{s}(\mathbf{x}_{s}) \cdot \bm{\sigma}'^{(j)}(\mathbf{x}_{s}; \mathbf{x}_{f}) \cdot \mathbf{\hat{n}} ) + \frac{\partial}{\partial {x}_{f,j}}(\mathbf{v}_{s}(\mathbf{x}_{s}) \cdot \bm{\sigma}'^{(i)} (\mathbf{x}_{s}; \mathbf{x}_{f}) \cdot \mathbf{\hat{n}} ) \right]_{\mathbf{x}_{f} = \mathbf{x}_{p}} \, dS, \end{equation} where $i$, $j$, and $k$ are elements of $\{x,y,z\}$, $x_{f,i}$ is the $i$ component of $\mathbf{x}_{f}$, and $\bm{\sigma}'^{(j)}(\mathbf{x}_{s}; \mathbf{x}_{f})$ is given by Eq. (\ref{}). We have now completely specified how to calculate the motion of a ``point-like'' particle advected by the flow in the volume of the solution. In turn, this flow is driven by the surface flow $\mathbf{v}_{s}(\mathbf{x}_{s})$ on a planar wall. At this point, for convenience, we introduce some new definitions and perform certain further reformulations. Concerning the rotational velocity, we define $\bm{\Omega}^{ws} \equiv \bm{\Omega}^{V} + \bm{\Omega}^{E}$, where we have distinguished the contributions from the vorticity (V) and the fluid strain rate (E), respectively. Given the assumed restriction of $\mathbf{\hat{d}}$ to the xy plane, we are interested only in the $\hat{z}$ component $\Omega_{z}^{ws} = \Omega_{z}^{V} + \Omega^{E}_{z}$. \MP{By expanding the vector and scalar products in the last term of Eq. (\ref{}), one arrives at} \begin{equation} \Omega^{E}_{z} = \Gamma \, \epsilon_{zjk} d_{j} E_{kl} d_{l}, \end{equation} where \begin{equation} \begin{split} \epsilon_{zjk} d_{j} E_{kl} d_{l} &= \epsilon_{zxy} d_{x} E_{yl} d_{l} + \epsilon_{zyx} d_{y} E_{xl} d_{l} \\ &= d_{x} E_{yl} d_{l} - d_{y} E_{xl} d_{l} \\ &= E_{yx} d_{x}^{2} + E_{yy} d_{x} d_{y} - E_{xx} d_{x} d_{y} - E_{xy} d_{y}^{2}. \end{split} \end{equation} Since $E_{xy} = E_{yx}$, one has \begin{equation} \Omega^{E}_{z} = \Gamma E_{xy} (d_{x}^{2} - d_{y}^{2}) + \Gamma (E_{yy} - E_{xx}) d_{x} d_{y}. \end{equation} It is convenient to introduce \begin{equation} \label{} \Omega^{E}_{z} \equiv \Omega^{E,{cross}}_{z} + \Omega^{E,{diag}}_{z}, \end{equation} where we have distinguished the contributions from the diagonal and the cross (i.e., off-diagonal) terms of the strain rate tensor. \subsection{Point-particle approach for the solute field} In order to be able to use Eqs. (\ref{})-(\ref{}), one needs analytical expressions for $\mathbf{v}_{s} = -b(\mathbf{x}_{s}) \nabla_{||} c(\mathbf{x})$. Here, we develop approximate, ``point-particle-like'' expressions for the solute number density field $c(\mathbf{x})$. The starting point is a multipole expansion for the solute number density field in \textit{unbounded solution}, i.e., with no confining boundaries. We consider the centroid of the particle (whether a sphere or a spherocylinder) to be located at the origin of a Cartesian coordinate system which is co-moving with the particle. The axis of symmetry of the particle is aligned with the $\hat{z}$ axis. The solute number density field is obtained, as defined in the beginning of Sect. II, by solving the Laplace equation subject to the boundary conditions noted there except for the one on the wall. The \textit{f}ree \textit{s}pace distribution $c^{fs}(\mathbf{x})$ of solute can be obtained numerically (e.g., by the BEM), or, for certain shapes of the particle, analytically. We seek to extract multipole expansion coefficients $\alpha_{i}$ from the result for $c^{fs}(\mathbf{x})$, where the coefficients $\alpha_{i}$ are determined by its expansion in terms of Legendre polynomials \cite{golestanian07}: \begin{equation} \label{} c^{fs}(r, \theta) = c^{\infty} + \frac{R_{1}}{D} \sum_{l=0}^{\infty} \frac{\alpha_{l}}{l + 1} \left(\frac{R_{1}}{r}\right)^{l+1} P_{l}(\cos(\theta)), \; r \geq R_{1}\,. \end{equation} This expansion, expressed in terms of spherical coordinates, is taken to be valid in the fluid solution outside a sphere of radius $R_{1}$, where $2 R_{1}$ is the largest dimension of the particle. For a \textit{spherical} particle of radius $R$, this expansion is valid over the entire exterior region, i.e., $R_{1} = R$. Furthermore, if the solute number density field around the sphere obeys the boundary condition $[\mathbf{\hat{n}} \cdot \nabla c^{fs}(\mathbf{r})]|_{R,\theta} = -\kappa/D$ over the lower hemisphere (i.e., the catalyst covered area) of the sphere, and zero elsewhere on the surface of the sphere, the multipole coefficients $\alpha_{i}$ are straightforwardly obtained by using the orthogonality of the Legendre polynomials: \begin{equation} \label{} \frac{\alpha_{l}}{\kappa} = \frac{2l + 1}{2} \int_{\pi/2}^{\pi} d\theta \sin(\theta) P_{l} (\cos(\theta))\,. \end{equation} For instance, the monopole and dipole terms are $\alpha_{0} = \kappa/2$ and $\alpha_{1} = -3 \kappa/4$, respectively. For other particle geometries, such as a spherocylinder, the expression for the coefficient $\alpha_{0}$ can be inferred from the following intuitive considerations. In the multipole expansion of Eq. \ref{}, the monopole term $\alpha_{0}$ encodes the net rate of solute production. For arbitrary geometries, we can write \begin{equation} \label{} \alpha_{0} = \frac{A_{catalyst}}{4 \pi R_{1}^{2}} \kappa, \end{equation} i.e., $\alpha_{0}$ is proportional to the ratio of catalyst area to the area of a sphere with diameter equal to $2 R_{1}$. The higher order coefficients require detailed calculations. In Appendix A, we discuss how these are obtained for an arbitrary shape of the particle from the numerically determined (e.g., via BEM) solute number density field $c^{fs}(\mathbf{x})$. Particularizing the general results to our specific spherocylinder geometry, we obtain (see Appendix A) $\alpha_0 = 0.2 \, \kappa$, $\alpha_1 = -\, 0.2262 \, \kappa$, etc. For a particle in the vicinity of the wall, the solute number density field can be approximated from the multipole expansion as follows. We restrict our consideration to the first two terms in the multipole expansion, and describe the particle as a point monopole plus a point dipole.\footnote{ \MP{We note that the strategy employed here is to keep the minimum number of terms in the multipole expansion based on which the approximate dynamics is in qualitative agreement with the numerical solution of the complete (exact) dynamics. In the sense of a perturbative approach, the approximation based on just the first two terms in the multipole expansion, combined with a point-particle solution of the hydrodynamics, is mathematically not justified when $h/R$ is not much larger than 1, as studied here. Thus, the range of validity is determined ``a posteriori'', similar to many other cases in which such approximations turn out to be accurate well beyond the mathematically rigorous limits of validity (see, e.g., Ref. \cite{spagnolie12}). For our system the approximate analysis seems to remain valid (at least quasi-quantitatively) down to very small values of $h/R-1$.} } The direction of the dipole (defined to be from the point source to the point sink) is $\mathbf{\hat{d}}$. We place an image monopole and an image dipole below the wall at the position $(x_p, y_p, -h)$. The image monopole has the same strength $\alpha_0$ as the actual monopole. Likewise, the image dipole has the same strength $\alpha_1$ and orientation $\mathbf{\hat{d}}$ as the actual dipole (recalling that the dipole is parallel to the wall surface, since we assumed $\mathbf{\hat{d}} \cdot \hat{z} = 0$). Accordingly, one has \begin{equation} c(\mathbf{x}_{s}) \approx c^{mp}(\mathbf{x}_{s}) + c^{dp}(\mathbf{x}_{s}), \end{equation} where $c^{mp}(\mathbf{x}_{s})$ is the solute number density field due to the monopole at the position $\mathbf{x}_{s}$ on the wall, plus its image, and $c^{dp}(\mathbf{x}_{s})$ is the field due to the dipole at the position $\mathbf{x}_{s}$ on the wall, plus its image. For the monopole term, we obtain \begin{equation} \label{} c^{mp}(\mathbf{x}_{s}) = \frac{2 \alpha_{0} R_{1}^{2}}{D} \frac{1}{|\mathbf{x}_{s} - \mathbf{x}_{p}|}\,, \end{equation} and for the dipole term, \begin{equation} \label{} c^{dp}(\mathbf{x}_{s}) = - \frac{|\alpha_{1}| R_{1}^{3}}{D} \frac{\mathbf{\hat{d}} \cdot \mathbf{r}_{s}}{|\mathbf{x}_{s} - \mathbf{x}_{p}|^{3}}, \end{equation} where $\mathbf{r}_{s} \equiv (x_s - x_p, y_s - y_p, 0)$. Consequently, the slip velocity $\mathbf{v}_{s}(\mathbf{x}_{s}) = - b(\mathbf{x}_{s}) \nabla_{||} c(\mathbf{x})|_{\mathbf{x}=\mathbf{x}_{s}}$ on the wall is decomposed as $\mathbf{v}_{s}(\mathbf{x}_{s}) \approx \mathbf{v}_{s}^{mp}(\mathbf{x}_{s}) + \mathbf{v}_{s}^{dp}(\mathbf{x}_{s})$. The contribution of the monopole to the wall slip velocity is \begin{equation} \label{} \mathbf{v}_{s}^{mp} = \frac{2 \alpha_{0} R_{1}^{2}}{D} \frac{b_{w}(\mathbf{x}_{s}) \, \mathbf{r}_{s}}{|\mathbf{x}_{s} - \mathbf{x}_{p}|^{3}}\,, \end{equation} and the contribution of the dipole is \begin{equation} \label{} \mathbf{v}_{s}^{dp} = \frac{b_w(\mathbf{x}_{s}) |\alpha_{1}| R_{1}^{3}}{D |\mathbf{x}_{s} - \mathbf{x}_{p}|^{3}} \,\; \mathbf{\hat{d}} \cdot \left( \mathcal{I} - \frac{3 \mathbf{r}_{s} \mathbf{r}_{s}}{|\mathbf{x}_{s} - \mathbf{x}_{p}|^{2}} \right) . \end{equation} \subsection{Monopole contribution to the particle velocity} The combination of Eqs. (\ref{}), (\ref{}), and (\ref{}) yields the monopole contributions to the fluid velocity at an arbitrary point $\mathbf{x}_{f}$ above a chemically patterned wall: \begin{equation} {u}_{x}^{mp}(\mathbf{x}_{f}) = \frac{3 h}{2 \pi} \frac{2 \alpha_{0} R^{2}}{D} \int \frac{b_{w}(\mathbf{x}_{s})(x_{s}-x_{f})[(x_{s}-x_{f})(x_{s}-x_{p}) + (y_{s} - y_{f})(y_{s}-y_{p})]}{((x_{f}-x_{s})^2 + (y_{f} - y_{s})^{2} + z_{f}^{2})^{5/2} ((x_{p} - x_{s})^{2} + (y_{p} - y_{s})^{2} + h^{2}))^{3/2}} \, dS \end{equation} and \begin{equation} {u}_{y}^{mp}(\mathbf{x}_{f}) = \frac{3 h}{2 \pi} \frac{2 \alpha_{0} R^{2}}{D} \int \frac{b_{w}(\mathbf{x}_{s})(y_{s}-y_{f})[(x_{s}-x_{f})(x_{s}-x_{p}) + (y_{s} - y_{f})(y_{s}-y_{p})]}{((x_{f}-x_{s})^2 + (y_{f} - y_{s})^{2} + z_{f}^{2})^{5/2} ((x_{p} - x_{s})^{2} + (y_{p} - y_{s})^{2} + h^{2}))^{3/2}} \, dS. \end{equation} Now we consider the derivatives of the fluid velocity. In order to obtain $E_{xy}^{mp}(\mathbf{x}_{p})$, we calculate $\frac{1}{2} \left( {\partial u_{x}^{mp}}/{\partial y_{f}} + {\partial u_{y}^{mp}}/{\partial x_{f}} \right)\big|_{\mathbf{x}_{f} = \mathbf{x}_{p}}$. The two partial derivatives are identical, yielding \begin{equation} E_{xy}^{mp}(\mathbf{x}_{p}) = \frac{3 h}{2 \pi} \frac{2 \alpha_{0} R^{2}}{D} \int \frac{b_{w}(\mathbf{x}_{s}) (x_{p} - x_{s})(y_{p} - y_{s}) [4 \, ((x_{p} - x_{s})^2 + (y_{p} - y_{s})^2) - h^{2}]}{((x_{p} - x_{s})^{2} + (y_{p} - y_{s})^2 + h^{2})^{5}} \, dS. \end{equation} Analogously one obtains $E_{xx}^{mp}(\mathbf{x}_{p})$ and $E_{yy}^{mp}(\mathbf{x}_{p})$: \begin{equation} \begin{split} & E_{xx}^{mp}(\mathbf{x}_{p}) = \frac{3 h}{2 \pi} \frac{2 \alpha_{0} R^{2}}{D} \\ & \, \, \times \int \frac{b_{w}(\mathbf{x}_{s}) [3 (x_p-x_s)^4 + 2(x_p-x_s)^2(y_p-y_s)^2 - (y_p-y_s)^4-(2(x_p-x_s)^2+(y_p-y_s)^2)h^2]}{((x_p - x_s)^2 + (y_p - y_s)^2 + h^2)^5} \, dS \end{split} \end{equation} and \begin{equation} \begin{split} & E_{yy}^{mp}(\mathbf{x}_{p}) = -\frac{3 h}{2 \pi} \frac{2 \alpha_{0} R^{2}}{D} \\ & \;\; \times \int \frac{b_{w}(\mathbf{x}_{s}) [((x_p-x_s)^2 - 3(y_p-y_s)^2)((x_p-x_s)^2+(y_p-y_s)^2)+((x_p-x_s)^2+2(y_p-y_s)^2)h^2]}{((x_{p} - x_{s})^{2} + (y_{p} - y_{s})^2 + h^{2})^{5}} \, dS\,. \end{split} \end{equation} The monopolar contribution to the vorticity component $\omega_{z}$ vanishes, because the two partial derivatives in $\omega_z = \left( {\partial u_{x}^{mp}}/{\partial y_{f}} - {\partial u_{y}^{mp}}/{\partial x_{f}} \right)\big|_{\mathbf{x}_{f} = \mathbf{x}_{p}}$ are identical so that \begin{equation} \omega^{mp}_{z} = 0. \end{equation} If the pattern exhibits translational symmetry in the $\hat{\mathbf{y}}$ direction, such that $b(\mathbf{x}_{s}) = b(x_{s})$, the particle velocity does not depend on $y_p$. Accordingly, in calculating the velocity, one can set $y_p = 0$ without loss of generality. Concerning $E_{xy}^{mp}({x}_{p})$, the resultant expression involves an integral over an odd function of $y_s$, and hence it is zero: \begin{equation} \label{} E_{xy}^{mp}({x}_{p}) = - \frac{3 h}{2 \pi} \frac{2 \alpha_{0} R^{2}}{D} \int dx_{s} \int dy_{s} \, \frac{b_{w}({x}_{s}) (x_{p} - x_{s}) y_{s} [4 \, ((x_{p} - x_{s})^2 + y_{s}^2) - h^{2}]}{((x_{p} - x_{s})^{2} + y_{s}^2 + h^{2})^{5}} = 0. \end{equation} Now we consider the specific case of a chemical step by substituting Eq. (\ref{}) for $b_{w}(\mathbf{x}_{s})$. According to Eq. (\ref{}), $E_{xy}^{mp}$ vanishes. This leaves four piecewise integrals to be performed: the contributions of the left ($x_s < 0$) and of the right ($x_s \,>\, 0$) side to $E_{xx}^{mp}$ and $E_{yy}^{mp}$. After straightforward but cumbersome analysis, we obtain the contribution of the monopole to particle rotation: \begin{equation} \label{} \Omega_{z}^{mp} = -\frac{3 \alpha_{0} \Gamma R_{1}^{2}}{128 D} \, (b_{w}^{l} - b_{w}^{r}) \, \frac{h (32 x_{p}^{3} + 7 x_{p} h^{2})}{(x_{p}^{2} + h^{2})^{7/2}} \sin(\phi) \cos(\phi)\,. \end{equation} \subsection{Dipole contribution to the particle velocity} By combining Eqs. (\ref{}), (\ref{}), and (\ref{}) one finds the dipolar contributions to the fluid velocity at $\mathbf{x}_{f}$: \begin{equation} \begin{split} \label{} u_{x}^{dp}(\mathbf{x}_{f}) = \frac{3 h}{2 \pi} \frac{|\alpha_{1}| R_{1}^{3}}{D} \left[\int \frac{ b_{w}(\mathbf{x}_{s})(x_s - x_f) ((x_s-x_f) d_x + (y_s - y_f) d_y)}{((x_f - x_s)^2 + (y_f - y_s)^2 + h^2)^{5/2} ((x_p - x_s)^2 + (y_p - y_s)^2 h^2)^{3/2}} \, dS \, + \, \right. \\ \left. - \int \frac{3 b_{w}(\mathbf{x}_{s}) (x_s-x_f)[(x_s-x_p) d_x + (y_s - y_p) d_y] [(x_s - x_f)(x_s - x_p) + (y_s - y_f)(y_s - y_p)]}{((x_f - x_s)^2 + (y_f - y_s)^2 + h^2)^{5/2} ((x_p - x_s)^2 + (y_p - y_s)^2 + h^2)^{5/2}} \, dS \right] \end{split} \end{equation} and \begin{equation} \begin{split} \label{} u_{y}^{dp}(\mathbf{x}_{f}) = \frac{3 h}{2 \pi} \frac{|\alpha_{1}| R_{1}^{3}}{D} \left[\int \frac{ b_{w}(\mathbf{x}_{s})(y_s - y_f) ((x_s-x_f) d_x + (y_s - y_f) d_y)}{((x_f - x_s)^2 + (y_f - y_s)^2 + h^2)^{5/2} ((x_p - x_s)^2 + (y_p - y_s)^2 h^2)^{3/2}} \, dS \, + \, \right. \\ \left. - \int \frac{3 b_{w}(\mathbf{x}_{s}) (y_s-y_f)[(x_s-x_p) d_x + (y_s - y_p) d_y] [(x_s - x_f)(x_s - x_p) + (y_s - y_f)(y_s - y_p)]}{((x_f - x_s)^2 + (y_f - y_s)^2 + h^2)^{5/2} ((x_p - x_s)^2 + (y_p - y_s)^2 + h^2)^{5/2}} \, dS \right], \end{split} \end{equation} where, in each equation, the two integrals correspond to the two terms in the parentheses in Eq. (\ref{}). As with the monopole contributions, one can construct surface integrals for the vorticity and the strain rate tensor associated with the dipole by taking partial derivatives of Eqs. (\ref{}) and (\ref{}). We omit these lengthy intermediate expressions, and provide only the final results, which are, somewhat surprisingly, rather compact expressions. The contribution of the dipole to rotation via vorticity (see the text below Eq. (\ref{})) is \begin{equation} \label{} \Omega_{z}^{V,dp} = - \frac{3 |\alpha_{1}| R_{1}^{3} }{64 D} \, (b_{w}^{l} - b_{w}^{r}) \, \frac{h}{(h^2 + x_p^{2})^{5/2}} \sin(\phi)\,. \end{equation} The dipolar contributions to rotation via the rate of strain (see Eq. (\ref{}) and the text below Eq. (\ref{})) are \begin{equation} \label{} \Omega_{z}^{E,{cross},dp}(\mathbf{x}_{p}) = \frac{3 |\alpha_{1}| R_{1}^{3} \Gamma }{512 D} (b_{w}^{l} - b_{w}^{r}) \, \frac{h (22 x_p^2 + 7 h^{2})}{(x_p^2 + h^2)^{7/2}} \sin(\phi) \, [\cos^{2}(\phi) - \sin^{2}(\phi)] \end{equation} and \begin{equation} \label{} \Omega_{z}^{E,{diag},dp}(\mathbf{x}_{p}) = - \frac{15 |\alpha_{1}|R_{1}^{3} \Gamma }{512 D} (b_{w}^{l} - b_{w}^{r}) \, \frac{h (20 x_p^4 + 2 x_{p}^{2} h^{2} + 3 h^{4})}{(x_p^2 + h^2)^{9/2}} \, \cos^{2}(\phi) \sin(\phi)\,. \end{equation} \begin{figure}[!htb] \includegraphics[scale=0.9]{./Fig7.pdf} \caption{\label{} (a) and (b) Angular velocity $\Omega_{z}^{ws}$ of a half-covered sphere due to chemi-osmotic slip at the wall as a function of orientation $\phi$ for several lateral distances $x_{p}$ from a chemical step. The step is characterized by $b_{w}^{r}/b_{w}^{l} = 4$ and $b_{w}^{l} < 0$ and is located at $x_{p} = 0$. In (a), the center of the sphere is at a height $h/R = 1.1$ above the surface. In (b), the height is $h/R = 3$. (c) and (d) Angular velocity of a half-covered spherocylinder as a function of orientation $\phi$ for several lateral distances from the same chemical step. In (c), the centroid of the spherocylinder is at height $h/c = 0.6$ above the surface, while in (d) it is at height $h/c = 3$, where $2 c$ is the tip-to-tip length of the spherocylinder. In (a) the theoretical and numerical data for $x_p/R = \pm 1$ \textit{de facto} coincide. In (b) the theoretical and numerical data for $x_p/R = \pm 1$ \textit{de facto} coincide, as do the theoretical and numerical data for $\pm 2$. } \end{figure} \subsection{Comparison of analytical and numerical results} Finally, by combining Eqs. (\ref{}), (\ref{}), (\ref{}), and (\ref{}) one obtains an approximation expression for the rotation of a particle in the vicinity of a patterned wall: \begin{equation} \label{} \Omega_{z}^{ws}(\mathbf{x}_{p}) \approx \Omega_{z}^{mp}(\mathbf{x}_{p}) + \Omega_{z}^{V,dp}(\mathbf{x}_{p}) + \Omega_{z}^{E,{cross},dp}(\mathbf{x}_{p}) + \Omega_{z}^{E,{diag},dp}(\mathbf{x}_{p}). \end{equation} We first consider the rotation of a sphere near a chemical step. For this shape of the particle, $\Omega_{z}^{mp}(\mathbf{x}_{p})$, $\Omega_{z}^{E,{cross},dp}(\mathbf{x}_{p})$, and $\Omega_{z}^{E,{diag},dp}(\mathbf{x}_{p})$ all vanish, because they originate from the strain rate tensor. The remaining term, $\Omega_{z}^{V,dp}(\mathbf{x}_{p})$, is a simple sinusoid (Eq. \ref{}), and hence crosses zero only at $\phi = 0$, $\phi = \pi$, and $\phi = 2 \pi$. Therefore, the only possible stable orientations of the sphere are docking states (i.e., pointing towards or away from the step). In Figs. \ref{ and (b) we compare this expression with \MP{the numerical solution of the exact dynamics, which is} obtained by BEM. In this figure, (a) corresponds to $h/R = 1.1$ and (b) to $h/R = 3$, and $\Omega_{z}^{ws}$ is plotted as a function of $\phi$ for various lateral distances $x_{p}/R$ from the step. We obtain excellent agreement for $h/R = 3$, for which we expect the truncation of the multipole expansion to the monopole and dipole terms to be a good approximation. Closer to the surface ($h/R = 1.1$), we obtain qualitative, but only semi-quantitative agreement. In particular, we note that while the numerical results show a stronger rotation by the surface than predicted by the approximate expressions, they retain the theoretically predicted (see above) sinusoidal dependence on $\phi$. For the spherocylinder, all terms in Eq. (\ref{}) contribute. The resultant expression for $\Omega_{z}^{ws}$ is complicated, and can cross $\Omega_{z}^{ws} = 0$ between $\phi = 0$ and $\phi = \pi$. This more complicated dependence enables the occurrence of the edge-following state. In Figs. \ref{ and (d), we show the theoretical and numerical results for a spherocylinder at $h/c = 0.6$ and $h/c = 3$, respectively. For $h/c = 3$, we again obtain excellent agreement. For $h/c = 0.6$, there is no \textit{a priori} reason to expect agreement, because our mapping of the particle solute field to the multipole expansion was performed for distances from the particle centroid larger than $R_{1} = c$. Nevertheless, the agreement remains surprisingly good. The strongest deviation occurs for $x_p = 0$, for which the numerical results exhibit a more complicated functional dependence on $\phi$ than the sinusoidal one predicted theoretically. Note that for $x_p/c = 1$ (brown circles and brown curve) and $x_p/c = -1$ (purple circles and purple curve), the disagreement between theory and numerics can hardly be distinguished by eye. \section{Conclusions} We have studied the effect of particle shape on the motion of a catalytically active particle near a chemically patterned surface. The active particle locally induces chemi-osmotic flows on the surface, which, in turn, drive flows in the volume of the solution that couple back to the particle. This chemi-osmotic contribution to particle motion is sensitive to the material identity of the surface, and it competes with the inherent self-diffusiophoretic motion of the particle. We have restricted our focus to ``inert-forward'' particles which tend to move away from their catalytic caps via self-diffusiophoresis. As first shown in previous work, near a chemical step ``inert-forward'' spherical particles exhibit a rich phenomenology \cite{uspal16}. A sphere can pass over the step, or stably ``dock'' at the edge between the two materials, depending on the system parameters and the initial position and orientation of the particle. A docked particle remains motionless with its axis of symmetry aligned perpendicular to the edge of the step. A particle which passes over the step can exhibit significant deviations of its orientation from the angle of approach. If the particle is non-spherical, such as a spherocylinder, this phenomenology is significantly enriched. In addition to the docking behavior, a rod-like particle can stably follow the edge of the step, with its axis of symmetry tilted slightly away from alignment with the edge. This ``edge-following'' state can coexist with the docking state for a broad range of parameters. Additionally, if the surface has a narrow chemical stripe (i.e., two neighboring step-like interfaces), the edge-following state is transformed into a stable ``stripe-following'' state in which the particle is located at the stripe center and is perfectly aligned with the stripe. These observations motivated the development of an approximate analytical approach in order to understand the underlying physical mechanisms. A central insight of our approach is a distinction between the orientational response of a particle to vorticity and to the fluid strain rate. All particles are driven to rotate by fluid vorticity, but only elongated (i.e., non-spherical) particles respond to the hydrodynamic strain rate. Our analytical expressions demonstrate that this significantly more complicated orientational behavior is the basis of the enriched phenomenology observed for spherocylinders. Therefore, our findings are not specific to spherocylinders, but generalize to other non-spherical shapes (e.g., prolate spheroids \cite{popescu10}), as detailed numerical calculations (omitted here) confirm. Our numerical and analytical calculations have considered a self-diffusiophoretic particle with a monopole term, i.e., a net producer of product molecules. However, our theoretical framework can also be applied to confined self-electrophoretic particles\MP{, under the additional assumption of the presence of a thin Debye layer required by a mapping to an osmotic slip velocity (which has been used in, e.g., Ref. \cite{liu16}),} with appropriate renaming of the variables: $c(\mathbf{x})$ as an electrical potential $\psi(\mathbf{x})$, the surface mobility $b(\mathbf{x}_{s})$ as a quantity proportional to the zeta potential $\zeta(\mathbf{x}_{s})$, etc. \cite{anderson89,golestanian07,chiang14,liu16}. A self-electrophoretic particle would have no monopole term ($\alpha_{0} = 0$), but otherwise our findings would carry over to this new setting. \MP{While the focus of the current work is on the deterministic dynamics, the effects of moderate thermal fluctuations can be assessed, to a certain extent, from the dynamical stability of the corresponding fixed points. The linear stability analysis of the deterministic equations is straightforward, requiring only evaluations of the Jacobians at the fixed points. Similarly with the observations in our previous work \cite{uspal15}, the attractors of the dynamics are ``stiff'', which prevents large excursions from the vicinity of the attractor in the phase plane. For instance, following the procedure of Ref. \cite{uspal15}, we obtain a longest relaxation time of $3.7\;\textrm{s}$ for the docking state in Fig. \ref{, which compares favorably with the characteristic rotational diffusion time of $D_{r}^{-1} = 95\;\textrm{s}$. However, one often encounters, as in Fig. \ref{, a combination of a fast dynamics along one coordinate (the translation) with a (very) slow dynamics along the other coordinate (i..e, the rotation of the particle axis). This implies that the small excursions in the phase plane may actually translate into extended periods of time that a particle, subject to stochastic dynamics, spends outside the fixed point (see also, e.g., Ref. \cite{mozaffari17}). } Our findings open new possibilities for programming the motion of catalytically active particles in lab-on-a-chip devices. The edge-following and stripe-following states can be used to direct rod-like particles from one location to another. If the ratio $U^{sd}/U_{0}$ (characterizing the relative contributions of self-diffusiophoresis and chemi-osmosis to motility) is sensitive to light or some other stimulus, a rod-like particle could be switched between edge-following and docking states on demand. This would be the case, for instance, if light could be used to evoke changes in the surface properties of either the particle or the wall. We note that we had previously predicted stripe-following behavior for spheres, but only for the ``catalyst-forward'' direction of motion, which is much less commonly seen in experiments. Additionally, rod-like particles quite naturally realize our assumption of quasi-2D motion, in which the particle is restricted to have a constant height above the wall and an orientation vector that always remains within a plane parallel to the wall. \WEU{Our findings concerning the role of the particle aspect ratio show that for $e < 0.85$ the edge-following states exist over a broad range of self-propulsion speeds.} Further work could also investigate more complicated shapes. L-shaped particles which are restricted to quasi-2D motion have a well-defined chirality \cite{kummel13}. The effect of fore-aft asymmetry (e.g., for an asymmetric dumbbell with a large ``head'') is an open question \cite{valadares10,uspal13}. Additionally, one could consider shapes with a more complicated topology, such as ring or toroidal shapes \cite{schmieding17,singh13}. \acknowledgments We thank C. Pozrikidis for making the \texttt{BEMLIB} library freely available \cite{pozrikidis02} \WEU{and an anonymous referee for suggesting Fig. \ref{}}. W.E.U., M.T., and M.N.P. acknowledge financial support from the German Science Foundation (DFG), grant no. TA 959/1-1. \WEU{M.T. acknowledges financial support from the Portuguese Foundation for Science and Technology (FCT) under Contracts no. IF/00322/2015.} \section*{Appendix A: numerical calculation of multipole coefficients} For many geometries of interest, such as the spherocylinder, there is no suitable system of coordinates in terms of which the Laplace equation is separable. For these geometries, instead a numerical method, such as BEM, can be used to obtain $c(\mathbf{r})$. We briefly outline how to obtain the multipole coefficients from a known solute density $c(\mathbf{r})$ surrounding an axisymmetric particle in unbounded solution. Any solution $c(\mathbf{r})$ of the Laplace equation can be expressed as an integral over the domain boundaries \cite{pozrikidis02}: \begin{equation} \label{} c(\mathbf{r}') = c_{\infty} - \int_{\cal D} G(\mathbf{r}', \mathbf{r}) [\mathbf{\hat{n}} \cdot \nabla c(\mathbf{r})] dS + \int_{\cal D} c(\mathbf{r}) [ \mathbf{\hat{n}} \cdot \nabla_{\mathbf{r}} G( \mathbf{r}', \mathbf{r}) ] dS. \end{equation} Here, ${\cal D}$ is the particle surface, $\mathbf{r}'$ is a point of interest in the fluid domain, and the local surface normal $\mathbf{\hat{n}}$ is defined to point from the particle surface into the fluid solution. The Green's function $G(\mathbf{r}, \mathbf{r}')$ is \begin{equation} \label{} G(\mathbf{r}, \mathbf{r}') \equiv \frac{1}{4 \pi |\mathbf{r} - \mathbf{r}'|}. \end{equation} Equations (\ref{}) and (\ref{}) indicate that one can interpret the solution ${c}(\mathbf{r})$ as arising from a distribution of point sources (monopoles) and point source - point sink pairs (dipoles) over the particle surface which, respectively, correspond to the first and the second integrals in Eq. (\ref{}). Now we replace the left hand side of Eq. (\ref{}) by Eq. (\ref{}): \begin{equation} \frac{R_{1}}{D} \sum_{l=0}^{\infty} \frac{\alpha_{l}}{l + 1} \left(\frac{R_{1}}{r'}\right)^{l+1} P_{l}(\cos(\theta')) = - \int_{\cal D} G(\mathbf{r}', \mathbf{r}) [\mathbf{\hat{n}} \cdot \nabla c(\mathbf{r})] dS + \int_{\cal D} c(\mathbf{r}) [ \mathbf{\hat{n}} \cdot \nabla_{\mathbf{r}} G( \mathbf{r}', \mathbf{r}) ] dS. \end{equation} We denote the spherical surface defined by $r' = R_{1}$ as ${{\cal D}_1}$. We multiply both sides by $P_{n}(\cos(\theta'))$ and integrate them as $\int_{{\cal D}_1} dS' $. Due to the orthogonality of the Legendre polynomials, one has \begin{multline} \label{} \frac{2 \pi R_{1}}{D} \frac{\alpha_{n}}{n + 1} \frac{2}{2n + 1} = - \int_{{\cal D}} dS \, [\mathbf{\hat{n}} \cdot \nabla c(\mathbf{r})] \int_{{\cal D}_1} dS' \, G(\mathbf{r}', \mathbf{r}) P_{n}(\cos(\theta')) \\ + \int_{{\cal D}} dS \, c(\mathbf{r}) \, \mathbf{\hat{n}} \cdot \nabla_{\mathbf{r}} \int_{{\cal D}_1} dS' \, G( \mathbf{r}', \mathbf{r}) P_{n}(\cos(\theta'))\,. \end{multline} In the next step we expand the Green's function in terms of spherical harmonics: \begin{equation} G(\mathbf{r}', \mathbf{r}) = \sum_{l=0}^{\infty} \frac{1}{2l + 1} \sum_{m=-l}^{l} (-1)^{m} \frac{r^{l}}{r'^{l+1}} Y_{l,-m}(\theta, \phi) Y_{l,m}(\theta', \phi')\,. \end{equation} This leads to \begin{equation} \label{} \begin{split} \int_{{\cal D}_1} dS' \, G(\mathbf{r}',\mathbf{r}) P_{n}(\cos(\theta')) & = \sum_{l=0}^{\infty} \frac{1}{2l + 1} \sum_{m=-l}^{l} \frac{r^{l}}{R_{1}^{l+1}} Y_{l,-m}(\theta, \phi) \int_{{\cal D}_1} dS'\, Y_{l,m}(\theta', \phi') \, P_{n}(\cos(\theta')) \\ & = \frac{1}{2n + 1} \frac{r^{n}}{R_{1}^{n+1}} Y_{n,0}(\theta, \phi) \sqrt{\frac{4 \pi}{2n + 1}} \\ &= \frac{1}{2n + 1} \frac{r^{n}}{R_{1}^{n+1}} P_{n}(\cos(\theta))\,, \end{split} \end{equation} where upon computing the integral on the right hand side above we have used $P_{n}(\cos(\theta')) = \sqrt{\frac{2n+1}{4 \pi}} ~ Y_{n,0}(\theta',\phi')$ and the orthogonality of the spherical harmonics. Substituting Eq. (\ref{}) into Eq. (\ref{}) renders \begin{equation} \label{} \frac{4 \pi R_{1}}{D} \frac{\alpha_{n}}{n + 1} = - \int_{{\cal D}} dS \, [\mathbf{\hat{n}} \cdot \nabla c(\mathbf{r})] \frac{r^{n}}{R_{1}^{n+1}} P_{n}(\cos(\theta)) + \int_{{\cal D}} dS \, c(\mathbf{r}) \, \mathbf{\hat{n}} \cdot \nabla_{\mathbf{r}} \left[ \frac{r^{n}}{R_{1}^{n+1}} P_{n}(\cos(\theta)) \right]. \end{equation} For $\alpha_{0}$, only the first integral contributes: \begin{equation} \alpha_{0} = - \frac{D}{4 \pi R_{1}^{2}} \int_{D} dS \, [\mathbf{\hat{n}} \cdot \nabla c(\mathbf{r})]. \end{equation} Recalling that $[\mathbf{\hat{n}} \cdot \nabla c(\mathbf{r})] = -\kappa/D$ over the catalyst covered area $A_{catalyst}$ and zero elsewhere (see the first paragraph in Sect. II), we have \begin{equation} \alpha_{0} = \frac{A_{catalyst}}{4 \pi R_{1}^{2}} \kappa, \end{equation} which we obtained earlier by simple arguments (see the text preceding Eq.(\ref{})). The dipole term $\alpha_{1}$ has contributions from both integrals in Eq. (\ref{}): \begin{equation} \frac{2 \pi R_{1}}{D} \alpha_{1} = - \int_{D} dS \, [\mathbf{\hat{n}} \cdot \nabla c(\mathbf{r})] \frac{r}{R_{1}^{2}} P_{1}(\cos(\theta)) + \int_{D} dS \, c(\mathbf{r}) \, \mathbf{\hat{n}} \cdot \nabla_{\mathbf{r}} \left[ \frac{r}{R_{1}^{2}} P_{1}(\cos(\theta)) \right]. \end{equation} Using $r \, P_{1}(\cos(\theta)) = z$, one finds \begin{equation} \label{} \frac{2 \pi R_{1}^{3}}{D} \alpha_{1} = - \int_{D} dS \, [\mathbf{\hat{n}} \cdot \nabla c(\mathbf{r})] z + \int_{D} dS \, c(\mathbf{r}) \, \mathbf{\hat{n}} \cdot \hat{z}\,. \end{equation} One can readily verify that the solution for a sphere (Eq. (\ref{})) satisfies Eq. (\ref{}).

\section{Introduction} \label{ref.170929.0} We have given in Ref.~\cite{oller.171114.1} a new perspective to the problem of the compositeness/elementariness of a bound state or a resonance by considering the expectation values in the state of the number operators of the free particle species. The essential problem is to discern whether this bound state is elementary or composite; because of the possible impact of underlying degrees of freedom this is not always a question of an easy answer. In many applications within effective field theory (EFT) the bare elementary discrete states are typically integrated out and do not appear explicitly in the Lagrangian of the theory. Nonetheless, one can still generate bound states after complementing the perturbative calculations in the corresponding EFT with non-perturbative techniques \cite{Weinberg.170929.1,oller.211116.5}. In particular, a near-threshold bare elementary discrete state can be mimicked by including a Castillejo-Dalitz-Dyson pole (CDD) \cite{Castillejo.171020.1}, which straightforward manifestation is an ``innocent'' zero in a partial wave amplitude of the free continuum states. Explicit examples are worked out in Refs.~\cite{oller.211116.5,kang.170930.1}. It is also the case that the Hamiltonian might be expressed in terms of degrees of freedom that are not asymptotically free, as it occurs in Quantum Chromodynamics in ordinary conditions. In all these cases it is not a priori clear whether an eigenstate of the full Hamiltonian is composite or elementary with regards to the asymptotic states in the continuum. Let us discuss first the case of a bound state within non-relativistic quantum mechanics (NRQM) \cite{ref.170928.1,ref.170928.2} and let us split the full Hamiltonian $H$ in an unperturbed free-particle part $H_0$ and an interaction $V$, $ H=H_0+V$. The spectrum of the full Hamiltonian consists of the continuum states and it might contain also discrete bound states $|\psi_{n}\rangle$, and similarly for $H_0$ (in which case the discrete states are bare elementary ones): \begin{align} \label{170928.2} \begin{array}{lllll} H|\psi_\alpha\rangle =E_\alpha|\psi_\alpha\rangle & & && H_0|\varphi_\alpha\rangle =E_{\alpha}|\varphi_\alpha\rangle \\ H|\psi_{n}\rangle =E_{n}|\psi_{n}\rangle & & && H_0|\varphi_n\rangle =E_{n}|\varphi_n\rangle \end{array} \end{align} Both $H$ and $H_0$ share the same spectrum \cite{ref.170928.3,ref.170928.4}. One should not confuse the masses of the eigenstates in $H_0$ with bare masses present in a Lagrangian \cite{ref.170928.3}; the difference if any must be included in $V$. Given a bound state $|\psi_B\rangle$ of $H$ we express it in terms of the eigenstates of $H_0$, which fulfill a completeness relation, as \begin{align} \label{170929.1b} |\psi_B\rangle&= \int d\alpha \langle \varphi_\alpha|\psi_B\rangle |\varphi_\alpha\rangle +\sum_n \langle \varphi_n|\psi_B\rangle |\varphi_n\rangle~,\\ \langle \psi_B|\psi_B\rangle&=1=\int d\alpha |\langle \varphi_\alpha|\psi_B\rangle|^2 +\sum_n |\langle \varphi_n|\psi_B\rangle|^2=Z+X~,\nonumber \end{align} where \begin{align} \label{170929.4} X&=\int d\alpha |\langle \varphi_\alpha|\psi_B\rangle|^2~,\\ \label{170929.4b} Z&=\sum_n |\langle \varphi_n|\psi_B\rangle|^2~. \end{align} These quantities are usually called compositeness ($X$) and elementariness ($Z$). \section{A different perspective on the compositeness of a bound state} \label{ref.170929.1} We first start with the non-relativistic (NR) case and later we move to the relativistic one. For definiteness, let us take two particle species $A$ and $B$ whose free-particle annihilation/creation operators are denoted by $a_\alpha/a_\alpha^\dagger$ and $b_\beta/b_\beta^\dagger$, respectively. The decomposition of the bound state in eigenstates of $H_0$, Eq.~\eqref{170929.1b}, reads now $ |\psi_B\rangle= \int d\gamma \langle AB_\gamma|\psi_B\rangle |AB_\gamma\rangle +\sum_n \langle \varphi_n|\psi_B\rangle |\varphi_n\rangle$. For a given particle species $A$ its number operator is denoted by $N_D^A$ and given by $N_D^A=\int d\alpha \,a_\alpha^\dagger a_\alpha.$ Here the subscript $D$ refers to the Dirac or interaction image. Notice that since $N_D$ and $H_0$ obviously commute then $N_D^A(t)=e^{i H_0 t} N_D^A(0) e^{-i H_0 t}=N_D$. Based on the number operators of $A$ and $B$ we define the compositeness $X$ of the bound state $|\psi_B\rangle$ as \begin{align} \label{170929.11} X&=\frac{1}{2}\langle \psi_B|N_D^A+N_D^B|\psi_B\rangle~. \end{align} That is, $X$ is the expectation value of the number operator of the free-particle constituents in the eigenstate $|\psi_B\rangle$ of $H$ divided by their nominal number, which in this case is 2. The new definition of $X$ is equivalent to the original one of Eq.~\eqref{170929.4} because $(N_D^A+N_D^B)|AB_\gamma\rangle=2 |AB_\gamma\rangle$ and the annihilation operators $a_\alpha$ and $b_\beta$ destroy the bare elementary discrete states. It then follows that $X$, as defined in Eq.~\eqref{170929.11}, reads $ X=\int d\gamma |\langle AB_\gamma|\psi_B\rangle|^2~,$ as in Eq.~\eqref{170929.4}. In general, if we are applying NRQM to a bound state $|\psi_B\rangle$ of $n$ particles corresponding to $m$ particle species $A_1$, $\ldots$, $A_m$, the compositeness is defined by a straightforward generalization of the two-body case of Eq.~\eqref{170929.11} as \begin{align} \label{170929.13} X&=\frac{1}{n}\langle \psi_B|\sum_{i=1}^m N_D^{A_i}|\psi_B\rangle~. \end{align} The evaluation of the expectation values for the number operators is amenable to a direct computation within NR Quantum Field Theory (QFT). At time $t$ the states $|\varphi(t)\rangle$ evolves in the Dirac picture by the time translation $|\varphi(t)\rangle=U_D(t,0)|\psi\rangle$, $U_D(t_2,t_1)=e^{iH_0 t}e^{-iHt}e^{-iH_0t}$. In particular, the bound state $|\psi_B\rangle=|\varphi_B(0)\rangle$ can be expressed by the time evolution from the asymptotic bare elementary discrete state $|\varphi_B\rangle$ as \begin{align} \label{170929.15} |\psi_B\rangle&=U_D(0,\pm \infty)|\varphi_B\rangle~. \end{align} This allows us to write $X$ in a time-ordered way by introducing an extra time evolution from 0 to $t$ as \cite{oller.171114.1} \begin{align} \label{170930.4} X&=\frac{1}{n}\lim_{T\to +\infty}\frac{1}{T}\int_{-T/2}^{+T/2} dt \langle \varphi_B|U_D(+\infty,t) N_D(t)U_D(t,-\infty)|\varphi_B\rangle~, \end{align} with $N_D=\sum_{i=1}^m N_D^{A_i}$. The factor $1/T$ cancels in the limit $T\to +\infty$ with the Dirac delta function of total energy conservation. It might be advantageous to express the number operator in terms of NR fields in Eq.~\eqref{170930.4}, e.g. in order to apply Feynman diagrams for its calculation. For a generic scalar particle species $A_i$ of physical mass $m_{A_i}$ we have the free field $\psi_{A_i}(x)$, $x=(t,\mathbf{x})$, and we can write $N_D=\sum_i \int d^3\mathbf{x}\, \psi^\dagger_{A_i}(x)\psi_{A_i}(x)~.$ Inserting this expression into Eq.~\eqref{170930.4} it reads \begin{align} \label{171110.3} X &=\frac{1}{n}\lim_{T\to +\infty}\frac{1}{T}\int d^4x \langle \varphi_B| P\left[ e^{-i\int_{-\infty}^{+\infty}dt' V_D(t')} \sum_i \psi_{A_i}^\dagger(x)\psi_{A_i}(x) \right] |\varphi_B\rangle~. \end{align} Here we denote the time-ordered product by $P$ and $V_D(t)$ is the interaction in the Dirac picture.\footnote{In Eq.~\eqref{171110.3} only the connected diagrams should be considered \cite{Low.171110.1}.} The extension to particles with other spin is straightforward. \begin{figure} \begin{center} \includegraphics[width=0.6\textwidth]{Basic_Skeleton.eps} \caption{Feynman diagrams for the calculation of $X$ within NR QFT for the two-particle case. The insertion of the number operators for the particles $A$ and $B$ is indicated by the double dot.} \label{fig.170930.1} \end{center} \end{figure} For two particles of types $A$ and $B$, the evaluation of $X$ corresponds to the calculation of the diagrams in Fig.~\ref{fig.170930.1}. In the $\ell S$ basis (with $\ell$ the orbital angular momentum and $S$ the total spin) one has, $ X=\sum_{\ell,S} X_{\ell S}$, \begin{align} \label{170930.6} X_{\ell S}&=\frac{1}{2\pi^2}\int_0^\infty dk k^2 \frac{g_{\ell S}^2(k^2)}{(k^2/2\mu-E_B)^2}~. \end{align} In this equation, $\mu$ is the reduced mass of particles $A$ and $B$, and $g_{\ell S}^2(k^2)$ is the coupling squared of the bound state, $\langle AB_{\ell S}|V|\psi_B\rangle^2$. The coupling can be calculated from the knowledge of the residue at $E=E_B$ of the the off-shell $T$ matrix, $T(E)(k',k)$, and it satisfies then the following integral equation (a matrix notation should be employed if appropriate) \cite{oller.171114.1} \begin{align} \label{170930.8} g(k)&=\frac{1}{2\pi^2}\int_0^\infty dk'\,{k'}^2V(k,k')\frac{1}{E_B-{k'}^2/2\mu}g(k')~. \end{align} From Eq.~\eqref{170930.8} and the fact that $V(-k,k')=(-1)^\ell V(k,k')$ (parity conservation) one concludes that the coupling squared only depends on $k^2$. The global normalization factor in Eq.~\eqref{170930.8} is fixed by the requirement that $g(k)$ matches the residue of the $T$ matrix at the pole position. From Eq.~\eqref{170930.8} it turns out that $g(k)$ is analytic in the $k$-complex plane without cuts \cite{oller.171114.1}. As a result, a non-constant $g(k)$ is not bounded for $k\to \infty$ because of the Liouville's theorem in complex analysis. It is worth stressing that for a given total Hamiltonian $H$ the compositeness $X$ is an observable in the sense that it is invariant under unitary transformations and field reparametrizations. This is clear from the decomposition of the bound state in terms of free-particle states. The absence of tad-pole like contributions within an appropriate regularization procedure in NR QFT drives to Eq.~\eqref{170930.6} as the final expression without any possible counterterm contributions \cite{oller.171114.1}. Thus, $X$ is a fully derived quantity from the knowledge of the (half) off-shell $T$ matrix (which contains all the spectroscopical information of the corresponding quantum system). We derive now a new closed expression for the compositeness $X$ in the case in which the wavelengths of the two scattering particles are large compared with the range of their interaction. In this case, $V(k',k)$ is a polynomial in its arguments and $|g(k)|$ grows only polinomically for $k\to \infty$. A simple way to deal with these power-like divergences is to regularize the potential as $ V(k',k)\to V(k',k) e^{i\epsilon (k+k')}$, with $\epsilon \to 0^+$. It is clear from Eq.~\eqref{170930.8} that this also implies that $g(k) \to g(k) e^{i\epsilon k}$~. Taking advantage of the fact that the integrand in Eq.~\eqref{170930.6} is an even function of $k$ we symmetrize it, extend the integration along the real axis from $-\infty$ to $+\infty$, and close the integration contour along the upper half plane of the $k$-complex plane with a semicircle of infinite radius. The calculation is straightforward by applying the Cauchy's integration theorem with the result \begin{align} X&=g^2(\varkappa^2)\frac{\mu^2}{2\pi\bar{\gamma}} +\frac{\mu^2}{2\pi}\left.\frac{\partial g^2(-\gamma^2)}{\partial \gamma}\right|_{\gamma=\bar{\gamma}}~. \label{170930.11} \end{align} where $\gamma=-ik$ and $\bar{\gamma}=-i\varkappa$, and $\varkappa=\sqrt{-2\mu E_B}$. The first term on the right hand side of this equation is a well-known contribution, although obtained with less generality \cite{ref.170928.2,hyodo.170930.1}. It is model independent because it is fixed once the pole position and the residue of the on-shell $T$ matrix at the pole position are known. The second term of Eq.~\eqref{170930.11} is an extra contribution, which cannot be fixed directly from the knowledge of the on-shell $T$ matrix and depends on the interaction $V(k',k)$. Let us apply Eq.~\eqref{170930.11} to some energy-independent potentials for which $X=1$. The first one is a pure $S$-wave potential $V(k',k)=\left[v_0+v_2 (k^2+{k'}^2)\right]e^{i\epsilon(k+k')}$, where $v_0$ and $v_2$ are constants. The scattering problem can be easily solved \cite{oller.171114.1} and the two terms that sum up $X=1$ in Eq.~\eqref{170930.11} are, in order, \begin{align} \label{171107.7} 1&=\frac{1-2\bar{\gamma}^2 v_2/v_0}{1-6\bar{\gamma}^2 v_2/v_0} -\frac{4\bar{\gamma}^2 v_2/v_0}{1-6\bar{\gamma}^2 v_2/v_0}~. \end{align} In this case, since $g^2(k^2)$ is not zero for $k=0$, the last contribution in Eq.~\eqref{171107.7} is suppressed by a factor $\bar{\gamma}^2 |v_2/v_0|\sim \bar{\gamma}R$. The last step is based on the relation between $v_0$ and $v_2$ with the effective range parameters \cite{oller.171114.1}.\footnote{The only exception to this rule might happen when there is a shallow CDD pole close enough to threshold \cite{kang.170930.1}. In such a case the $s$-channel exchange of the bare elementary state cannot be traded by a perturbative expansion in momenta through contact interactions employing the NR equations of motion \cite{oller.171112.1}. The compositeness $X$ is then smaller than 1. \label{footnote.171107.1}} The Weinberg's formula for the compositeness of an $S$-wave shallow bound state state is recovered \cite{ref.170928.2} once the subleading contribution is neglected. As a specific example for which $g^2(k^2)=0$ at $k^2=0$, let us take a projected potential with orbital angular momentum $\ell$, $V(k',k)= v_\ell {k'}^\ell k^\ell e^{i\epsilon (k+k')}$, with $v_\ell$ being a constant. The two terms in Eq.~\eqref{170930.11} are now, in order, \begin{align} \label{171107.12} 1&=\frac{1}{2\ell+1}+\frac{2\ell}{2\ell+1}~. \end{align} Both contributions count on the same footing, but as $\ell$ increases the 2nd one becomes indeed dominant. If the range of the interactions between the non-relativistic particles is explicitly resolved, the formula for $X$ of Eq.~\eqref{170930.11}, derived in the limit of large wavelengths compared to $R$, is not valid in general. Of course, Eq.~\eqref{170930.6} is still applicable. The reason is that closing the integration contour to end with Eq.~\eqref{170930.11} is not always possible \cite{oller.171114.1}. Up to the best of our knowledge there is no a general criterion for a relativistic bound state to be qualified as elementary. In the relativistic case one generally relies on the study of the wave function renormalization $Z$ \cite{amado.170930.1,salam.170930.1,lurie.170930.1}. The straightforward extrapolation of the definition of $X$ in Eq.~\eqref{170929.13} cannot be given because contributions of multi-particle eigenstates of $H_0$. In this way, Eq.~\eqref{170929.1b} now generalizes to \begin{align} \label{170930.12} |\psi_B\rangle &= \int d\gamma C_\gamma |AB_\gamma\rangle +\int d\eta D_\eta |AAB_\eta\rangle+\int d\mu \,\delta_\mu |ABB_\mu\rangle+\ldots\\ &+\int d\eta_\nu F_\nu|CD_\nu\rangle+\ldots +\sum_n C_n|\varphi_n\rangle+\sum_n \int d\alpha C_{n\alpha} |A_\alpha\varphi_n\rangle +\ldots+\sum_{n,m} C_{nm} |\varphi_n\varphi_m\rangle+\ldots \nonumber \end{align} with quite an obvious notation. Nonetheless, we can still take advantage of the use of the number operators which are defined in the relativistic case as in NR QFT. E.g. the average number of asymptotic particles of type $A$ in $|\psi_B\rangle$ as given by the decomposition in Eq.~\eqref{170930.12} is $\langle \psi_B|N_D^A|\psi_B\rangle=\int d\gamma |C_\gamma|^2 +2\int d\eta |D_\eta|^2 + \int d\mu |\delta_\mu|^2+\ldots+\sum_n \int d\alpha |C_{n\alpha}|^2+\ldots $ In this way we can deduce the following universal criterion for a bound state to be considered as elementary, applicable both in the relativistic and NR cases: \begin{align} \label{170930.14} \langle \psi_B|N_D^A|\psi_B\rangle&=0~~,~~\forall A~;~~ \langle \psi_B|N_D^E|\psi_B\rangle = 1~. \end{align} where $N_D^E$ is the sum of the number operators for the bare elementary discrete states ($N_D^n$). The last condition in Eq.~\eqref{170930.14} avoids the possibility that $|\psi_B\rangle$ had components of states made by several bare elementary discrete states. If for particle species $A$ one has that $ \langle \psi_B|N_D^A|\psi_B\rangle=x_A$, with $x_A\geq m$ and $m\geq 0$ a natural number, then the free-particle states containing $m$ or more asymptotic particles of type $A$ are relevant in the bound state $|\psi_B\rangle$. The expression for the expectation value $\langle \psi_B|N_D^A|\psi_B\rangle $ in relativistic QFT can be obtained following similar steps as for NR QFT \cite{oller.171114.1}, cf. Eq.~\eqref{170930.4}. One can also express the number operator $N_D^A$ in terms of free fields, analogously as done in the non-relativistic case. Let us a consider a scalar particle $A$, we can write $ N_D^A(t)=-2i\int d^3 \mathbf{x} \dot{\psi}^{(-)}(x)\psi^{(+)}(x)$, \begin{align} \label{171110.4} \psi^{(+)}(x)&=\int\frac{d^3\mathbf{q}}{(2\pi)^3}a(\mathbf{q})e^{-i\tilde{q} x}~,\nonumber\\ \langle \psi_B|N_D^A|\psi_B\rangle &=-2i\lim_{T\to +\infty}\frac{1}{T}\int d^4x \langle \varphi_B| P\left[ e^{-i\int d^4 x' {\cal H}_D(x')} \dot{\psi}^{(-)}(x)\psi^{(+)}(x) \right] |\varphi_B\rangle~. \end{align} where $\psi^{(-)}={\psi^{(+)}}^\dagger$ and the interaction have been written in terms of an interaction-Hamiltonian density ${\cal H}_D(x)$ in the Dirac picture. The set of Feynman diagrams involved can be schematically represented as in Fig.~\ref{fig.171110.1}, where the shaded circle represents any set of connected vertices without any insertion of the number operator which is indicated by the double dot. However, in order to apply Feynman rules to the calculation of Eq.~\eqref{171110.4} one has to take into account that any of the two internal lines in Fig.~\ref{fig.171110.1} ending in the double dots is not a standard Feynman propagator but $ \frac{i}{2E_k(k^0-E_k+i\varepsilon)}$. If one can conclude that only two-body channels dominate then one expects that the more important Feynman diagrams are those of Fig.~\ref{fig.170930.1}. It is worth stating that for a given total Hamiltonian $H$ the expectation values $N_D^A$ are observable in the sense that they are invariant under unitary transformations and field reparametrizations, analogously as in NR QFT, see Ref.~\cite{oller.171114.1} for further details \begin{figure} \begin{center} \includegraphics[width=0.25\textwidth]{Sum_Rules.eps} \caption{Schematic representation of the Feynman diagrams for the calculation of $\langle \psi_B|N_D^A|\psi_B\rangle$ in QFT.} \label{fig.171110.1} \end{center} \end{figure} \section{Resonances} \label{sec.170930.4} In this section we discuss the generalization of many of the results given for bound states to the case of resonances. The latter correspond to poles in an unphysical RS that can be reached by the analytical extrapolation in energy of the $T$ matrix. An approximate way to afford the problem of the evaluation of $Z$ for an unstable particle in the non-relativistic case near a two-body threshold was considered in Ref.~\cite{markusin.170930.2}. These results have also a clear connection with the counting pole rule of Morgan \cite{morgan.170930.1} and with the presence of near Castillejo-Dalitz-Dyson poles \cite{kang.170930.1}. \begin{figure} \begin{center} \includegraphics[width=0.3\textwidth]{Continuation_Contour.eps} \caption{Deformation of the integration contour along the physical energy needed to reach the resonance pole at $E_R=M_R-i\Gamma/2$.} \label{fig.171001.1} \end{center} \end{figure} Our definition of compositeness $X$ is based on Eq.~\eqref{170929.13} in NRQM. We derive our results for resonance states by the analytical continuation in energy of $\langle \psi^-_\alpha|N_D|\psi^+_\alpha\rangle$ extrapolated to the resonance pole position, with $|\psi_\alpha^{\pm}\rangle $ an in/out state, respectively. The previous matrix element has a double pole at the resonance pole whose residue divided by the coupling squared is the expectation value of the operator $N_D$ in the resonance state \cite{albaladejo.171001.1}. This limit process can be avoided if we use an analogous formalism to that explained in Sec.~\ref{ref.170929.1}, but now for resonance states. We express the in/out resonance state $|\psi^\pm_R\rangle$ by evolving the bare one $|\varphi_R\rangle$ from asymptotic times, $ |\psi^\pm_R\rangle =U_D(0,\mp\infty)|\varphi_R\rangle$. Thus an equation analogous to Eq.~\ref{170930.4} results \cite{oller.171114.1}. We can also re-express $N_D$ in terms of bilinears of the NR fields $\psi_i(x)$. Therefore, for scalar particles $A_i$ we have $N_D=\sum_i \int d^3\mathbf{x} \psi_{A_i}^\dagger(x)\psi_{A_i}(x)$ and we end with an expression for $X$ analogous to Eq.~\eqref{171110.3} \cite{oller.171114.1}. In the case of two-particle asymptotic states the calculation of $X$ can be done by evaluating the Feynman diagrams of Fig.~\ref{fig.170930.1}. Performing the corresponding partial-wave decomposition one has the following expression for $X_{\ell S}$ \begin{align} \label{171001.7} X_{\ell S}&=\frac{1}{2\pi^2}\int_0^\infty dk k^2\frac{g_{\ell S}^2(k^2)}{(k^2/2\mu-E_R)^2} +\frac{i\mu^2}{\pi\varkappa}\left.\frac{\partial kg_{\ell S}^2(k^2)}{\partial k}\right|_{k=\varkappa}~. \end{align} Compared with Eq.~\eqref{170930.6} there is an extra term due to the deformation of the analytical contour for integration, as shown in Fig.~\ref{fig.171001.1}. Because of the same reason, the homogeneous integral equation satisfied by $g(k)$ has an extra term compared to Eq.~\eqref{170930.8}, as derived in Ref.~\cite{oller.171114.1}. Although Eq.~\eqref{171001.7} is not explicitly real and positive, it is shown in Ref.~\cite{ref.171001.2} that for an energy-independent regular potential $X=1$, which implies that a resonance in ordinary QM is a composite state. The case of a separable potential $V(k,k';E)=v(E) f(k^2)f({k'}^2)$ is explicitly worked out in Ref.~\cite{oller.171114.1}. There it is shown that if $\partial v(E)/\partial E=0$ then $X=1$, but for $\partial v(E)/\partial E\neq 0$ the compositeness $X$ is in general a complex number. We can also derive a new closed formula for $X$ in the case of large wavelengths of the two scattering particles compared with the range $R$ of their interactions. Since the resonance pole lies in the 2nd RS, with ${\rm Im}k<0$, now we regularize the potential as $ V(k',k)\to V(k',k) e^{-i\epsilon (k+k')}$. It is straightforward to obtain \begin{align} \label{171001.9} X_{\ell S}&=g^2(\varkappa^2)\frac{i\mu^2}{2\pi\varkappa} +\frac{i\mu^2\varkappa}{\pi}\left.\frac{\partial g^2(k^2)}{\partial k^2}\right|_{k=\varkappa}~. \end{align} The first term is already well-known while the latter is a new contribution. The result for $X$ in Eq.~\eqref{171001.9} can also be expressed as \begin{align} \label{171001.10} X_{\ell S}&=\frac{2\mu^2}{\pi^2}\int_0^\infty dk^2 \sqrt[{\rm II}]{k^2+i\varepsilon}\frac{g_{\ell S}^2(k^2)}{(k^2-\varkappa^2)^2}~, \end{align} where $\sqrt[II]{z}$ is $\sqrt{z}$ in the 2nd RS with ${\rm arg}z\in [2\pi,4\pi[$. In this equation it is clear the appearance of the square of the resonance wave function $g_{\ell S}^2(k^2)/(k^2-\varkappa^2)^2$ as it corresponds to a Gamow state \cite{ref.171001.2}. For the relativistic case we can evaluate the matrix elements of the operator numbers $N_D^A$ between resonances states, and end with a similar expression to Eq.~\eqref{170930.4} \cite{oller.171114.1}. We can also express the number operators $N_D^A(t)$ as bilinear of relativistic fields, cf. Eq.~\eqref{171110.4}. The set of Feynman diagrams is represented in Fig.~\ref{fig.171110.1}, with the obvious replacement of $|\psi_B\rangle$ by $|\psi_R^{\pm}\rangle$ to the right and left, respectively. We have to keep in mind the meaning of the internal lines joining the field bilinear associated to the number operator, as discussed in Sec.~\ref{ref.170929.1}. Since the annihilation operators in the number operators $N_D^A$ kill the bare elementary discrete states we can state that a necessary condition in relativistic (NR) QFT for a resonance being elementary is that \begin{align} \label{171002.1} \langle \psi_R^-|N_D^A|\psi_R^+\rangle&=0 ~~,~~\forall A~. \end{align} It cannot be qualified as sufficient condition as well because given a decomposition of a resonance state as in Eq.~\eqref{170930.12}, one should not expect to have the sum of the modules squared of the coefficients in the linear decomposition but rather coefficients squared \cite{oller.171114.1}. Let us consider energy-dependent transformations in the partial-wave projected in/out states. These are driven by a function $\eta_i(E)$ for the $i_{\rm th}$ partial wave, which at least has a unitarity cut and satisfy the Schwarz reflection principle $\eta_i(E\pm i\varepsilon)=\eta_i(E\mp i\varepsilon)^*$. Namely, \begin{align} \label{171001.16b} |\psi_\alpha^+\rangle&\to e^{\eta_i(E_\alpha+i\varepsilon)}|\psi_\alpha^+\rangle~,\\ \langle \psi_\alpha^-|&\to \langle \psi_\alpha^-| e^{\eta_i(E_\alpha-i\varepsilon)^*}=\langle \psi_\alpha^-|e^{\eta_i(E_\alpha+i\varepsilon)}~. \end{align} When performing the analytical extrapolation to the 2nd RS to reach the resonance pole at $E_R$ one has to cross the unitarity cut and end with the value $\eta^{{\rm II}}(M_R-i\Gamma/2)$. As a result the couplings change as $g_i^2(k^2)\to g_i^2(k^2) e^{2\eta_i^{{\rm II}}(E_R)}$. For the case of a narrow resonance we can write a plausible dispersion relation for the smooth function $\eta_i(E)$ around the resonance region as \begin{align} \label{171001.19} \eta_i(E)&=\frac{1}{\pi}\int dE'\frac{{\rm Im}\eta_i(E')}{E'-M_R-i\Gamma/2} \approx \frac{1}{\pi} \Xint- dE'\frac{{\rm Im}\eta_i(E')}{E'-M_R} +i\,{\text{Im}\,}\eta_i(M_R)~. \end{align} Since ${\rm Im}\eta_i(E')$ is uniform around the narrow-resonance mass, its Cauchy principal value around should be very small and Eq.~\eqref{171001.19} is quite a pure imaginary number. The couplings then change as $g_i^2(k^2)\to g_i^2(k^2) e^{2i{\rm Im}\eta_i^{{\rm II}}(E_R)}$. This derivation also shows that for a finite width resonance is not so clear that $\eta_i(E)$ is just a purely imaginary number. However, if the resonance is manifest on the physical real energy axis, for some physical process, the modules of its residues can be interpreted as physical couplings and the corrections on them (if any) would be relatively small. In this way, we can then properly choose the phase factor $\eta_i^{{\rm II}}(E_R)$ of the coupling to a partial wave so that its compositeness is $|X_{\ell S}|$. As a result, $|X|\ll 1$ is the criterion for the elementariness of a NR resonance (qualified in the above sense) with respect to the explicit open channels. In the relativistic case the situation is a priori less clear since one cannot exclude contributions from closed channels containing particles of type $A$. Therefore, the change of phase in the couplings of only the open channels is not enough in general to end with real and positive expectation values of $N_D^A$. Nonetheless, in practical applications within models that incorporate only a few coupled channels one could still apply these changes of phase in the couplings for the open channels and give physically reasonable results. The phase transformation here discussed is closely connected to the transformations in the partial-wave projected $S$-matrix introduced in Ref.~\cite{guo.170930.1}. \begin{figure}[ht] \begin{center} \includegraphics[angle=0.0,width=.3\textwidth]{Resonance_X_2poles.eps} \end{center} \caption{Double-pole virtual state case in the $s$-complex plane . The pole positions are indicated by the full circles and the convergence radius for the Laurent series around each of them is the radius of every circle. \label{fig.171124.1}} \end{figure} If a resonance has a finite width and its mass is smaller than threshold, which can happen for an $S$-wave \cite{pelaez.171125.1}, then the resonance signal is not directly manifest apart from a cusp effect. In such case the condition above to apply the phase transformations does not apply. If the width is small we are near the realization of a double-pole virtual state, as represented in Fig.~\ref{fig.171124.1}, where the poles lie in the 2nd RS and the circles correspond to the radius of convergence of the Laurent series around the poles. Note that such a situation is not a common one and gives rise to two quite peculiar effects. The first one is that along the real axis the resonance signal is purely real because $g^2/(s-s_R)+{g^*}^2/(s-s_R^*)=2 {\Re g^2/(s-s_R)}$ (with $s$ the usual Mandelstam variable). The second effect is that if $s_{{\rm th}}$ is the threshold represented in the figure then the maximum at the cusp is given by the sum of the two poles contributions $g^2/(s_{{\rm th}}-s_R)+{g^*}^2/(s_{{\rm th}}-s_R^*)$. Therefore, if $s_R\to s_{{\rm th}}$ the cusp is enhanced compared to standard cases with only one pole for the virtual state. This situation could be well that corresponding to the $X(3872)$, as shown in Ref.~\cite{kang.170930.1}, for which case the state whose threshold is depicted in Fig.~\ref{fig.171124.1} by the crossing of the two perpendicular lines is the $\bar{D}^0{D^*}^0$. Furthermore, the $X(3872)$ has a tiny decay width to this state. As analyzed in Ref.~\cite{kang.170930.1} the most recent data on spectral lines of the $X(3872)$ are compatible with the resonance being a double- and even a triple-pole virtual state (in the limit of vanishing ${D^*}^0$ width). Thus, this resonance might be the first one in particle phenomenology that manifests as a higher degree pole. In all the cases the poles emerge from the interplay between the presence of a bare state and the direct interactions between the $\bar{D}^0{D^*}^0$ mesons. Reference \cite{kang.170930.1} proceeds by making use of general principles of $S$-matrix theory. Instead of considering explicitly the inclusion of a bare pole it derives the consequences of the presence of a near-threshold CDD pole, as an extra source for a strong energy signal in the partial wave beyond the presence of the threshold. This is a more general approach than including a bare pole \cite{baru.171125.1} because it can also reproduce positive values for the effective range (as well as for other shape parameters in the effective range expansion) \cite{kang.170930.1}.

\section{Introduction} Heavy quarkonia, bound states of charm or bottom quarks, constitue a prominent probe of the quark-gluon plasma produced in ultra-relativistic heavy ion collisions, and are the object of many investigations, both theoretically and experimentally. Recent data from the LHC provide evidence for a sequential suppresion, with the most fragile (less bound) states being more strongly suppressed \cite{}, while there are indications that charm quarks are sufficiently numerous to recombine, an effect that is seen to counterbalance the expected suppression \cite{}. These findings are in line with general expectations. The dissociation of quarkonium was suggested long ago \cite{} as resulting from the screening of the binding forces by the quark gluon plasma. Recombination is a natural phenomenon to expect \cite{} whenever the number of heavy quarks is sufficiently large, which seems to be the case of charm quarks at the LHC. However, in order to go beyond these qualitative remarks, and extract precise information about the dynamics, we have to address a rather complicated many-body problem. Even leaving aside the production mechanisms of heavy quarks in hadronic collisions, which is not fully understood yet (see e.g. \cite{}), the description of the interactions of these heavy quarks with an expanding quark-gluon plasma is indeed complicated for a number of reasons. Many effects can contribute, among which: screening affecting the binding potential, collisions with the plasma constituents, absorption of gluons of the plasma by the bound states. It should be added to this that the bound states do not exist as objects ``deposited'' in the plasma: it takes time before a newly created quark-antiquark pair can be considered as a bound state, and during this time it is interacting with the plasma. This is a feature that is often forgotten in many models that attempts to describe the data (for representatives of recent phenomenological analyses, see e.g. \cite{} and references therein). Thus, most models emphasize static or stationary aspects (even when the expansion of the plasma is take into account): this is the case of potential models \cite{}, spectral function calculations \cite{}, or kinetic approaches based on rate equations \cite{}. Clearly a fully time-dependent, out of equilibrium treatment is called for. Such a treatment should also establish contact between the dynamics of heavy quark-antiquark pairs and their bound state, and that of isolated heavy quarks in a quark-gluon plasma (see e.g. \cite{}). In short, there is a need for a general, simple, and robust formalism, where all the relevant effects can be treated within the same framework. In this respect, the observation that the collisions could be taken into account by an imaginary potential is a significant one \cite{}. In recent years it has been recognized that techniques from the theory of open quantum systems (see e.g. \cite{OQS,Weiss}) could offer a fruitful perspective on this problem. A system of heavy quarks in a quark gluon plasma falls indeed in the category of typical problems addressed by this theory: a small system, weakly interacting with a large ``reservoir'', the quark-gluon plasma. This point of view has emerged explicitly or implicitly in a number of recent works: derivation of a master equation, and corresponding rate equations \cite{}, use of the influence functional method \cite{}, solution of a stochastic Schr\"{o}dinger equation \cite{}, or direct computation of the evolution of the density matrix \cite{}. The present work follows similar lines. Its initial motivation was to generalize the results of \cite{} to the non-Abelian case (QCD). Part of that generalization is straightforward, and relies on the same approximations as those used in the case of the abelian plasma. To some extent, this program has already been considered in the recent work by Akamatsu \cite{}, albeit using a formalism slightly different from that used in \cite{} and in the present paper. However, color degrees of freedom modifies the picture in a very substantial way. The reason is that, in a collision involving one gluon exchange for instance, color changes in a discrete way, in contrast to position or momentum which vary continuously. Thus, while we can treat the motion of the heavy quarks within a semi-classical approximation, there is no such semi-classical limit for the color dynamics (except perhaps in the large $N_c$ limit). It follows that the derivation of Fokker-Planck or Langevin equations made in the abelian case needs to be reconsidered, which we do in this paper. We shall see that the complete dynamics, including the color degrees of freedom, can still be described by Fokker-Plack and Langevin equations, but only in very specific circumstances. This paper focusses on conceptual issues. It is organized as follows. In Sect.~\ref{} we derive the quantum master equation for the reduced density matrix of a system of heavy quarks and antiquarks immersed in a quark-gluon plasma, in thermal equilibrium. This equation, whose structure is close to that of a Lindblad equation, is used as a starting point of all later developments. In Sect.~\ref{} we rederive from it the results that we had previously obtained for the abelian plasma \cite{} using a path integral formalism. In particular we recover, after performing a semi-classical approximation, the Fokker-Planck and Langevin equations that describe the random walks of center of mass and relative coordinates of a quark-antiquark pair. This section on the abelian plasma paves the way for the treatment of the non abelian case discussed in Sect.~\ref{}. The equations that we present there, before we do the semi-classical approximation, are fully quantum equations. But they are difficult to solve in general. Thus, in Sect.~\ref{numerics} we look for additional approximations that allow us to obtain solutions in some particular regimes, in order to start getting insight into the general solution. In particular, we explore two ways of implementing the semi-classical approximation. In the first case, we restrict the dynamics to stay close to a maximum entropy color state, where the colors of the heavy quarks are random. In this case the dynamics is described by a Langevin equation with a new random color force. The method used in this case is easily extended to the case of an arbitrary number of quark-antiquark pairs, and allows us to address the question of recombination. However, it is based on a perturbative approach that breaks down for some values of the parameters. Another strategy focuses on the case of a single quark-antiquark pair. The transition between singlets and octets are treated as ``collisions'' in a kinetic equation that we solve using Monte Carlo techniques. The last section summarizes our main results, and presents a brief outlook. Several appendices at the end gather various technical material. \section{Equation for the density matrix of heavy quarks in a quark-gluon plasma} \label{} Our description of the heavy quark dynamics in a quark-gluon plasma is based on the assumption that the interaction between the heavy quarks and the quark-gluon plasma is weak, and can be treated in perturbation theory (with appropriate resummations). The generic hamiltonian for such a system reads \beq H=H_Q+H_1+H_{\rm pl}, \eeq where $H_Q$ describes the dynamics of the heavy quarks in the absence of the plasma, $H_{\rm pl}$ is the hamiltonian of the plasma in the absence of the heavy quarks, and $H_1$ is the interaction between the heavy quarks and the plasma constituents. The heavy quarks are treated as non relativistic particles, and the spin degree of freedom is ignored: the state of a heavy quark is then entirely specified by its position and color. As we have mentioned already, we shall consider $H_1$ to be small and treat it as a perturbation. In Coulomb gauge, and neglecting magnetic interactions, this interaction takes the form \beq\label{H1} H_1=-g\int_{\boldsymbol r} A_0^a({\boldsymbol r}) n^a({\boldsymbol r}), \eeq where $n^a$ denotes the color charge density of the heavy particles. For a quark-antiquark pair, for instance, this is given by\footnote{We denote here the position operator by $\hat r$, but most often the symbol \^{ } will be omitted, the context being enough to recognize the operators.} \beq\label{colordensity} n^a({\boldsymbol x})=\delta({\boldsymbol x}-\hat {\boldsymbol r})\, t^a\otimes \mathbb{I} -\mathbb{I}\otimes\delta({\boldsymbol x}-\hat{\boldsymbol r}) \, \tilde t^a, \eeq where we use the first quantization to describe the heavy quark and antiquark, and the two components of the tensor product refer respectively to the Hilbert spaces of the heavy quark (for the first component) and the heavy antiquark (for the second component). In Eq.~(\ref{colordensity}), $t^a$ is a color matrix in the fundamental representation of SU(3) and describes the coupling between the heavy quark and the gluon. The coupling of the heavy antiquark and the gluon is described by $-\tilde t^a$, with $\tilde t^a$ the transpose of $t^a$. We are looking for an effective theory for the heavy quark dynamics, obtained by eliminating the plasma degrees of freedom. In previous works, this was achieved explicitly by constructing the Feynman-Vernon influence functional \cite{FeynmanVernon}, using the path integral formalism (see e.g. \cite{}). In the present paper, we shall proceed differently, by writing directly the equations of motion for the reduced density matrix of the heavy quarks. Although the derivations presented here are self-contained, we emphasize that the main approximations that are implemented in the present section are quite common in various fields, and belong to what is commonly referred to as the theory of open quantum systems (see e.g. \cite{OQS}). We assume that the system contains a fixed number, $N_Q$, of heavy quarks (and, in general, an equal number of antiquarks). We call ${\cal D}$ the density matrix of the full system, and ${\cal D}_Q$ the reduced density matrix for the heavy quarks. The latter is defined as the partial trace of the full density matrix over the plasma degrees of freedom, that is \beq {\cal D}_Q={\rm Tr}_{\rm pl} {\cal D}. \eeq In order to make contact with the work of Ref.~\cite{}, we recall that a typical question addressed there was the following: Given a set of heavy quarks at position ${\boldsymbol X}_i$ at time $t_i$, where ${\boldsymbol X}$ denotes collectively the set of coordinates of the quarks and antiquarks (temporarily ignoring color), what is the probability $P({\boldsymbol X}_f, t_f|{\boldsymbol X}_i, t_i)$ to find them as position ${\boldsymbol X}_f$ at time $t_f$? This probability is given by \beq\label{probability0} P({\boldsymbol X}_f, t_f|{\boldsymbol X}_i, t_i) =\left| \langle{\boldsymbol X}_f,t_f|{\boldsymbol X}_i,t_i\rangle\right|^2=\bra{{\boldsymbol X}_f}{\cal D}_Q\ket{{\boldsymbol X}_i}, \eeq that is, it can be obtained as a specific element of the reduced density matrix. In \cite{} a representation of this quantity was obtained in terms of a path integral which is remains difficult to evaluate in general.\footnote{The analogous path integral for a single heavy quark in an abelian plasma has been evaluated in \cite{}. However, this evaluation was performed in Euclidean space. An analytic continuation is needed to recover the real time information, and procedures to do so numerically are not without ambiguities.} However, in the regime where a semi-classical approximation is valid, the dynamics that it describes is equivalent to that of a Fokker-Planck equation which can be easily solved numerically, in particular by solving the associated Langevin equation. Two approximations are involved in the construction of the influence functional such as presented in \cite{}. The first one is the weak coupling approximation for the interaction of the heavy quarks with the plasma, the second assumes that the response of the plasma to the perturbation caused by the heavy quarks is fast compared to the characteristic time scales of the heavy quark motion. An additional approximation, to which we refer to as a semi-classical approximation, leads, as we have just mentioned, to Fokker Planck and Langevin equations. The last two approximations exploit the fact that the mass of the heavy quark is large, i.e., $M\gg T$. Thus, when the heavy quark is not too far from thermal equilibrium, its thermal wavelength $\lambda_{\rm th}\sim 1/\sqrt{MT}$ is small compared to the typical microscopic length scale $\sim 1/T$. Under such condition, the density matrix can be considered as nearly diagonal (in position space), motivating a semi-classical approximation: indeed the off-diagonal matrix elements $\bra{{\boldsymbol X}}{\cal D_Q}\ket{{\boldsymbol X}'}$ die off when $|{\boldsymbol X}-{\boldsymbol X}'|\gtrsim \lambda_{\rm th}$. The typical heavy quark velocity is of the order of the thermal velocity $\sim\sqrt{T/M}\ll 1$, and the dynamics of the heavy fermions is much slower than that of the plasma. The typical frequency for the plasma dynamics is the plasma frequency which, for ultra-relativistic plasmas, is of the order of the Debye screening mass $m_D$. During a time $t\sim m_D^{-1}$, the heavy quark moves a distance which is small compared to the size of the screening cloud, $\sim m_D^{-1}$. Thus, over a time scale characteristic of the plasma collective dynamics, the heavy quark positions are almost frozen (they are completely frozen in the limit $M\to\infty$). One can also recognize that the collisions of the heavy particles with the light constituents of the plasma involve the exchange of soft gluons, with typical momenta $q\lesssim m_D\ll M$. The corresponding energy transfer $\sim q^2/M\sim m_D^2/M$ is small on the scale of the plasma frequency, $m_D^2/M\ll m_D$. \subsection{Equation for the density matrix} The density matrix obeys the general equation of motion \beq i\frac{{\rm d} {\cal D}}{{\rm d} t}=[H,{\cal D}]. \eeq I order to treat the interaction between the plasma and the heavy particle using perturbation theory, we move to the interaction representation. We set $H=H_0+H_1$, with $H_0=H_Q+H_{\rm pl}$ and define \beq {\cal D}(t)=U_0(t,t_0) {\cal D}^I(t) U_0^\dagger(t,t_0), \eeq where ${\cal D}^I(t) $, the interaction representation of the density matrix, satisfies the equation \beq\label{exacteqintrepres0} \frac{{\rm d} {\cal D}^I}{{\rm d} t}&=&-i [H_1(t),{\cal D}^I(t)],\qquad H_1(t)=U_0(t,t_0)^\dagger H_1 U_0(t,t_0). \eeq Here, $H_1(t)$ denotes the interaction representation of $H_1$. The evolution operator in the interaction representation, $U_I(t_0,t)=U_0^\dagger(t,t_0) U(t,t_0)$, can be expanded in powers of $H_1(t)$ in the usual way \beq U_I(t,t_0)={\rm T}\exp\{ -i\int_{t_0}^t {\rm d} t' H_1(t') \}, \eeq where the symbol ${\rm T}$ denotes time ordering. Similarly, Eq.~(\ref{exacteqintrepres0}) can be integrated formally using the Schwinger-Keldysh contour \cite{Schwinger61,Keldysh64}: \beq\label{exprhoSK} {\cal D}^I(t)&=& U_I(t,t_0) {\cal D}(t_0) U_I^\dagger(t,t_0)\nonumber\\ &=&{\rm T}_{\cal C}\left[ \exp\left\{ -i\int_{\cal C} {\rm d} t_{_{\cal C}} H_1(t_{_{\cal C}}) \right\}{\cal D}(t_0) \right], \eeq where the operator ${\rm T}_{\cal C}$ orders the operators $ H_1(t_{_{\cal C}}) $ along the contour parameterized by the contour time $t_{_{\cal C}}$, with the operators carrying the largest $t_{_{\cal C}}$ coming before those with smaller $t_{_{\cal C}}$ (see Fig.~\ref{}). The upper branch of the contour, with time running from $t_0$ to $t$, represents the evolution operator $U_I(t,t_0)$, the lower branch of the contour, with time running from $t$ to $t_0$, represents the operator $U_I^\dagger (t,t_0)$. As can be seen in Eq.~(\ref{exprhoSK}), in the expansion of ${\cal D}^I(t)$ in powers of $H_1(t)$, the operators $H_1(t)$ that sit on the left of ${\cal D}(t_0)$ live on the upper branch of the contour, while those that appear on the right of ${\cal D}(t_0)$ live on the lower branch (they come later along the contour). \vspace{-0.0cm} \begin{figure}[!hbt] \begin{center} \includegraphics[width=1.1\textwidth]{SK-contour.pdf} \vspace{-7cm} \caption{The Schwinger-Keldysh contour} \vspace{-0.25in} \label{} \end{center} \end{figure} To proceed further, we assume that, at the initial time $t_0$, the density matrix factorizes \beq\label{factorizedD} {\cal D}^I(t_0)= {\cal D}_Q^I(t_0)\otimes {\cal D}_{\rm pl}^I(t_0). \eeq We also assume that at time $t_0$, the plasma is in a state of thermal equilibrium, so that its density matrix ${\cal D}^I_{\rm pl}(t_0)={\cal D}_{\rm pl}(t_0)$ is a canonical density matrix, \beq {\cal D}_{\rm pl}(t_0)=\frac{1}{Z_{\rm pl}} \sum_m {\rm e}^{-\beta E_m}, \eeq where $\beta=1/T$, with $T$ the equilibrium temperature. This factorization of the density matrix allows for a simple calculation of the trace over the plasma degrees of freedom. Let us then examine perturbation theory at second order in $H_1$, with $H_1$ given by Eq.~(\ref{H1}). Performing the trace over the plasma degrees of freedom is immediate, thanks to the specific structure of $H_1$ and the factorization of the density matrix at $t=t_0$. One obtains \beq\label{exprhoSK2b} && {\cal D}_Q^I(t)= {\cal D}_Q(t_0) -i\int_{t_0}^t {\rm d} t' \int_{\boldsymbol x} \langle A_0^a({\boldsymbol x})\rangle [n^a({\boldsymbol x},t), {\cal D}_Q(t_0)]\nonumber\\ &&-\frac{1}{2}\int_{t_0}^t {\rm d} t_1 \int_{t_0}^{t} {\rm d} t'_1\int_{{\boldsymbol x}\x'}\,{\rm T}[n^a(t_1,{\boldsymbol x}) n^b(t_1',{\boldsymbol x}')] {\cal D}_Q(t_0) \,\langle {\rm T}[A_0^a(t_1,{\boldsymbol x}) A_0^b(t_1',{\boldsymbol x}')] \rangle_0\nonumber\\ &&-\frac{1}{2}\int_{t_0}^t {\rm d} t_2 \int_{t_0}^{t} {\rm d} t_2'\int_{{\boldsymbol x}\x'}\, {\cal D}_Q(t_0)\tilde{\rm T}[n^a(t_2,{\boldsymbol x}) n^b(t'_2,{\boldsymbol x}')] \,\langle \tilde{\rm T}[A_0^a(t_2,{\boldsymbol x}) A_0^b(t'_2,{\boldsymbol x}')] \rangle_0\nonumber\\ &&+\int_{t_0}^t {\rm d} t_1 \int_{t_0}^{t} {\rm d} t_2\int_{{\boldsymbol x}\x'}\, [n^a(t_1,{\boldsymbol x}){\cal D}_Q(t_0) n^b(t_2,{\boldsymbol x}')] \,\langle A_0^a(t_2,{\boldsymbol x}')A_0^b(t_1,{\boldsymbol x}) \rangle_0 ,\nonumber\\ \eeq where, in the last three lines, we have used the convention that $t_1, t'_1$ run on the upper part of the contour, while $t_2,t'_2$ run on the lower branch. Note that the linear term vanishes since the plasma is color neutral (so that $ \langle A_0^a({\boldsymbol x})\rangle_0=0$). Here the notation $\langle\cdots \rangle_0$ stands for the average with the plasma equilibrium density matrix, that is \beq \langle \cdots\rangle_0={\rm Tr}_{\rm pl}\left[ \frac{1}{Z_{\rm pl}} {\rm e}^{-\beta H_{\rm pl}} \cdots\right]. \eeq Similarly the correlators of the gauge fields are diagonal in color, i.e. they are proportional to $\delta^{ab}$. These correlators are the exact correlators in the plasma (the fields are in the interaction representation, which corresponds to the Heisenberg representation when considering the plasma alone). They are written as \beq\label{correlators} \langle {\rm T}[A_0^a(t_1,{\boldsymbol x}) A_0^b(t_1', {\boldsymbol x}')] \rangle_0&=&-i\delta^{ab}\Delta(t_1-t_1',{\boldsymbol x}-{\boldsymbol x}')\nonumber\\ \langle\tilde {\rm T}[A_0^a(t_2,{\boldsymbol x}) A_0^b(t_2' ,{\boldsymbol x}')] \rangle_0&=&-i\delta^{ab}\tilde\Delta(t_2-t_2',{\boldsymbol x}-{\boldsymbol x}')\nonumber\\ \langle {\rm T}_C A_0^a(t_2,{\boldsymbol x}')A_0^b(t_1,{\boldsymbol x}) \rangle_0&=&\delta^{ab}\Delta^>(t_2-t_1,{\boldsymbol x}'-{\boldsymbol x})\nonumber\\ &=&\delta^{ab}\Delta^<(t_1-t_2,{\boldsymbol x}-{\boldsymbol x}'). \eeq The apparent inversion of the order of times in the last correlator results from the relation ${\rm Tr}_{\rm pl}A_0^b(t_1,{\boldsymbol x}) {\cal D}_{\rm pl}(t_0)A_0^a(t_2,{\boldsymbol x}')= \langle A_0^a(t_2,{\boldsymbol x}')A_0^b(t_1,{\boldsymbol x}) \rangle_0$ which follows from the cyclic invariance of the trace. It is convenient to represent the evolution of the density matrix by a diagram such as that in Fig.~\ref{}, where the upper and lower parts of the diagram may be associated to the corresponding upper and lower parts of the Schwinger-Keldysh contour. The explicit expression that this diagram represents is \beq \bra{\alpha_f} {\cal D}_Q(t)\ket{\beta_f}=\sum_{\alpha_i\beta_i}\bra{\alpha_f}U_I(t,t_0)\ket{\alpha_i}\bra{\alpha_i} {\cal D}_Q(t_0) \ket{\beta_i}\bra{\beta_i} U_I^\dagger(t,t_0)\ket{\beta_f}, \eeq where $\alpha$ of $\beta$ represent the set of quantum numbers that are necessary to specify the state of the heavy particles (essentially the position and color). \vspace{-0cm} \begin{figure}[!hbt] \begin{center} \includegraphics[width=0.8\textwidth]{evol_densitymatrix.pdf} \vspace{-2.5cm} \caption{Graphical representation of the evolution of the density matrix from $t_0$ to $t$. The horizontal lines represent the evolution operators $U_I(t,t_0)$ (upper branch) or $U_I^\dagger(t,t_0)$ (lower branch). When ${\cal D}_Q $ is the density matrix of a single heavy quark, these horizontal lines may be interpreted as the associated propagators of the heavy particle. When ${\cal D}_Q $ is the density matrix of a heavy quark-antiquark pair, a single horizontal line is replaced by a pair of lines, associated with the propagator of the pair (see Fig.~\ref{} below).} \vspace{-0.25in} \label{} \end{center} \end{figure} The diagrammatic interpretation of Eq.~(\ref{exprhoSK2b}) is then given in Fig.~\ref{} (for the case of a single particle density matrix). \vspace{-0cm} \begin{figure}[!hbt] \begin{center} \includegraphics[width=1\textwidth]{prederivative.pdf} \vspace{-5cm} \caption{These diagrams are in one-to-one correspondence with the terms in the last three lines of the right-hand side of Eq.~(\ref{exprhoSK2b}) for the single particle density matrix ${\cal D}_Q^I(t)$. } \vspace{-0.25in} \label{} \end{center} \end{figure} \\ In order to implement our further approximations, it is convenient to consider the time derivative of the density matrix. This can be obtained by taking the derivative of Eq.~(\ref{exprhoSK2b}) above (see Eq.~(\ref{exprhoSK2bder}) in \ref{}). But it is more instructive to return to Eq.~(\ref{exacteqintrepres0}), and rewrite it as \beq\label{exacteqintrepres1} \frac{{\rm d} {\cal D}^I}{{\rm d} t}=-i [H_1(t),{\cal D}^I(t_0)]-\int_{t_0}^t{\rm d} t' [H_1(t),[H_1(t'),{\cal D}^I(t')]]. \eeq This exact equation is obtained by formally integrating Eq.~(\ref{exacteqintrepres0}) and inserting the solution back into the equation. Perturbation theory at second order in $H_1$ is recovered by replacing ${\cal D}^I(t')$ by ${\cal D}^I(t_0)$ in the double commutator in the right hand side. One may then proceed to the average over the plasma degrees of freedom, as we did before, and get the following equation for the reduced density matrix ${\cal D}_Q$ \beq\label{eqrhoAt0b2b} \frac{{\rm d} {\cal D}^I_Q(t)}{{\rm d} t} &=&-\int_{t_0}^t {\rm d} t'\int_{{\boldsymbol x}\x'}\left( [n^a(t,{\boldsymbol x}), n^a(t',{\boldsymbol x}'){\cal D}^I_Q(t_0)]\Delta^>(t-t',{\boldsymbol x}-{\boldsymbol x}')\right.\nonumber\\ &&\qquad\qquad\quad \left. +[{\cal D}^I_Q(t_0) n^a(t',{\boldsymbol x}'),n^a(t,{\boldsymbol x})]\Delta^<(t-t',{\boldsymbol x}-{\boldsymbol x}')\right).\nonumber\\ \eeq We have used the fact that the linear term vanishes in a neutral plasma, and the sum over the color index $a$ is implicit. At this point, we can improve on strict perturbation theory. To do so we notice that the integral over $t'$ in Eq.~(\ref{exacteqintrepres1}) is in fact limited to a region near $t'\lesssim t$: this is because $\Delta(t-t')$ dies out when $t-t'\gtrsim m_D^{-1}$, and we are interested on the evolution of the density matrix over time scales that are much larger than $m_D^{-1}$. Thus, noticing also that the difference ${\cal D}^I(t)-{\cal D}^I(t')$ involves in any case an extra power of $H_1$, we replace ${\cal D}^I(t')$ by ${\cal D}^I(t)$ in the right hand side of Eq.~(\ref{exacteqintrepres1})\footnote{An alternative procedure, which leads to slightly different equations, is presented in \ref{}}, turning the equation into an equation for ${\cal D}(t)$ which is now local in time. We shall furthermore exploits the fact that the density matrix approximately factorizes at all times, as does the density matrix at the initial time $t_0$. Again, this is consistent with the weak coupling approximation since the correction to the factorized from necessarily involves additonal powers of $H_1$. The latter approximation allows us to perform the trace over the plasma degree of freedom, in the same way as we did earlier for strict perturbation theory. The resulting equation is in fact identical to Eq.~(\ref{eqrhoAt0b2b}) in which we replace in the right hand side ${\cal D}^I_Q(t_0)$by ${\cal D}^I_Q(t)$. It is convenient for the following to write this equation in the Schr\"{o}dinger picture. A simple calculation yields \beq &&\!\!\!\!\!\!\!\!\!\! \frac{{\rm d} {\cal D}_Q}{{\rm d} t}+i[H_Q,{\cal D}_Q(t)]=\nonumber\\ &&-\int_{{\boldsymbol x}\x'} \int_{0}^{t-t_0} {\rm d} \tau \,[n^a_{\boldsymbol x} ,U_Q(\tau)n^a_{{\boldsymbol x}'}U_Q^\dagger(\tau){\cal D}_Q(t)] \, \Delta^>(\tau;{\boldsymbol x}-{\boldsymbol x}'))\nonumber\\ &&-\int_{{\boldsymbol x}\x'}\int_{0}^{t-t_0} {\rm d} \tau \, [ {\cal D}_Q(t)U_Q(\tau)n^a_{{\boldsymbol x}'}U_Q^\dagger(\tau), n^a_{\boldsymbol x}] \,\Delta^<(\tau;{\boldsymbol x}-{\boldsymbol x}').\nonumber\\ \eeq where we have set $t-t'=\tau$. This equation has the same physical content as the influence functional derived in \cite{}, and it is based on analogous approximations. It relies on a weak coupling approximation, but goes beyond strict second order perturbation theory, in particular by resumming secular terms. This equation still contains a memory integral that we shall simplify thanks to our last approximation: In line with the fact that only small values of $\tau $ are relevant, it consists in replacing ${\rm e}^{-iH_Q\tau}\simeq 1-iH_Q\tau$, and keep terms up to linear order in $\tau$ in the integrals. More precisely, we write \beq U_Q(\tau)n_{{\boldsymbol x}'}^a U_Q^\dagger(\tau)=U_Q^\dagger(-\tau)n_{{\boldsymbol x}'}^a U_Q(-\tau)=n_{{\boldsymbol x}'}^a(-\tau) \eeq and \beq n_{{\boldsymbol x}'}^a(-\tau)=n_{{\boldsymbol x}'}^a-\tau \dot n_{{\boldsymbol x}'}^a,\qquad \dot n_{{\boldsymbol x}'}^a=i\left[H_Q,n_{{\boldsymbol x}'}^a\right], \eeq the time-dependence of $n_{{\boldsymbol x}'}(t)$ being given by the Heisenberg representation, $n_{{\boldsymbol x}'}^a(t)={\rm e}^{iH_Qt} n_{{\boldsymbol x}'}^a{\rm e}^{-iH_Qt}$. We get \beq\label{premain} \frac{{\rm d} {\cal D}_Q}{{\rm d} t}+i[H_Q,{\cal D}_Q(t)]&\approx& -\int_{{\boldsymbol x}\x'} [n^a_{\boldsymbol x},n^a_{{\boldsymbol x}'} {\cal D}_Q ] \, \int_{0}^{\infty} {\rm d} \tau \Delta^>(\tau;{\boldsymbol x}-{\boldsymbol x}'))\nonumber\\ &-&\int_{{\boldsymbol x}\x'} [ {\cal D}_Q n^a_{{\boldsymbol x}'}, n^a_{\boldsymbol x}] \,\int_{0}^{\infty} {\rm d} \tau \,\Delta^<(\tau;{\boldsymbol x}-{\boldsymbol x}')\nonumber\\ &+&\int_{{\boldsymbol x}\x'}[n^a_{\boldsymbol x}, \dot n^a_{{\boldsymbol x}'}{\cal D}_Q] \, \int_{0}^{\infty} {\rm d} \tau\,\tau\, \Delta^>(\tau;{\boldsymbol x}-{\boldsymbol x}'))\nonumber\\ &+&\int_{{\boldsymbol x}\x'}[ {\cal D}_Q \dot n^a_{{\boldsymbol x}'}, n^a_{\boldsymbol x}] \,\int_{0}^{\infty} {\rm d} \tau \,\tau\,\Delta^<(\tau;{\boldsymbol x}-{\boldsymbol x}').\nonumber\\ \eeq At this point we use the values of the time integrals given in \ref{}. These involve the zero frequency part of the time-order propagator $\Delta(\omega=0)=\Delta^R(\omega=0,{\boldsymbol r})+i\Delta^<(\omega=0,{\boldsymbol r})$, which we identify with the real and imaginary part of a ``potential''. More precisely, we set \beq V({\boldsymbol r})=-\Delta^R(\omega=0,{\boldsymbol r}), \qquad W({\boldsymbol r})=-\Delta^<(\omega=0,{\boldsymbol r}). \eeq This terminology stems from the fact that $V({\boldsymbol r})+iW({\boldsymbol r})$ plays the role of a complex potential in a Schr\"{o}dinger equation describing the relative motion of a quark-antiquark pair: the real part represents the screening corrections, and adds to the familiar interaction arising in leading order from one-gluon exchange, the imaginary part accounts effectively for the collisions between the heavy quarks and the plasma constituents \cite{}. After a simple calculation that uses the properties $V({\boldsymbol x}-{\boldsymbol x}')=V({\boldsymbol x}'-{\boldsymbol x})$ and $W({\boldsymbol x}-{\boldsymbol x}')=W({\boldsymbol x}'-{\boldsymbol x})$, we get \beq\label{main} \frac{{\rm d} {\cal D}_Q}{{\rm d} t}+i[H_Q,{\cal D}_Q(t)]&\approx&-\frac{i}{2}\int_{{\boldsymbol x}\x'} V({\boldsymbol x}-{\boldsymbol x}') [n_{\boldsymbol x}^a n^a_{{\boldsymbol x}'},{\cal D}_Q ],\nonumber\\ &+&\frac{1}{2}\int_{{\boldsymbol x}\x'} W({\boldsymbol x}-{\boldsymbol x}') \left( \{n^a_{\boldsymbol x} n^a_{{\boldsymbol x}'}, {\cal D}_Q\}-2 n^a_{{\boldsymbol x}} {\cal D}_Q n^a_{{\boldsymbol x}'}\right)\nonumber\\ &+&\frac{i}{4T}\int_{{\boldsymbol x}\x'}\, W({\boldsymbol x}-{\boldsymbol x}') \left( [n^a_{\boldsymbol x} , \dot n^a_{{\boldsymbol x}'}{\cal D}_Q] +[n^a_{\boldsymbol x} ,{\cal D}_Q\dot n^a_{{\boldsymbol x}'} ] \right) \nonumber\\ \eeq Note that first line of the right hand side of this equation describes a hamiltonian evolution, that is, it can be written as the commutator in the left hand side, with $H_Q $ replaced by $\frac{1}{2}\int_{{\boldsymbol x}\x'} V(x-{\boldsymbol x}') n_{\boldsymbol x}^an_{{\boldsymbol x}'}^a$. It follows that we can shift the ``direct'', one-gluon exchange potential initially contained in $H_Q$ into $V$, and keep in $H_Q$ only the kinetic energy of the heavy quarks. This is the strategy that was followed in \cite{} and that we shall adopt in this paper. In this way the potential $V({\boldsymbol r})$ becomes the screened Coulomb potential, and $H_Q$ represents only the kinetic energy of the heavy particles (see also the discussion after Eq.~(\ref{phaserealVqed}) below). The equation (\ref{main}) is our main equation. It is a fully quantum mechanical equation. It is a Markovian equation for the reduced density matrix ${\cal D}_Q(t)$. We shall write this equation in the following way \beq\label{main2} \frac{{\rm d} {\cal D}_Q}{{\rm d} t}={\cal L}\,{\cal D}_Q, \eeq with ${\cal L}={\cal L}_0+{\cal L}_1+{\cal L}_2+{\cal L}_3$, and \beq\label{main3} {\cal L}_0\,{\cal D}_Q&\equiv &-i[H_Q,{\cal D}_Q],\nonumber\\ {\cal L}_1\,{\cal D}_Q&\equiv& -\frac{i}{2}\int_{{\boldsymbol x}\x'} V({\boldsymbol x}-{\boldsymbol x}') [n^a_{\boldsymbol x} n^a_{{\boldsymbol x}'},{\cal D}_Q ], \nonumber\\ {\cal L}_2\,{\cal D}_Q&\equiv&\frac{1}{2}\int_{{\boldsymbol x}\x'} W({\boldsymbol x}-{\boldsymbol x}') \left( \{n^a_{\boldsymbol x} n^a_{{\boldsymbol x}'}, {\cal D}_Q\}-2 n^a_{{\boldsymbol x}} {\cal D}_Q n^a_{{\boldsymbol x}'}\right), \nonumber\\ {\cal L}_3\,{\cal D}_Q&\equiv&\frac{i}{4T}\int_{{\boldsymbol x}\x'}\, W({\boldsymbol x}-{\boldsymbol x}') \left( [n^a_{\boldsymbol x} , \dot n^a_{{\boldsymbol x}'}{\cal D}_Q] +[n^a_{\boldsymbol x} ,{\cal D}_Q\dot n^a_{{\boldsymbol x}'} ] \right) . \eeq The structure of Eq.~(\ref{main2}) is close to that of a Lindblad equation \cite{Lindblad}, but Eq.~(\ref{main2}) is not quite a Lindblad equation: while the operator ${\cal L}_2$ can be put in the Lindblad form, this is not so of the operator ${\cal L}_3$, unless one adds extra, subleading terms (see the discussion in \ref{}). For a recent discussion of the Lindblad equation for an abelian plasma, in a formalism not too different from the present one, see \cite{}). The notation is, at this stage, symbolic and just expresses the fact that the right hand side of Eq.~(\ref{main}) is a linear functional of the density matrix. It will acquire a more precise meaning as we proceed. We may however make the following observation. When taking matrix elements between localized states, specified by the coordinates of the heavy particles, the density operators $n_{\boldsymbol x}$ play the role of projection operators, and are diagonal in the coordinate representation. Thus the same matrix elements, as far as the coordinates are concerned, will appear on the left and the right. The operator ${\cal L}$ will then appear as a differential operator acting on this matrix element (in fact ${\cal L}_1$ and ${\cal L}_2$ are simply multiplicative, as we shall see). It is convenient to associate a diagrammatic representation of the various contributions that we shall calculate. The relevant diagrams will preserve the topological structures of those already introduced, but because of the various approximations that we have performed, they cannot be calculated with standard rules. As an illustration, we display in Fig.~\ref{} diagrams corresponding to the time derivative of the single particle density matrix (diagrams corresponding to the two particle density matrix are displayed in Fig.~\ref{} below). All interactions in Eq.~(\ref{main}) have become instantaneous. For this reason, we draw these as vertical gluon lines, or as tadpole insertions, located anywhere between $t-\tau$ and $t$. Note that terms where the two densities sit close together in Eq.~(\ref{main2}), like in $[n^a_{\boldsymbol x} n^a_{{\boldsymbol x}'}, {\cal D}_Q]$, correspond to diagrams where the two ends of the gluon is hooked on the upper (or lower) part of the diagram, while a term such as $n^a_{{\boldsymbol x}} {\cal D}_Q n^a_{{\boldsymbol x}'}$ corresponds to a gluon joining the upper and lower parts of the diagram. Since, as we shall see, in QCD these two types of terms correspond also to different color structures, we shall find convenient to split the operators ${\cal L}_i$ into two components, ${\cal L}_i={\cal L}_{ia}+{\cal L}_{ib}$, with for instance ${\cal L}_{2a}\propto \{n^a_{\boldsymbol x} n^a_{{\boldsymbol x}'}, {\cal D}_Q\}$ and ${\cal L}_{2b}\propto n^a_{{\boldsymbol x}} {\cal D}_Q n^a_{{\boldsymbol x}'}$. Note that ${\cal L}_1$ has only contributions of type $a$, i.e., ${\cal L}_1={\cal L}_{1a}$.\\ \vspace{-0.0cm} \begin{figure}[!hbt] \begin{center} \includegraphics[width=0.9\textwidth]{derivative_rho_one1a.pdf} \vspace{-3.5cm} \caption{These diagrams illustrate generic processes taken into account in Eq.~(\ref{main3}), in the case of the single particle density matrix. Note that there is another diagram with a tadpole insertion, on the lower line, not drawn. Depending on the operator considered, the propagator of the gluon corresponds to $V$, $W$ or involves spatial derivatives of $W$. Note that since we treat the heavy quarks and antiquarks as non relativistic particles, the direction of the arrows in such a diagram does not refer to the nature (quark or antiquark) of the heavy particle: rather, it is correlated to the SK contour time, forward (to the right) in the upper branch, backward (to the left) in the lower branch.} \vspace{-0.25in} \label{} \end{center} \end{figure} In the rest of this paper, we shall deal only with the heavy quark reduced density matrix. We shall then drop the subscript $Q$ in order to simplify the notation and write simply ${\cal D}$ in place of ${\cal D}_Q$. \section{Semi-classical approximation for abelian plasmas} \label{} The equation (\ref{main}) is quite general. It holds for any system of heavy quarks and antiquarks. Depending on the system considered, the color density $n^a({\boldsymbol x})$ and the density matrix ${\cal D}$ take different forms. In this section, we study the specific form of Eq.~(\ref{main}), and the associated operators ${\cal L}_i$ in Eq.~(\ref{main3}), for the single particle and the two particle density matrices, in the case of an electromagnetic (abelian) plasma. For simplicity, we shall continue to refer to the charged particles as quarks (positive charge) and antiquarks (negative charge). The interaction hamiltonian reads as in Eq.~(\ref{H1}), with $n^a({\boldsymbol x})$ replaced by the density of charged particles. Our goal here is twofold: i) This section is a preparation for the more complicated case of non abelian plasmas presented in the next section. Some of the results obtained here will indeed extend trivially to QCD, to within multiplicative color factors. ii) We wish to establish the relation with results obtained previously for the influence functional obtained in the path integral formalism. In particular, we shall show that we obtain, after a semi-classical approximation, the same Fokker-Planck equations, and the associated Langevin equations, as derived earlier in the path integral formalism in Ref.~\cite{}. \subsection{Single particle density matrix} In first quantization, the charge density $n({\boldsymbol x})$ of a single heavy quark is the operator $n({\boldsymbol x})=\delta({\boldsymbol x}-\hat{\boldsymbol r})$, with matrix elements \beq \bra{{\boldsymbol r}}n({\boldsymbol x})\ket{{\boldsymbol r}'}=\delta({\boldsymbol x}-{\boldsymbol r})\delta({\boldsymbol r}-{\boldsymbol r}'). \eeq We also need the matrix elements of the time derivative of the density. These can be easily obtained from the continuity equation, $\dot n({\boldsymbol x})=-\mbox{\boldmath$\nabla$}_{\boldsymbol x} {\boldsymbol j}({\boldsymbol x})$, where the matrix elements of the current ${\boldsymbol j}({\boldsymbol x})$ are given by \beq \bra{{\boldsymbol r}}{\boldsymbol j}({\boldsymbol x})\ket{{\boldsymbol r}'}=\frac{1}{2iM} [\mbox{\boldmath$\nabla$}_{\boldsymbol r} \delta({\boldsymbol r}-{\boldsymbol r}')][\delta({\boldsymbol x}-{\boldsymbol r}')+\delta({\boldsymbol x}-{\boldsymbol r})]. \eeq One then gets \beq\label{divjx} \bra{{\boldsymbol r}}\dot n({\boldsymbol x})\ket{{\boldsymbol r}'}=-\frac{1}{2iM}\left\{\left[\mbox{\boldmath$\nabla$}_{\boldsymbol r} \delta({\boldsymbol r}-{\boldsymbol r}')\right] \cdot\mbox{\boldmath$\nabla$}_{\boldsymbol x} [\delta({\boldsymbol x}-{\boldsymbol r}')+\delta({\boldsymbol x}-{\boldsymbol r})]\right\}. \eeq \\ We can then proceed to the evaluation of the various contributions ${\cal L}_i$ in Eq.~(\ref{main3}), first in the case of the single particle density matrix. It is easy to show that the first line of Eq.~(\ref{main3}) yields \beq \bra{{\boldsymbol r}} {\cal L}_1{\cal D}\ket{{\boldsymbol r}'}=-\frac{i}{2} \int_{{\boldsymbol x},{\boldsymbol x}'} V({\boldsymbol x}-{\boldsymbol x}')\bra{{\boldsymbol r}}\left[n({\boldsymbol x})n({\boldsymbol x}'), {\cal D}\right]\ket{{\boldsymbol r}'}=0. \eeq Thus, the real part of the potential does not contribute to the evolution of the single particle density matrix. In terms of diagrams, this results from the cancellation of the tadpole insertions in the upper and lower branches (see the second diagram in Fig.~\ref{}), which represent here (unphysical) self-interactions. \\ Taking the matrix element of the second line of Eq.~(\ref{main3}), one obtains \beq \bra{{\boldsymbol r}} {\cal L}_2{\cal D}\ket{{\boldsymbol r}'}=[W(0)-W({\boldsymbol r}-{\boldsymbol r}') ]\bra{{\boldsymbol r}}{\cal D}\ket{{\boldsymbol r}'}=-\Gamma({\boldsymbol r}-{\boldsymbol r}') \bra{{\boldsymbol r}}{\cal D}\ket{{\boldsymbol r}'}, \eeq where we have set \beq\label{} \Gamma({\boldsymbol r})\equiv W({\boldsymbol r})-W(0). \eeq This equation illustrates the role of the collisions, captured here by the imaginary part of the potential, in the phenomenon of decoherence (the damping of the off-diagonal matrix elements of the density matrix). In contrast to what happens with the real part of the potential that we have just discussed, in the present case the two tadpole contributions add up, instead of cancelling. They are in fact needed to properly define the damping rate $\Gamma$, and insure in particular that it cancels when ${\boldsymbol r}\to 0$, so that the density (the diagonal part of the density matrix) is not affected by the collisions. A useful estimate of $\Gamma({\boldsymbol r})$ is obtained in the Hard Thermal Loop approximation \cite{} which gives \beq \Gamma({\boldsymbol r})=\alpha T \phi(m_D r), \eeq where $T$ is the temperature, and $\phi(x)$ a monotonously increasing function such that $\phi(x=0)=0$ and $\phi(x\to \infty)=1$ \cite{}. The same formula holds in the case of QCD, with $\alpha$ replaced by $\alpha_s$, the strong coupling constant, and the multiplication by appropriate color factors (see Sect.~\ref{numerics}). In the limit of a large separation, $\Gamma({\boldsymbol r})\simeq 2\gamma_Q$, where $\gamma_Q=\alpha_sT/2$ is the so-called damping factor of a heavy quark (or antiquark) \cite{}. At small separation, interferences cancel the effect of collisions: the heavy quark pair is seen then as a small electric dipole, i.e., an electrically neutral object on the scale of the wavelengths of the typical modes of the plasma. Note that at large separation, the imaginary part of the potential itself vanishes, so that $W(0)=-2\gamma_Q$. \\ Considering finally the third line of Eq.~(\ref{main3}) one gets\footnote{Here, and throughout this paper, we use the shorthand $\nabla W(0)$ for $\left.\nabla_x W(x)\right|_{x=0}$, and similarly for $\nabla^2 W(0)$.} \beq\label{eqforrhoQEDM1} \bra{{\boldsymbol r}} {\cal L}_3{\cal D}\ket{{\boldsymbol r}'}&=&\frac{1}{4MT}\left[ \nabla^2 W(0) -\nabla^2W({\boldsymbol r}-{\boldsymbol r}') \right]\bra{{\boldsymbol r}}{\cal D}\ket{{\boldsymbol r}'}\nonumber\\ &-&\frac{1}{4MT} \nabla_{\boldsymbol r} W({\boldsymbol r}-{\boldsymbol r}')\cdot (\nabla_{\boldsymbol r} -\nabla_{{\boldsymbol r}'}) \bra{{\boldsymbol r}}{\cal D}\ket{{\boldsymbol r}'}. \eeq The spatial derivatives originate from the time derivatives of the density (see Eq.~(\ref{divjx})), which involve the velocity of the heavy quark (hence the factor $1/M$). In fact, there is a close correspondence between ${\cal L}_3$ and ${\cal L}_2$. Observe indeed that ${\cal L}_3$ can be obtained from ${\cal L}_2$ by multiplying the latter by the overall factor $1/(4MT)$, and performing the following substitutions: $W(0)\to \nabla^2 W(0)$, $W({\boldsymbol r}-{\boldsymbol r}')\to \nabla_{\boldsymbol r} W({\boldsymbol r}-{\boldsymbol r}')\cdot (\nabla_{\boldsymbol r} -\nabla_{{\boldsymbol r}'})$. We shall see that analogous correspondences also exist in the more complicated case of the 2 particle density matrix. \\ At this point, we make the following change of variables \beq \label{newvariables0} {\mbox{\boldmath${\cal R}$}}=\frac{{\boldsymbol r}+{\boldsymbol r}'}{2},\quad {\boldsymbol y}={\boldsymbol r}-{\boldsymbol r}', \eeq and set \beq \bra{{\boldsymbol r}}{\cal D}(t)\ket{{\boldsymbol r}'}={\cal D}({\mbox{\boldmath${\cal R}$}},{\boldsymbol y},t). \eeq The equation (\ref{main}) becomes then $\frac{{\rm d} }{{\rm d} t}{\cal D} ({\mbox{\boldmath${\cal R}$}},{\boldsymbol y},t)={\cal L} {\cal D} ({\mbox{\boldmath${\cal R}$}},{\boldsymbol y},t)$, with ${\cal L}$ appearing now explicitly as a differential operator acting on the function ${\cal D}({\mbox{\boldmath${\cal R}$}},{\boldsymbol y},t)$: \beq\label{eqforrhoQEDM2} {\cal L}= \frac{i}{M}\nabla_{\mbox{\boldmath${\cal R}$}}\cdot\nabla_{\boldsymbol y} -\Gamma({\boldsymbol y})+\frac{1}{4MT}\left[ \nabla^2 W(0) -\nabla^2_{\boldsymbol y} W({\boldsymbol y})-2\nabla_{\boldsymbol y} W({\boldsymbol y})\cdot\nabla_{\boldsymbol y} \right]. \eeq The first term arises from the kinetic energy, i.e., it represents ${\cal L}_0$. Note that the other terms, which represent the effect of the collisions, vanish for $y=0$, in particular thanks to the property $\nabla W(0)=0$. As already mentioned, this reflects the fact that the collisions do not change the local density of heavy quarks. \\ The equation (\ref{eqforrhoQEDM2}) above is the explicit form of the operators ${\cal L}_i$ in Eq.~(\ref{main3}) for the density matrix of a single heavy quark (in the abelian case). It has been obtained without any additional approximation beyond those leading to Eq.~(\ref{main3}). We may proceed further and simplify Eq.~(\ref{eqforrhoQEDM2}) by performing a small ${\boldsymbol y}$ expansion. The variable ${\boldsymbol y}$ measures by how much the density matrix deviates from a diagonal matrix, a situation which is reached in the classical limit. Thus, the small ${\boldsymbol y}$ expansion may be viewed as a semi-classical expansion. We have \beq W({\boldsymbol y})=W(0)+\frac{1}{2} {\boldsymbol y}\cdot {\cal H}(0) \cdot {\boldsymbol y} +\cdots \eeq where ${\cal H}(0)$ is the (positive definite) Hessian matrix of $W$, \beq {\cal H}_{ij}({\boldsymbol y})\equiv \frac{\partial^2 W({\boldsymbol y})}{\partial y_i\partial y_j}, \eeq evaluated at ${\boldsymbol y}=0$, and we have used $\partial_{\boldsymbol y} W(y)|_{{\boldsymbol y}=0}=0$. Note that if we stop the expansion of $W({\boldsymbol y})$ at quadratic order, $\nabla^2 W(0) -\nabla^2_{\boldsymbol y} W({\boldsymbol y})=0$. The differential operator (\ref{eqforrhoQEDM2}) reads then \beq \label{eqforrhoQEDM4} {\cal L}= \frac{i}{M} \nabla_{\cal R}\cdot\nabla_{\boldsymbol y} -\frac{1}{2} {\boldsymbol y}\cdot {\cal H}(0) \cdot {\boldsymbol y} -\frac{1}{2MT}{\boldsymbol y} \cdot {\cal H}(0) \cdot\nabla_{\boldsymbol y}. \eeq At this point some comments on the order of magnitude of the various terms are appropriate. It is convenient to measure the time in terms of the inverse temperature, setting for instance $\tau=T\, t$. Dividing both members of the equation by $T$, on gets on the left hand side $\partial_\tau$, and the operator ${\cal L}/T$ on the right hand side is dimensionless. We shall assume in this paper that the heavy particles are initially close to rest. In interacting with the medium they ultimately thermalize, their velocity reaching values of order $\sqrt{T/M}$, so that $\nabla_{\cal R}\lesssim \sqrt{MT}$. The variable ${\boldsymbol y}$ measures the non locality of the density matrix. When the heavy quark is not too far from equilibrium, this non locality is of the order of the thermal wavelength, that is ${\cal D}({\cal R},{\boldsymbol y},t)$ dies out when $y\gtrsim \lambda_{\rm th}\sim 1/\sqrt{MT}$. Thus in the first term, typically $\nabla_y\sim \sqrt{MT}$, so that $\nabla_{\cal R}\cdot\nabla_{\boldsymbol y}\sim MT$. It follows that the term ${\cal L}_0/T$, where ${\cal L}_0$ represents the kinetic energy of the heavy quark, is of order unity, while the other two terms are both of order $\Gamma({\boldsymbol y})$. The range of variation of $\Gamma({\boldsymbol y})$ is controlled by the Debye mass, i.e., it varies little on the scale of the thermal wavelength of the heavy particles. More precisely, using the HTL estimate $\Gamma({\boldsymbol r})\approx \alpha T (m_D r)^2$, we get $\Gamma({\boldsymbol y})/T\approx \alpha m_D^2/(MT)\ll 1$, the inequality following from our assumption $M\gg T$, and the fact that $m_D\lesssim T$ (in strict weak coupling $m_D^2\approx \alpha T^2$). In summary, the ratio of the last two terms in Eq.~(\ref{eqforrhoQEDM4}) to the kinetic term is of order $\alpha m_D^2/(MT)\ll 1$, which justifies the semi-classical expansion. \\ To see better the physical content of Eq.~(\ref{eqforrhoQEDM4}), we take its Wigner transform with respect to ${\boldsymbol y}$. We define, with a slight abuse of notation, \beq {\cal D} ({\mbox{\boldmath${\cal R}$}},{\boldsymbol p},t)=\int {\rm d}^3{\boldsymbol y} \,{\rm e}^{-i{\boldsymbol p}\cdot{\boldsymbol y}}\, {\cal D} ({\mbox{\boldmath${\cal R}$}},{\boldsymbol y},t), \eeq and obtain \beq \label{eqforrhoQEDM4W} {\cal L}= -\frac{{\boldsymbol p}}{M} \cdot \mbox{\boldmath$\nabla$}_{\mbox{\boldmath${\cal R}$}} +\frac{1}{2} \mbox{\boldmath$\nabla$}_{\boldsymbol p}\cdot {\cal H}(0) \cdot \mbox{\boldmath$\nabla$}_{\boldsymbol p} +\frac{1}{2MT}\mbox{\boldmath$\nabla$}_{\boldsymbol p} \cdot {\cal H}(0) \cdot{\boldsymbol p} . \eeq The corresponding equation for ${\cal D} ({\cal R},{\boldsymbol p},t)$ may be interpreted as a Fokker-Planck equation. The second term in Eq.~(\ref{eqforrhoQEDM4W}), proportional to $\mbox{\boldmath$\nabla$}_{\boldsymbol p}^2$ can be viewed as a diffusion term (in momentum space), and is associated with a noise term in the corresponding Langevin equation (see below). It originates from the contribution ${\cal L}_2$. The last term, steming from the opeartor ${\cal L}_3$, can be associated with friction. This can be made more transparent by introducing the following definitions \beq\label{noisedef} {\cal H}_{ij}(0)=\frac{1}{3} \nabla^2 W(0)\,\delta_{ij}\equiv \eta\,\delta_{ij},\qquad \eta=2\gamma T. \eeq Then we operator above yields the followwing Fokker-Planck equation \beq \left( \frac{\partial}{\partial t}+{\boldsymbol v}\cdot\mbox{\boldmath$\nabla$}_{\mbox{\boldmath${\cal R}$}} \right) {\cal D} ({\mbox{\boldmath${\cal R}$}},{\boldsymbol p},t)=\frac{1}{2}\eta \,\nabla_{\boldsymbol p}^2{\cal D} ({\mbox{\boldmath${\cal R}$}},{\boldsymbol p},t)+\gamma \,\mbox{\boldmath$\nabla$}_{\boldsymbol p}\cdot \left({\boldsymbol v} \, {\cal D} ({\mbox{\boldmath${\cal R}$}},{\boldsymbol p},t) \right), \eeq where ${\boldsymbol v}\equiv{\boldsymbol p}/M$ is the velocity of the particle. It is easily shown that this equation can be obtained from the following Langevin equation \beq M\ddot {\mbox{\boldmath${\cal R}$}} =-\gamma \dot {\mbox{\boldmath${\cal R}$}} +\mbox{\boldmath$\xi$}(t), \qquad \langle \xi_i(t)\xi_j(t')\rangle=\eta\, \delta_{ij}\delta(t-t'). \eeq The relation $\eta=2\gamma T$ between the diffusion constant $\eta$ and the fiction coefficient $\gamma$ can be viewed as an Einstein equation and expresses the fact that both noise and friction have the same origin, as can be made obvious by rewriting Eq.~(\ref{eqforrhoQEDM4W}) as follows \beq {\cal L}= -{\boldsymbol v} \cdot \mbox{\boldmath$\nabla$}_{\cal R} +\frac{1}{2} \mbox{\boldmath$\nabla$}_{\boldsymbol p}\cdot {\cal H}(0) \cdot \left( \mbox{\boldmath$\nabla$}_{\boldsymbol p} +\frac{{\boldsymbol v}}{T}\right). \eeq \subsection{The two particle density matrix} We consider now a heavy quark-antiquark pair. The charge density operator is written as \beq\label{rhoadef0qed2} n({\boldsymbol x})=\delta({\boldsymbol x}-\hat{\boldsymbol r})\otimes \mathbb{I} -\mathbb{I}\otimes\delta({\boldsymbol x}-\hat{\boldsymbol r}), \eeq where the first term refers to the quark and the second to the antiquark, the minus sign reflecting the fact that the antiquark has a charge opposite to that of the quark. The matrix elements of $n({\boldsymbol x})$ are given by \beq \bra{{\boldsymbol r}_1{\boldsymbol r}_2} n({\boldsymbol x})\ket{{\boldsymbol r}_1'{\boldsymbol r}_2'}=\delta({\boldsymbol r}_1-{\boldsymbol r}_1')\delta({\boldsymbol r}_2-{\boldsymbol r}_2')\left[\delta({\boldsymbol x}-{\boldsymbol r}_1)-\delta({\boldsymbol x}-{\boldsymbol r}_2)\right]. \eeq Similarly, the matrix elements of the time derivative of the density are given by \beq &&\bra{{\boldsymbol r}_1,{\boldsymbol r}_2}\dot n({\boldsymbol x})\ket{{\boldsymbol r}_3,{\boldsymbol r}_4}\nonumber\\ &&\qquad=-\frac{1}{2iM} [\mbox{\boldmath$\nabla$}_{{\boldsymbol r}_1}\delta({\boldsymbol r}_{13})]\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol x}[\delta({\boldsymbol x}-{\boldsymbol r}_3)+\delta({\boldsymbol x}-{\boldsymbol r}_1)] \delta({\boldsymbol r}_{24})\nonumber\\ &&\qquad\quad+\frac{1}{2iM} [\mbox{\boldmath$\nabla$}_{{\boldsymbol r}_2}\delta({\boldsymbol r}_{24})]\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol x}[\delta({\boldsymbol x}-{\boldsymbol r}_4)+\delta({\boldsymbol x}-{\boldsymbol r}_2)]\delta({\boldsymbol r}_{13}), \eeq which is easily obtained from Eq.~(\ref{divjx}). Note that we have introduced here a short notation, ${\boldsymbol r}_{ij}\equiv {\boldsymbol r}_i-{\boldsymbol r}_j$, that will be used often in the following. We shall also occasionally write $\nabla_1$ for $\nabla_{{\boldsymbol r}_1}$, and introduce similar other shorthands, in order to reduce the size of some formulae. \vspace{-0.cm} \begin{figure}[!hbt] \begin{center} \includegraphics[width=0.8\textwidth]{coordinates} \vspace{-2.5cm} \caption{The various coordinates that are used in the evaluation of the two particle density matrix elements $\bra{{\boldsymbol r}_1{\boldsymbol r}_2} {\cal D}_Q\ket{{\boldsymbol r}_1'{\boldsymbol r}_2'}$.} \vspace{-0.25in} \label{} \end{center} \end{figure} \vspace{+0.5cm} It will be also convenient at a later stage to change variables. Thus, we define the center of mass and relative coordinates, \beq\label{changeofvar1} {\boldsymbol R}=\frac{{\boldsymbol r}_1+{\boldsymbol r}_2}{2},\quad {\boldsymbol s}={\boldsymbol r}_1-{\boldsymbol r}_2,\quad {\boldsymbol R}'=\frac{{\boldsymbol r}_1'+{\boldsymbol r}_2'}{2},\quad {\boldsymbol s}'={\boldsymbol r}_1'-{\boldsymbol r}_2', \eeq as well as the set of coordinates that generalize those introduced in Eq.~(\ref{newvariables0}) for the single particle case, \beq\label{changeofvar5} {\mbox{\boldmath${\cal R}$}}=\frac{{\boldsymbol R}+{\boldsymbol R}'}{2},\qquad {\boldsymbol Y}={\boldsymbol R}-{\boldsymbol R}',\qquad {\boldsymbol y}={\boldsymbol s}-{\boldsymbol s}',\qquad {\boldsymbol r}=\frac{{\boldsymbol s}+{\boldsymbol s}'}{2}. \eeq The latter are useful to derive the semi-classical approximation. In this approximation, $Y\to 0, y\to 0$, and ${\mbox{\boldmath${\cal R}$}}$ and ${\boldsymbol r}$ become respectively the center of mass and the relative coordinates. These various coordinates are illustrated in Fig.~\ref{ We now turn to the specific evaluation of the matrix elements of Eq.~(\ref{main}) in the case of a quark-antiquark pair. Consider first the matrix element of the free evolution, governed by the hamiltonian \beq H_Q=\frac{p_1^2}{2M}+\frac{p_2^2}{2M}. \eeq We have \beq -i\bra{{\boldsymbol r}_1 {\boldsymbol r}_2} [H_Q,{\cal D}]\ket{{\boldsymbol r}_1'{\boldsymbol r}_2'}=i\left(\frac{\nabla_{\cal R}\cdot \nabla_{\boldsymbol Y}}{2M}+\frac{2\nabla_{\boldsymbol r}\cdot\nabla_{\boldsymbol y}}{M} \right) \bra{{\boldsymbol r}_1 {\boldsymbol r}_2} {\cal D}\ket{{\boldsymbol r}_1'{\boldsymbol r}_2'}, \eeq that is \beq {\cal L}_0=i\left(\frac{\nabla_{\cal R}\cdot \nabla_{\boldsymbol Y}}{2M}+\frac{2\nabla_{\boldsymbol r}\cdot\nabla_{\boldsymbol y}}{M} \right) . \eeq Turning now to the operator ${\cal L}_1$, a simple calculation yields \beq\label{phaserealVqed} \bra{{\boldsymbol r}_1{\boldsymbol r}_2}{\cal L}_1\,{\cal D}\ket{{\boldsymbol r}_1'{\boldsymbol r}_2'}&=&i [V({\boldsymbol r}_{12})-V({\boldsymbol r}_{1'2'})]\bra{{\boldsymbol r}_1{\boldsymbol r}_2}{\cal D}\ket{{\boldsymbol r}_1'{\boldsymbol r}_2'}. \eeq Note the cancellation of the self-interaction terms, as was the case for the single particle density matrix. The real part of the potential produces just a phase in the evolution of the density matrix. This can be understood as a hamiltonian evolution, ${\cal L}\,{\cal D}=-i [H,{\cal D}]$, with here $H\to -V$, the minus sign resulting from the fact that the two interacting heavy particles have opposite charges. As we have mentioned earlier (see the discussion after Eq.~(\ref{main})), the structure of the equation makes it possible to include in the potential $V$ both the ``direct'' interaction between the heavy quarks, by which we mean the interaction that exists in the absence of the plasma, as well as the ``induced'' interaction that results from the interaction of the heavy quarks with the plasma constituents. The latter is responsible for the screening phenomenon. In the HTL approximation, we have \beq V(r)=\alpha m_D +\alpha \frac{e^{-m_Dr}}{r}, \eeq where the first term cancels the constant contribution hidden in the screened Coulomb potential (the second term) at short distance. Thus as $r\to 0$, $V(r)$ reduces to the Coulomb potential, $V(r)\sim \alpha/r$. Note that the function $V({\boldsymbol r})$ thus defined corresponds to the interaction potential of two particles with identical charges. \\ \vspace{-0.0cm} \begin{figure}[!hbt] \begin{center} \includegraphics[width=1.0\textwidth]{derivative_rho1.pdf} \vspace{-3cm} \caption{Graphical illustrations for typical contributions to the operators ${\cal L}_i$ for the two-particle density matrix. In the first two diagrams, the gluon line represents either $V$ or $W$, while in the last two, only $W$ and its spatial derivatives are involved (the hamiltonian evolution, involving the real part of the potential, does not connect the upper and lower parts of the diagrams). In the last diagram, the gluon line connect two particles with the same charge, and contribute to the quantity called $W_a$. In the third diagram, the gluon line connect two particles with opposite charges, and contributes to the quantity called $W_a$. When $W$ is involved in the first diagram, it represents a contribution to $W_c$, and finally the tadpole insertion in the second diagram is associated to $V(0)$, or to $W(0)$ and its spatial drivatives. The defiitions of $W_{a,b,c}$ are given in Eq.~(\ref{Wabcdef}). } \vspace{-0.0cm} \label{} \end{center} \end{figure} By taking the matrix element of the second line of Eq.~(\ref{main3}), we obtain \beq\label{eqWlinear} \bra{{\boldsymbol r}_1{\boldsymbol r}_2}{\cal L}_2\,{\cal D}\ket{{\boldsymbol r}_1'{\boldsymbol r}_2'} = [2W(0)-W_{12}-W_{1'2'}- W^-]\bra{{\boldsymbol r}_1{\boldsymbol r}_2}{\cal D}\ket{{\boldsymbol r}_1'{\boldsymbol r}_2'}. \eeq The various terms in this expression correspond to the various ways the exchanged gluon can be hooked on the upper and lower lines. To simplify the bookkeeping of the various contributions, and the writing of the equations, we define the following quantities, which will often appear in forthcoming formulae: \beq\label{Wabcdef} &&W_a\equiv W_{11'}+W_{22'}, \qquad W_b\equiv W_{21'}+W_{12'}\nonumber\\ &&W_c\equiv W_{12}+W_{1'2'},\qquad W^\pm \equiv W_a\pm W_b \eeq These quantities correspond to the diagrams in Fig.~\ref{}, where $W$ plays the role of the propagator: $W_a$ connects particles with the same charge in the bra and the ket, while $W_b$ connect particles with opposite charges; $W_c$ connects particles within the bra, or within the ket. In the infinite mass limit, ${\boldsymbol r}_1={\boldsymbol r}_1'$, ${\boldsymbol r}_2={\boldsymbol r}_2'$ and $W_a=2W(0)$, $W_b=W_c=2W({\boldsymbol r})$, and $W_-({\boldsymbol r})=-2\Gamma({\boldsymbol r})$. Note that $2W(0)-W_{12}-W_{1'2'}=-\Gamma_{12}-\Gamma_{1'2'}$, where $\Gamma({\boldsymbol r})$ is defined in Eq.~(\ref{}). As was the case for the single particle density matrix, the collisions tend to equalize the coordinates (here the relative coordinates) in the ket and in the bra, bringing the density matrix to the diagonal form. The structure of the entire damping factor takes actually a form more complicated than in the case of the single particle density matrix. The combination of terms in the right hand side of Eq.~(\ref{eqWlinear}) can indeed be written \beq 2W(0)-W_{12}-W_{1'2'}- W^-= -\left( \Gamma_{12}+ \Gamma_{1'2'}+ \Gamma_{11'}+ \Gamma_{22'}- \Gamma_{12'}- \Gamma_{21'} \right). \eeq Note that the entire contribution vanishes when ${\boldsymbol r}_1'={\boldsymbol r}_1$ and ${\boldsymbol r}_2'={\boldsymbol r}_2$: $\Gamma_{11'}\to \Gamma(0)=0$, and similarly for $\Gamma_{22'}$ while the other terms mutually cancel. This is of course related to the fact that the collisions do not change the local density of heavy particles, as we have already discussed. For future reference, we write ${\cal L}_2$ as a sum of two contributions (as explained at the end of Sect.~\ref{}), and write \beq\label{L2aQED} {\cal L}_{2a}^{\rm QED}=2W(0)-W_c, \qquad {\cal L}_{2b}^{\rm QED}=-W^- . \eeq The diagonal elements (${\boldsymbol r}_1'={\boldsymbol r}_1,{\boldsymbol r}_2'={\boldsymbol r}_2$) of ${\cal L}_{2a}$ and ${\cal L}_{2b}$ mutually cancel, as we have seen. \\ Finally, we turn to the $1/M$ corrections, which are more involved. The calculations are straightforward, but lengthy. However, as we shall see, the results are simply related to those obtained for ${\cal L}_2$. Again, we split ${\cal L}_3$ into two contributions, ${\cal L}_3={\cal L}_{3a}+{\cal L}_{3b}$. We obtain \beq\label{QEDanticom} {\cal L}_{3a}^{\rm QED}&=&-\frac{i}{8T}\int_{{\boldsymbol x}\x'}\, W({\boldsymbol x}-{\boldsymbol x}') \left( 2{\cal D}\dot n_{{\boldsymbol x}'} n_{\boldsymbol x} -2n_{\boldsymbol x} \dot n_{{\boldsymbol x}'}{\cal D} \right)\nonumber\\ &=&\frac{1}{4MT}\,[2\mbox{\boldmath$\nabla$}^2 W(0)-\nabla^2 W_c -\mbox{\boldmath$\nabla$} W_c\cdot \mbox{\boldmath$\nabla$}_c] , \eeq where we have used $\nabla W(0)=0$, and introduced the following shorthand notation \beq \mbox{\boldmath$\nabla$} W_c\cdot \mbox{\boldmath$\nabla$}_c\equiv \mbox{\boldmath$\nabla$}_{1} W_{12}\cdot \mbox{\boldmath$\nabla$}_{12}+ \mbox{\boldmath$\nabla$}_{1'} W_{1'2'}\cdot \mbox{\boldmath$\nabla$}_{1'2'}. \eeq with analogous definitions for $W_a$, $W_b$, $W_\pm$ (to be used later). In this formula, \beq \mbox{\boldmath$\nabla$}_{12}\equiv \mbox{\boldmath$\nabla$}_1-\mbox{\boldmath$\nabla$}_{2}, \eeq and recall that $\mbox{\boldmath$\nabla$}_1$ stands for $\mbox{\boldmath$\nabla$}_{{\boldsymbol r}_1}$ and $W_{11'}$ for $W({\boldsymbol r}_1-{\boldsymbol r}_{1'})$. The second contribution to ${\cal L}_3$ reads \beq\label{QEDndotn1} {\cal L}_{3b}^{\rm QED}&=& -\frac{i}{4T}\int_{{\boldsymbol x},{\boldsymbol x}'} W({\boldsymbol x}-{\boldsymbol x}')\,\left[ \dot n({\boldsymbol x}) {\cal D} n({\boldsymbol x}')-n({\boldsymbol x}){\cal D}\dot n({\boldsymbol x}') \right]\nonumber\\ &=&-\frac{1}{4MT}\left\{ \nabla^2 W_a +\mbox{\boldmath$\nabla$} W_a\cdot \mbox{\boldmath$\nabla$}_a-\nabla^2 W_b -\mbox{\boldmath$\nabla$} W_b\cdot \mbox{\boldmath$\nabla$}_b\right\}\nonumber\\ &=&-\frac{1}{4MT}\left\{ \nabla^2 W^- +\mbox{\boldmath$\nabla$} W^-\cdot \mbox{\boldmath$\nabla$}^-\right\}. \eeq Note the analogy between Eqs.~(\ref{L2aQED}) and Eqs.~(\ref{QEDanticom}) and (\ref{QEDndotn1}). The latter follow from the former after the replacements $W(0)\to \nabla^2 W(0)$, $W_c\to \nabla^2 W_c +\mbox{\boldmath$\nabla$} W_c\cdot \mbox{\boldmath$\nabla$}_c$, $W_-\to \nabla^2 W_- +\mbox{\boldmath$\nabla$} W_-\cdot \mbox{\boldmath$\nabla$}_-$, and the multiplication by the overall factor $1/(4MT)$. This correspondence is analogous to that already observed after Eq.~(\ref{eqforrhoQEDM1}), and its origin is the same. \\ Until now, the expressions that we have obtained are an exact transcription of the operators ${\cal L}_i$ in Eq.~(\ref{main3}) to the case of a (abelian) quark-antiquark pair.. At this point it is useful to go to the coordinates (\ref{changeofvar5}), i.e., $\bra{{\boldsymbol r}_1{\boldsymbol r}_2}{\cal D}\ket{{\boldsymbol r}_1'{\boldsymbol r}_2'}\to {\cal D}({\cal R},{\boldsymbol Y},{\boldsymbol r},{\boldsymbol y})$, and perform a semi-classical expansion similar to that which leads to Eq.~(\ref{eqforrhoQEDM4}) for the single particle density matrix. We obtain then \beq \label{eqforrhoQEDM4b} &&\frac{{\rm d} }{{\rm d} t}{\cal D}({\cal R},{\boldsymbol Y},{\boldsymbol r},{\boldsymbol y})=[{\cal L}_0+{\cal L}_1+{\cal L}_2+{\cal L}_3]\,{\cal D}({\cal R},{\boldsymbol Y},{\boldsymbol r},{\boldsymbol y}), \eeq where \beq\label{L123QEDa} &&{\cal L}_{1}\approx i {\boldsymbol y}\cdot\mbox{\boldmath$\nabla$} V({\boldsymbol r}),\nonumber\\ &&{\cal L}_{2}\approx {\boldsymbol Y}\cdot({\cal H}({\boldsymbol r})-{\cal H}(0))\cdot {\boldsymbol Y}-\frac{1}{4} {\boldsymbol y}\left( {\cal H}({\boldsymbol r})+{\cal H}(0)\right){\boldsymbol y},\nonumber\\ &&{\cal L}_{3}\approx-\frac{1}{2MT}\left\{ {\boldsymbol Y}\cdot\left({\cal H}(0)-{\cal H}({\boldsymbol r}) \right)\cdot \nabla_{\boldsymbol Y} + {\boldsymbol y}\cdot \left( {\cal H}(0)+ {\cal H}({\boldsymbol r})\right) \cdot\nabla_{\boldsymbol y} \right\}. \nonumber\\ \eeq After performing the Wigner transform with respect to the variables ${\boldsymbol Y}$ and ${\boldsymbol y}$, \beq {\cal D}({\cal R},{\boldsymbol P}, {\boldsymbol r},{\boldsymbol p})=\int_{\boldsymbol Y}\int_{\boldsymbol y} \,{\rm e}^{-i{\boldsymbol P}\cdot{\boldsymbol Y}} {\rm e}^{-i{\boldsymbol p}\cdot{\boldsymbol y}}\, {\cal D}({\cal R},{\boldsymbol Y},{\boldsymbol r},{\boldsymbol y}), \eeq we obtain \beq &&{\cal L}_0=-\left(\frac{{\boldsymbol P}\cdot\mbox{\boldmath$\nabla$}_{\cal R}}{4M}+\frac{2{\boldsymbol p}\cdot\mbox{\boldmath$\nabla$}_{\boldsymbol r}}{M} \right),\nonumber\\ && {\cal L}_1=-\mbox{\boldmath$\nabla$} V({\boldsymbol r})\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol p} ,\nonumber\\ &&{\cal L}_2=\left[\mbox{\boldmath$\nabla$}_{\boldsymbol P}\cdot ({\cal H}(0)-{\cal H}({\boldsymbol r}))\cdot\mbox{\boldmath$\nabla$}_{\boldsymbol P}+\frac{1}{4} \nabla_{\boldsymbol p}\cdot \left( {\cal H}({\boldsymbol r})+{\cal H}(0)\right)\cdot \nabla_{\boldsymbol p}\right] ,\nonumber\\ &&{\cal L}_3=\frac{1}{2MT} \mbox{\boldmath$\nabla$}_{\boldsymbol P}\cdot ({\cal H}(0)-{\cal H}({\boldsymbol r}))\cdot {\boldsymbol P}+\frac{1}{2MT}\mbox{\boldmath$\nabla$}_{\boldsymbol p}\cdot\left[ {\cal H}({\boldsymbol r})+{\cal H}(0)\right] \cdot {\boldsymbol p}. \nonumber\\ \eeq We note that the operators for the relative coordinates are independent of the center of mass coordinates. It is then easy to identify the operators for the relative coordinates, and determine the elements of the corresponding Langevin equation. The relative velocity is given by $2{\boldsymbol p}/M={\boldsymbol p}/(M/2)$. Thus \beq\label{FPrel0} {\cal L}_0^{\rm rel}=-\frac{2{\boldsymbol p}\cdot\mbox{\boldmath$\nabla$}_{\boldsymbol r}}{M}=-{\boldsymbol v}\cdot\mbox{\boldmath$\nabla$}_{\boldsymbol r}. \eeq Similarly, \beq\label{FPrel1} {\cal L}_1^{\rm rel}=-\mbox{\boldmath$\nabla$} V({\boldsymbol r})\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol p}=-\frac{2}{M}\mbox{\boldmath$\nabla$} V({\boldsymbol r})\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol v}. \eeq The term ${\cal L}_2$ corresponds to the noise term. We write it as \beq\label{FPrel2} {\cal L}_2^{\rm rel}= \frac{1}{4} \nabla_{\boldsymbol p}\cdot \left( {\cal H}({\boldsymbol r})+{\cal H}(0)\right)\cdot \nabla_{\boldsymbol p} =\frac{2}{M^2}\eta_{ij}({\boldsymbol r}) \nabla_{\boldsymbol v}^i \nabla_{\boldsymbol v}^j, \eeq with \beq \eta_{ij}({\boldsymbol r})=\frac{1}{2} \left( {\cal H}_{ij}({\boldsymbol r})+{\cal H}_{ij}(0)\right). \eeq Finally ${\cal L}_3$ corresponds to the friction, and we write it as \beq\label{FPrel3} {\cal L}_{3}^{\rm rel}=\frac{1}{2MT}({\cal H}_{ij}(0)+{\cal H}_{ij}({\boldsymbol r})) \nabla^i_{\boldsymbol p} {\boldsymbol p}^j=\frac{2}{M}\gamma_{ij}({\boldsymbol r})\nabla_{\boldsymbol v}^i v^j. \eeq Friction and noise are related by an Einstein relation \beq \gamma_{ij}({\boldsymbol r})=\frac{1}{2T} \eta_{ij}({\boldsymbol r}). \eeq The Langevin equation associated with the relative motion is then of the form \beq\label{langevinrelative} \frac{M}{2} \ddot {\boldsymbol r}^i=-\gamma_{ij} {\boldsymbol v}^j-\mbox{\boldmath$\nabla$}^i V({\boldsymbol r}) +\xi^i({\boldsymbol r},t),\quad \langle\xi^i({\boldsymbol r},t)\xi^i({\boldsymbol r},t')\rangle=\eta_{ij}({\boldsymbol r})\delta(t-t'). \eeq Note that for an isotropic plasma, we have (cf. Eq.~(\ref{noisedef})) \beq \eta_{ij}({\boldsymbol r})=\delta_{ij}\eta({\boldsymbol r}),\qquad \eta({\boldsymbol r})=\frac{1}{6}\left(\nabla^2 W(0)+\nabla^2 W({\boldsymbol r}) \right). \eeq One can repeat the same for the center of mass coordinates. Here we set ${\boldsymbol v}={\boldsymbol P}/(2M)$. We get \beq &&{\cal L}_0^{\rm CM}=-\frac{{\boldsymbol P}\cdot\mbox{\boldmath$\nabla$}_{\cal R}}{2M} = -{\boldsymbol v}\cdot\mbox{\boldmath$\nabla$}_{\cal R}\nonumber\\ &&{\cal L}_1^{\rm CM}=0\nonumber\\ &&{\cal L}_2^{\rm CM}= \nabla_{\boldsymbol P}\cdot ({\cal H}(0)-{\cal H}({\boldsymbol r}))\cdot\nabla_{\boldsymbol P}=\frac{1}{8 M^2} \eta_{ij}({\boldsymbol r})\nabla_{\boldsymbol v}^i \nabla_{\boldsymbol v}^j \nonumber\\ &&{\cal L}_{3}^{\rm CM}=\frac{1}{2MT}({\cal H}_{ij}(0)-{\cal H}_{ij}({\boldsymbol r}))\nabla_{\boldsymbol P}^i P^j=\frac{1}{2M}\gamma_{ij}({\boldsymbol r})\nabla_{\boldsymbol v}^i {\boldsymbol v}^j, \nonumber\\ \eeq with \beq \eta_{ij}({\boldsymbol r})=2 \left( {\cal H}_{ij}(0)-{\cal H}_{ij}({\boldsymbol r})\right) \eeq and the Einstein relation \beq \gamma_{ij}({\boldsymbol r})=\frac{1}{2T}\eta_{ij}({\boldsymbol r}). \eeq The Langevin equation associated with the center of mass motion is then \beq\label{langevincm} 2M\ddot {\cal R}^i=-\gamma_{ij} {\boldsymbol v}^j+\xi^i({\boldsymbol r},t) \eeq These two equations (\ref{langevinrelative}) and (\ref{langevincm}) are identical to those obtained in \cite{} (see Eqs.~(4.69) there). Note that while the Langevin equation for the relative motion does not depend on the center of mass, this is not so for the Langevin equation describing the center of mass motion, which depends on the relative coordinate: this is because, as we have already emphasized, the effect of the collisions on the system depends on its size, with small size dipoles being little affected by the typical plasma field fluctuations. \section{QCD} \label{} We turn now to QCD. Much of the calculations are similar to those of the QED case, with however the obvious additional complications related to the color algebra. As we did in QED, we shall consider successively the case of the single particle density matrix, and that of a quark-antiquark pair. \subsection{Single quark density matrix} For a single quark, the color charge density can be written as (see Eq.~(\ref{colordensity})) \beq\label{rhoadef0a1} n^a({\boldsymbol x})=\delta({\boldsymbol x}-\hat{\boldsymbol r})\, t^a, \eeq with matrix elements \beq \bra{{\boldsymbol r},\alpha}n^a({\boldsymbol x})\ket{{\boldsymbol r}',\alpha'}=\delta({\boldsymbol r}-{\boldsymbol r}')n({\boldsymbol r})\bra{\alpha}t^a\ket{\alpha'}, \eeq where $n({\boldsymbol r})$ is the density of heavy quarks, that is, the number of heavy quarks located at point ${\boldsymbol r}$ irrespective of their color state. The reduced density matrix of a single quark can be written as follows (see ~\ref{}) \beq\label{Dcolor1} {\cal D}=D_0\, \mathbb{I}+ D_1\cdot{\boldsymbol t} . \eeq It depends on 9 real parameters, and contains a scalar as well as a vector (octet) contributions. In fact, since we assume the plasma to be color neutral, we need consider only the scalar part of the density matrix (see however ~\ref{}), that is the quantity $\bra{{\boldsymbol r}}D_0\ket{{\boldsymbol r}'}$ . With ${\cal D}$ having only a scalar component, i.e., ${\cal D}=D_0\, \mathbb{I}$, one can perform immediately the sum over the color indices in Eq.~(\ref{main}), using \beq\label{CasimirCF} t^a t^a=C_F=\frac{N_c^2-1}{2N_c}. \eeq The result is then identical to that obtained in QED, to within the multiplicative factor $C_F$: there is no specific effect of the color degree of freedom on the color singlet component of the density matrix, aside from this multiplicative color factor. The resulting Fokker-Planck equation is then essentially identical to that first derived by Svetitsky long ago \cite{}, which has been used in numerous phenomenological applications. \subsection{Density matrix of a quark-antiquark pair} The color density of a quark-antiquark pair is given by Eq.~(\ref{colordensity}). The color structure of the reduced density matrix ${\cal D}$ is discussed in ~\ref{}. We shall use two convenient representations. In the first one, to which we refer as the $(D_0, D_8)$ basis, the density matrix takes the form \beq {\cal D}=D_0 \,\mathbb{I}\otimes \mathbb{I} +D_8 \, t^a\otimes \tilde t^a, \label{} \eeq where $D_0$ and $D_8$ are matrices in coordinate space (product of coordinate spaces of the quark and the antiquark), e.g. $\bra{{\boldsymbol r}_1,{\boldsymbol r}_2}D_0\ket{{\boldsymbol r}_1',{\boldsymbol r}_2'}$, and similarly for $D_8$. The second representation is in terms of components $D_{\rm s}$ and $D_{\rm o}$ defined by (see ~\ref{}) \beq {\cal D}=D_{\rm s} \ket{\rm s}\bra{\rm s}+D_{\rm o} \sum_c\ket{{\rm o}^c}\bra{{\rm o}^c} \eeq where $\ket{\rm s}\bra{\rm s}$ denotes a projector on a color singlet state, and $\ket{{\rm o}^c}\bra{{\rm o}^c}$ a projector on a color octet state with given projection $c$. We shall refer to this basis as the ($D_{\rm s},D_{\rm o}$) basis, or as the singlet-octet basis. The relation between the two basis is given by the following equations \beq &&D_{\rm s}=D_0+C_F D_8,\qquad D_{\rm o}=D_0-\frac{1}{2N_c} D_8\nonumber\\ &&D_0=\frac{1}{N_c^2}(D_{\rm s}+(N_c^2-1) D_{\rm o}),\qquad D_8=\frac{2}{N_c}(D_{\rm s}-D_{\rm o}). \eeq The advantage of the singlet-octet basis is that it involves states of the quark-antiquark pair with well defined color (singlet or octet), which is not the case in the $(D_0,D_8)$ basis. The latter will play a role when we address the issue of equilibration of color. Then the matrix $D_0$, which represents a completely unpolarized color state, or a maximum (color) entropy state, plays an essential role. Because of the existence of two independent functions of the coordinates to describe the density matrix of the quark-antiquark pair, the equation of motion for ${\cal D}$ takes the form of a matrix equation \beq \frac{\partial}{\partial t}{\cal D}={\cal L}\,{\cal D}, \eeq where ${\cal D} $ can be viewed as a two dimensional vector, e.g. \beq {\cal D}=\left( \begin{array}{c} D_0 \\ D_8 \end{array} \right),\qquad {\rm or}\qquad {\cal D}=\left( \begin{array}{c} D_{\rm s} \\ D_{\rm o} \end{array} \right), \eeq and ${\cal L}$ is a $2\times 2$ matrix in the corresponding space. The derivation of the equations for the various components of the reduced density matrix in the two basis is straightforward, but lengthy. The results are listed in ~\ref{}. In this section we shall give brief indications on how to obtain the equations in the singlet-octet basis: the color algebra is then transparent, and most of the equations can be related to the corresponding ones of the abelian case. \subsection{The equations in the singlet-octet basis}\label{} \noindent{\bf The contribution to ${\cal L}_1$. } When written in terms of $D_{\rm s}$ and $D_{\rm o}$ the two equations decouple. This is because the product $n^a_{\boldsymbol x} n^a_{{\boldsymbol x}'} $ is a scalar in color space, and hence does not change the color state of the pair. In other terms, the one-gluon exchange does not change the color state (singlet or octet) of the heavy quark-antiquark pair. The color algebra is then trivial, and yields \beq\label{DsDoForce} &&\frac{\rm d D_{\rm s}}{{\rm d} t}=iC_F [V_{12}-V_{1'2'}]\,D_{\rm s}\nonumber\\ &&\frac{{\rm d} D_{\rm o}}{{\rm d} t}=-\frac{i}{2N_c}[V_{12}-V_{1'2'}]D_{\rm o} . \eeq The diagrams contributing here are the first two diagrams in Fig.~\ref{}, and the equations above are analog to that obtained in QED. In fact we have \beq {\cal L}_1^{\rm ss}=C_F {\cal L}_1^{\rm QED},\qquad {\cal L}_1^{\rm oo}=-\frac{1}{2N_c} {\cal L}_1^{\rm QED}, \eeq where ${\cal L}_1^{\rm QED}$ is given by Eq.~(\ref{phaserealVqed}). Note the well known property that the interaction is attractive (and proportional to $C_F$) in the singlet channel, and repulsive (and proportional to $1/(2N_c)$) in the octet channel. \\ \noindent{\bf The contribution to ${\cal L}_2$.} In this case we have two types of contributions. The first ones involve products of the color charges, making up a color scalar. These contribute to ${\cal L}_{2a}$, which is diagonal in color. The second type of contribution involves transitions from singlet to octet or from octet to octet. We denote these contributions by ${\cal L}_{2b}$. Consider first ${\cal L}_{2a}$. We have, for the singlet \beq\label{eqforrhoA4css} {\cal L}_{2a}^{\rm ss} =C_F [2W(0)-W_c]D_{\rm s}=C_F{\cal L}_{2a}^{\rm QED} . \eeq with $ {\cal L}_{2a}^{\rm QED}$ given in Eqs.~(\ref{L2aQED}). For the octet \beq\label{eqforrhoA4coo} {\cal L}^{\rm oo}_{2a}= 2C_FW(0)+\frac{1}{2N_c}W_c. \eeq Consider next ${\cal L}_{2b}$. The corresponding contributions involve transitions from singlet to octet intermediate states, or, when considering the octet density matrix, transitions from octet to singlet and also octet to octet transitions that generate an additional diagonal contribution. We have \beq\label{L2bsoQCD} {\cal L}_{2b}^{\rm so}=-C_F W^- =C_F\, {\cal L}_{2b}^{\rm QED} ,\qquad {\cal L}_{2b}^{\rm os}=-\frac{1}{2N_c} W^-=\frac{1}{2N_c} {\cal L}_2^{\rm QED} , \eeq where $ {\cal L}_{2b}^{\rm QED}$ is given in Eqs.~(\ref{L2aQED}). Similarly, for the octet to octet transition, we have \beq \label{L2booQCD} {\cal L}_{2b}^{\rm oo}=-\left(\frac{N_c^2-4}{4N_c}W^- +\frac{N_c}{4} W^+ \right)=-\left( \frac{N_c^2-2}{2N_c}W_a+\frac{1}{N_c} W_b \right), \eeq which has no counterpart in QED. \\ \noindent{\bf The contribution to ${\cal L}_3$.} The calculation of these contributions is more involved, but we can relate simply the results to those obtained earlier for the operators ${\cal L}_2$. Let us first list the results. We have \beq &&{\cal L}_{3a}^{\rm ss}=\frac{C_F}{4MT}\,[2\mbox{\boldmath$\nabla$}^2 W(0)-\nabla^2 W_c- \mbox{\boldmath$\nabla$} W_c\cdot \mbox{\boldmath$\nabla$}_c ]=C_F {\cal L}_{3a}^{\rm QED},\\ &&{\cal L}_{3a}^{\rm oo}=\frac{C_F}{2MT}\,[\mbox{\boldmath$\nabla$}^2 W(0)] +\frac{1}{4MT}\frac{1}{2N_c}\left(\nabla^2 W_c+ \mbox{\boldmath$\nabla$} W_c\cdot \mbox{\boldmath$\nabla$}_c \right),\\ && {\cal L}_{3b}^{\rm so}=C_F\,{\cal L}_{3b}^{\rm QED},\qquad {\cal L}_{3b}^{\rm os}=\frac{1}{2N_c}\,{\cal L}_{3b}^{\rm QED}, \eeq with \beq\label{L3bQED2} {\cal L}_{3b}^{\rm QED}= -\frac{1}{4MT}\left\{\mbox{\boldmath$\nabla$}^2 W^- +\mbox{\boldmath$\nabla$} W^-\cdot \mbox{\boldmath$\nabla$}^- \right\}, \eeq and \beq {\cal L}_{3b}^{\rm oo} =-\frac{1}{4MT}\left\{\frac{N_c^2-4}{4N_c}\left[\nabla^2 W^- +\mbox{\boldmath$\nabla$} W^-\cdot \nabla^- \right] + \frac{N_c}{4} \left[\nabla^2 W^+ +\mbox{\boldmath$\nabla$} W^+\cdot \nabla^+ \right] \right\}\nonumber\\ \eeq It is easy to verify that the equations giving ${\cal L}_2$ and the corresponding equations for ${\cal L}_3$ are related via the same substitutions that are discussed after Eq.~(\ref{QEDndotn1}). \subsubsection{Semiclassical approximation} The formulae listed in the previous subsection are an exact transcription of the main equation, Eq.~(\ref{main}) for a quark-antiquark pair. Analogous formulae can be written for a pair of quarks. They are given in ~\ref{}. We shall be mostly concerned in this paper, in particular for the numerical studies presented in the next section, by the semi-classical limits of these equations. These can be obtained easily by using the formulae given in ~\ref{}. In this subsection, we just reproduce the few equations that will be used in the next section. We consider first the equation for the component $D_{\rm s}$, which we write as \beq \frac{{\rm d} D_{\rm s}}{{\rm d} t}=(D_{\rm s}|{\cal L}|{\cal D}). \eeq We obtain \beq\label{eqfinDs} &&(D_{\rm s}|{\cal L}|{\cal D})= \left( 2i \frac{\nabla_{\boldsymbol r}\cdot \nabla_{\boldsymbol y}}{M}+ i \frac{\nabla_{\cal R}\cdot \nabla_{\boldsymbol Y}}{2M} +iC_F {\boldsymbol y}\cdot \mbox{\boldmath$\nabla$} V({\boldsymbol r})\right) D_{\rm s}\nonumber\\ &&\qquad\qquad-2C_F \Gamma({\boldsymbol r}) (D_{\rm s}-D_{\rm o})\nonumber\\ &&\qquad\qquad-\frac{C_F}{4}\left({\boldsymbol y}\cdot {\cal H}({\boldsymbol r})\cdot{\boldsymbol y}\, D_{\rm s}+{\boldsymbol y}\cdot {\cal H}(0)\cdot{\boldsymbol y} \,D_{\rm o} \right)\nonumber\\ &&\qquad\qquad-C_F {\boldsymbol Y}\cdot [{\cal H}(0)-{\cal H}({\boldsymbol r})]\cdot {\boldsymbol Y} D_{\rm o}\nonumber\\ &&\qquad\qquad+\frac{C_F}{2MT}\left[ \nabla^2 W(0)-\nabla^2 W({\boldsymbol r})-\mbox{\boldmath$\nabla$} W({\boldsymbol r})\cdot\mbox{\boldmath$\nabla$}_{\boldsymbol r} \right](D_{\rm s}-D_{\rm o})\nonumber\\ &&\qquad\qquad-\frac{C_F}{2MT}\left({\boldsymbol y}\cdot {\cal H}({\boldsymbol r})\cdot\mbox{\boldmath$\nabla$}_{\boldsymbol y} \,D_{\rm s}+{\boldsymbol y}\cdot {\cal H}(0)\cdot\mbox{\boldmath$\nabla$}_{\boldsymbol y} \,D_{\rm o} \right)\nonumber\\ &&\qquad\qquad-\frac{C_F}{2MT}{\boldsymbol Y}\cdot [{\cal H}(0)-{\cal H}({\boldsymbol r})]\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol Y} D_{\rm o}. \eeq One reason why we display this equation is that it reduces to the QED equation when $D_{\rm s}=D_{\rm o}$. Thus, if color quickly equilibrates, an assumption that we shall exploit in the next section, the dynamics becomes analogous to that of the abelian case. In this case, color degrees of freedom play a minor role, and the motion of the heavy particles can be described by a Fokker-Planck equation or the associated Langevin equation. As we have already emphasized, the component $D_0$ corresponds to the maximum (color) entropy state, where all colors are equally probable. This state plays an important role in the calculations to be presented in the next section. Thus, it is useful to write the corresponding equations of motion, or equivalently the operators ${\cal L}$ of the $(D_0,D_8)$ basis, in the semi-classical limit. We have: \beq\label{L00m} {\cal L}^{00}&=&-C_F\left\{{\boldsymbol Y}\cdot {\cal H}(0)\cdot {\boldsymbol Y}+\frac{1}{4} {\boldsymbol y}\cdot {\cal H}(0)\cdot {\boldsymbol y} \right\}\nonumber\\ &&-\frac{C_F}{2MT}\left\{{\boldsymbol Y}\cdot {\cal H}(0)\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol Y}+{\boldsymbol y}\cdot {\cal H}(0)\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol y} \right\}, \eeq and \beq\label{L08m} {\cal L}^{08}&=&i\frac{C_F}{2N_c} {\boldsymbol y}\cdot\mbox{\boldmath$\nabla$} V({\boldsymbol r})\nonumber\\ &&-\frac{C_F}{2N_c}\left\{ \frac{1}{4} {\boldsymbol y}\cdot {\cal H}({\boldsymbol r})\cdot {\boldsymbol y}-{\boldsymbol Y}\cdot {\cal H}({\boldsymbol r})\cdot {\boldsymbol Y} \right\}\nonumber\\ &&-\frac{C_F}{2N_c}\frac{1}{2MT}\left\{ {\boldsymbol y}\cdot {\cal H}({\boldsymbol r})\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol y} -{\boldsymbol Y}\cdot {\cal H}({\boldsymbol r})\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol Y}\right\}.\nonumber\\ \eeq We shall also need the operators ${\cal L}^{08}$ and ${\cal L}^{88}$ at leading order in ${\boldsymbol y}$. These are given by \beq\label{L80m} {\cal L}^{80}=i\ {\boldsymbol y}\cdot\mbox{\boldmath$\nabla$} V({\boldsymbol r}),\qquad {\cal L}^{88}=-N_c \Gamma({\boldsymbol r}). \eeq \section{Numerical studies}\label{numerics} The equations for the time evolution of the reduced density matrix that we have obtained in the previous sections are difficult to solve in their original form, that is, for the operator ${\cal L}$ given in Sect.~\ref{}, or ~\ref{}, for a quark-antiquark pair. We shall not attempt to solve them directly in the present paper. In the case of QED, we have seen that an additional approximation, the semi-classical approximation, allows us to transform these equations into Fokker-Planck, or equivalently, Langevin equations, describing the relative and center of mass motions of the heavy particles as simple random walks. In QCD, the presence of transitions between singlet and octet color states complicates the situation, since such transitions are a priori not amenable to a semi-classical description. The purpose of this section is to present numerical studies that illustrate two possible strategies to cope with this problem, namely preserve as much as possible the simplicity of the semi-classical description of the heavy particle motions, while taking into account the effects of color transitions. To simplify the discussion we shall ignore the center of mass motion in most of this section. The new feature in QCD, as compared to QED, namely the transitions between distinct color states, is best seen in the infinite mass limit, where the relative motion is entirely frozen. Then the only dynamics is that of color: as a result of the collisions with the plasma constituents the colors of the heavy quarks and antiquarks can change in time. The corresponding equations of motion for the density matrix are easily obtained from the formulae listed in ~\ref{}. They read, for a quark-antiquark pair, \beq\label{eqDsDoMinfty} &&\frac{{\rm d} D_{\rm s}}{{\rm d} t}=-2C_F \Gamma({\boldsymbol r}) (D_{\rm s}-D_{\rm o}),\nonumber\\ &&\frac{{\rm d} D_{\rm 0}}{{\rm d} t}=-\frac{1}{N_c} \Gamma({\boldsymbol r}) (D_{\rm o}-D_{\rm s}), \eeq where ${\boldsymbol r} $ is the (fixed) relative coordinate. These equations exhibit the decay widths in the singlet ($2C_F \Gamma({\boldsymbol r})$) and the octet ($(1/N_c)\Gamma({\boldsymbol r})$) channels, respectively. These two coupled equations acquire a more transparent physical interpretation in the $(D_0,D_8)$ basis, where they take a diagonal form \beq\label{eqD0D8Minfty} &&\frac{\partial D_0}{\partial t}=0,\nonumber\\ &&\frac{\partial D_8}{\partial t}=- N_c \Gamma({\boldsymbol r}) D_8. \eeq The first equation merely reflects the conservation of the trace of the density matrix. Recall also that $D_0$ is associated with the maximum color entropy state, where all colors are equally probable (see Eq.~(\ref{Dunpolar})): this component of the density matrix represents an equilibrium state that remains unaffected by the collisions. The second equation indicates that $D_8\propto D_{\rm s}-D_{\rm o}$ decays on a time scale $(N_c\Gamma({\boldsymbol r}))^{-1}$. Thus, at large times only $D_0$ survives, that is, the collisions drive the system to the maximum entropy state. Note that the distance between the quark and the antiquark enters the rate $\propto \Gamma({\boldsymbol r})$ at which equilibrium is reached. When $|{\boldsymbol r}|\gtrsim m_D$, $\Gamma({\boldsymbol r})\approx 2\gamma_Q$, where $\gamma_Q $ is the damping factor of one heavy quark (or antiquark): at large separation, the quark and the antiquark equilibrate their color independently (with a rate $N_c \gamma_Q$ -- see ~\ref{}). On the other hand, when $|{\boldsymbol r}|\lesssim m_D$ the two quarks screen their respective colors, hindering the effect of collisions in equilibrating color.\\ The first strategy that we shall explore in order to treat the relative motion of the heavy quarks semi-classically and at the same time take into account the color transitions just discussed, rests on the following observation. There is one instance where one can recover a situation analogous to that of QED: this is the regime where the color exchanges are fast enough to equilibrate the color on a short time scale (short compared to the typical time scale of the relative motion). We have for instance observed in the previous section that the component $D_{\rm s}$ of the density matrix, Eq.~(\ref{eqfinDs}), obeys the same equation as the QED density matrix when $D_{\rm o}=D_{\rm s}$ (to within the multiplicative color factor $C_F$), which corresponds indeed to the maximum entropy state. In this case, one can explore the dynamics in the vicinity of this particular color state, treating the color transitions in perturbation theory. One can then derive Langevin equations which contain an additional random force arising from the fluctuations of the color force between the heavy particles. This perturbative approach is easily generalized to the case of a large number of quark-antiquark pairs. The second strategy is based on an analogy between the equations (\ref{eqDsDoMinfty}), and their generalizations to include the semi-classical corrections, and a Boltzmann equation, with the right hand side being viewed as a collision term. That is, the changes of color that accompany the singlet-octet transitions are then treated as collisions rather than an additional random force in a Langevin equation. This strategy allows us to overcome some of the limitations of the perturbative approach. \\ \subsection{Langevin equation with a random color force: single quark-antiquark pair} \label{} We shall now examine the corrections to Eq.~(\ref{eqD0D8Minfty}) that arise in the semi-classical approximation, i.e. taking into account corrections to the infinite mass limit. Note first that the kinetic energy of the heavy quarks leaves ${\cal L}$ as a diagonal operator. In the $(D_0,D_8)$ basis, this operator reads \begin{equation} {\cal L}={\cal L}_0+\Gamma({\bf r})\left(\begin{array}{cc} 0 & 0 \\ 0 & -N_c \end{array}\right),\qquad {\cal L}_0= \frac{2i}{M}{\mbox{\boldmath$\nabla$}}_{\boldsymbol r}\cdot{\mbox{\boldmath$\nabla$}}_{\boldsymbol y}. \end{equation} The semi-classical corrections brings ${\cal L}$ to the form \beq\label{Lpert} \partial_t\left( \begin{array}{c} D_0\\ D_8 \end{array} \right) =\left( \begin{array}{cc} {\cal L}_0+y a^{(1)}_{00}+y^2 a^{(2)}_{00}&y a^{(1)}_{08}+y^2 a^{(2)}_{08}\\ y a^{(1)}_{80}+y^2 a^{(2)}_{80} & {\cal L}_0+a^{(0)}_{88}+y a^{(1)}_{88}+y^2 a^{(2)}_{88} \end{array} \right) \left( \begin{array}{c} D_0\\ D_8 \end{array} \right),\nonumber\\ \eeq where the various coefficients can be read off the equations recalled in the previous section (see Eqs.~(\ref{L00m}) and (\ref{L08m})): \beq\label{aij} &&a^{(1)}_{00}=0,\qquad a^{(2)}_{00}=-\frac{C_F}{4}{\boldsymbol y}\cdot{\cal H}(0)\cdot{\boldsymbol y}-\frac{C_F}{2MT}{\boldsymbol y}\cdot{\cal H}(0)\cdot \mbox{\boldmath$\nabla$}_y,\nonumber\\ &&a^{(1)}_{08}=-i\frac{C_F}{2N_c}{\boldsymbol y}\cdot{\boldsymbol F}({\boldsymbol r}),\qquad a^{(1)}_{80}=-i{\boldsymbol y}\cdot{\boldsymbol F}({\boldsymbol r}),\qquad a^{(0)}_{88}= -N_c\Gamma({\boldsymbol r}),\nonumber\\ \eeq and we have set \beq {\boldsymbol F}({\boldsymbol r})\equiv -\mbox{\boldmath$\nabla$} V({\boldsymbol r}). \eeq One can diagonalize this new operator ${\cal L}$, and in particular find the eigenvalue that corresponds to the maximum entropy state in the limit where $y\to 0$. This is given by usual perturbation theory \beq\label{pertformula} {\cal L}_0+y a^{(1)}_{00}+y^2\left( a^{(2)}_{00}-\frac{a^{(1)}_{08}a^{(1)}_{80}}{a^{(0)}_{88}} \right). \eeq By using the explicit expressions given above for the various coefficients, one deduces that the corresponding eigenvector, $D_0'$, fulfills the equation \begin{eqnarray} &&\partial_t D_0'=\left(\frac{2i}{M}\mbox{\boldmath$\nabla$}_{\boldsymbol r}\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol y}-\frac{C_F}{4}{\boldsymbol y}\cdot{\cal H}(0)\cdot{\boldsymbol y}-\frac{C_F({\boldsymbol y}\cdot {\boldsymbol F}({\bf r}))^2}{2N_c^2\Gamma({\bf r})}\right.\nonumber\\ &&\qquad\qquad\qquad\left.-\frac{C_F}{2MT}{\boldsymbol y}\cdot{\cal H}(0)\cdot \mbox{\boldmath$\nabla$}_y\right)D_0' \equiv {\cal L}'\, D_0'. \label{} \end{eqnarray} Performing the Wigner transform we obtain \begin{align} {\cal L}'=-\frac{2{\boldsymbol p}\cdot\mbox{\boldmath$\nabla$}_{\boldsymbol r}}{M}+\frac{C_F}{4}\mbox{\boldmath$\nabla$}_{\boldsymbol p}\cdot {\cal H}(0)\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol p} +\frac{C_F ({\boldsymbol F}({\bf r})\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol p})^2}{2N_c^2\Gamma({\bf r})} +\frac{C_F}{2MT}\mbox{\boldmath$\nabla$}_{\boldsymbol p}\cdot {\cal H}(0)\cdot {\boldsymbol p}. \end{align} The comparison with the Fokker-Planck operators given in Eqs.~(\ref{FPrel0}) to (\ref{FPrel3}), allows us to write the corresponding Langevin equation: \beq\label{langevinrelqcd} {\boldsymbol v}=\dot{\boldsymbol r}=\frac{2{\boldsymbol p}}{M},\qquad \frac{M}{2} \ddot {\boldsymbol r}=-\gamma^{ij} v^j+\xi^i(t)+\Theta^i(t,{\bf r})\,, \eeq where \begin{equation}\label{noisePsi} \langle \xi^i(t)\xi^j(t')\rangle=\delta(t-t')\eta_{ij},\qquad \eta_{ij}=\frac{C_F}{2}{\cal H}^{ij}(0)=2T \gamma_{ij}, \end{equation} and \begin{equation} \langle \Theta^i(t,{\bf r})\Theta^j(t',{\bf r})\rangle=\delta(t-t')\frac{C_FF^i({\bf r})F^j({\bf r})}{N_c^2\Gamma({\bf r})}\,. \label{} \end{equation} As compared to the QED case, Eq.~(\ref{langevinrelative}), we note a number of new features. The most noteworthy is the presence of two random forces. The force $\xi$ is the familiar stochastic force and is related to the drag force as indicated in Eq.~(\ref{noisePsi}). The second random force, $\Theta$, has a different nature: it originates from the fact that the force between a quark and an antiquark is a function of their color state. Now, $D_0'$ represents a state close to the maximum entropy state, that is, a state in which the probability to find the pair in an octet is approximately $N_c^2-1$ times bigger than the probability to find it in a singlet. At the same time, the force between heavy quarks in a singlet state is $N_c^2-1$ times bigger than that in an octet state, and it has opposite sign. The net result is that, on average, the force between the heavy quark and the heavy antiquark is zero. But this is true only in average. There are fluctuations, and these give rise to $\Theta$. The vanishing of the average force between the quark and the antiquark explains the absence of the force term in the Langevin equation, as compared to the QED case. Note that this picture is valid as long as transitions between singlet and octet states are fast compared with the rest of the dynamics. This is no longer the case when the size of the pair becomes too small: then, $\Gamma({\boldsymbol r})$ becomes small, reducing the energy denominator in Eq.~(\ref{}), i.e., amplifying the effect of the random color force. This, as we shall see, can lead to unphysical behavior. \subsection{Many heavy quarks and antiquarks} \label{} The discussion of the previous subsection can be generalized to a system containing many heavy quarks and antiquarks. We call $N_Q$ the number of heavy quarks, and for simplicity we assume that it is equal to the number of heavy antiquarks. The density matrix can be written as a product of density matrices of the individual quarks and antiquarks, generalizing the construction of Eq.~(\ref{Dtensorproduct}) for the quark-antiquark density matrix. One can then write \beq {\cal D}=D_0\, (\mathbb{I})^{2N_Q}+\cdots \eeq where the dots represent all the scalar combinations that can be formed with products of $n\leq 2N_Q$ color matrices $t^a$, with coefficients corresponding to the components of ${\cal D}$. We shall not need the explicit form of these extra components. As for $D_0$, this is clearly the maximum entropy state, where all colors of individual quarks and antiquarks are equally probable and uncorrelated. It corresponds to the following density matrix (cp. Eq.~(\ref{Dunpolar})): \beq {\cal D}=\sum_{\alpha_i, \bar\alpha_i} \ket{\alpha_1,\cdots\bar\alpha_{N_Q}}\bra{\alpha_1,\cdots,\bar\alpha_{N_Q}}, \eeq where the sum runs over all the colors of the quarks ($\alpha_1,\cdots\alpha_{N_Q}$) and the antiquarks ($\bar\alpha_1,\cdots,\bar\alpha_{N_Q}$). We want to construct for this system the analog of Eq.~(\ref{}) which describes how the maximum entropy state is modified by the semiclassical corrections. Our starting point is the main equation, Eq.~(\ref{main}). To proceed, it is useful to have in mind the diagrammatic representation of the density matrix that we have introduced earlier. As compared to the diagrams displayed in Fig.~\ref{}, in the present case, the diagrams will contain $2N_Q$ lines in the upper part, and $2N_Q$ lines in the lower part. All the interactions that we are dealing with involve a single gluon exchange, represented by one gluon line joining quark or antiquark lines in various ways. The evolution equation for ${\cal D} $ is still described by an operator ${\cal L}$, which is a matrix in the space of all the independent components. For our perturbative calculation, we need only to consider diagonal (to order $y^2$) and non diagonal (to order $y$) elements of this matrix, the non diagonal elements involving the maximum entropy state as one of their entries. Consider first the diagonal elements. We have first the kinetic energy, trivially given by \begin{equation} \sum_{j\in \{N_Q,N_{\bar{Q}}\}}\left(\frac{2i}{M}\mbox{\boldmath$\nabla$}_{r_j}\cdot \mbox{\boldmath$\nabla$}_{p_j}\right)\,. \end{equation} The leading order diagonal element that involves the interaction can be obtained in the infinite mass limit. It represents the decay of the components of ${\cal D}$ that are connected to $D_0$ by one gluon exchange. It is given by \begin{equation} -N_c\Gamma({{\boldsymbol r}}_{kl})\,. \end{equation} where ${\boldsymbol r}_{kl} ={\boldsymbol r}_k-{\boldsymbol r}_l$, ${\boldsymbol r}_k$ and ${\boldsymbol r}_l$ denoting the coordinates of the quark or the antiquarks to which the gluon is attached. The factor $N_c$ can be understood from the same argument as that given after Eq.~(\ref{eqD0D8Minfty}): when the separation ${\boldsymbol r}_{kl}$ is large, the color of the quark (or antiquark) at ${\boldsymbol r}_k$ and ${\boldsymbol r}_l$ relax independently at a rate $N_c \gamma_Q$. At the order of interest, we need also the diagonal element for the maximum entropy state, including the semi-classical corrections up to order $y^2$ and $\frac{y}{M}$. This is given by \begin{equation} -\frac{C_F}{4} \sum_{j\in \{N_Q,N_{\bar{Q}}\}}\left({\boldsymbol y}_j\cdot {\cal H}(0)\cdot \left({\boldsymbol y}_j+\frac{2\mbox{\boldmath$\nabla$}_{{\boldsymbol y}_j}}{MT}\right)\right)\,. \end{equation} We turn now to the non-diagonal elements. To leading order, these involve solely the real part of the potential. Diagrammatically, the corresponding exchanged gluon connects only either the upper or the lower lines among themselves, but does not connect upper with lower lines. Consider first the non-diagonal elements that bring the maximum entropy state to another state. If the pair connected by the exchanged gluon is formed by quark $k$ and antiquark $l$ then the element is (cp. Eq.~(\ref{aij})) \begin{equation} -\frac{iC_F}{2N_c}{\boldsymbol y}_{kl}\cdot F({\bf r}_{kl})\,, \end{equation} while if it is formed by quark (antiquark) $l$ and quark (antiquark) $k$ then it is (cp. Eq.~(\ref{D0QQ})) \begin{equation} \frac{iC_F}{2N_c}{\boldsymbol y}_{kl}\cdot F({\bf r}_{kl})\,. \end{equation} We also need the non-diagonal elements that bring the system back to the maximum entropy state. If this pair is formed by quark $k$ and antiquark $l$ then the element is (cp. Eq.~(\ref{aij})) \begin{equation} -i\,{\boldsymbol y}_{kl}\cdot F({\bf r}_{kl})\,, \end{equation} while if it is formed by quark (antiquark) $l$ and quark (antiquark) $k$ then it is (cp. Eq.~(\ref{D8QQ})) \begin{equation} i{\boldsymbol y}_{kl}\cdot F({\bf r}_{kl})\,. \end{equation} With this information, it is straightforward to construct the generalization of Eq.~(\ref{}) for an arbitrary number of quark-antiquark pairs. The corresponding equations will be presented later. \subsection{Simulation of a Langevin equation with a random color} \label{} In this subsection we present numerical results obtained by simulating the equations of the previous subsections. We examine successively the evolution of the relative coordinate of a single heavy quark-antiquark pair, and then that of fifty pairs initially produced in a thin layer. The first case will help us to understand the range of applicability of the perturbative method, while the second will illustrate how the Langevin equation may account for recombination. In these calculations, we use the standard QCD running coupling constant $\alpha_s$ determined at one loop order for $N_f=3$ massless flavors and $\Lambda_{QCD}=250$ MeV. The screening Debye mass is given by its HTL approximation, $m_D^2=(2\pi/3)(6+N_f)\alpha_s T^2$, with $\alpha_s$ evaluated at the scale $2\pi T$, with $T$ the temperature. Further details on the parameters of the calculation will be given as we proceed. We should emphasize here that the numerical simulations to be presented in this section are meant to illustrate the main physical content, as well as the limitations, of the equations obtained in this paper. Although the numbers that go into the calculations are appropriate for the physics of quarkonia in a quark-gluon plasma, we make no attempt to develop a phenomenological discussion. \subsubsection{A single heavy quark-antiquark pair} The Langevin equation for the relative motion is given in Eq.~(\ref{langevinrelqcd}). The information about the medium is encoded in the functions $ V({\bf r})$ and $W({\bf r})$ which we estimate using the HTL approximation. Note that the resulting value of $\Delta W({\bf r})$ and $ V({\bf r})$ diverge as $r\to 0$ (for different reasons though, see e.g. \cite{}). In \cite{} the divergence of $\Delta W({\boldsymbol r})$ was regulated with the help of an ultraviolet cut-off. Here, we shall use a simpler procedure and choose \begin{equation} \gamma=\frac{\Delta W({\bf 0})}{6T} \end{equation} as a free parameter of our simulation, for which we choose the value $\gamma=0.19 T^2$. For the real part of the potential, we proceed as in \cite{}, that is we define it as the Fourier transform of a Yukawa potential integrated in momentum space up to a cutoff $\Lambda=4$ GeV. The coupling constant that appears in $V({\boldsymbol r})$ is evaluated at a scale corresponding to the inverse Bohr radius of the bottomonium (1348 MeV). The spatial dependence of $W({\boldsymbol r})$ is obtained, as already mentioned, from the HTL approximation, and is of the form $W({\boldsymbol r})=W(0)+\alpha_s T\phi(m_D r)$ \cite{}, with $\alpha_s$ evaluated at the scale $2\pi T$, and $\phi(x)$ a monotonously increasing function such that $\phi(x=0)=0$ and $\phi(x\to \infty)=1$. At small separation, i.e., for $m_D r\ll 1$, $\phi(x)$ can be approximated by $ \phi(x)=\frac{x^2}{3}(-\log x+\frac{4}{3}-\gamma_E). $ \begin{figure} \begin{center} \includegraphics[scale=0.5]{./10traj.eps} \caption{Example of ten random trajectories for the relative distance of a bottom quark-antiquark pair prepared in a $1S$ bound state. About half of these trajectories are unphysical, since they correspond to supraluminal velocity, $r(t)\>t$.} \label{} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[scale=0.5]{./histini.eps} \caption{Histogram representing the initial distribution of relative distances given by the square of the $1S$ wave-function of the bottomonium.} \label{} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[scale=0.5]{./histfinal_new.eps} \caption{Histogram representing the final distribution of relative distances after a time $t=5\,\rm{fm/c}$ assuming the initial distribution of Fig.~\ref{}. Note the change in horizontal scale with respect to Fig.~\ref{}.} \label{} \end{center} \end{figure} In Fig.~\ref{} we show a set of ten random trajectories produced for bottomonium (with mass $M_b=4881\,\text{MeV}$) at a temperature of $T=\,350\, \text{MeV}$. We take as initial distribution of relative distances that obtained from the wave function of the $1S$ state (which we plot as an histogram with $1000$ events in Fig.~\ref{} for further comparison). We see that many of the trajectories are clearly unphysical since they involve supraluminal velocities: this is because some of the random kicks can occasionally be very strong due to the amplification produced by the small denominator in Eq.~(\ref{}) at small ${\boldsymbol r}$. A more systematic comparison can be done by looking at the distribution of relative distances after some time $t$. This is shown in Fig.~\ref{} for $t=5\,\rm{fm}$. The histogram reveals that there remains at this time only a tiny probability to find the pair within a relative distance corresponding to the size of the bound state: the random color force is clearly too efficient in suppressing the bound state! \subsubsection{Many heavy quark-antiquark pairs} \label{} In spite of its shortcomings, the perturbative method remains interesting as it allows us to treat an assembly with an arbitrary number of quark-antiquark pairs, and address in particular the issue of recombination. The relevant equations can be constructed along the lines developed in the previous subsection. We need to take into account not only the relative coordinates but also the center of mass motions. Moreover the random force does not only act between heavy quarks and antiquarks but also between two heavy quarks and two heavy antiquarks (the equations for the density matrix of a pair of two heavy quark are given in ~\ref{}). The resulting equations are given by \begin{equation} M\ddot{\bf r}_a=-C_F\gamma\dot{\bf r}_a+{\bf\Xi}_a(t)+\sum_{b\neq a}^{N_Q}{\bf \Theta}_{ab}({\boldsymbol r}_{ab})+\sum_{\hat{b}}^{N_Q}{\bf\Theta}_{a\hat{b}}({\boldsymbol r}_{a\hat{b}},t)\,, \end{equation} \begin{equation} M\ddot{{{\bf r}}}_{\hat{a}}=-C_F\gamma\dot{{{\bf r}}}_{\hat{a}}+{\bf \Xi}_{\hat{a}}(t)+\sum_{\hat{b}\neq \hat{a}}^{N_Q}{\bf \Theta}_{\hat{a}\hat{b}}({\boldsymbol r}_{\hat{a}\hat{b}},t)+\sum_j^{N_Q}{\bf\Theta}_{\hat{a}b}({\boldsymbol r}_{\hat{a}b},t)\,, \end{equation} where the noises have vanishing means and correlators given by \beq \langle \Xi_{ia}(t) \Xi_{jb}(t')\rangle=\frac{C_F}{6}\Delta W({\bf 0})\delta_{ab}\delta_{ij}\delta(t-t')\,, \eeq \beq \langle \Theta^{i}_{ab}(t)\Theta^{j}_{cd}(t')\rangle=\frac{C_F}{N_c^2\Gamma_{ab}} F^i_{ab}F^j_{cd}\,\delta_{ac}\delta_{bd}\delta(t-t')\nonumber\\ \eeq \begin{eqnarray} \langle \Theta^{i}_{a\hat{b}}(t)\Theta^{j}_{c\hat{d}}(t')\rangle=\frac{C_F}{N_c^2\Gamma_{a\hat{b}}} F^i_{a\hat{b}}F^j_{c\hat{d}}\,\delta_{ac}\delta_{\hat{b}\hat{d}}\delta(t-t')\nonumber\\ \end{eqnarray} \begin{eqnarray} \langle \Theta^{i}_{\hat{a}\hat{b}}(t)\Theta^{j}_{\hat{c}\hat{d}}(t')\rangle= \frac{C_F}{N_c^2\Gamma_{\hat{a}\hat{b}}}F^i_{\hat{a}\hat{b}}F^j_{\hat{c}\hat{d}}\,\delta_{\hat{a}\hat{c}}\delta_{\hat{b}\hat{d}}\delta(t-t').\nonumber\\ \end{eqnarray} In these equations $N_Q$ is the number of heavy quarks, equal to the number of heavy antiquarks. The indices $a$ or $b$ are color indices in the $3$ representation while the same with a hat are a color indices in the $\bar{3}$ representation. The nature of the color index specifies whether a given quantity refers to a quark or an antiquark. Thus, ${\boldsymbol r}_a$ represents the coordinate of a quark, while ${\boldsymbol r}_{\hat a}$ represents the coordinate of an antiquark. We use also compact notation, such as ${\boldsymbol r}_{ab}={\boldsymbol r}_a-{\boldsymbol r}_b$ to denote the relative distance between a quark of color $a$ and a quark of color $b$, or ${\boldsymbol r}_{a\hat{b}}={\boldsymbol r}_a-{\boldsymbol r}_{\hat b}$ to denote the relative distance between a quark of color $a$ and an antiquark of color $\hat b$. Finally, for functions of coordinates, we set $F_{\hat a b}=F(\hat {\boldsymbol r}_a-{\boldsymbol r}_b)$, $\Gamma_{ab}=\Gamma({\boldsymbol r}_a-{\boldsymbol r}_b)$, and so on. \begin{figure} \begin{center} \includegraphics[scale=0.6]{./50traj.eps} \caption{Example of random trajectories of fifty heavy quark-antiquark pairs. The pairs are prepared as explained in the text, and evolve during a time $t=5\,\rm{fm/c}$.} \label{} \end{center} \end{figure} In Fig.~\ref{} we plot some random trajectories of fifty pairs of quarks and antiquarks. The parameters are different from the ones used in the previous section, now the temperature is $T=250\,\text{MeV}$ and the cut-off for $V({\bf r})$ is $\Lambda=1500\,\text{MeV}$. We keep the same value of $\gamma$. This new choice of parameters makes the problem of the violent hard kicks less severe, at the cost of having a cutoff $\Lambda$ unrealistically small (it is of the order of the inverse of the Bohr radius of the ground state). The system is prepared in the following way: in a square of size $2.5\,\rm {fm}$ we chose fifty random points; in each point we put a quark-antiquark pair following a probability for the relative coordinate given by \begin{equation} \frac{2}{\pi r}\sin(\Lambda r)\,, \end{equation} where $\Lambda$ is the same cut-off as for the real part of the potential (this distribution becomes a Dirac delta as $\Lambda\to\infty$). The fifty quark-antiquark pairs then evolve for a time $t=5\,\text{fm/c}$, according to the stochastic equations displayed above. As can be observed by looking at Fig.~\ref{} some of the trajectories remain close enough to allow for ``recombinations'' into bound state. \begin{figure} \begin{center} \includegraphics[scale=0.5]{./hist_new.eps} \caption{Histogram of the number of bound states formed in $300$ simulations with the same initial conditions as in Fig. \ref{}} \label{} \end{center} \end{figure} To quantify the phenomenon, we perform a statistical analysis of how many bound states are observed at the end of the evolution, starting from the previous initial condition. To define a bound state, we follow the procedure of Ref.~\cite{}, but with slightly different parameters. We declare a heavy quark-antiquark pair to be bound if the quark and the antiquark remain at a distance smaller than $0.5\,\rm{fm}$ during a time bigger than $0.1\,\rm{fm/c}$. This procedure can become ambiguous when many quarks and antiquarks are found in a small region, for example in the case in which two quarks and two antiquarks are contained in a sphere of radius smaller than $0.25\,\rm{fm}$. In this situation we count the number of bound states by choosing the combination that yields the maximum number. The results obtained after $300$ simulations are shown in Fig. \ref{}. We can see that from the $50$ initial bound states, on average $17$ remain after a time of $5\,\rm{fm/c}$ spent inside the medium. This is to be contrasted for Fig.~\ref{} which would suggest that all pairs should become unbound if they were evolving independently of each other. Of course, we should recall that different parameters have been used in the two calculations. However, repeating the simulation of the previous subsection with the present parameters, one finds that about 10 to 15\% of the bound states would survive after a time $t=5\,$fm/c. This is about half of what the present calculation suggests. We may therefore take this result as evidence for recombination, in line with what was found in Ref.~\cite{}. \subsection{Langevin equations coupled to random collisions} \label{sec:mc} We now turn to the second strategy presented in the introduction of this section. This will allow us, in particular, to bypass the limitations of the perturbative diagonalization of the matrix in Eq.~(\ref{Lpert}), caused by the vanishing of $\Gamma({\boldsymbol r})$ at small ${\boldsymbol r}$. Let us then go back to the small $y$ expansion of Eq.~(\ref{eqfinDs}), and let us temporarily neglect the terms that go like $\frac{y}{M}$ or $y^2$, that is we keep only the kinetic term and the force term. We get \begin{equation}\label{DsMC} \partial_t D_s=\frac{2i}{M}{\bf \nabla_r}\cdot{\bf \nabla_y}D_s-2C_F\Gamma({\bf r})(D_s-D_o)-iC_F{\bf y}\cdot{\boldsymbol F}({\boldsymbol r})D_s\,, \end{equation} \begin{equation}\label{DoMC} \partial_t D_o=\frac{2i}{M}{\bf \nabla_r}\cdot{\bf \nabla_y}D_o-\frac{1}{N_c}\Gamma({\bf r})(D_o-D_s)+i\frac{1}{2N_c}{\boldsymbol y}\cdot {\boldsymbol F}({\boldsymbol r})D_o\,. \end{equation} The equation for $D_{\rm o}$ is not given explicitly in Eq.~(\ref{eqfinDs}), but it is easily derived from the material presented in ~\ref{}. We note that only the terms proportional to $\Gamma({\bf r})$ mix singlets and octets, i.e. the terms involving the force preserve the color state of the pair. We now perform a Wigner transform with respect to the variable ${\boldsymbol y}$, and define $P_{\rm s}=D_{\rm s}$ and $P_{\rm o}=(N_c^2-1)D_{\rm o}$, the probabilities for the pair to be in a singlet or octet state, respectively. We get \begin{equation} \left[ \partial_t +\frac{2{\bf p}\cdot {\bf \nabla_r}}{M}-C_F {\boldsymbol F}({\boldsymbol r})\cdot{\mbox{\boldmath$\nabla$}_{\boldsymbol p}}\right]P_{\rm s}=-2C_F\Gamma({\boldsymbol r})\left(P_{\rm s}-\frac{P_{\rm o}}{N_c^2-1}\right)\,, \label{} \end{equation} \begin{equation} \left[ \partial_t +\frac{2{{\boldsymbol p}}\cdot{\mbox{\boldmath$\nabla$}_{\boldsymbol r}}}{M}+\frac{1}{2N_c}{\boldsymbol F}({\boldsymbol r})\cdot{\mbox{\boldmath$\nabla$}_p}\right]P_{\rm o}=-\frac{1}{N_c}\Gamma({\bf r})(P_{\rm o}-(N_c^2-1)P_{\rm s})\,. \label{} \end{equation} The right hand sides of these two equations can be interpreted as a ``collision term'' in a Boltzmann equation, with gain and loss terms. Note that these collision terms are opposite in the singlet and octet channels, as expected: \beq 2C_F \left(P_s-\frac{P_o}{N_c^2-1}\right)=-\frac{1}{N_c}(P_o-(N_c^2-1)P_s). \eeq The left hand sides of the equations (\ref{}) and (\ref{}) describe the relative motion of the pair under the influence of the color force ${\boldsymbol F}({\boldsymbol r})$. The corresponding classical equations of motion read \beq\label{classeqmotion} \frac{d{\bf r}}{dt}=\frac{2{\bf p}}{M},\qquad \frac{d{\bf p}}{dt}=-C_F{\bf F}({\bf r}),\qquad \frac{d{\bf p}}{dt}=\frac{1}{2N_c}{\bf F}({\bf r}), \eeq where the last two equations refer to the singlet and octet channels, respectively. Thus, instead of treating the singlet-octet transitions as an additional color force in a Langevin equation, we can treat these transitions as ``collisions''. In practice we can solve the set of equations (\ref{}) and (\ref{}) using a Monte Carlo method, deciding at each time step, according to a probability proportional to the respective decay widths, whether a transition takes place or not, and then evolve the system through the time step according to the classical equations of motion (\ref{classeqmotion}). This is somewhat analogous to the Monte Carlo Wave Function method applied to a 2-level problem in Ref.~\cite{}. The equations (\ref{}) and (\ref{}) capture some of the important physics but they miss the drag forces and the stochastic forces that have to go with them in order to fulfill the fluctuation-dissipation theorem. These come from the semi-classical corrections that we have left out in writing Eqs.~(\ref{DsMC}). However, if we were to include these corrections as they appear for instance in Eq.~(\ref{eqfinDs}), we would introduce extra couplings between $D_{\rm s }$ and $D_{\rm o}$ that would lead in particular to a collision term involving derivatives of $P_{\rm s,o}$ and we do not know of any efficient numerical tools to solve the resulting equation. However, if the system is not too far from the maximum entropy state, terms that go like $y^2\left(P_s-\frac{P_o}{N_c^2-1}\right)$ or ${\frac{y}{M}}\left(P_s-\frac{P_o}{N_c^2-1}\right)$ are small and can be safely neglected. We again rely on the assumption that color relaxes faster than the relative motion. Under these conditions, and after performing the Wigner transform, we obtain for the singlet \beq\label{FPPs} &&\left\{\partial_t +{\boldsymbol v} \cdot \mbox{\boldmath$\nabla$}_{\boldsymbol r}-C_F{\boldsymbol F}({\boldsymbol r})\cdot\mbox{\boldmath$\nabla$}_{\boldsymbol p} -\frac{1}{2}\mbox{\boldmath$\nabla$}_{\boldsymbol p}\cdot \mbox{\boldmath$\eta$}_{\rm s}({\boldsymbol r}) \cdot\left( \mbox{\boldmath$\nabla$}_{\boldsymbol p}+\frac{{\boldsymbol v}}{T} \right)\right\}P_{\rm s}\nonumber\\ &&\qquad=-2C_F\Gamma({\bf r})\left(P_{\rm s}-\frac{P_{\rm o}}{N_c^2-1}\right). \eeq with \beq \mbox{\boldmath$\eta$}_{\rm s}({\boldsymbol r})=\frac{C_F}{2}\left( {\cal H}(0)+{\cal H}({\boldsymbol r})\right), \eeq Comparing the operator in the first line of this equation with that given in Eqs.~(\ref{FPrel0}) to (\ref{FPrel3}), we easily derive the following stochastic equations: \beq {\boldsymbol v}=\frac{{\rm d}{{\boldsymbol r}}}{{\rm d} t}=\frac{2{{\boldsymbol p}}}{M},\qquad \frac{{\rm d}{\boldsymbol p}}{{\rm d} t}=-C_F{\boldsymbol F}({{\boldsymbol r}})-\mbox{\boldmath$\gamma$}_{\rm s}\cdot {\boldsymbol v}+\mbox{\boldmath$\xi$}({{\boldsymbol r}},t)\,, \eeq where \beq \langle \xi^i({\bf r},t)\xi^j({\bf r},t')\rangle=\delta(t-t')\eta^{ij}_{\rm s}({\boldsymbol r})),\qquad \mbox{\boldmath$\gamma$}_{\rm s}=\frac{1}{2T}\mbox{\boldmath$\eta$}_{\rm s}. \eeq These equations are the analogs of Eq.~(\ref{langevinrelative}) for the QED case. Performing completely analogous manipulations, we obtain a similar result for the octet. The Fokker-Planck equation reads \beq\label{FPPo} &&\left\{\partial_t +{\boldsymbol v}\cdot \mbox{\boldmath$\nabla$}_{\boldsymbol r}+\frac{1}{2N_c}{\boldsymbol F}({\boldsymbol r})\cdot\mbox{\boldmath$\nabla$}_{\boldsymbol p} -\frac{1}{4}\mbox{\boldmath$\nabla$}\cdot\mbox{\boldmath$\eta$}_{\rm o}\cdot\left( \mbox{\boldmath$\nabla$}_{\boldsymbol p}+\frac{{\boldsymbol v}}{T} \right)\right\}P_{\rm o}\nonumber\\ &&\qquad=-\frac{1}{N_c}\Gamma({\bf r})\left(P_{\rm o}-(N_c^2-1)P_{\rm s}\right). \eeq and the corresponding Langevin equation is \begin{equation} \frac{{\rm d}{{\boldsymbol p}}}{{\rm d} t}=\frac{1}{2N_c}{\boldsymbol F}({{\boldsymbol r}})-\mbox{\boldmath$\gamma$}_{\rm o}\cdot {\boldsymbol v}+\mbox{\boldmath$\xi$}({{\boldsymbol r}},t)\,, \end{equation} with now \begin{equation} \mbox{\boldmath$\eta$}_{\rm o}({\boldsymbol r})=\frac{1}{2}\left(C_F{\cal H}(0)-\frac{1}{2N_c}{\cal H}({\boldsymbol r})\right),\qquad \mbox{\boldmath$\gamma$}_{\rm o}=\frac{1}{2T}\mbox{\boldmath$\eta$}_{\rm o}. \end{equation} We shall now present the results of the simulation of these equations, for the case of a single quark-antiquark pair. We consider first the static limit, and then turn to the full equations including the semi-classical corrections. \subsubsection{The static limit} \begin{figure} \begin{center} \includegraphics[scale=0.5]{./colorro.eps} \caption{Probability of having a static quarkonium in a color state as a function of $\tau={2C_F\alpha_sTt}/{10}$ assuming that the initial probability is ${1}/{2}$ and the temperature $T=250\,\textit{MeV}$. We compare the analytic results with the average of different number of simulations} \label{} \end{center} \end{figure} The study of the static limit (or infinite mass limit) offers us the possibility to test the numerical method, since the exact solution can be obtained analytically in this case. In particular, this will give us an idea of the number of iterations that are needed in order to get a good estimate. We consider a heavy quark-antiquark pair, whose relative distance is $r=0.1\,\rm{fm}$, and in a well-defined color state, in a quark-gluon plasma at temperature $T=250\,\rm{MeV}$. The equations to be solved are Eqs.~(\ref{eqDsDoMinfty}). If the initial conditions are such that singlet or and octet states are equally probable, i.e., $P_{\rm s}(0)=P_{\rm o}(0)$, the probability to be in an octet state at time $t$ is \begin{equation} P_{\rm o}(t)=\frac{N_c^2-1}{N_c^2}-\frac{N_c^2-2}{2N_c^2}e^{-N_c\Gamma({\boldsymbol r})\,t}\,, \end{equation} and that to be in a singlet state is $P_{\rm s}(t)=1-P_{\rm o}(t)$. We can compare this result to that of a simulation using the Monte Carlo method described above. The results are plotted in Fig. \ref{}, as a function of $\tau\equiv (2C_F\alpha_sTt)/{10}$, with a time step $\Delta\tau=0.02$. We see that for 100 events the results of the simulation match relatively well the analytic result, although sizeable fluctuations remain. The simulations to be presented next involve 1000 events. \subsubsection{Simulation with dynamical quarks} \begin{figure} \begin{center} \includegraphics[scale=0.5]{jpsi.eps} \caption{Comparison of the evolution of a pair of heavy quarks initially prepared in a $J/\Psi$ state with or without considering the transition into octet states. The screening radius is $r_D=m_D^{-1}$.} \label{} \end{center} \end{figure} \begin{figure} \begin{center} \includegraphics[scale=0.5]{colora.eps} \caption{Probability to find a heavy quark-antiquark pair in an octet state at time $t$, after it has been prepared at time $t=0$ as a $J/\Psi$ state.} \label{} \end{center} \end{figure} We consider now the full Eqs.~(\ref{FPPs}) and (\ref{FPPo}). In Fig.~\ref{} we plot the average mean distance $\langle r_{c\bar{c}}\rangle$ of a pair of charm quarks prepared in a $J/\Psi$, according to the prescriptions used in Ref.~\cite{}. That is, the radius of the pair is chosen randomly between 0 and $1/m_D$, and the relative momentum is chosen according to a Maxwell distribution with most probable velocity given by $v^2=0.3$. Finally one retains only pairs with binding energy bigger than 500 MeV and radius bigger than 0.1 fm. The temperature is taken to be $T=160\,\rm{MeV}$, the charm quark mass $M_c=1460 $ MeV, and $\gamma/M_c=0.2$ fm. These conditions differ slightly from those used earlier: the reason is that we want to check our results against those of \cite{} in a domain where they could be compared (see Fig.~\ref{}). Thus, we also use a different running coupling than above, $\alpha_s=0.5/ (1+0.76 \ln (T/160))$, and $m_D=(16\pi\alpha_s/3)T^2$ (for two massless flavors). The cutoff on $V({\boldsymbol r})$ is 4GeV, while that on $W({\boldsymbol r})$ is $4.58 \,m_D$. The unit of time in Fig.~\ref{} is the physical unit fm/c. We compare the results of the simulation of Eqs.~(\ref{FPPs}) and (\ref{FPPo}) with those obtained by neglecting color rotation, i.e., the singlet-octet transitions. The latter case is equivalent to a QED simulation, and indeed our result in that case reproduce those obtained in \cite{} (cp. the corresponding result in Fig.~\ref{} with Fig.~5 in \cite{}). As expected, we see that the bound state tends to remain bound longer if the transition to octet is not taken into account. The effect of color rotation is clearly to accelerate the melting of the bound state, although, according to the criterion used in \cite{}, $ \langle r_{c\bar{c}}\rangle\lesssim r_D=m_D^{-1}$, we may consider the system to remain bound at time $t=4$ fm/c. This is to be contrasted with the result obtained with the Langevin equation with a color random force: in the present case, the disappearance of the bound state is a more gradual phenomenon, not amplified by unphysical violent kicks of a random color force. This gradual transition can be visualized by looking at the evolution of the probability to find the pair in an octet state, which is plotted in Fig.~\ref{}. We can see that it takes a non negligible time to lose the information that the system was initially in a singlet state. \\ \section{Summary and outlook} In this paper we have obtained a set of equations for the time evolution of the reduced density matrix of a collection of quark-antiquark pairs immersed in a quark-gluon plasma in thermal equilibrium. These equations are fairly general (they are valid for an arbitrary number of heavy particles), and rely on two major approximations: weak coupling between the heavy quarks and the quark-gluon plasma, small frequency approximation for the plasma response. In the weak coupling approximation, the plasma sees the heavy quarks as a perturbation, and responds linearly to it. This response is characterized by a set of correlators, expectations values of gauge fields in the equilibrium state of the plasma, and because the heavy quark motion is slow on the typical scale of the plasma dynamics, only static, or nearly static response functions are needed. These functions account for some of the dominant effects of the plasma on the dynamics of the heavy quarks: the screening of the instantaneous Coulomb interaction between the heavy quarks, and the effect of soft, low momentum transfer, collisions of the heavy quarks with the plasma constituents taken into account by an imaginary potential. The main equations that result from these two approximations alone generalize the equations that were obtained for an abelian plasma in the path integral formalism, using the Feynman-Vernon influence functional method \cite{}. Their structure is close to that of a Lindblad equation, and they are essentially equivalent to the equations obtained for QCD by Akamatsu \cite{}, although the present formulation differs from his in several aspects. Recently a Lindblad equation was obtained for the evolution of the density matrix of a quark-antiquark pair, using similar approximations, but formulated in the context of a non relativistic effective theory (pNRQCD, \cite{}). This formalism puts the emphasis on the singlet-octet transitions, and the validity of the employed effective theory requires specific conditions, viz. $1/r\gg T\sim \gg E$, with $E$ the typical binding energy. The corresponding Lindblad equation keeps the quantum features of the problem, however at the price of a high computational cost. In the case of abelian plasmas, a further approximation, the semi-classical approximation, leads to a Fokker-Planck equation, and a corresponding Langevin equation, which are relatively easy to solve numerically. When trying to extend this semi-classical approximation to QCD, we have to face new features related to color dynamics. In the particular case of a quark-antiquark pair, this involves the transitions between the singlet and the octet color configurations of the pair. Taking these transitions into account yields coupled equations for the two independent components of the density matrix, that are not easily solved, even when the motion of the heavy particles is treated semi-classically. We have then explored numerically two strategies to solve approximately these coupled equations. In the first one, we assume that the color dynamics is fast compared to the motion of the heavy quarks. In this case, the collisions drive the systen quickly to a maximum entropy state where all colors are equally probable and uncorrelated. One can then use the Langevin equations to describe the dynamics in the vicinity of this maximum color entropy state, using a perturbative approach. This is sufficiently simple that it can be generalized to a system of an arbitrary number of quarks and antiquarks. However, the perturbative approach is limited by the fact that the color relaxation is slow when the size of the quark-antiquark pairs is small, which may lead to unphysical behavior for a physically relevant choice of parameters. To overcome this limitation, we have explored another strategy, which appear more promising. It consists in treating the singlet-octet transitions as collisions, viewing the corresponding equations as Boltzmann equations that we solved using Monte Carlo techniques. Although they are fairly general, the equations that we have obtained so far do not yet capture all the relevant physics. For instance, in the particular case of a single quark-antiquark pair, the transitions between singlet and octet color states cause rapid changes in the heavy quark hamiltonian. These are not properly handled, and may be in conflict with the assumption that the dynamics of the heavy particles is slow compared to that of the plasma. In addition we have left aside the possibility of absorption or emission of real gluons from the plasma, that are responsible in particular for gluo-dissociation, known to be an important mechanism in some temperature range. These shortcomings, and further aspects of the problem, will be addressed in a forthcoming publication. \\ \noindent {\bf Acknowledgements} This work has been supported in part by the European Research Council under the Advanced Investigator Grant ERC-AD-267258. The work of M.A.E. was supported by the Academy of Finland, project 303756.

\section{Introduction} Since Sadi Carnot proposed an ultimate quantitative bound \citep{carnot} on the performance of thermal devices, our understanding of microscopic physics has been revolutionized over the last two centuries. In contradistinction with the grounding of classical thermodynamics on a deterministic microscopic dynamical model, small scale thermal machines with constituents obeying laws of quantum theory have only recently come within our purview \citep{huberreview, kosloffreview}. However, these quantum machines still run up against the Carnot bound and barring non-thermal reservoirs like coherent \citep{nonthermal1} or squeezed baths \citep{nonthermal2,nonthermal3,nonthermal4,nonthermal5}, generally fail to surpass it. Yet, the Carnot bound, attainable only through an infinitely slow process, is of limited practical utility compared to the performance of thermal devices at finite power. Classically, a bound on the efficiency at maximum power for cyclic thermal engines\citep{reitlinger,novikov,ca} and refrigerators \citep{yan,yanchen,velasco} was obtained in several works, providing impetus to a growing body of research on finite time and non-equilibrium thermodynamic processes \citep{wubook,prigoginebook}. Investigations on finite time cyclic thermodynamic processes for quantum thermal machines \citep{abah, kurizkireview,raamuzdin} have already provided crucial physical insight on topics ranging from the origin of friction\citep{friction1,friction2,friction3,friction4} in thermal processes to the third law of thermodynamics \citep{thirdlaw1,thirdlaw2}. \\ Apart from cyclic thermal machines; one may also conceive of self-contained thermal machines where the energy required to perform a task, e.g. cooling a cold bath, is provided by a third bath, thus requiring no external control. Motivated by algorithmic cooling, such a quantum absorption refrigerator (QAR) was proposed in \citep{smallest}. Numerous other proposals of QAR have already been put forward \citep{abs1,abs5, adesso1, rc_silva}. Experimental realization of QAR has been recently achieved in ion trap systems \citep{iontrap} . Experimental proposals for various alternate systems, viz. cavity QED systems \citep{abs4} and quantum dots \citep{abs3}, have also been made. In the case of autonomous three-qubit refrigerators, the constituent qubits interact among themselves via an energy conserving unitary operation as well as thermalize with their corresponding heat baths. Depending upon the initial conditions, the setup at its steady state may be shown to act as a refrigerator which cools down the cold bath by extracting heat at a constant rate. These machines have the advantage of both simplicity as well as practical usefulness. \\ In addition to finding the steady state performance of such refrigerators, one should also consider how long it takes for the system to reach the steady state. If the system cools very reliably at steady state but only gets there very slowly, it is of limited utility. One can study the finite time characteristics of the QAR \citep{brask15, sreetama, Mitchison} for cooling the cold qubit, however this approach may necessitate precise time control mechanism for decoupling the cold qubit. In addition, our goal is steady cooling of the cold bath rather than cooling the qubit by itself. In this paper, we seek to understand the behaviour of QAR conditioned through the intrinsic restriction to evolution of a quantum system through its state space. The latter feature, known as the quantum speed limit is a fundamental feature of quantum dynamics and finds several applications in quantum computation, control theoretic settings and in the study of shortcut towards adiabaticity (See, for example, Ref. \citep{deffnerreview}. and references therein for an overview). To this end, we analyze a figure of merit encompassing this feature of the QAR described in \citep{smallest}. This figure of merit, called \emph{``bounding second order cooling rate" (BSOCR)} is defined as the product of the equilibrium cooling rate and the maximum possible speed of attaining equilibrium. This is of significant practical importance, since even if a refrigerator cools the cold bath very well in the steady regime, it will be of limited use if it takes a very long time to reach the steady regime. We link the proposed figure of merit with the transient features of the refrigerator and go on to illustrate the dependence of BSOCR on various system parameters. We show that BSOCR increases linearly with the coupling strength. We note that the presence of quantumness in the form of initial coherence in the system can boost the value of BSOCR. This is followed up by an investigation of the figure of merit vis-a-vis the refrigeration efficiency. We show that at a high temperature limit and subject to other specific conditions, we can recover an expression for efficiency at maximum value of BSOCR which scales similarly to the Curzon-Ahlborn bound. The paper is organized as follows. In section II, we briefly recapitulate the QAR model used in our work. In section III, we mention the form of quantum speed limit used. Our novel results are presented in section IV. In section V, we conclude through a discussion of our results and possible future extensions thereof. The detailed calculations are to be found in the appendix. \section{Absorption refrigerator at steady state} We consider the model of three qubit absorption refrigerator introduced in Ref. \citep{smallest}. In what follows we briefly discuss about the model and its working principle. The three qubits consisting the refrigerator are coupled to three different baths at different temperatures. The first qubit which is the object to be cooled, is coupled to the coldest bath at temperature $T_c$. The second qubit which takes energy from the first qubit and disposes into the environvent, is coupled to a hotter bath at temperature $T_r$. The third and final qubit which provides the free energy for refrigeration is coulped to the hottest bath at temperature $T_h$. Here $T_c\leq T_r\leq T_h.$ Without loss of generality, the ground state energy of all the qubits are considered to be zero and the excited state energy of the $ i$-th qubit is $ E_{i} $, where $i\in \{c,r,h\}$. The free Hamiltonian of the combined system is $ H_{0}=\sum_{i\in \{c,r,h\}} E_{i} |1\rangle_i\langle1|$. In thermal equilibrium the qubits are in the corresponding thermal states $\tau_{i}=r_{i}\Ket{0}\Bra{0}+\bar{r}_{i}\Ket{1}\Bra{1} $, where \begin{equation} r_{i}=(1+e^{-\beta_i E_{i}})^{-1} \label{popul} \end{equation} is the probability of the $i$-th qubit to be in the ground state and $\bar{r}_{i} = 1 - r_{i}$. Here $\beta$ is the inverse temperature $1/T$. {\color{black}The populations are thus dependent on both the temperature and energy spacing.} The Boltzmann costant $k_B$ is set to unity and this convention is followed through out the paper. The total system is initially in a product state of locally thermal qubits $\rho_{0}=\tau_{c}\otimes\tau_{r} \otimes \tau_{h}$. The qubits interact via the interacting hamiltonian $H_{int}=g(\Ket{101}\Bra{010}+\Ket{010}\Bra{101})$. Here the order of the qubits is $c,r,h$ which is maintained throughout the paper unless otherwise mentioned. The interaction strength $g$ is taken weak enough compared to the the energy level $E_i$-s, i.e., $ g<<E_{i} $, so that the energy levels and the energy eigenstates of the combined system are almost unaltered and the temperature of the each qubit can be defined neglecting the interaction energy \citep{smallest}. The total Hamiltonian of the combined system is given by \begin{equation} H=\sum_{i\in \{c,r,h\}}E_i |1\rangle_i\langle1|+ g(\Ket{101}\Bra{010}+\Ket{010}\Bra{101}). \end{equation} As the qubits are coupled with heat baths at each time step there is a finite probability that it will thermalize. Suppose $ p_{i} $ is the probability density per unit time that the $ i$-th qubit will thermalize back. Then the evolution of the combined system is given by the following master eqaution \begin{equation} \label{master} \frac{\partial \rho}{\partial t} = -i[H,\rho]+\sum_{i\in \{c,r,h\}} p_{i} (\tau_{i}\otimes \tr_{i}\rho - \rho). \end{equation} It is necessary to mention that this master equation is valid only in the perturbative regime where the simultaneous thermalization of more than one qubit can be neglected. The steady state refrigeration with the aforementioned model has been demonstrated in great detail in Ref. \cite{smallest, jpa}. The steady state for this master equation, as obtained in Ref. \citep{jpa}, is given as \begin{equation} \label{final-state} \rho_{f} = \tau_{c} \otimes \tau_{r} \otimes \tau_{h} + \gamma \sigma_{crh}, \end{equation} where \begin{eqnarray} \sigma_{crh} &=&\Big(Q_{rh} Z_c\tau_r\tau_h + Q_{ch} \tau_c Z_r \tau_h + Q_{cr} \tau_c \tau_r Z_h\ \nonumber \\ && + q_c \tau_c Z_{rh} + q_r \tau_r Z_{ch} + q_h Z_{cr}\tau_h + Z_{crh} + \frac{q}{2g}Y_{crh}\Big). \end{eqnarray} Here, $Y_{crh}~=~i|101\rangle\langle010|-i|010\rangle\langle101|$ and $Z_{crh} = |010\rangle\langle010|-|101\rangle\langle101|$, $Z_{ij} = \tr_k Z_{ijk}$, where $\{i,j,k\}\in\{c,r,h\}$ and $q_i$ and $Q_{jk}$, are given as follows $ q_i = \frac{p_i}{q-p_i}$, $Q_{jk} = \frac{p_jq_k + p_kq_j}{q-p_j-p_k}$, where $q = p_c + p_r + p_h$. The parameter $\gamma$ in Eq. \ref{final-state} is given by \begin{equation}\label{e:gamma} \gamma = \frac{-\Delta}{2 + \frac{q^2}{2g^2} + \sum_i q_i + \sum_{jk} Q_{jk}\Omega_{jk}} \end{equation} where $\Delta = r_1 (1-r_2) r_3 - (1-r_1) r_2 (1-r_3), \quad \Omega_{jk} = r_j'(1-r_k') + (1-r_j')r_k.$ Here $r'_i = (1-r_i)$ for $i=r$, otherwise $r'_i = {r_i}$. The cooling rate for the cold bath is given as \begin{equation} {Q_{c} = q\gamma E_{c}, \label{qc}}\end{equation} which clearly shows that the machine acts as a refrigerator only when $\gamma > 0$. It has been shown that the efficiency $\eta$ of this refrigerator is equal to $E_c / E_h $ \citep{jpa}. \section{Quantum speed limit} Quantum evolution for a closed system is a unitary map. It has been shown that the fluctuation \citep{mandelstam,anandan} or the average value \citep{margolus, lloyd} of the generator of such maps determines the maximum rate of unitary evolution of a quantum system through the corresponding state space, giving rise to the concept of a limiting speed for dynamical evolution. For pure quantum states, this speed of evolution was introduced by Anandan and Aharonov \citep{anandan} utilizing the Fubini-Study metric, with subsequent works \citep{pati, vaidman} building on the concept. Generalizations for mixed states in the case of unitary evolutions were proposed \citep{giovanetti} and tightness of bounds found earlier were proved in some cases \citep{toffoli}. For a generalized quantum evolution characterized by CPTP maps, it is possible to find similar speed limits \citep{debamixed, deffnerlutz,Debasis, taddei, diego}. In particular, for a Markovian channel on an open quantum system expressed via a dynamical subgroup $\mathcal{L}$, the following lower bound on the time $t_{evolution}$ required for evolution of a quantum system from initial state $\rho_{0}$ to a state $\rho_{f}$ was given in Ref. \citep{delcampo} as \begin{equation} \label{delcampo} t_{evolution} \geq \frac{|\cos \theta -1| \tr \rho_{0}^{2}}{\sqrt{ \tr \left(\mathcal{L}^{\dagger} \rho_{0} \right)^{2}}} = \tau, \end{equation} \noindent where $\theta = \cos^{-1} \frac{\tr \left(\rho_{0} \rho_{f} \right)}{ \tr \left(\rho_{0}^{2} \right)}$ is expressed in terms of relative purity between the initial and the final state. Thus, $1/\tau$ can be interpreted as the maximum speed of the evolution. In this paper, using this maximum speed of evolution, we show how the same constrains the performance of the QAR. \section{Effects on the performance of quantum refrigerator from quantum speed limit} In this section, we establish a link between the cooling rate of the QAR at its steady state and the minimum time that it takes to reach the steady state. We define the novel figure of merit, i.e., the product of the steady cooling power of the QAR and the maximum speed of evolution to the steady state as \begin{equation} \label{coolingrate1} \chi = \frac{Q_{c}}{\tau}. \end{equation} Note that, for better performance, we need higher $\chi$, i.e., higher cooling rate as well as faster evolution to the steady state. Interestingly, we observe that there lies a trade-off between these two desired criteria. It will be interesting to explore in detail how the performance of steady QAR depends on $\chi$, as well as how $\chi$ itself depends on the system parameters. Before doing so, we would like to digress a bit towards transient refrigeration by QAR. The efficiency and cooling rate in the transient regime by QAR have been extensively studied in Ref. \cite{sreetama}. The figure of merit $\chi$ is the upper bound on the temporal average of time derivatives of the instantaneous cooling rate. The latter quantity, say $P(t)$, is defined as $P(t) = \frac{dQ_{c}(t)}{dt}$, where $Q_{c}(t)$ is the instantaneous cooling rate. Now, the time average \begin{equation} \overline{P (t)} = \frac{1}{t_{evolution}} \int_{0}^{t_{evolution}} \frac{dQ_{c}(t')}{dt'} dt' = \frac{Q_{c}}{t_{evolution}} \leq \frac{Q_{c}}{\tau} = \chi, \end{equation} justifying our assertion. {\color{black} Thus the time-derivative of the transient cooling rate, averaged over the entire duration of dynamics, is upper bounded by our figure of merit. The transient cooling rate is, of course, the amount of heat drawn from the bath per unit time during the transient regime. Thus, the time derivative of cooling rate may be argued as a second order time-derivative of the amount of heat drawn from the bath during transient dynamics. Since our figure of merit represents an upper bound on the average time-derivative of the transient cooling rate, we coin the term \emph{bounding second order cooling rate (BSOCR)} for the figure of merit $\chi$. To avoid misleading the reader, we emphasize that the exact expression for $\chi$ is homogeneous to $Q_c$, i.e., the steady cooling rate, and not to the time derivative of the transient cooling rate. } \begin{figure} \includegraphics[scale=0.25]{motivation_plot.pdf} \caption{ {\color{black}Demonstration that high equilibrium cooling power $Q_c$ can come at the cost of slow evolution to equilibrium state. When all parameters other than $g$ are fixed, steady cooling rate $Q_c$ and the inverse of evolution time $1/ \tau$, both functions of $g$, are parametrically plotted by varying the interaction strength $g$ from $10^{-4}$ to $10^{-1}$. The parameters are chosen as $E_c = 1$, $p_c = p_r = p_h = 0.1, \eta = 1, T_c =1,T_r = 5,$ $ T_h = 10$.}} \label{motivation} \end{figure} Let us now see how the BSOCR controls the steady refrigeration. In terms of the steady state parameters, the BSOCR can be expressed in terms of the initial state $\rho_0$ via \eqref{final-state}, \eqref{e:gamma}, and \eqref{delcampo} as \begin{equation} \label{trade-off} \frac{Q_{c}}{\tau} = \frac{q E_{c}}{ \left[\tr (\rho_{0} \sigma_{crh})\right]/ \left[\sqrt{ \tr ([H_{int}, \rho_{0} ])^{2} } \right]} \end{equation} Using the equation above, we demonstrate the trade-off between the steady cooling power and the maximum speed of evolution. As $Q_c$ and $1/\tau$ both are function of the interaction strength $g$, we plot these quantities by varying the interaction strength $g$. Fig. \ref{motivation} clearly demonstrates the trade-off between steady cooling power $Q_c$ and maximum speed of evolution $1/\tau$, implying one can only get better cooling rate by compromising on the speed of evolution. This sets up the main crux of our work. It is also surprising that the performance of an autonomous quantum thermal machine at the steady state depends on the minimum time taken to reach the steady state. In the rest of the work, we shall study how the BSOCR $\chi$ depends on the parameters of the QAR. We shall also investigate whether initial quantumness can enhance the performance of the steady QAR. \subsection{Interaction strength vs BSOCR} We know that the three-body interaction cools the cold qubit and draws heat from the cold bath. Hence, the more the interaction strength, the stronger the biasing facilitating the refrigeration. However, the cooling power depends in a complicated way on the strength $g$ of the three-body interaction. On the contrary, from Eq. \ref{trade-off} (see the appendix for a detailed expression) when the resetting probabilities $\{p_i\}$ are equal), we observe that \textit{the BSOCR grows linearly with the strength $g$ of the three-body interaction $H_{int}$ when the qubits start from their respective thermal states.} \begin{figure} \centering \includegraphics[width=0.23\textwidth, keepaspectratio]{g_with_coherence.pdf} \label{with_coh_g_1} \quad \includegraphics[width=0.23\textwidth, keepaspectratio]{with_g_different_subspace_coherence.pdf} \label{with_coh_g_2} \caption{ Variation of BSOCR with interaction strength $g$. \textit{Left:} for different strengths of initial coherence $\kappa$ applied to the $|000\rangle\langle111|$ subspace as well as \textit{Right : }for same amount of coherence in different subspaces viz. $36$, i.e., $|101\rangle\langle 010|$, and $18$, i.e., $|000\rangle\langle 111|$ subspaces. The reset probabilities have been taken as $p_c = p_r = p_h = p$ throughout. The other parameters are taken as $T_c = 1, T_r = 2, T_h = 10 , \eta = 0.5 ,$ $ p = 0.05, E_c = 1$.} \label{g_with_coherence} \end{figure} In Fig. \ref{g_with_coherence}, we demonstrate the linear dependence of the BSOCR on the interaction strength. It will be interesting to inquire whether some initial coherence in the three qubit QAR can boost the BSOCR beyond the linear increase with the interaction strength $g$ seen above. We add an additional real off diagonal part to the $|000\rangle\langle 111|$ and the corresponding adjoint element of the diagonal initial density matrix $\rho_{0}$ with the following magnitude \begin{equation} = \kappa \sqrt{\prod_{i=c,r,h} r_{i} \bar{r_{i}}}\quad ; \quad 0 \leq \kappa \leq 1. \label{inputcoh}\end{equation} Fig. \ref{g_with_coherence} shows that this initial coherence can significantly increase the BSOCR. Thus, quantum coherence, already identified as a useful resource in quantum information theory and quantum thermodynamics \citep{adessoreview, spekkens, avijit, lostaglio}, can enhance the performance of the QAR for a fixed interaction strength by reducing the minimum time taken to reach the steady state. It is worth mentioning that in Ref. \citep{Mitchison}, it has been pointed out that coherence can enhance the cooling of the cold qubit in the transient regime. In this work, we show that coherence can also enhance the performance of the steady heat machine. This is consistent with the assertion made in Ref. \cite{Debasis} that quantum coherence can serve to augment the speed limit for general dynamics. {\color{black} At this juncture, we want to emphasize that the present treatment of adding initial coherence to the system is different from having a bath which is coherent. The choice of adding coherence only to the $|000\rangle\langle 111|$ subspace may seem restrictive, but it can be shown that the nature of functional dependence of the BSOCR on the interaction strength or reset probabilities does not change whether we add the coherence in any other density matrix element, say, $|010\rangle\bra101|$. However, the numerical value of the BSCOR depends on the subspace to which coherence is added. In Fig. \ref{g_with_coherence}, we show that as far as the efficacy of applying coherence to facilitate cooling is concerned, applying coherence to $|000\rangle\bra111| $ subspace is far better than applying the same amount of coherence to $|010\rangle\bra101|$ subspace.} \subsection{Thermalization strength vs BSOCR} \begin{figure} \centering \includegraphics[width=0.23\textwidth, keepaspectratio]{Zeno_with_coherence.pdf} \label{with_coh_p_1} \quad \includegraphics[width=0.23\textwidth, keepaspectratio]{with_p_different_subspace_coherence.pdf} \label{with_coh_p_2} \caption{ Variation of BSOCR with reset probability $p_{c} = p_{r} $ = $p_{h}$ $= p$. \emph{Left :} for different strengths of initial coherence $\kappa$ in the $|000\rangle\langle 111|$ subspace as well as \emph{Right :} for same amount of coherence in different subspaces viz. $36$, i.e., $|101\rangle\langle 010|$, and $18$, i.e., $|000\rangle\langle 111|$ subspaces. The other parameters are taken as $T_c = 1, T_r = 2, T_h = 10, \eta = 0.5, g = 0.05, E_c = 1$.} \label{zeno_with_coherence} \end{figure} In this section we study the dependence of the BSOCR on the thermalization strengths, i.e., the reset probabilities. In case the reset probabilities are different, the expression for BSOCR is a bit cumbersome. Thus we confine ourselves to the case of equal reset probabilities. If the reset probabilities vanish, there is no energy exchange and consequently no cooling either. \emph{If the reset probabilities $p_{c},p_{r}, p_{h}$ are all equal to, say $p$, then the BSOCR is linearly proportional to $p$.} Again, one can verify that this simple linear relation does not hold for the heat extracted from the cold bath. Now, let us start with a slightly more general case where the density matrix corresponding to the initial state $\rho_{0}$ possesses some non-local coherence, the latter typified by non-zero elements as in Eq. \ref{inputcoh} in the earlier section. It is readily verified that the reduced states are still Gibbsian in their respective energy eigenbases. Fig. \ref{zeno_with_coherence} shows that although the BSOCR increases rapidly with increasing reset probability if the $p$ is very low, it decays considerably as we go on increasing $p$. Of course we cannot indefinitely go on increasing the reset probability due to the weak coupling assumption used in the derivation of this master equation but the decrease of the BSOCR with increasing reset probability is observed even in the regime depicted in the figure, where $g, p \ll E_{1} $. \subsection{Efficiency vs BSOCR} \begin{figure} \includegraphics[scale=0.4]{Power_vs_eta_unequal_p.pdf} \caption{Variation of BSOCR with steady state efficiency $\eta$ for fixed parameter values $p_c = 0.01, p_r = 0.02, p_h = 0.05, T_c = 1, T_r = 2, T_h = 10, E_c = 1, g = 0.01$.} \label{pow_vs_eta} \end{figure} For finite time thermodynamic cycles, finding the efficiency at maximum power is a well-motivated pursuit. For QAR models, in addition to finding the efficiency at maximum power \citep{adesso1}, it makes sense to find the efficiency of the QAR when the BSOCR, the figure of merit encompassing both steady power characteristics and speed of evolution, is maximal. Let us now focus on how the BSOCR depends on the efficiency of the machine. For this model, the efficiency $\eta$ is expressed as $E_{c}/E_{h}$ . This allows us to explicitly compute the BSOCR for any efficiency. We demonstrate in Fig. \ref{pow_vs_eta} that in a generic case, the BSOCR vanishes when there is no cooling as well as at the Carnot point, and attains its maximal value at some intermediate point. Exactly calculating the efficiency at maximal BSOCR in the most general case is quite cumbersome, therefore we restrict ourselves to the case of equal reset probabilities and use the expression for BSOCR derived in the supplementary material \citep{supple}. In the high temperature limit and with the following assumptions, this allows us to derive the following crucial result, the proof of which can be found in the supplementary material \citep{supple}. \emph{\textbf{Theorem} - If the reset probabilities $p_{c},p_{r}, p_{h}$ are equal and $T_{h} \gg T_{r} \gg T_{c} \gg E_{c}$ along with the condition that $\frac{T_{c}}{T_{r}} \approx \frac{T_{r}}{T_{h}}$, the efficiency at maximal BSOCR equals \begin{equation} \frac{T_{c}}{T_{r}}\left(1-\sqrt{\frac{T_{c}}{T_{h}}}\right). \label{etamax} \end{equation}} \noindent Let us mention that in the limit considered above, the Carnot efficiency for this refrigerator is given by \begin{equation} \eta_{Carnot} = \frac{T_{c}}{T_{r}} \left( 1 - \frac{T_{c}}{T_{h}} \right). \label{etacarnot} \end{equation} We immediately see that the relation between Eq. \ref{etamax} and Eq. \ref{etacarnot} is remarkably similar to the relation between the expressions for Carnot bound and Curzon-Ahlborn (CA) bound for efficiency heretofore derived for various cyclic heat pumps \citep{abah, yan}. \section{conclusion} We have introduced a figure of merit, which we call the BSOCR, defined as the product of the steady cooling power and the maximum speed of evolution towards the steady state, and analyzed it for a specific model of QAR. We have shown that the BSOCR scales linearly with the interaction strength. We have also observed that the BSOCR increases with the value of reset probabilities. Consequently, we have observed how the relation of the BSOCR with interaction strength and reset probabilities change with the introduction of initial quantumness in the system. Interestingly, we have demonstrated that the initial quantumness is helpful to increase the magnitude of BSOCR. Strikingly, we have shown that the efficiency at maximum BSOCR has a simple expression reminiscent of the CA type expressions in a limiting case. That the steady state performance parameter depends on the dynamics which generates the steady state is perhaps worth exploring further for more generic heat machines. Further work elaborating on the link between cycle based and self contained thermal machines may elucidate this connection even more. In place of the speed limit used in this work, one may put alternate expressions for the speed limit to derive alternate forms of the BSOCR. Further work on power characteristics of general quantum thermal devices as well as quantum biological processes using the procedure outlined here may be performed. {\color{black} Another potentially interesting avenue of future investigation is to consider non-Markovian baths, e.g., Fermionic central spin baths, and see whether they impart any advantage as far as the BSOCR is concerned. The seminal paper by Deffner and Lutz \citep{deffnerlutz} indeed demonstrates that the time taken to reach a target state may be significantly less for non-Markovian processes. Thus, if a steady state exists for a non-Markovian model, the value of our figure of merit may be significantly enhanced. One particular obstacle towards pursuing this idea in a straighforward way is the fact that for such baths - there may not even exist any unique steady state in general. We are optimistic that engaging non-equilibrium steady states in the discussion of quantum autonomous refrigators may be a very fruitful approach (see, for example, Ref. \cite{ness} for such a recent work). The treatment of such situations lies outside the scope of the present paper, but may be considered as important future projects. } At the time of completion of the first draft of the present manuscript, two recent preprints \citep{csl1,csl2} appeared which proposed the existence of speed limits to evolution in the classical phase space. It may be interesting to perform a similar analysis to ours for classical heat machines. The authors acknowledge financial support from Department of Atomic Energy, Govt. of India. \section*{Appendix} In the appendix, we present the calculations for the results in the main work. The initial state of the three-qubit refrigerator is the product of locally thermal states $ \rho_{0} = \tau_{c} \otimes \tau_{r} \otimes \tau_{h} $ and the steady state is given by $\rho_{f} = \rho_{0} + \gamma \sigma_{crh} $, as described in the main paper. First, let us note that the cooling rate $Q_c$ is expressible in the form \begin{equation} Q_c = \frac{\xi_1}{\Upsilon_1 + \frac{\Upsilon_2}{g^2} } \ \end{equation} and the inverse speed limit, i.e. minimum time required for evolution to the steady state is similarly expressible in the form \begin{equation} \tau = \frac{\xi_2 /g }{\Upsilon_1 + \frac{\Upsilon_2}{g^2} } , \ \end{equation} where the parameters $\Upsilon_1, \Upsilon_2, \xi_1, \xi_2$ all depend on the system parameters other than g. All these parameters are explicitly expressed as $\xi_1 = q E_c \Delta, \xi_2 = \frac{\Delta \tr (\tau_c \tau_r \tau_h \sigma_{crh})}{\sqrt{2} (r_1 r_3 + r_2 (r_1 + r_3 -2 r_1 r_3 -1))}, \Upsilon_1 = 2 +\sum_i q_i + \sum_{jk} Q_{jk} \Omega_{jk}, \Upsilon_2 = q^2 /2 $. Thus, if other system parameters are kept fixed and only the interaction strength $g$ is tuned, then the above parametric relation yields the following link between $Q_c$ and $\tau$, which is valid for any interaction strength. \begin{equation} \tau^2 = \left(\frac{\xi_2 ^2}{\xi_1} \right) Q_c - \Upsilon_1 \frac{\xi_2^2}{\xi_1^2} Q_c^2\end{equation} In case the first term of the RHS dominates over the second term, this immediately reveals that there exists a trade-off between the speed of evolution and the steady state cooling rate, which provides the rationale for investigating this tradeoff through our figure of merit. In this sense, the existence of this tradeoff is not obvious from the outset. However, in general, if the cooling rate $Q_c$is small, the quadratic term is subleading, and the tradeoff is observed. More specifically, in the scenarios we illustrated in the paper, for example in Fig. 1 of the main text, the first term indeed dominates and this trade-off is observed. The BSOCR $\chi$ is the ratio of the cooling power at steady state and the minimal time of evolution from $\rho_{0}$ to $\rho_{f}$, i.e., \begin{equation} \chi = q \gamma E_{c} \frac{\sqrt{ \tr \left(\mathcal{L}^{\dagger} \rho_{0} \right)^{2}}} {|\cos \theta -1| \tr \rho_{0}^{2}} \label{bsocr_appendix} \end{equation} The relative purity angle $\theta$ between initial and final states is given by $\theta = \cos^{-1} \frac{\tr \left(\rho_{0} \rho_{f} \right)}{ \tr \left(\rho_{0}^{2} \right)}$. Consequently, $|\cos \theta - 1| \tr (\rho_{0}^{2})$ equals $|\gamma \tr (\rho_{0} \sigma_{crh})| $. The conjugate of Lindbladian acting on the initial state is simplified as \begin{eqnarray} \mathcal{L}^\dagger \rho_{0} = -i [H_{int}, \rho_{0}] + \sum_{i} p_{i} \left( \tau_{i} \otimes \tr_{i} \rho_{0} - \rho_{0} \right). = -i [H_{int}, \rho_{0}] \end{eqnarray} \\ The expression for BSOCR thus equals $\frac{q E_{c} \sqrt{\tr [H_{int} , \tau_{c} \tau_{r} \tau_{h}]^{2}}}{\tr \left(\tau_{c} \tau_{r} \tau_{h} \sigma_{crh} \right) } $. For equal reset probabilities, i.e., on putting $p_{c} = p_{r} = p_{h} = p$, the expression for BSOCR written in terms of the excited state probabilities $\lbrace \bar{r}_i \rbrace$ reduces to \begin{widetext} \footnotesize \begin{eqnarray} \frac{Q_{c}}{\tau} = 3\sqrt{2}g p E_{c}\frac{\bar{r}_{c} r_{h} - \bar{r}_{r} \left(1-\bar{r}_{c} - \bar{r_{h}} + 2 \bar{r}_{c} \bar{r}_{h} \right)}{\frac{3}{2} - 3 \bar{r}_{h} + \bar{r}_{r} (-1 + 3 \bar{r}_{h} + \bar{r}_{h}^{2} ) + \bar{r}_{r}^{2} (5 - 9 \bar{r}_{h} + 4 \bar{r}_{h}^{2}) + \bar{r}_{c}^{2} (\bar{r}_{r} + \bar{r}_{r}^{2} (4-8 \bar{r}_{h}) + 8 \bar{r}_{r} \bar{r}_{h}^{2} - \bar{r}_{h} (1+ 4 \bar{r}_{h})) - \bar{r}_{c} (3 - 5 \bar{r}_{h} + \bar{r}_{h}^{2} + 6 \bar{r}_{r} \bar{r}_{h} - 3 \bar{r}_{r} + \bar{r}_{r}^{2} (9-16 \bar{r}_{h} + 8 \bar{r}_{h}^{2})) }. \nonumber \\ \label{damra} \end{eqnarray} \end{widetext} \normalsize \subsection{BSOCR in presence of initial coherence} We add an additional real positive off diagonal part the $|000\rangle\langle 111|$ and the corresponding adjoint element of the diagonal initial density matrix $\rho_{0}$ with the following magnitude \[ = \kappa \sqrt{\prod_{i=c,r,h} r_{i} \bar{r_{i}}}\quad ; \quad 0 \leq \kappa \leq 1. \] Now, the corresponding BSOCR is given by \begin{equation} \frac{Q_{c}}{\tau} = \frac{qE_{c} \sqrt{\tr \left(-i [H,\rho_{0} + \mu] - q \mu \right)^{2}}}{\left|\tr\left(\rho_{0} \sigma + \mu \sigma \right) - \frac{1}{\gamma} \tr \left(\rho_{0}\mu + \mu^{2} \right) \right|}, \end{equation} where $\mu =\kappa \sqrt{\prod_{i=c,r,h} r_{i} \bar{r_{i}}} \left( |0_{c}0_{r}0_{h}\rangle\langle 1_{c}1_{h}1_{r}| + h.c. \right) $. Explicitly computing this expression yields the following expression for BSOCR \begin{equation} = 3p E_{c} \sqrt{\frac{N_{1} p^{2} + N_{2} p + N_{3}}{D_{1} p^{4} + D_{2} p^{2} + D_{3}}}, \end{equation} where \begin{widetext} \begin{equation} N_{1} = 9\tr \mu^{2}, N_{2} = -6 \tr M\mu, N_{3} = \tr M^{2}, D_{1} = 81 \frac{\Pi_{2}^{2}}{4 g^{4} \Delta^{2}} , D_{2} = 18 \left(\Pi_{1} + \frac{\lambda \Pi_{2}}{\Delta} \right) \frac{\Pi_{2}}{2 g^{2} \Delta}, D_{3} = \left(\Pi_{1} + \frac{\lambda \Pi_{2}}{\Delta} \right)^{2}, \lambda = 2+ \sum_i q_i + \sum_{jk} Q_{jk}\Omega_{jk}. \end{equation} \end{widetext} Here, \begin{equation} M = -i [H,\rho_{0} + \mu], \gamma = \frac{-\Delta}{\lambda + \frac{9p^{2}}{2g^{2}}}, \Pi_{1} = \tr \left( \rho_{0} \sigma + \mu\sigma \right), \Pi_{2} = \tr (\rho_{0} \mu + \mu^{2}). \end{equation} \noindent Clearly only $N_{3}$ and $D_{3}$ coefficients survive in the absence of initial coherence, i.e., $\kappa = 0$, thus giving rise to the linearity with $p$ in that case. One may easily check now for $\mu = 0$ that $\frac{N_{3}}{D_{3}}$ is proportional to $g^{2}$, thus confirming the linearity of BSOCR with interaction strength. In fact it may be shown that the BSOCR can be expressed in the following alternate way \begin{equation} = g E_{c} \sqrt{\frac{N'_{1} g^{2} + N'_{2} g + N'_{3}}{D'_{1} g^{4} + D'_{2} g^{2} + D'_{3}}}, \end{equation} with only the coefficients $N'_{3}, D'_{3}$ surviving in the special case of $\kappa = 0$, i.e., no initial coherence. \subsection{Derivation of efficiency at maximum BSOCR} \emph{\textbf{Efficiency at maximal BSOCR in the high temperature limit-} If the reset probabilities $p_{c},p_{r}, p_{h}$ are equal and $T_{h} \gg T_{r} \gg T_{c} \gg E_{c}$ along with the condition that $\frac{T_{c}}{T_{r}} \approx \frac{T_{r}}{T_{h}} $ , the efficiency at maximal BSOCR equals $\frac{T_{c}}{T_{r}}\left(1-\sqrt{\frac{T_{c}}{T_{h}}} \right)$. } \textit{\textbf{ Proof :}} In the high temperature limit, the occupancies of the excited states of qubits are approximated by \begin{equation} \bar{r}_{i} \cong \frac{1}{2} - \frac{x_{i}}{4} - \frac{x_{i}^{2}}{8} + \frac{x_{i}^{3}}{4} \hspace{0.25in};\hspace{0.2in} x_{i} = \frac{E_{i}}{T_{i}}. \end{equation} Putting them in the expression for BSOCR and weeding out higher order terms using the fact that $T_{h} \gg T_{r} \gg T_{c} \gg E_{c}$ leads to the following reasonably simple expression for the BSOCR \begin{equation} \frac{Q_{c}}{\tau} \cong \frac{3\sqrt{2}g p E_{c}}{3 - \frac{x_{c} x_{r} (x_{c} - x_{r})}{3(x_{c} - x_{r} + x_{h}) }} = \frac{3\sqrt{2}g p E_{c}}{3 - \frac{1}{3}F}. \end{equation} \noindent Now optimizing the BSOCR reduces to maximizing $F$. Putting $x_{c} = \frac{E_{c}}{T_{c}} , x_{r} = \frac{E_{c}}{T_{r}} \left(1+\frac{1}{\eta} \right) , x_{h} = \frac{E_{c}}{\eta T_{h}}$ , we differentiate $F$ with respect to $\eta$ to obtain the efficiency at maximal BSOCR. The solution to the equation $\frac{\partial F }{\partial \eta} = 0$ yields the following nontrivial solutions for efficiency at extremal BSOCR - \begin{equation} \eta_{opt} = \frac{-T_{c}^{2}T_{r} + T_{c}^{2} T_{h} - T_{c} T_{r} T_{h} \pm \sqrt{T_{c}^{4} T_{r}^{2} + T_{c}^{3} T_{r}^{2} T_{h}}}{2 T_{c}^{2} T_{r} - T_{c}T_{r}^{2} - T_{c}^{2} T_{h} + 2 T_{c}T_{r} T_{h} - T_{r}^{2} T_{h} }. \end{equation} \noindent Applying the condition $T_{h} \gg T_{r} \gg T_{c}$ , we now have the expression for efficiency at extremal BSOCR as \begin{equation} \eta_{opt} \cong \frac{T_{c} T_{r} T_{h} - T_{c}^{2} T_{h} \mp \sqrt{T_{c}^{3} T_{r}^{2} T_{h}}}{(T_{r} - T_{c})^{2} T_{h}} = \frac{T_{c}}{T_{r} - T_{c}} \mp \frac{T_{c}T_{r}}{(T_{r} - T_{c})^{2}} \sqrt{\frac{T_{c}}{T_{h}}}, \end{equation} \noindent Noting $T_{r} - T_{c} \approx T_{r} $, we arrive at the following expression for efficiency at optimal BSOCR \begin{equation} \eta_{opt} = \frac{T_{c}}{T_{r}} \left( 1 \mp \sqrt{\frac{T_{c}}{T_{h}}} \right). \end{equation} \noindent Now one can show that the minus sign corresponds to the \emph{maximal} BSOCR (taking into account the fact $\frac{T_{c}}{T_{r}} \sim \frac{T_{r}}{T_{h}} $, the other solution lies beyond the Carnot efficiency) and consequently the efficiency at maximal BSOCR is given by $\frac{T_{c}}{T_{r}}\left(1-\sqrt{\frac{T_{c}}{T_{h}}} \right)$ - thus completing the proof. \qed \\ One can check that the Carnot efficiency $\eta_{Carnot}$ is given by \[ \eta_{Carnot} = \frac{T_{c}}{T_{r}} \left[ 1 - \frac{T_{c}}{T_{h}} + \left( \frac{T_{c}}{T_{r}} - \frac{T_{r}}{T_{h}} \right)\right]. \] If $\frac{T_{c}}{T_{r}} \approx \frac{T_{r}}{T_{h}} $ , the Carnot efficiency is given by \[ \eta_{Carnot} = \frac{T_{c}}{T_{r}} \left(1- \frac{T_{c}}{T_{h}} \right). \] \noindent One can compare this with the efficiency at maximal BSOCR obtained here , which is given by - \[ \eta_{opt} = \frac{T_{c}}{T_{r}} \left(1- \sqrt{\frac{T_{c}}{T_{h}}} \right) . \] \bibliographystyle{apsrev4-1}

\section{Introduction} \label{} Due to advances in miniature technologies and demands from industrial and biomedical applications, microfluidic devices, such as valves, pumps, and mixers have been developed and studied recently. More information can be found from the review paper by Stone \& Kim \cite{stone2001} and the references therein. Tweezers, which capture, transport and deliver particles in the microfluids, have drawn attention. A wide range of methods have been developed to achieve the tweezers' trapping and transportation functions, utilizing forces based on optical \cite{Liu2016}, electrical \cite{Cohen2006Suppressing}, magnetic \cite{Lee2004Manipulation}, and acoustic \cite{li2013chip} potential fields, or simply fluidic forces \cite{Petit2012,Zhou2017Dumbbell,Karimi2013}. Due to the non-contact and gentle features, the fluidic tweezers are particularly promising for biomedical applications. To better design and utilize fluidic tweezers, it is essential to understand the interaction between a particle and the flow structure, which is a fundamental fluid mechanics problem \cite{hjelmfelt1966motion,kaftori1995particle}. One of the fluid structures used to interact and to manipulate particles is the microfluidic vortex. By generating a vortex on a micro-scale, researchers have invented micro-fluidic devices for centrifuging \cite{mach2011automated}, mixing \cite{lee2010rapid}, particle sorting \cite{wang2013vortex}, and separating particles from fluids \cite{Haller2015Microfluidic}. Using corners in channels is one way to generate vortices. For example, Tanyeri \& Schroeder \cite{tanyeri2013manipulation} built a single particle manipulation and confinement device based on perpendicularly crossing channels. Such devices often utilize the inertial effect of the particle to deviate its trajectory from the fluid. Another way to generate vortices is to rotate an object controlled by an external magnetic field or ultrasound. These methods have been shown to be useful for propulsion and manipulation of particles \cite{balk2014kilohertz,Petit2012}. Qualitative force analysis has been done for rotational tweezers by Ye {\it et al.} \cite{ye2012micro,ye2014dynamic}. However, the mechanism for capture has not been fully explained. Although inertial effect and non-Newtonian effects have been considered as the mechanisms for trapping in the previous studies, and such effects could be responsible for some tweezers, particle manipulations are possible in Newtonian flows in the Stokes regime \cite{Petit2012}. Furthermore, the precise capture range and the critical translation speed of the manipulator have not yet been systematically investigated. In this paper, we study a fluidic actuator based on a rotating nanowire. We build a mathematical model, find the exact velocity field for the flow generated by a tumbling spheroid in free space, numerically solve the Stokes flow induced by a tumbling wire that interacts with a spherical particle, and examine the mechanism and condition in which the particle is captured and moving with the wire. We also investigate the relations between the translation speed, geometric parameters, and the capture region. \section{Model}\label{} \begin{figure} \begin{center} \includegraphics[width=.50\textwidth]{fluidtweezer_config_v3} \caption{Schematic of the fluid tweezers. The cylindrical nanowire with length $\ell$ and cross-section radius $r$ is rotating in a horizontal plane with an angular velocity $\omega \mathbf{e}_z$ while it also is translating with a constant velocity $U\mathbf{e}_x$. The sphere represents the target particle with a radius $R$ in the flow. The black dots along the nanowire's center-line and on the surface of the sphere illustrate the locations of the Stokeslets. }\label{} \end{center} \end{figure} \subsection{Configuration of the problem} The fluidic tweezers are made of a magnetic nanorod or nanowire controlled by magnetic fields. In our numerical model, the nanowire is modeled as a long cylinder with a length $\ell$ and a radius $r$. The target particle is a rigid sphere with radius $R$ in the fluid (illustrated in figure~\ref{}). The wire is translating with a uniform velocity $U \mathbf{e}_x$ and rotating with an angular velocity $\omega\,\mathbf{e}_ z$. Without loss of generality, the wire is initially placed horizontally in the $x-y$ plane. The position of the centerline of the wire is prescribed as \begin{eqnarray} \mathbf{x}=R_{\omega} s \, \mathbf{e}_ x + U\, t \,\mathbf{e}_ x, \label{rod_centerline} \end{eqnarray} where \begin{eqnarray} R_{\omega}=\left(\begin{array}{crc} \cos(\omega t)& -\sin(\omega t)& 0 \\ \sin(\omega t)&\cos(\omega t)& 0 \\ 0 & 0 & 1 \\ \end{array}\right) \end{eqnarray} is the rotation transformation matrix , and $-\ell/2 \leq s \leq \ell/2$. The velocity of the rod can be computed as \begin{eqnarray} \mathbf{U}_{w} =\omega s \left[ -\sin(\omega t) \,\mathbf{e}_ x + \cos(\omega t) \,\mathbf{e}_ y \right]+ U \mathbf{e}_ x. \label{rod_vel} \end{eqnarray} We consider the particle as neutrally buoyant in the fluid and that Brownian motion is negligible. Assuming the particle is centered at $\mathbf{x}_p$, the translation velocity at the center is $\mathbf{v}_p$, and the angular velocity about its center is $\boldsymbol{\omega}_p$, the surface velocity $\mathbf{U}_{p}$ can be computed as \begin{equation}\label{protate} \mathbf{U}_{p}= \mathbf{v}_p+(\mathbf{x}_0-\mathbf{x}_p) \times \boldsymbol{\omega}_p, \end{equation} for any $\mathbf{x}_0$ on the surface of the particle $\partial \Omega_p$. For flows induced by micro-fluid tweezers, the characteristic Reynolds number is $\mbox{\textit{Re}}=U\,\ell\, \rho/\mu$, based on the translation speed, or $\mbox{\textit{Re}}=\rho\, \omega \,\ell^2/\mu$, based on the rotational speed. For tweezers and particle size below 100$\mu$m, $Re$ is usually much less than one and the inertial effect is negligible. Therefore, we assume the flow is in the Stokes regime. Also, we assume that the fluid is incompressible. Then, the governing equations are reduced from the Navier-Stokes equations to the Stokes equations \begin{eqnarray}\label{Stokes} \mu \nabla^2 \mathbf{u}-\nabla p+ \mathbf{f} = 0, \quad \nabla\cdot \mathbf{u}=0, \end{eqnarray} where $\mathbf{u}$ is the fluid velocity, $p$ denotes the fluid pressure, $\mathbf{f} $ is the external force exerted on the fluid, and $\mu$ is the dynamic viscosity. The flow is at rest without the wire and therefore the boundary conditions satisfy \begin{eqnarray}\label{BD} \begin{array}{rcl} \mathbf{u}(\mathbf{x}_0)&=&\mathbf{U}_{w,p}(\mathbf{x}_0),\\ \displaystyle{\lim_{|\mathbf{x}|\to\infty}}\mathbf{u}(\mathbf{x})&=&0, \end{array} \end{eqnarray} where $\mathbf{x}_0$ is on the boundary of the wire $\partial \Omega_w$ or the boundary of the particle $\partial \Omega_p$. \subsection{Exact solution for flow induced by a tumbling spheroid nanowire} For a tumbling rod with a prolate spheroid shape, the exact solution for the induced flow can be constructed from fundamental singularity solutions \cite{Blake1971, Pozrikidis1997, Zhao2010, Camassa2011JFM}. Here we consider two reference frames. In the lab frame $\mathbf{x}$, the rod is tumbling with uniform translating velocity $U \mathbf{e}_x$ and rotating with angular velocity $\omega\mathbf{e}_ z$. The centerline of the rod is prescribed as \eqref{rod_centerline} and its velocity is \eqref{rod_vel}. In the body frame $\mathbf{x}_b$, the rod is fixed. Without loss of generality, we assume the body is along the $x$-axis. Then, the centerline of the rod is $(x_b,y_b,z_b)=(s,0,0)$, where $-\ell/2\leq s\leq \ell/2$. The relation between the body frame $\mathbf{x}_b$ and the lab frame $\mathbf{x}$ is \begin{eqnarray} \label{body-lab-vel} \mathbf{x}=R_{\omega}\mathbf{x}_b+ U\, t\, \mathbf{e}_x, \mbox{ or } \mathbf{x}_b=R_{\omega}^{T} \left( \mathbf{x}-U\, t\, \mathbf{e}_x \right). \end{eqnarray} The relation of the body frame velocity $\mathbf{v}_b$ and the lab frame velocity $\mathbf{u}$ is \begin{eqnarray*} \mathbf{v}_b = R_{\omega}^T \left( \mathbf{u}- U \mathbf{e}_x \right) +\omega \left(\begin{array}{ccc} 0& 1& 0 \\ -1&0& 0 \\ 0 & 0 & 0 \\ \end{array}\right) \mathbf{x}_b. \end{eqnarray*} Since $\mathbf{u}(\mathbf{x})=0$ as $|\mathbf{x}|\to\infty$, the equation above gives the background flow in the body frame: \begin{eqnarray}\label{backgroudV} \mathbf{v}_b(\mathbf{x}_b)&=& U\left(- \cos(\omega t+\phi)\mathbf{e}_x +\sin(\omega t+\phi) \mathbf{e}_y\right) \nonumber\\ &&+\omega \left(y_b \mathbf{e}_x -x_b\mathbf{e}_y \right), \end{eqnarray} which is the combination of a uniform flow and a rotation flow. The boundary conditions become \begin{eqnarray} \mathbf{v}(\mathbf{x}_{b,0})&=&0,\\ \lim_{|\mathbf{x}_b|\to\infty}\mathbf{v}(\mathbf{x})&=&\mathbf{v}_b(\mathbf{x}_b), \end{eqnarray} where $\mathbf{x}_{b,0}$ is the on the surface of the rod in the body frame. The surface of the spheroid rod is described by \begin{equation} \frac{x^2}{a^2}+\frac{y^2+z^2}{b^2}=1, \end{equation} where $a$ is the major semi-major axis and $b$ is the semi-minor axes of the spheroid. The related focal length $2c$ and the eccentricity $e$ are \begin{equation} c=\sqrt{{a^2}-{b^2}}= e a. \end{equation} Based on the results of Chwang \& Wu \cite{Wu1975}, we employ a line distribution of singularities between the foci $x=-c$ and $c$. The detailed derivation and a list of singularities used for this study are provided in the Appendix. For a background flow \eqref{backgroudV} past a spheroid rod, the velocity field is \begin{eqnarray}\label{exact_vel} &&\mathbf{u}(\mathbf{x})=U_1\mathbf{e}_x+U_2 \mathbf{e}_y+\omega y \mathbf{e}_x-\omega x \mathbf{e}_y -\alpha _1\int _{-c}^c\mathbf{u}_{\rm S}\left(\mathbf{x}-\bdxi ;\mathbf{e}_x\right)d\xi -\alpha _2\int _{-c}^c\mathbf{u}_{\rm S}\left(\mathbf{x}-\bdxi ,\mathbf{e}_y\right)d\xi \nonumber \\ &&+\beta _1\int _{-c}^c\left(c^2-\xi ^2\right)\mathbf{u}_{\rm D}\left(\mathbf{x}-\bdxi ;\mathbf{e}_x\right)d\xi +\beta _2\int _{-c}^c\left(c^2-\xi ^2\right)\mathbf{u}_{\rm D}\left(\mathbf{x}-\bdxi ,\mathbf{e}_y\right)d\xi \\ &&+\chi _1\int _{-c}^c\left(c^2-\xi ^2\right)\mathbf{u}_{\rm{SS}}(\mathbf{x}-\bdxi ;\mathbf{e}_x,\mathbf{e}_y)d\xi +\chi _2\int _{-c}^c\left(c^2-\xi ^2\right)^2\frac{\partial }{\partial y} \mathbf{u}_{\rm D}\left(\mathbf{x}-\bdxi ;\mathbf{e}_x\right)d\xi \nonumber \\&& +\chi _3\int _{-c}^c\left(c^2-\xi ^2\right)\mathbf{u}_{\rm R}\left(\mathbf{x}-\bdxi ;\mathbf{e}_z\right)d\xi ,\nonumber \end{eqnarray} where \begin{eqnarray*} U_1= - U \cos\left(\omega t+\phi\right),&\quad& U_2= U\sin\left(\omega t+\phi\right),\\ \chi _1=\frac{e^2 \omega }{-2 e+\left(1+e^2\right) L_e},&\quad& \chi _2=\frac{\omega -e^2 \omega }{-8 e+4 \left(1+e^2\right) L_e},\qua \chi _3=\omega \frac{2-e^2}{-2 e+\left(1+e^2\right)L_e}, \end{eqnarray*} and \begin{eqnarray*} L_e&=&\log \left(\frac{1+e}{1-e}\right). \end{eqnarray*} The singularities are on the location $\boldsymbol{\xi}=\xi\mathbf{e}_x$. All integrals in the velocity field \eqref{exact_vel} can be integrated explicitly and we can use the relation \eqref{body-lab-vel} to convert velocities between the body frame and the lab frame. In other words, we have found the explicit expression of the velocity field. \subsection{Numerical method for the fluidic tweezers and coupled system} To solve the nanowire coupled with non-zero volume particles, we implement the regularized Stokeslet method developed by Cortez \cite{Cortez2001}. The idea of regularized Stokeslets is to replace the singular force for Stokeslet with a small, smooth, and concentrated force ($\mathbf{f}=\mathbf{f}_0\phi_\epsilon(\mathbf{x}-\mathbf{x}_o)$) so that the velocity is defined everywhere, including the location of the force $\mathbf{x}_o$. The regularization parameter $\epsilon$ approximately represents the radius of the blob. When $\epsilon \to 0$, $\phi_\epsilon(\mathbf{x}-\mathbf{x}_o)\to \delta(\mathbf{x}-\mathbf{x}_o)$. We choose the regularized Stokeslet given by Cortez \cite{Cortez2001} and denote the regularized Green's function as $\mathbf{G}(\mathbf{x},\mathbf{x}_o,\epsilon)$. The blob is \begin{equation*} \phi_\epsilon(\mathbf{x})=\frac{15\epsilon^4}{8\pi ({|\mathbf{x}|}^2+\epsilon^2)^{7/2}}. \end{equation*} The velocity of the flow with a single regularized Stokeslet at $\mathbf{x}_o$ is \begin{eqnarray*} && \mathbf{U}_{\rm s}(\mathbf{x};\mathbf{f}_o)=\mathbf{G}(\mathbf{x},\mathbf{x}_o,\epsilon)\cdot \mathbf{f}_o \\ &&=\frac{1}{8\pi}\left[ \frac{\mathbf{f}_o}{(|\mathbf{x}-\mathbf{x}_o|^2+\epsilon^2)^{1/2}}+\frac{\epsilon^2\mathbf{f}_o}{(|\mathbf{x}-\mathbf{x}_o|^2+\epsilon^2)^{3/2}} + \frac{(\mathbf{f}_o\cdot(\mathbf{x}-\mathbf{x}_o))(\mathbf{x}-\mathbf{x}_o)}{(|\mathbf{x}-\mathbf{x}_o|^2+\epsilon^2)^{3/2}} \right]. \end{eqnarray*} More details of these formulae can been found in previous references~\cite{Cortez2001,Buchmann2015,Zhao2015}. To solve the problem with the proper boundary conditions, we distribute the regularized Stokeslets evenly along the center-line of the wire (see figure \ref{}). The regularization parameter on the centerline of the wire, $\epsilon_w$, is chosen as the radius of the wire $r$. For the particle, regularized Stokeslets are distributed evenly on the surface of the particle and locations of the Stokeslets are obtained by using Spherical Centroidal Voronoi Tessellation (SCVT) and the package STRIPACK \cite{Renka1997,Du2003}. At any instantaneous time, the strength $\mathbf{F}(\mathbf{x})$ of a regularized Stokeslet is determined by the no-slip boundary conditions on the wire and the spherical particle. The total velocity induced by all the regularized Stokeslets in the flow is given by \begin{eqnarray} \mathbf{u}(\mathbf{x},t) &=& \sum_{m=1}^{N_w} \mathbf{G}(\mathbf{x}, \mathbf{x}_m;\epsilon_w) \cdot \mathbf{F}(\mathbf{x}_m)+ \sum_{k=1}^{N_p} \mathbf{G}(\mathbf{x}, \mathbf{x}_k;\epsilon_p) \cdot \mathbf{F}(\mathbf{x}_k) , \label{u_F} \end{eqnarray} where $\mathbf{x}_m$ is the position of the $m$th regularized Stokeslet on the wire, $\mathbf{x}_k$ is the position of the $k$th regularized Stokeslet on the particle, and $\epsilon_p$ is the regularization parameter on the particle. Without the finite-sized particle embedded in the flow, the boundary conditions \eqref{BD} are applied on the wire only and yield \begin{equation} \sum_{m=1}^{N_w} \mathbf{G}(\mathbf{x}_i, \mathbf{x}_m;\epsilon_w) \cdot \mathbf{F}(\mathbf{x}_m) =\mathbf{U}_w(\mathbf{x}_i), \label{Fonwire} \end{equation} where $\mathbf{x}_i\in \partial \Omega_w$, $i=1,\cdots, N_w$. These equations can be rewritten in matrix form with $3N_w$ unknown variables. The linear system is then solved numerically to obtain the regularized Stokeslet strengths $\mathbf{F}_m=\mathbf{F}(\mathbf{x}_m(t))$. Consequently, the fluid velocity field $\mathbf{u}(\mathbf{x},t)$ can be reconstructed everywhere using \eqref{u_F}. When a finite-sized particle presents in the system, besides the boundary condition on the wire \begin{eqnarray} &&\sum_{m=1}^{N_w} \mathbf{G}(\mathbf{x}_i, \mathbf{x}_m;\epsilon_w) \cdot \mathbf{F}(\mathbf{x}_m)+ \sum_{k=1}^{N_p} \mathbf{G}(\mathbf{x}_i, \mathbf{x}_k;\epsilon_p) \cdot \mathbf{F}(\mathbf{x}_k) =\mathbf{U}_w(\mathbf{x}_i), i=1,2,\cdots, N_w, \label{Fonwire2} \end{eqnarray} additional boundary conditions $\mathbf{u}(\mathbf{x}_k) = \mathbf{U}_p(\mathbf{x}_k)$ ($k=1,2,\cdots,N_p$) should also be satisfied. However, the velocities on the particle $\mathbf{U}_p(\mathbf{x}_k)$ are unknowns. By substituting $\mathbf{U}_p$ with equation \eqref{protate}, we obtain relations \begin{eqnarray} &&\sum_{m=1}^{N_w} \mathbf{G}(\mathbf{x}_j, \mathbf{x}_m;\epsilon_w) \cdot \mathbf{F}(\mathbf{x}_m)+ \sum_{k=1}^{N_p} \mathbf{G}(\mathbf{x}_j, \mathbf{x}_k;\epsilon_p) \cdot \mathbf{F}(\mathbf{x}_k) -\mathbf{v}_p-(\mathbf{x}_0-\mathbf{x}_p) \times \boldsymbol{\omega}_p=0, \label{Fonsphere} \end{eqnarray} where $\mathbf{x}_i\in \partial \Omega_p$, $i=1,\cdots, N_p$. The unknown variables in equations \eqref{Fonwire2} and \eqref{Fonsphere} are $\mathbf{F}$, $\mathbf{v}_p$, and $\boldsymbol{\omega}_p$, which add up to $3(N_w+N_p)+6$ unknowns. Since equations \eqref{Fonwire2} and \eqref{Fonsphere} correspond to only $3(N_w+N_p)$ scalar equations, 6 additional equations are required to determine the system. They can be derived from the non-inertia approximation of the sphere. Because the inertia is negligible, the total force and total torque on the particle are both zero. These imply \begin{eqnarray}\label{FTzero} \sum_{k=1}^{N_p} \mathbf{F}(\mathbf{x}_k)&=&0 , \\ \sum_{k=1}^{N_p} (\mathbf{x}_k-\mathbf{x}_p) \times \mathbf{F}(\mathbf{x}_k)&=&0. \end{eqnarray} After the force strength $\mathbf{F}(\mathbf{x})$ and the rigid body motion for the spherical particle $\mathbf{v}_p$ and $\boldsymbol{\omega}_p$ are determined, the velocity of the flow can be evaluated with the equation \eqref{u_F}. \begin{figure*}[tbh!] \begin{center} \includegraphics[width=.8\textwidth]{./fluidTracer_snapshot_v2} \caption{Snapshots of fluid tracers in the Stokes flow generated by the tumbling wire. The translation speed of the wire is $U=0.08$. These fluid tracers are initialized on a flat plane above the wire at $z=0.3$. The fluid tracers are colored according to their initial distance from the $z$-axis. (a) and (b): top view and side view of the initial setup; (c)-(e): top views at $t=5$, $10$, and $15$, respectively; (f) and (g): top view and side view at $t=20$. The red thick lines represent the centerline of the wire. (Multimedia view)} \label{} \end{center} \end{figure*} \subsubsection*{Nondimensionalization, parameters and errors} In our simulations, we use dimensionless parameters so that the length of the wire $\ell =1 $, rotation rate $\omega= -2\pi$, and the period of the rotation $T=1$. As such, the value of $U$ actually represents the relative time scale between rotation and translation. In order to compare with the results of the experiment in Petit {\it et al.} \cite{Petit2012}, the radius of the wire is set to 0.0067 and the radius of the sphere is set to 0.2. The numbers of regularized Stokeslet on the particle and on the wire are $N_p=150$ and $N_w =94$, respectively. Since the wire and its motion have reflection symmetry in the $z=0$ plane, the behavior of a particle below the tweezers is the same as those in its image position above the tweezers. Therefore, we only show the results for the particles or tracers initialized above the wire. Since the boundary conditions are enforced only at those discretized points where Stokeslets are centered, relatively larger errors may occur between the Stokeslet points. Increasing the regularization parameter can decrease such errors, but might introduce more errors in the computed forces \cite{Cortez2001}. The regularization parameter on the surface of particle is chosen as 0.4 times the estimated average distance between two points, that is $\epsilon_p=0.4\sqrt{4\pi R^2/N_p}$, where the factor 0.4 is a numerical optimal value for the discretized spherical particle \cite{cortez2005method, Martindale2016}. For a sphere at typical positions in the region ($-0.5<x<0.5$, $-0.5<y<0.5$, $0.25<z<0.6$), the velocity error on a random point on the sphere is about 4\% of the average magnitude of the velocity on the sphere. A similar test shows that the velocity error along the wire's center-line is less than 3\%. To test whether the velocity of the sphere can be sufficiently described by the Stokeslets, we increased the number of the points on the sphere to 400 and found only 0.2\% difference from the 150-point case. \section{Results} With our model, we examine the flow induced by the fluid tweezers, focus on the spherical particle captured by the fluid tweezers, and attempt to explain the capture mechanism. Before we couple non-zero volume particles with the tweezers in the system, we examine the fluid tracers in the flow induced by the wire, for which the exact velocity field is obtained. The movement of the fluid tracers in the flow is shown in figure \ref{} (Multimedia view). We find that some particles starting in the vicinity of the wire follow the motion of the wire and some fall behind. Little motion is observed in the vertical ($z$) direction. \begin{figure}[b!] \begin{center} \includegraphics[width=.5\textwidth]{trajectory} \caption{Top view (a) and side view (b) of the trajectories of a spherical particle in the Stokes flow generated by a tumbling wire at different translation speeds $U=0.08, 0.10, 0.12$ and $0.15$. The center of the sphere is initialized at (0.2, 0.05, 0.3) and its radius is $R=0.2$. The total time for the simulations is 30. The dashed line indicates the position of the wire with the smallest $U=0.08$ at $t=30$. The solid markers (red diamond, orange dot, and green square) represent the center of the wire at $t=30$ for corresponding trajectories. The consistent color scheme is used to represent different velocities for the trajectories and the centers of the wire. The center of the wire for the largest translation speed $U=0.15$ is outside the plotting window. Different scales are adopted for the vertical axes in (a) and (b) to make the trajectories observable. (Multimedia view)} \label{} \end{center} \end{figure} By examining the trajectories of individual spherical particles, we find that their trajectories are similar to those of the tracer particles (figure~\ref{}, multimedia view). The capture depends on the initial position of the particle and also on the translation speed. Note that since we normalize the angular velocity of the rotation of the wire, the effect of the wire's rotation speed is represented by the effect of the translation speed. The typical trajectory of a captured particle exhibits a trochoid-like curve with small wiggles, as the red, orange and green curves show in figure~\ref{} (Multimedia view). The shape of the curves depends on the particle's initial position and the translation speed of the wire. When the translation speed is large, the particle may deviate from the prescribed route of the wire. The trajectories of the trapped particles are almost planar, while the untrapped particle moves along a trajectory (the blue curve in figure~\ref{) with a relatively large movement in the $z$ direction, perpendicular to the plane of the motion of the wire. \begin{figure}[t!] \begin{center} \includegraphics[width=.450\textwidth]{disp_vs_v} \caption{The displacement of the particle in the $x$ direction (red circles) as a function of the translation speed of the wire when the center of the wire reaches $x=8$. The particle is initially placed at (0,\,0,\,0.3). The dashed red line is used to guide the eyes, the black dotted line indicates the displacement of the wire, and the green line indicates the speed at which capture has been achieved experimentally by Petit {\it et al.} \cite{Petit2012}. }\label{} \end{center} \end{figure} To distinguish the successful capture cases of a particle from the no-capture cases and quantify the capture region, we measure the horizontal movement of the spheres or tracer particles when the wire translates over a specified distance. We choose a total distance of 8, which corresponds to a total time $t=8/U$. If the displacement of the particle in the forward ($x+$) direction relative to the rod center is less than 1, we consider the particle has been captured (figure~\ref{}). The results with a larger displacement are qualitatively the same and will be discussed in the Discussion section. As shown in figure~\ref{}, a particle with a radius of 0.2 and starting 0.3 above the center of the wire is captured when the translation speed is below about 0.13. This is consistent with the experiment reported in the previous study by Petit {\it et al.}\cite{Petit2012} that a 6$\mu$m diameter particle can be captured by a 15$\mu$m long wire when the wire is rotating at 54\,Hz and translating at as high as 87\,$\mu$m/s, which corresponds to a lower limit of nondimensionalized maximal capture speed 0.11. The speed is marked as the green line in figure 4. \begin{figure*} \begin{center} \includegraphics[width=1.0\textwidth]{capture_regions_v2} \caption{Capture regions of different dimensions of the wire and particle (rows) and the translation speeds (columns) of the wire. The boundaries of the capture region are shown above $z=0.25$ as colored surface for a rigid sphere and the boundaries above $z=0.05$ are shown for a tracer particle. The color (i.e. (R,G,B) value) on the surface depends on the local position (i.e. (x,y,z)). }\label{} \end{center} \end{figure*} To quantify the condition in which the particles are captured, we systematically vary the initial condition of the spherical particle for three translation speeds of the wire and two particle sizes. Since the flow is symmetric, we only examine the initial position in the $z+$ half space, i.e., above the plane where the wire is, for figure 5. For the tracer particle ($R=0$), the initial positions are considered in a box $(-0.65\leq x \leq 0.9, -0.55\leq y \leq 1.5,0.05 \leq z \leq 0.75)$, with a grid resolution of 0.05 in all three dimensions. For the particle with a radius of 0.2, the box with $z\geq 0.25$ is considered to avoid overlapping between the wire and particles at the initial state. The function {\it isosurface} in MATLAB (ver. 8.0.0.783, Mathworks, Inc.) is used to find the isosurface with a horizontal movement between 7 and 9 in the $x+$ direction. Then, inside the surface is the capture region. The space in figure~\ref{} is color coded so that the position of the surfaces can be compared point by point. We find that overall the capture regions have pear-like shapes, as shown in figure~\ref{}. The cross sections in the $x$-$y$ plane are pointier in the $y+$ side. The green pointy region connects to a thin and scattered region, which phases out between the $y+$ and $x+$ direction. The capture region shrinks as the translation speed $U$ increases. In the region both accessible by the tracer particles and spherical particles with radius $R=0.2$, the capture regions are nearly identical. Because a larger range of $z$ is used for the tracer particle, the capture region for the tracer particle in the second row may look bigger than that for the sphere in row 1 in figure~\ref{}. Besides comparing the capture region for the passive tracer particle, we also examined the capture region of spheres by a wire with the wire's radius doubled (i.e., $r=0.0134$). From row 1 and row 3 in figure~\ref{}, we find that the capture region expands in all three dimensions slightly \begin{figure*} \begin{center} \includegraphics[width=1.0\textwidth]{forceballwire2} \caption{Stresses and forces on the sphere and wire. (a) The force on the sphere and the wire exerted by the fluid at an example time $t=4.2$. The color on the sphere represents the magnitude of the total stress. The red arrow represents the rotation vector $\omega_p$. (b) The average magnitude of the stress $\sigma$, the normal component $p$ of the stress, the parallel component $\tau$ of the stress, and the three components of the rotation of the sphere as a function of time. $r=0.0067$, $U=0.08$. (c) The magnitude of the total torque $T^t$ and the total force in the $x$ direction $F^t_x$ as a function of time for different configurations. ``2r'' refers to the case where $r=0.0134$. $R=0.2$ and $U=0.08$. }\label{} \end{center} \end{figure*} To further explore the influence of the capture on the particle and wire, we examine the rotation and the stress of the sphere and the force distribution on the tweezers (figure~\ref{}). We find that the rotation of the sphere about its center is mainly in the $-z$ direction, which is the same as the rotation of the wire. There are two frequencies in the rotation (figure~\ref{). The higher frequency is twice the frequency of rotation of the wire. This high frequency is due to the periodic motion of the two ends of the wire relative to the sphere. The lower frequency corresponds to the periodic motion of the sphere relative to the center of the wire. For the stress $\sigma$ on the sphere, we decompose it into the normal component $p$, which is normal to the surface, and the shear stress $\tau$, which is parallel to the surface. The normal stress $p$ is negligible compared to the shear street $\tau$. Taking the average over the surface of the sphere, the magnitude of the shear stress is about 1.5 in dimensionless units. The flow exerts the largest shear stress on the sphere near the close encounter point for the wire and the sphere. Even though the wire and sphere are coupled in the system, the total torque $T^t$ and net force $F^t$ on the wire are nearly independent of the presence of the sphere (figure \ref{). The magnitude of the torque is about 1.5 in dimensionless units and is about $3\times10^{-18}$N$\cdot$m using experimental parameters. The component of the total force in the translation direction ($\text{F}_x^t$) is about 0.15 in dimensionless unit or $2.5\times10^{-14}$N in physical units. For a wire with a doubled diameter, the torque and net force on the wire are both increased by about 25\%. Such an increase is consistent with the slender body theory that the force on a slender rod is proportional to the slenderness parameter $\mathrm{ln}^{-1}(\ell/r)$ with a fixed velocity \cite{Batchelor1970}. \section{Discussion} \begin{figure*} \begin{center} \includegraphics[width=.90\textwidth]{displacementfield} \caption{Displacement field analysis for tracer particles. (a)-(c) are displacement fields in the horizontal plane, in a moving frame translating at the same speed of the center of the wire. (a) and (b) are at $z=0.25$ with the translation speed $U=0.08$ and $U=0.12$, respectively. (c) is with $U=0.08$ and at $z=0.45$, a different height to (a). Blue arrows represent the displacement. The red dots indicate the stagnation points. The magenta, green and black trajectories in (a)-(c) are three representative tracer particles trajectories over a time period of 100. (d) is the capture diagram predicted by displacement field analysis. The red curve is the maximal displacement in the $x$ direction generated by a rotating wire with $r=0.0067$ at different heights $z$. Bounded by the red curve, the shaded region is the predicted region at which height $z$ and translating speed $U$ no capture occurs. The green dotted curve represents the maximal displacement to height relation for a thicker wire with $r=0.0134$. The blue asterisks mark the observed maximal $z$ coordinate of the capture regions and the respective $U$ in the first row in figure \ref{}.} \label{} \end{center} \end{figure*} To reveal the capture mechanism and understand the capture region, we examine the displacement for tracer particles in the moving frame within one period of rotation of the wire (i.e., the vector field $\mathbf{d}=\mathbf{x}_p(t=1)-\mathbf{x}_p(t=0)$ in the moving frame). Since the capture is insensitive to the particle size, we consider the displacement of a tracer particle in the body frame translating with the center of the wire. The displacement field can be approximated as the combination of the translation displacement and a vortex. When the particle is far from the wire and the influence of the wire's rotation on the particle is weak, the displacement field is simply the uniform displacement $-UT$. A rotating wire generates a vortex in the fluid with maximal linear velocity of the fluid near the end points and the velocity decreases with increasing $z$. Note that the vortex has some detailed but complex features: (1) Due to the non-slip boundary condition, the fluid near the surface of the wire always moves with the wire and therefore the displacement is nearly zero near the surface. (2) The displacement field is not exactly axis-symmetric, since the initial orientation of the wire defines a direction. Since the motion of the wire is in the $x$-$y$ plane and the diameter of the wire is very small compared to the length, the displacement field is also nearly planar. The displacement in the $z$ direction is negligible ($< 0.007$ in the $z=0.25$ plane) in determining the overall capture region. Therefore, we focus on the displacement field in the $x$-$y$ plane at different heights $z$. When the translation speed $U$ is small compared to the linear velocity of the wire and $z$ is small, the displacement field has two stagnation points along the $y+$ axis. As the linear shear velocity generated by the rotation of the wire increases from zero from the center of the wire and then decreases to zero as the distance goes to infinity, the two stagnation points appear where the forward ($x+$) velocity is canceled out by the uniform backward displacement. Since the only region with forward velocity is between these two stagnation points, the stagnation points indicate the size and the position of the capture region. The stagnation point with a smaller $y$ value is a stable stagnation point. A particle starting between these two stagnation points encircles this stable stagnation point with an oval shape (green curves in figure~\ref{). The stagnation point with a greater $y$ value is an unstable saddle point. A particle that starts from the stable manifold in the upstream $(x+,y+)$ takes an infinite time to reach the saddle point for a finite distance. Thereafter, the particle appears to be captured. Because the trajectory of a point near the stable manifold is sensitively dependent on the initial position, the apparent capture is sensitive to initial positions and numerical errors. Therefore, the particle captured is scattered in this region. When the starting point is sufficiently far from the saddle point in the $y+$ direction, the uniform translation dominates and no capture can occur (the magenta curves in figure~\ref{} (a-c)). A similar simplified 2D kinematic model of atmospheric or oceanic flow was analyzed and similar capture regions were found by Flierl (1981)\cite{flierl1981particle}. As shown from the analysis, the particles in the scattered region near the stable manifold of the saddle point are not truly captured. They gradually fall behind the rod as the system evolves for a longer time and such spurious capture regions shrink. Figure~\ref{} shows a comparison of the capture regions when the rod travels with different distances, i.e., with a different total time for the simulations. \begin{figure} \begin{center} \includegraphics[width=.50\textwidth]{./longdistance} \caption{Comparison of the capture regions when the wire travels a distance of 8 (a) and 16 (b). $R=0$, $r=0.0067$, and $U=0.05$.}\label{} \end{center} \end{figure} When $U$ or $z$ increases, the two stagnation points move closer, and eventually merge and disappear (figure \ref{}). Since the distance between the two stagnation points roughly determines the radius of the capture region in a $z$ cross sectional plane, the decrease of the distance as $z$ increases or $U$ increases explains the drop-like shape of the capture region and the loss of volume as $U$ increases. Essentially, canceling the displacement in the forward direction $\mathbf{d}_x$ in the vortex by $-UT$ determines the existence of the capture region. We examine the maximal displacement in the $x$ direction in cross-sections of the vortex as a function of $z$ (figure \ref{). Such a plot indicates the height of the capture region at a given $U$ and outlines the capture region and the non-capture region. As shown in figure \ref{, the heights of the capture regions (asterisks) lie closely on this curve. With a thicker wire, the curve (green dotted) is shifted slightly upward, which is consistent with the larger capture regions by the thicker wire. As the fluidic tweezers often intend to gently manipulate biological cells, we compare the rough magnitudes of resultant stresses and forces on the spherical particle with the biological values. With the parameters in Petit {\it et al.} \cite{Petit2012}, the forces and stresses on the sphere are in general gentle for biological cells. The magnitude of the stress on average is about 0.08\,Pa on the sphere. Such a magnitude of stress is sufficient to induce intracellular $\mathrm{Ca}^{2+}$ or $\mathrm{Ka}^{+}$ channel activation (0.02-0.4 Pa and 0.02-1.65 Pa, respectively), but unlikely to induce other biological response (typically $>$0.5 Pa) if the sphere is an endothelial cell \cite{davies1993mechanical}. The stretching force is around 10\,pN; such a force is not sufficient to generate a large deformation of cells such as blood cells, who require a force in the magnitude of 100\,pN to deform significantly ($>$10\% diameter) \cite{dao2003mechanics}. This study is limited to purely fluid mechanical effects of tweezers in free space. Since there are no other forces to overcome for the particle, the kinematic effects are dominant and thoroughly examined. Future studies might introduce other effects from real-life applications, such as forces from the contact between the particle and the surface, the gravity forces, buoyancy, boundaries of the fluid, and Brownian motion for particles with smaller sizes (typically $\le 1 \mu m $). When the immersed particle is no longer a sphere, the trajectory of a particle may depend on both its shape and its initial orientation. Our preliminary study suggests that the capture regions are qualitatively similar when the sphere is replaced by an oblate or prolate spheroid. However, the details of the capture region depends on the particle's shape and initial orientation in a non-trivial way. When the inertial force plays a role in the flow beyond the Stokes regime, the flow fields change qualitatively. Due to centrifugal forces, outward motions are expected as the sphere and fluid rotate. As such, the flow may become fully three-dimensional and the stagnation points in the planar displacement field are not sufficient to describe the capture region. Thorough investigations on these effects will be carried out in the near future. \section{Conclusions} In conclusion, we have studied a tumbling nanowire that manipulates micro-objects. For the spheroid tweezers, we derive an exact velocity field for the flow with the singularity method \cite{Wu1975}; in the presence of a neutrally buoyant spherical particle, we study the flow with the regularized Stokeslet method. For a particle in a viscous/micro fluid, the kinematic effect of the vortex generated by the fluidic tweezers is responsible for the capture of the particles. The trajectories of the particles can be understood from the combination of a vortex and a uniform translation. The stagnation points in the combined displacement field provide key information on the shape and size of the capture regions. We found that the behavior of a finite-size sphere is nearly the same as a tracer particle for typical sizes reported in the experiments. Furthermore, the external forces and torques to drive the nanowire vary little for different particles. \begin{acknowledgments} Funding for Y.D. was provided by NSAF-NSFC grant No. U1530401, and the Recruitment Program of Global Young Experts. \end{acknowledgments}

\section{Result} \section{Introduction} Assume $G$ is a graph. We associate to each vertex $v$ of $G$ a variable $x_v$. The graph polynomial $P_G(\vec{x})$ of $G$ is defined as $$P_G(\vec{x}) = \prod_{ u \sim v, u < v}(x_v-x_u),$$ where $\vec{x} = \{ x_v: v \in V(G)\}$ and $``<"$ is an arbitrary fixed ordering of the vertices of $G$. It is easy to see that a mapping $\phi: V \to R$ is a proper colouring of $G$ if and only if $P_G(\phi) \ne 0$, where $P_G(\phi)$ means to evaluate the polynomial at $x_v=\phi(v )$ for $v \in V(G)$. Thus to find a proper colouring of $G$ is equivalent to find an assignment of $\vec{x}$ so that the polynomial evaluated at this assignment is non-zero. The Combinatorial Nullstellensatz gives a sufficient condition for the existence of such an assignment. Assume $P(\vec{x})$ a polynomial with variable set $X$. An {\em index function } $\eta$ for $P({\vec{x}})$ is a mapping which assigns to each variable $x$ a non-negative integer $\eta(x)$. Given an index function $\eta$, we denote by ${\vec{x}}^{\eta}$ the monomial $\prod_{x \in X} x^{\eta(x )}$, and denote by $c_{P, \eta}$ the coefficient of ${\vec{x}}^{\eta}$ in the expansion of $P(\vec{x})$. The Combinatorial Nullstellensatz asserts that if $\eta$ is an index function with $\sum_{x \in X}\eta(x) $ equals the degree of $P(\vec{x})$ and $c_{P, \eta} \ne 0$, and $A_x$ is a set of $\eta(x)+1$ real numbers ( or elements of a field) for each $x \in X$, then there is an assignment $\phi$ such that $\phi(x ) \in A_x$ for each $x \in X$ and $P(\phi) \ne 0$. In particular, if $c_{P_G, \eta} \ne 0$ and $\eta(x_v) < k$ for all $v \in V$, then $G$ is $k$-choosable. This method developed by Alon and Tarsi is now a powerful tool in the study of list colouring of graphs. Jensen and Toft \cite{JT} defined the {\em Alon-Tarsi number} of $G$ as $$AT(G) = \min \{k: \text{$c_{P_G, \eta} \ne 0$ for some index function $\eta$ with $\eta(x_v) < k$ for all $v \in V(G)$}\}.$$ As observed in \cite{HE}, $AT(G)$ has some distinct features and it is of interest to study $AT(G)$ as a separate graph invariant. Let $ch(G)$ be the choice number of $G$ and $\chi_P(G)$ be the paint number (or the online choice number) of $G$ (cf. \cite{Scha} and \cite{Zhu}). It follows from a result of Alon- and Tarsi \cite{AT} and a generalization of this result by Schauz \cite{Sch} $ch(G) \le \chi_P(G) \le AT(G)$ for any graph $G$. There are graphs for which both inequalities are strict. However, upper bounds for the choice number of many natural classes of graphs are also upper bounds for their Alon-Tarsi number. Thomassen \cite{TH} proved that every planar graph $G$ has $ch(G) \le 5$, and Schauz \cite{Scha} showed that every planar graph $G$ has $\chi_P(G) \le 5$. A natural question (cf. \cite{HE}) is whether $AT(G) \le 5$ for every planar graph $G$. In this note we answer this question in affirmative. \begin{theorem} \label{thm-main0} If $G$ is a planar graph, then $AT(G) \le 5$. \end{theorem} \section{A proof of Theorem \ref{thm-main0}} The proof of Theorem \ref{thm-main0} is parallel to Thomassen's proof of the $5$-choosability of planar graphs in \cite{TH}. For simplicity, we write $c_{G,\eta}$ for $c_{P_G, \eta}$, and say $\eta$ is an index function of $G$ instead of an index function for $P_G({\vec{x}})$. \begin{definition} \label{def-nice} Assume $G$ is a plane graph and $e=v_1v_2$ is a boundary edge of $G$. An index function $\eta$ of $G-e$ is a {\em nice } for $(G,e)$ if the following hold: \begin{itemize} \item $c_{ {G-e}, \eta} \ne 0$. \item $\eta(v_1) = \eta(v_2)=0$, $\eta(v) \le 2$ for every other boundary vertex $v$, and $\eta(v) \le 4$ for every interior vertex $v$. \end{itemize} \end{definition} If $\eta$ is a nice index function for $(G,e)$, then let $\eta'(x)=\eta(x)$ except that $\eta'(x_{v_1})=1$. As $P_G(\vec{x}) = (x_{v_1}-x_{v_2})P_{G-e}(\vec{x})$ and $\eta'(v_2)=0$, we know that $c_{G,\eta'} =c_{G-e, \eta} \ne 0$. Note that $\eta'(x_{v}) < 5$ for each vertex $v$. Thus Theorem \ref{thm-main0} follows from Theorem \ref{thm-2} below. \begin{theorem} \label{thm-2} Assume $G$ is a plane graph and $e=v_1v_2$ is a boundary edge of $G$. Then there exists a nice index function $\eta$ for $(G,e)$. \end{theorem} A variable $x$ is a {\em dummy variable} in $P({\vec{x}})$ if $x$ does not really occur in $P({\vec{x}})$, or equivalently, $\eta(x)=0$ for each monomial ${\vec{x}}^{\eta}$ in the expansion of $P$ with a nonzero $c_{P, \eta}$. We shall frequently need to consider the summation and the product of polynomials. By introducing dummy variables, we assume the involved polynomails in the sum or the product have the same set of variables. For example, we may view $x_2^2$ be the same as $x_1^0x_2^2x_3^0\ldots x_n^0$, i.e, $x_2^2={\vec{x}}^{\eta}$, where the variable set is $X=\{x_1, x_2, \ldots, x_n\}$ and $\eta(x_2)=2$, $\eta(x_i)=0$ for $i \ne 2$. We denote by $X$ the set of variables for polynomials in concern. For two index functions $\eta_1, \eta_2$, we write $\eta_1 \le \eta_2$ if $\eta_1(x) \le \eta_2(x)$ for all $x \in X$, and $\eta= \eta_2-\eta_1$ means that $\eta(x) = \eta_2(x)-\eta_1(x)$ for all $x \in X$. \begin{observation} \label{obs} \mbox{} \begin{enumerate} \item If $P(\vec{x})=\alpha P_1(\vec{x})+ \beta P_2(\vec{x})$, then $c_{P, \eta}=\alpha c_{P_1, \eta} + \beta c_{P_2, \eta}.$ \item If $P({\vec{x}}) = {\vec{x}}^{\eta'}P_1({\vec{x}})$, then $c_{P, \eta}= c_{P_1, \eta-\eta'}.$ \item If $P({\vec{x}}) = {\vec{x}}^{\eta'}P_1({\vec{x}})$ and $\eta' \not\le \eta$, then $c_{P, \eta}=0$. \item If $P({\vec{x}}) = P_1({\vec{x}})P_2({\vec{x}})$ and for any $\eta'$ with $c_{P_2,\eta'}\ne 0$, there is a dummy variable $x$ of $P_1({\vec{x}})$ such that $\eta'(x) \ne \eta(x)$, then $c_{P, \eta}=0$. \item If $G$ is a graph and $c_{ G, \eta} \ne 0$, then $\sum_{x \in X} \eta(x) = |E(G)|$. \end{enumerate} \end{observation} \noindent {\bf Proof of Theorem \ref{thm-2}} Assume the theorem is not true and $G$ is a minimum counterexample. First we consider the case that $G$ has a chord $e'=xy$. Let $G_1, G_2$ be the two $e'$-components (i.e., $G_1, G_2$ are induced subgraphs of $G$ with $V(G) = V(G_1) \cup V(G_2)$ and $V(G_1) \cap V(G_2) = \{x,y\}$) with $e \in G_1$. By the minimality of $G$, there exist a nice index function $\eta_1$ for $(G_1,e)$, and a nice index function $\eta_2$ for $(G_2,e')$. Let $\eta=\eta_1+\eta_2$. Note that $$ P_{G-e}(\vec{x})= P_{G_1-e}(\vec{x}) P_{G_2-e'}(\vec{x}).$$ Let $R(\vec{x}) = c_{ {G_2-e}, \eta_2} {\vec{x}}^{\eta_2}$, $Q(\vec{x}) = P_{G_2-e'}({\vec{x}}) -R(\vec{x})$, $P_1({\vec{x}}) = R(\vec{x}) P_{G_1-e}(\vec{x})$ and $P_2({\vec{x}}) = Q(\vec{x})P_{G_1-e}(\vec{x}) $. Then $P_{G-e}({\vec{x}}) = P_1({\vec{x}})+P_2({\vec{x}})$. By (5) of Observation \ref{obs}, for any index function $\eta'$ with $c_{Q,\eta'} \ne 0$, we have $\eta' \ne \eta_2$ and hence there is a vertex $v \in V(G_2) -\{x,y\}$ such that $\eta'(x_v) \ne \eta_2(x_v) = \eta(x_v)$. As $x_v$ is a dummy variable in $P_{G_1-e}({\vec{x}})$, by (4) of Observation \ref{obs}, we have $c_{P_2,\eta}=0$. By (1) and (2) of Observation \ref{obs}, $$c_{ {G-e}, \eta} = c_{ P_1, \eta} = c_{ {G_2-e'}, \eta_2} c_{ {G_1-e}, \eta_1}\ne 0.$$ So $\eta$ is a nice index function for $(G,e)$. Assume $G$ has no chord and assume $B(G)=(v_1, v_2, \ldots, v_n)$. Let $G'=G-v_n$. Let $v_1, u_1, u_2, \ldots, u_k, v_{n-1}$ be the neighbours of $v_n$. Let $$S({\vec{x}}) = (x_{v_n}-x_{v_1})(x_{v_{n-1}}-x_{v_n })( x_{u_1}-x_{v_n }) \ldots ( x_{u_k}-x_{v_n }).$$ Then $P_{G-e}(\vec{x})=S({\vec{x}})P_{G'-e}(\vec{x}) .$ If $n=3$, then let $\eta' $ be nice for $(G',e)$. Let $\eta(x_v)=\eta'(x_v)$ for $v \notin \{u_1,u_2,\ldots, u_k\}$ and $\eta(x_v) = \eta'(x_v)+1$ for $v \notin \{u_1,u_2,\ldots, u_k\}$ and $\eta(x_{v_n})=2$. Let $\eta''(x_{v_n})=2$, $\eta''(x_{u_i})=1$ for $i=1,2,\ldots,k$ and $\eta''(x)=0$ for other $x$. Then $$S({\vec{x}}) = - {\vec{x}}^{\eta''} +x_{v_1}A(\vec{x}) + x_{v_2}B(\vec{x})+x_{v_n}^3C({\vec{x}})$$ for some polynimals $A(\vec{x})$, $B(\vec{x})$ and $C({\vec{x}})$. Let $P_1({\vec{x}}) = x_{v_1}A(\vec{x})P_{G'-e}({\vec{x}})$, $P_2({\vec{x}}) = x_{v_2}B(\vec{x})P_{G'-e}({\vec{x}})$ and $P_3({\vec{x}}) = x^3_{v_n}C(\vec{x})P_{G'-e}({\vec{x}})$. As $\eta(x_{v_1})=\eta(x_{v_2})=0$ and $\eta(x_{v_n})=2$, it follows from (3) of Observation \ref{obs} that $c_{P_1,\eta}=c_{P_2,\eta}=c_{P_3,\eta}=0$. By (1) and (2) of Observation \ref{obs}, $c_{G-e, \eta} = -c_{G'-e, \eta'} \ne 0$. Hence $\eta$ is nice for $(G,e)$. Assume $n \ge 4$. We say an index function $\eta'$ for $G'-e$ {\em special} if $\eta'(v_{n-1}) \le 1$, $\eta'(v_1) = \eta'(v_2)=0$, $\eta'(u_j) \le 3$ for $j=1,2,\ldots, k$. $\eta'(v) \le 2$ for each other boundary vertex $v $ and $\eta'(v) \le 4$ for each interior vertex $v$. \bigskip \noindent {\bf Case 1.} $c_{ {G'-e}, \eta'} \ne 0$ for some special index function $\eta'$ for $G'-e$. Let $\eta(v) = \eta'(v)$ for $v \notin \{u_1, u_2, \ldots, u_k, v_{n-1}\}$ and $\eta(v) = \eta'(v)+1$ for $v \in \{u_1, u_2, \ldots, u_k, v_{n-1}\}$ and $\eta(v_n)=1$. Let $\eta''$ be the index function defined as $\eta''(x_{v_n}) = \eta''(x_{v_{n-1}}) = \eta''(x_{u_1}) = \ldots = \eta''(x_{u_k})=1$ and $\eta''(x)=0$ for other variables $x$. Then $$S({\vec{x}}) = {\vec{x}}^{\eta''} +x_{v_1}A(\vec{x}) + x_{v_n}^2 B(\vec{x})$$ for some polynimals $A(\vec{x})$ and $B(\vec{x})$. Let $P({\vec{x}}) = {\vec{x}}^{\eta''}P_{G'-e}({\vec{x}})$, $P_1(\vec{x})= x_{v_1}A(\vec{x})P_{G'-e}(\vec{x})$ and $P_2(\vec{x})= x^2_{v_n}B(\vec{x})P_{G'-e}(\vec{x}) $. Then $$P_{G,e}({\vec{x}}) = P({\vec{x}})+P_1({\vec{x}})+P_2({\vec{x}}).$$ As $\eta(v_1)=0$ and $\eta(v_n)=1$, it follows from (3) of Observation \ref{obs} that $c_{P_1, \eta } = c_{P_2, \eta }=0.$ By (1) and (2) of Observation \ref{obs}, we have $c_{ {G-e}, \eta}=c_{ P, \eta} = c_{ {G'-e}, \eta'} \ne 0.$ Hence $\eta$ is nice for $(G,e)$. \bigskip \noindent {\bf Case 2.} $c_{ {G'-e}, \eta'} = 0$ for every special index function $\eta'$ for $G'-e$. By the minimality of $G$, there is an index function $\eta''$ nice for $(G',e)$. Let $\eta(x_v)=\eta''(x_v)$ for $v \notin \{u_1,u_2,\ldots, u_k\}$ and $\eta(x_v) = \eta''(x_v)+1$ for $v \notin \{u_1,u_2,\ldots, u_k\}$ and $\eta(x_{v_n})=2$. Let $\eta'''(x_{v_n})=2$, $\eta'''(x_{u_i})=1$ for $i=1,2,\ldots,k$ and $\eta''(x) = 0$ for other $x$. Then $$S({\vec{x}})= {\vec{x}}^{\eta'''} +x_{v_1}A(\vec{x}) + x_{v_{n-1}} B(\vec{x})+x_{v_n}^3C({\vec{x}})$$ for some polynimals $A(\vec{x})$, $B(\vec{x})$ and $C(\vec{x})$. Let $P(\vec{x}) = {\vec{x}}^{\eta'''} P_{G'-e} (\vec{x})$, $P_1(\vec{x}) = x_{v_1}A(\vec{x}) P_{G'-e}(\vec{x})$, $P_2(\vec{x}) =x_{v_{n-1}}B(\vec{x}) P_{G'-e}(\vec{x}) $ and $P_3(\vec{x}) = x_{v_n}^3C(\vec{x}) P_{G'-e}(\vec{x})$. As $\eta(x_{v_1})=0$ and $\eta(x_{v_n})=2$, it follows from (3) of Observation \ref{obs} that $c_{ P_1, \eta } =c_{P_3, \eta}=0.$ As $\eta(v_{n-1})\le 2$, $\eta(u_i) = \eta''(u_i)+1 \le 3$, it follows from (2) of Observation \ref{obs} that $c_{P_2, \eta } = c_{ {G'-e}, \eta'}$ for a special index function $\eta'$ for $G'-e$ (note that $\eta'(x_{v_{n-1}}) = \eta(x_{v_{n-1}})-1$ and $\eta'(x) \le \eta(x)$ for other $x$). By our assumption, $c_{ {G'-e}, \eta'} = 0$. Therefore $c_{P_2, \eta } =0 $. As $$P_{G-e}({\vec{x}}) = P({\vec{x}}) + P_1({\vec{x}})+P_2({\vec{x}})+P_3({\vec{x}}),$$ by (1) and (2) of Observation \ref{obs}, $c_{ {G-e}, \eta}=c_{P, \eta} = c_{ {G'-e}, \eta''} \ne 0.$ \qed \section{An alternate proof} A digraph $D$ is {\em Eulerian} if $d_D^+(v)=d_D^-(v)$ for every vertex $v$. Assume $G$ is a graph and $D$ is an orientation of $G$. Let $EE(D)$ (respectively, $OE(D)$) be the set of spanning Eulerian sub-digraphs of $D$ with an even (respectively, an odd) number of edges. Alon and Tarsi \cite{AT} showed that for an index function $\eta$ of $G$, $c_{G, \eta} = \pm ( |EE(D)|-|OE(D)|)$ for an orientation $D$ with $d_D^+(v) = \eta(v)$ for every $v \in V(G)$. Thus to prove that $c_{G, \eta} \ne 0$ is equivalent to show that there is an orientation $D$ of $G$ with $d_D^+(v) = \eta(v)$ for every $v \in V(G)$ for which $|EE(D)| \ne |OE(D)|$. \begin{definition} \label{def-nice} Assume $G$ is a plane graph and $e=v_1v_2$ is a boundary edge of $G$. An orientation $D$ of $G-e$ is a {\em nice } for $(G,e)$ if the following hold: \begin{itemize} \item $|EE(D)| \ne |OE(D)|$. \item $d_D^+(v_1) = d_D^+(v_2)=0$, $d_D^+(v) \le 2$ for every other boundary vertex $v$, and $d_D^+(v) \le 4$ for every interior vertex $v$. \end{itemize} \end{definition} The following theorem is just a restatement of Theroem \ref{thm-2}, and its proof is essentially the same as the proof of Theorem \ref{thm-2}. However, the translation from calculating the coefficients of a polynomial to counting Eulerian subgraphs is not completely trivial. We include a proof of this statement for pedagogical reason. \begin{theorem} \label{thm-main} Assume $G$ is a plane graph and $e=v_1v_2$ is a boundary edge of $G$, then $(G,e)$ has a nice orientation. \end{theorem} \begin{proof} Assume the theorem is not true and $G$ is a minimum counterexample. First we consider the case that $G$ has a chord $e'=xy$. Let $G_1, G_2$ be the two $e'$-components with $e \in G_1$. By the minimality of $G$, $(G_1,e)$ has a nice orientation $D_1$, and $(G_2,e')$ has a nice orientation $D_2$. Let $D=D_1 \cup D_2$. Edges in $D_2$ incident to $x,y$ are not contained in any directed cycles, and hence are not contained in any Eulerian sub-digraph of $D$. Therefore \begin{eqnarray*} EE(D) &=& EE(D_1) \times EE(D_2) + OE(D_1) \times OE(D_2), \\ OE(D) &=& EE(D_1) \times OE(D_2) + OE(D_1) \times EE(D_2). \end{eqnarray*} So $|EE(D)|-|OE(D)| = (|EE(D_1)|-|OE(D_1)|)(|EE(D_2)|-|OE(D_2)|) \ne 0$. Hence $D$ is a nice orientation of $(G,e)$. Assume $G$ has no chord. Assume $B(G)=(v_1, v_2, \ldots, v_n)$. Let $G'=G-v_n$. If $n=3$, i.e. $B(G)$ is a triangle, then let $D'$ be a nice orientation for $(G',e)$, and let $D$ be obtained from $D'$ by adding arcs $(v_3,v_1)$ and $(v_3, v_2)$. As $v_1, v_2$ are sinks, no edge incident to $v_n$ is contained in a directed cycle and hence $EE(D)=EE(D')$ and $OE(D)=OE(D')$. Therefore $D$ is a nice orientation for $(G,e)$. Assume $n \ge 4$. Let $v_1, u_1, u_2, \ldots, u_k, v_{n-1}$ be the neighbours of $v_n$. We call an orientation $D$ of $(G',e)$ {\em special} if the following hold: \begin{itemize} \item $v_1, v_2$ has out-degree $0$, $v_{n-1}$ has out-degree at most $1$, and each of $u_1,u_2, \ldots, u_k$ has out-degree at most $3$, each other boundary vertex has out-degree at most $2$. \item every interior vertex has out-degree at most $4$. \end{itemize} \bigskip \noindent {\bf Case 1.} $(G',e)$ has a special orientation $D'$ with $|EE(D')| \ne |OE(D')|$. Let $D$ be the orientation of $G-e$ which is obtained from $D'$ by adding arcs $$(v_n, v_1), (v_{n-1}, v_n), (u_1, v_n), \ldots, (u_k, v_n).$$ Then $D$ is a nice orientation of $(G,e)$, as $EE(D')=EE(D)$ and $OE(D')=OE(D)$. \bigskip \noindent {\bf Case 2.} For any special orientation $D'$ of $(G',e)$, $|EE)D')| = |OE(D')|$. By the minimality of $G$, $(G',e)$ has a nice orientation $D''$. Let $D$ be the orientation of $G-e$ obtained from $D''$ by adding arcs $(v_n, v_1), ( v_n, v_{n-1}), (u_1, v_n), \ldots, (u_k, v_n)$. For $i=1,2,\ldots, k$, let $$EE_i(D) = \{H \in EE(D): (u_i,v_n) \in H\}, \ OE_i(D) = \{H \in OE(D): (u_i,v_n) \in H\}.$$ For $i=1,2,\ldots, k$, if $ EE_i(D) \cup OE_i(D) \ne \emptyset$, then let $C_i$ be a directed cycle in $D$ containing $(u_i,v_n)$. Note that every directed edge of an Eulerian digraph is contained in a directed cycle, and $C_i$ must contain $(v_n,v_{n-1})$. Let $D'_i$ be the orientation of $G'$ which is obtained from $D''$ by reversing the direction of edges in $C_i \cap D''$ (note that $C_i \cap D'' =C_i- \{(u_i,v_n), (v_n, v_{n-1})\}$ is a directed path from $v_{n-1}$ to $u_i$). Observe that $D'_i$ is a special orientation of $(G',e)$. Hence $|EE(D'_i)| = |OE(D'_i)|$. Now we show that $|EE_i(D)| = |OE_i(D)|$ for $i=1,2,\ldots,k$. If $ EE_i(D) \cup OE_i(D) = \emptyset$, then this is trivially true. Assume $ EE_i(D) \cup OE_i(D) \ne \emptyset$. For each $H \in EE_i(D) \cup OE_i(D)$, $H \vartriangle C_i^{-1} \in EE(D'_i) \cup OE(D'_i)$. Here $C_i^{-1}$ is the reverse of $C_i$ and $H \vartriangle C_i^{-1}$ is the symmetric difference of $H$ and $C_i^{-1}$, i.e., the digraph obtained from the edge disjoint union of $H$ and $C^{-1}_i$ by deleting digons. Note that the symmetric difference of any two Eulerian digraphs is an Eulerian digraph. Moreover, any edge of $C_i$ contained in $H$ will form a digon with the corresponding edge in $C_i^{-1}$ and hence is deleted. In particular, $(u_i,v_n), (v_n, v_{n-1})$ are edges of $C_i$ contained in $H$ and are deleted. So $H \vartriangle C_i^{-1}$ is a sub-digraph of $D''$. Similarly, for each $H \in EE(D'_i) \cup OE(D'_i)$, $H \vartriangle C_i \in EE_i(D) \cup OE_i(D)$. As $(H \vartriangle C_i ) \vartriangle C_i^{-1} = H$, $\phi(H) := H \vartriangle C_i^{-1}$ is a one-to-one correspondence between $EE(D'_i) \cup OE(D'_i)$ and $EE_i(D) \cup OE_i(D)$. If $C_i$ is of even length, then $|E(\phi(H))|$ and $|E(H)|$ have the same parity; if $C_i$ is of odd length, then $|E(\phi(H))|$ and $|E(H)|$ have different same parities. So if $C_i$ is of even length, then $|EE_i(D)| = |EE(D'_i)|, |OE_i(D)| = |OE(D'_i)|$; if $C_i$ is of odd length, then $|EE_i(D)| = |OE(D'_i)|, |OE_i(D)| = |EE(D'_i)|$. In any case, $|EE_i(D)| = |OE_i(D)|$. Now $$EE(D) = EE(D'') \cup \bigcup_{i=1}^k EE_i(D), \ \ OE(D) = OE(D'') \cup \bigcup_{i=1}^k OE_i(D).$$ The unions above are disjoint unions. So $|EE(D)|-|OE(D)| = |EE(D'')|-|OE(D'')| \ne 0$. \end{proof} I would like to thank Grzegorz Gutowski for bringing this problem to my attention, and thank Jaroslaw Grytczuk for pointing out that the problem was contained in \cite{HE}. \bibliographystyle{unsrt}

\section{Introduction} The goal of this paper is to study the weak tracial Rokhlin property for finite group actions on simple (not necessarily unital) C*-algebras. We propose a definition of the weak tracial Rokhlin property for finite group actions on simple not necessarily unital C*-algebras, which coincides with the existing definition for the simple unital case. Studying the weak tracial Rokhlin property in the simple nonunital case is more challenging due to the lack of enough traces and the fact that one can not expect to work with the unitization or multiplier algebra as the simplicity will be missed. This leads us to use the machinery of Cuntz subequivalence to resolve these difficulties. It is argued that the weak tracial Rokhlin property can be considered as a noncommutative freeness. We prove its permanence properties such as stability under restriction to invariant hereditary subalgebras, taking minimal tensor products and direct limits of actions. Even in the unital case, some of these properties were asked as open problems in \cite{Ph.free}. We present some examples of finite group actions on simple stably projectionless C*-algebras. Furthermore, stability of tracial rank zero under taking crossed products by finite group actions with the weak tracial Rokhlin property is obtained. For this, we give a new look to simple (not necessarily uintal) C*-algebras of tracial rank zero. The Rokhlin property for actions on C*-algebras appeared in \cite{HO84, BSKR93, Ki95, Ki96}, and elsewhere. Izumi gave a modern definition of the Rokhlin property for finite group actions on unital C*-algebras \cite{IZ04a, IZ04b}. This property is useful to understand the structure of the crossed product of C*-algebras and properties passing from the underlying algebra to the crossed product \cite{OP12, Ph15}. However, actions with the Rokhlin property are rare and many C*-algebras admit no finite group actions with the Rokhlin property. Indeed, the Rokhlin property imposes some K-theoretical obstructions on C*-algebras. Phillips introduced the tracial Rokhlin property for finite group actions on simple unital C*-algebras \cite{Ph11} for the purpose of proving that every simple higher dimensional noncommutative torus is an AT~algebra \cite{Ph06}, and proving that certain crossed products of such algebras by finite cyclic groups are AF~algebras \cite{Echterhoff-Phillips} (see \cite{OP06} for $\mathbb{Z}$ actions with this property). The tracial Rokhlin property is generic in many cases (see \cite{Ph12} and \cite[Chapter~4]{Wang13}), and also can be used to study properties passing from the underlying algebra to the crossed product \cite{Ph11, Echterhoff-Phillips, Archey}. Weak versions of the tracial Rokhlin property in which one uses orthogonal positive contractions instead of orthogonal projections were studied for actions on simple unital C*-algebras with insufficiently many projections \cite{Ph12, MS12, Hirshberg, Wang13, GHS, Wang18} (see Definitions~\ref{unitalgar} and \ref{unitalghs}). As an example, the flip action on the Jiang-Su algebra $\mathcal{Z} \cong \mathcal{Z} \otimes \mathcal{Z}$ has the weak tracial Rokhlin property but it does not have the tracial Rokhlin property \cite{Hirshberg, AGJP17}. The Rokhlin property was extended to the case of actions on nonunital C*-algebras \cite{Nawata, Sa15}, and there are actions with the Rokhlin property on stably projectionless C*-algebras, in particular on the Razak-Jacelon algebra $\mathcal{W}$ \cite{Nawata}. However, there has been no work on the extension of the (weak) tracial Rokhlin property to the simple nonunital case. (As far as we know, a suitable definition of the tracial Rokhlin property for actions on nonsimple C*-algebras is not known.) Moreover, actions on simple nonunital C*-algebras naturally appear, for instance, the restriction of an action on a simple unital C*-algebra to an invariant nonunital hereditary subalgebra (see \autoref{propwtrpher}). Also, there are many examples of finite group actions on simple nonunital C*-algebras without the Rokhlin property, but they have the weak tracial Rokhlin property (see Parts~\eqref{thm_ex_intro_it3}, \eqref{thm_ex_intro_it4}, and \eqref{thm_ex_intro_it5} of Theorem~\ref{thm_ex_intro} below). In fact, the problem of finding a right definition of the tracial Rokhlin property for actions on simple nonunital C*-algebras was asked by Phillips \cite{Ph.free}. This urged us to investigate the weak tracial Rokhlin property for finite group actions on simple not necessarily unital C*-algebras. We give the following definition: \begin{defx}\label{weak.t.R} Let $\alpha \colon G \rightarrow {\mathrm{Aut}} (A)$ be an action of a finite group $G$ on a simple not necessarily unital C*-algebra $A$. We say that $\alpha$ has the \emph{weak tracial Rokhlin property} if for every $\varepsilon > 0$, every finite subset $F \subseteq A$, and all positive elements $x,y \in A$ with $\|x\|=1$, there exists a family of orthogonal positive contractions $(f_{g})_{g \in G}$ in $A$ such that, with $f=\sum_{g \in G} f_{g}$, the following hold: \begin{enumerate} \item\label{weak.t.R_it1} $\|f_{g}a-af_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{weak.t.R_it2} $\|\alpha_{g}(f_{h})-f_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{weak.t.R_it3} $(y^{2}-yfy - \varepsilon)_{+} \precsim_{A} x$; \item\label{weak.t.R_it4} $\|fxf\| > 1-\varepsilon$. \end{enumerate} \end{defx} In~\autoref{wtr=tr}, we show that if $\alpha \colon G \rightarrow {\mathrm{Aut}}(A)$ is an action of a finite group $G$ on a simple separable C*-algebra $A$ with tracial rank zero and $\alpha$ has the weak tracial Rokhlin property, then we can always choose $(f_{g})_{g \in G}$ in Definition~\ref{weak.t.R} to be mutually orthogonal {projections}. This result can be considered as a generalization of \cite[Theorem~1.9]{Ph15} to the nonunital case. This motivates us to say that $\alpha \colon G \rightarrow {\mathrm{Aut}}(A)$ has the \emph{tracial Rokhlin property} if we can arrange $(f_{g})_{g \in G}$ in Definition~\ref{weak.t.R} to be mutually orthogonal projections (see \autoref{t.R.p}). It turns out that this definition extends Phillips's definition of the tracial Rokhlin property \cite[Definition~1.2]{Ph11} to the nonunital simple case. We recall that in Phillips's definition of the tracial Rokhlin property for finite group actions on simple unital C*-algebras, Condition~\eqref{weak.t.R_it3} is formulated as follows: \begin{enumerate} \item[$(3)'$]\label{t.R_it3} $1-f \precsim_{A} x$ (or equivalently, $1-f$ is Murray von~Neumann equivalent to a projection in the hereditary subalgebra generated by $x$), \end{enumerate} where $(f_{g})_{g \in G}$ is a family of orthogonal projections in $A$. Phillips in \cite{Ph.free} asked for a correct analogue of Condition~$(3)'$ in the simple nonunital case. In this paper, we give an answer to his question by proposing Condition~\eqref{weak.t.R_it3}. Indeed, Condition~\eqref{weak.t.R_it3} in Definition~\ref{weak.t.R} contains our main idea for a suitable notion of the weak tracial Rokhlin property (as well as tracial Rokhlin property) in the nonunital case. This condition---which may seem strange at the first glance---says that $1-f$ is small with respect to the Cuntz subequivalence relation. The rationale behind this condition is that since $y\in A_{+}$ is arbitrary, we can take it to be arbitrarily large (i.e., close to $1$) and so $y^{2}-yfy=y(1-f)y$ is close to $1-f$. The $\varepsilon$ gap in this condition is a technical condition needed, e.g., when applying a key lemma in the Cuntz semigroup (\autoref{lemkr}). Phillips in \cite[Problem~3.2]{Ph.free} asked whether there is a reasonable formulation of the tracial Rokhlin property for finite group actions on simple unital C*-algebras in terms of the central sequence algebra. We give an answer to this problem in the \emph{not necessarily unital} simple case. Indeed, it turns out that if moreover $A$ is separable and one works with the central sequence algebra $A_{\infty}$, then Condition~\eqref{weak.t.R_it3} can be replaced by $y^{2}-yfy\precsim_{A_{\infty}} x$ (see Propositions~\ref{propcs} and \ref{propcs.tr}). Also, in the setting of the Kirchberg central sequence algebra $F(A)$, it can be replaced by $1_{F(A)} - f \precsim_{F(A)} x$ \cite{G19}. The Rokhlin property can be viewed as a form of noncommutative freeness \cite{Ph.free}. Following this point of view, we prove that an action with the weak tracial Rokhlin property is pointwise outer (\autoref{outer}), and hence the resulting crossed product is simple. Moreover, we prove several permanence properties for finite group actions with the weak tracial Rokhlin property on simple not necessarily unital C*-algebras, e.g., passing to restriction to invariant hereditary C*-algebras, and minimal tensor products and direct limits of actions. In particular, our result concerning tensor products gives an affirmative answer to a problem of Phillips \cite[Problem~3.18]{Ph.free} (see Theorems~\ref{proptensor} and \ref{thmtensor} and the discussion before \autoref{thmtensor}). One of the difficulties to prove such results is the lack of enough traces in this nonunital setting. We use the machinery of Cuntz subequivalence to deal with this difficulty (see, e.g., the proof of Theorems~\ref{proptensor} and \ref{wtr=tr}). We present examples of finite group actions with the (weak) tracial Rokhlin property on simple nonunital C*-algebras (see Sections~\ref{sec_per}, \ref{sec_trp}, and \ref{subsec_examples} for the proof): \begin{thmx}\label{thm_ex_intro} Let $n \geq 2$ and let $S_{n}$ denote the group of all permutations of $\{1,2,\ldots, n\}$. \begin{enumerate} \item\label{thm_ex_intro_it1} (\cite{AGJP17}) Let $\mathcal{W}$ be the Razak-Jacelon algebra. Then the permutation action of $S_{n}$ on $\mathcal{W}^{\otimes n} \cong \mathcal{W}$ has the weak tracial Rokhlin property. \item\label{thm_ex_intro_it2} Let $A$ be a nonelementary simple (not necessarily unital) C*-algebra with tracial rank zero. Then the permutation action of $S_{n}$ on $A^{\otimes n}$ has the tracial Rokhlin property. \item\label{thm_ex_intro_it3} Let $A=\mathcal{Z}\otimes \mathcal{K}$. Then the flip action on $A\otimes A \cong A$ has the weak tracial Rokhlin property but it does not have the Rokhlin property. (We do not know whether this action has the tracial Rokhlin property.) \item\label{thm_ex_intro_it4} Let $A$ be a simple $\mathcal{Z}$-absorbing C*-algebra. Then for every finite nontrivial group $G$ there is an action $\alpha\colon G \to \mathrm{Aut}(A)$ with the weak tracial Rokhlin property. If moreover, $A$ is separable with either $K_{0}(A)=\mathbb{Z}$ or $K_{1}(A)=\mathbb{Z}$, then $\alpha$ does not have the Rokhlin property. \item\label{thm_ex_intro_it5} Let $\alpha \colon G\to \mathrm{Aut}(A)$ and $\beta \colon G\to \mathrm{Aut}(B)$ be actions of a finite group $G$ on simple C*-algebras $A$ and $B$. If $\alpha$ has the weak tracial Rokhlin property then so does $\alpha \otimes \beta \colon G \to \mathrm{Aut}(A \otimes_{\min} B)$. If $\alpha$ has the tracial Rokhlin property then so does $\alpha \otimes \beta \colon G \to \mathrm{Aut}(A \otimes_{\min} B)$ whenever $B^{\beta}$ has an approximate identity (not necessarily increasing) consisting of projections. In particular, if $\alpha$ is Blackadar's action of $\mathbb{Z}_{2}$ on $M_{2^{\infty}}$ (see \cite{Bla90} and \cite{Ph15}), then the action $\alpha \otimes {\mathrm{id}} \colon {\mathbb{Z}}_{2} \to M_{2^{\infty}}\otimes \mathcal{K}$ has the tracial Rokhlin property but not the Rokhlin property. Also, if $B$ is any simple stably projectionless C*-algebra, then the action $\alpha \otimes {\mathrm{id}} \colon {\mathbb{Z}}_{2} \to M_{2^{\infty}}\otimes B$ has the weak tracial Rokhlin property but not the tracial Rokhlin property. \end{enumerate} \end{thmx} That the action in Part~\eqref{thm_ex_intro_it1} of the preceding theorem has the weak tracial Rokhlin property appeared in \cite{AGJP17} for the first time. We do not know whether this action has the Rokhlin property. Parts~\eqref{thm_ex_intro_it2}, \eqref{thm_ex_intro_it3}, and \eqref{thm_ex_intro_it4} are mainly based on a significant result proved in \cite{AGJP17} saying that the permutation action of $S_{n}$ on $A^{\otimes n}$ has the weak tracial Rokhlin property, where $A$ is any simple (not necessarily unital) tracially $\mathcal{Z}$-absorbing C*-algebra such that $A^{\otimes n}$ is finite. We remark that this result from \cite{AGJP17} is used only in Section~\ref{subsec_examples}. Part~\eqref{thm_ex_intro_it5} is based on Theorems~\ref{proptensor} and \ref{thmtensor}. It is proved in \cite{Ph15} that Blackadar's action used in this part has the tracial Rokhlin property but not the Rokhlin property. Phillips proved that the crossed product of a simple unital C*-algebras with tracial rank zero by a finite group action with the tracial Rokhlin property, is again simple with tracial rank zero \cite{Ph11}. Part~\eqref{thm_cross_wtrp_intro_it1} of the following theorem generalizes this result to the nonunital case. \begin{thmx}\label{thm_cross_wtrp_intro} The following classes of not necessarily unital C*-algebras are preserved under taking crossed products and fixed point algebras by finite group actions with the weak tracial Rokhlin property: \begin{enumerate} \item\label{thm_cross_wtrp_intro_it1} simple separable C*-algebras of tracial rank zero; \item\label{thm_cross_wtrp_intro_it2} simple purely infinite C*-algebras; \item\label{thm_cross_wtrp_intro_it3} (\cite{AGJP17, Hirshberg}) simple tracially $\mathcal{Z}$-absorbing C*-algebras. \end{enumerate} \end{thmx} We will prove in \autoref{thmclass1} that more properties are preserved under taking crossed products and fixed point algebras by finite group actions with the tracial Rokhlin property. Part~\eqref{thm_cross_wtrp_intro_it3} of Theorem~\ref{thm_cross_wtrp_intro} is proved in \cite{Hirshberg} for the unital case, and in \cite{AGJP17} for the nonunital case. In \cite{AGJP17}, the second author and others define a tracial version of $\mathcal{Z}$-absorption for simple (not necessarily unital) C*-algebras, extending the same notion defined in \cite{Hirshberg} for (simple) unital C*-algebras. Note that, for simple separable nuclear unital C*-algebras, tracial $\mathcal{Z}$-absorption is equivalent to $\mathcal{Z}$-absorption \cite{Hirshberg}. In \cite{AGJP17}, a simple separable unital purely infinite C*-algebra is given which is tracially $\mathcal{Z}$-absorbing but is not $\mathcal{Z}$-absorbing. The main use so far of tracial $\mathcal{Z}$-absorption has been to prove $\mathcal{Z}$-absorption, because it is easier to check; see for example \cite{Kerr}. This is due to the local nature of tracial $\mathcal{Z}$-absorption. One of the other motivation to study tracial $\mathcal{Z}$-absorption is that it implies several (regularity) properties like having unperforated Cuntz semigroup; see \cite{Hirshberg} and \cite{AGJP17}. By Part~\eqref{thm_cross_wtrp_intro_it3}, tracial $\mathcal{Z}$-absorption is preserved under taking crossed products by finite group actions with the weak tracial Rokhlin property. Moreover, in \cite{AGJP17}, a class of finite group actions with the weak tracial Rokhlin property is introduced by showing that the permutation action of $S_{n}$ on $A^{\otimes n}$ has the weak tracial Rokhlin property provided that $A$ is simple tracially $\mathcal{Z}$-absorbing and $A^{\otimes n}$ is finite. These results suggest that our definition of the weak tracial Rokhlin property on simple not necessarily unital C*-algebras behaves well. To prove Part~\eqref{thm_cross_wtrp_intro_it1} of Theorem~\ref{thm_cross_wtrp_intro}, we need to work with simple nonunital C*-algebras of tracial rank zero. Recall that Lin in \cite{Lin01} first gave the definition of tracial rank for unital C*-algebras and then he defined the tracial rank of a nonunital C*-algebra to be the tracial rank of its minimal unitization. However, working with the unitization of C*-algebras is not always convenient. Moreover, the unitization of a simple nonunital C*-algebra is not simple and so one can not use techniques which are applicable only to simple C*-algebras. To deal with this difficulty, we develop an approach which unifies the concept of tracial rank for both unital and nonunital simple C*-algebras, see~\autoref{main.trk}. This approach helps us to study crossed products of simple nonunital C*-algebras with tracial rank zero by finite group actions with the weak tracial Rokhlin property. We also need some results about simple nonunital C*-algebras with tracial rank zero, such as Morita invariance and having real rank zero and stable rank one (the last two results in the simple \emph{unital} case are proved in \cite{Lin01}). We did not find any reference proving these results (in the nonunital case), however, they may be known to some researchers. So we prove them in \autoref{app}. This paper is organized as follows. Section~\ref{secpre} is devoted to some basic results concerning Cuntz subequivalence which are needed in the subsequent sections. In Section~\ref{sec_wtRp}, we give a definition of the weak tracial Rokhlin property for finite group actions on simple not necessarily unital C*-algebras and study its basic properties. In Section~\ref{sec_per}, we study permanence properties of this notion. Section~\ref{sec_trp} is devoted to the tracial Rokhlin property for actions on simple not necessarily unital C*-algebras. In Section~\ref{subsec_examples}, we present several examples of finite group actions on simple nonunital C*-algebras with the weak tracial Rokhlin proeprty. In Section~\ref{sec_cross}, we mainly study the behavior of tracial rank under taking crossed products by finite group actions with the weak tracial Rokhlin property. We also consider simple purely infinite C*-algebras and other structural properties. In \autoref{app}, we first briefly review some basic results on tracial rank from \cite{Lin01} and then we introduce a unified approach to study tracial rank for both unital and nonunital simple C*-algebras. We show that the class of simple (not necessarily $\sigma$-unital) C*-algebras of tracial rank zero is invariant under Morita equivalence, and every C*-algebra in this class has real rank zero and stable rank one. \section{Cuntz subequivalence}\label{secpre} In this section, we recall some results on Cuntz subequivalence relation and provide some lemmas which will be used in the subsequent sections. We refer the reader to \cite{APT11} and \cite{Ph14} for more information about the Cuntz subequivalence. \begin{nota} We use the following notation in this paper. \begin{enumerate} \item For a C*-algebra $A$, $A_{+}$ denotes the positive cone of $A$. Also, $A^{+}$ denotes the unitization of $A$ (adding a new identity even if $A$ is unital), while $A^{\sim} = A$ if $A$ is unital and $A^{\sim} = A^{+}$ if $A$ is nonunital. \item If $p $ and $q$ are projections in a C*-algebra $A$, then we write $p\sim_{\mathrm{MvN}}q$ if $p$ is Murray-von Neumann equivalent to $q$. \item If $E$ and $F$ are subsets of a C*-algebra $A$ and $\varepsilon>0$, then we write $E\subseteq_{\varepsilon} F$ if for every $a\in E$ there is $b\in F$ such that $\|a-b\|<\varepsilon$. \item We write $\mathcal{K}=K(\ell^{2})$ and $M_{n}=M_{n}(\mathbb{C})$. \item Let $A$ be a C*-algebra. For $a,b\in A_{+}$, we say that $a$ is \emph{Cuntz subequivalent} to $b$ in $A$ and we write $a\precsim_{A} b$, if there is a sequence $(v_n)_{n \in \mathbb{N}}$ in $A$ such that $\|a-v_{n}bv_{n}^{*}\|\to 0$. We write $a\sim_{A} b$ if both $a\precsim_{A} b$ and $b\precsim_{A} a$. If $a,b\in (A\otimes \mathcal{K})_{+}$, we write $a\precsim_{A} b$ if $a$ is {Cuntz subequivalent} to $b$ in $A\otimes \mathcal{K}$. \item Let $a$ be a positive element in a C*-algebra $A$ and let $\varepsilon>0$. Let $f\colon [0,\infty) \to [0,\infty)$ be defined by $f=0$ on $[0,\varepsilon]$ and $f(\lambda)=\lambda-\varepsilon$ on $(\varepsilon,\infty)$. We denote $(a-\varepsilon)_+=f(a)$. \item We use the notation $\mathbb{Z}_{n}$ for the group $\mathbb{Z}/n\mathbb{Z}$. Moreover, $\mathbb{N}$ denotes the set of natural numbers. In particular, $0\not\in\mathbb{N}$. \end{enumerate} \end{nota} The following key lemma will be used several times throughout the paper. \begin{lma}[\cite{KR02}, Lemma~2.2]\label{lemkr} Let $A$ be a C*-algebra, let $a,b\in A_{+}$, and let $\varepsilon >0$. If $\|a-b\|<\varepsilon$ then there is a contraction $d\in A$ such that $(a-\varepsilon)_{+}=dbd^{*}$. In particular, $(a-\varepsilon)_{+}\precsim_{A} b$. \end{lma} In the preceding lemma, if instead of $\|a-b\|<\varepsilon$ we assume that $\|a-b\|\leq\varepsilon$, then again we get $(a-\varepsilon)_{+}\precsim_{A} b$. In fact, for any $\delta>0$ we have $\|a-b\|< \varepsilon + \delta$ and so $(a-\varepsilon-\delta)_{+}\precsim_{A} b$. Letting $\delta \to 0$, we get $(a-\varepsilon)_{+}\precsim_{A} b$. We need the following lemma to work with Condition~\eqref{w.t.R.p_it3} in \autoref{w.t.R.p}. \begin{lma}[\cite{AGJP17}]\label{lemmkey} Let $A$ be a C*-algebra, let $x\in A $ be a nonzero element, and let $b\in A_{+}$. Then for any $\varepsilon>0$, \[ (xbx^{*}-\varepsilon)_{+}\precsim_{A} x \left(b- {\varepsilon}/{\|x\|^{2}}\right)_{+}x^{*}. \] In particular, if $\|x\|\leq 1$ then $(xbx^{*}-\varepsilon)_{+}\precsim_{A} x (b-\varepsilon)_{+}x^{*}\precsim_{A} (b-\varepsilon)_{+}$. \end{lma} \begin{proof} We have \begin{align*} \| xbx^{*} - x \left( b - {\varepsilon}/{\|x\|^{2}}\right)_{+} x^{*}\| \leq \|x\|^{2}\| b - \left( b - {\varepsilon}/{\|x\|^{2}}\right)_{+} \| \leq \|x\|^{2}\frac{\varepsilon}{\|x\|^{2}} = \varepsilon. \end{align*} Using the remark following \autoref{lemkr}, we get \[ (xbx^{*}-\varepsilon)_{+}\precsim_{A} x \left(b- {\varepsilon}/{\|x\|^{2}}\right)_{+}x^{*}. \] If $ \|x\|\leq 1 $ then $ \tfrac{\varepsilon}{\|x\|^{2}} \leq \varepsilon $ and so $(xbx^{*}-\varepsilon)_{+}\precsim_{A} x (b-\varepsilon)_{+}x^{*} \precsim_{A} (b - \varepsilon)_{+}$. \end{proof} Observe that if $a\precsim_{A} b$ then, by definition, there is a sequence $(v_{n})$ in $A$ such that $\|a-v_{n}bv_{n}^{*}\|\to 0$. But it is not the case that there exists always a \emph{bounded} sequence with this property. However, we have the following lemma. \begin{lma}[\cite{AGJP17}]\label{lembdd1} Let $A$ be a C*-algebra, let $a,b\in A_{+}$, and let $\delta>0$. If $a\precsim_{A} (b-\delta)_{+}$ then there exists a bounded sequence $(v_{n})$ in $A$ such that $\|a-v_{n}bv_{n}^{*}\|\to 0$. We can take this sequence such that $\|v_{n}\|\leq \|a\|^{\frac{1}{2}}\delta^{-\frac{1}{2}}$ for every $n\in \mathbb{N}$. \end{lma} \begin{proof} Let $n\in \mathbb{N}$. Since $a\precsim_{A} (b-\delta)_{+}$, by definition, there exists $w_{n}\in A$ such that $\|a-w_{n} (b-\delta)_{+}w_{n}^{*}\|<\frac{1}{n}$. By \autoref{lemkr}, there exists $d_{n}\in A$ such that $(a-\frac{1}{n})_{+}=d_{n}w_{n} (b-\delta)_{+}w_{n}^{*}d_{n}^{*}$. By \cite[Lemma~2.4(i)]{KR02}, there exists $v_{n}\in A$ such that $(a-\frac{1}{n})_{+}=v_{n} bv_{n}^{*}$ and $\|v_{n}\|\leq \|(a-\frac{1}{n})_{+}\|^{\frac{1}{2}}\delta^{-\frac{1}{2}}$. Therefore, $v_{n} bv_{n}^{*}\to a$ and $\|v_{n}\|\leq \|a\|^{\frac{1}{2}}\delta^{-\frac{1}{2}}$. \end{proof} The following lemma is known, however, we did not find a reference. We provide a proof for the convenience of the reader. \begin{lma}\label{lemmvn} Let $A$ be a C*-algebra, let $a\in A_{+}$, and let $p\in A$ be a projection. The following statements are equivalent: \begin{enumerate} \item\label{lemmvn_it1} $p\precsim_{A} a$; \item\label{lemmvn_it2} there exists $v\in A$ such that $p=vav^{*}$; \item\label{lemmvn_it3} $p\sim_{\mathrm{MvN}} q$ in $A$ for some projection $q$ in $\overline{aAa}$. \end{enumerate} \end{lma} \begin{proof} \eqref{lemmvn_it1}$\Rightarrow$\eqref{lemmvn_it2}: Let $p\precsim_{A} a$. Choose $\delta$ with $0<\delta<1$. Then there exists $w\in A$ such that $\|p-waw^{*}\|<\delta$. Then by \autoref{lemkr}, $(p-\delta)_{+}=dwaw^{*}d^{*}$ for some $d\in A$. Note that $(p-\delta)_{+}=(1-\delta)p$. Put $v=(1-\delta)^{-\frac{1}{2}}wd$. Then $p=vav^{*}$. \eqref{lemmvn_it2}$\Rightarrow$\eqref{lemmvn_it3}: Let $p=vav^{*}$ for some $v\in A$. Then $q=a^{\frac{1}{2}}v^{*}va^{\frac{1}{2}}$ is a projection in $\overline{aAa}$ and is Murray-von~Neumann equivalent to $p$. \eqref{lemmvn_it3}$\Rightarrow$\eqref{lemmvn_it1}: Let $p \sim_{\mathrm{MvN}} q$ for some projection $q$ in $\overline{aAa}$. Then $p\sim_{A} q\precsim_{A} a$. Alternatively, we give another argument (which avoids \autoref{lemkr}) to prove that \eqref{lemmvn_it1} implies \eqref{lemmvn_it3}. Let $p\precsim_{A} a$. Then there is $z\in A$ such that $\|p-zaz^{*}\|<\frac{1}{2}$. By \cite[Lemma~2.5.4]{Lin01}, there exists a projection $r$ in $C^{*}(zaz^{*})$, the C*-subalgebra of $A$ generated by $zaz^{*}$, such that $\|p-r\|<1$. Thus $p\sim_{\mathrm{MvN}} r$. We have \[ r\in C^{*}(zaz^{*})\subseteq \overline{zaz^{*}Azaz^{*}} \subseteq \overline{zaAaz^{*}}. \] By \cite[Lemma~3.8]{Pas}, there is a projection $q$ in $\overline{az^{*}Az^{*}a}\subseteq \overline{aAa}$ such that $q \sim_{\mathrm{MvN}} r$. Therefore, $p \sim_{\mathrm{MvN}} q$ in $A$, and $q\in\overline{aAa}$. \end{proof} Note that the statements in \autoref{lemmvn} are also equivalent to $[p]\leq [a]$ in the sense of \cite[Definition~2.2]{Lin01} (see \autoref{app}). In general, there is no upper bound for the norm of $v$ in the previous lemma, unless there is a gap between $p$ and $a$; see the following lemma (which may be considered as a special case of \cite[Lemma~2.4]{KR02}). \begin{lma}\label{lembddp} Let $A$ be a C*-algebra, let $a\in A_{+}$, let $\varepsilon>0$, and let $p\in A$ be a projection. If $p\precsim_{A} (a-\varepsilon)_{+}$, then there exists $v\in A$ such that $p=vav^{*}$ and $\|v\|\leq \varepsilon^{-\frac{1}{2}}$. \end{lma} \begin{proof} By \autoref{lemmvn}, there exists $w\in A$ such that $p=w(a-\varepsilon)_{+}w^{*}$. Then \cite[Lemma~2.4(i)]{KR02} implies that there is $v\in A$ such that $p=vav^{*}$ and $\|v\|\leq \varepsilon^{-\frac{1}{2}}$. \end{proof} Part~\eqref{lembdd_it1} of the following lemma is a variant of \cite[Lemma~2.4]{KR02}. We shall use this lemma in the proof of \autoref{lemtrp}. \begin{lma}\label{lembdd} Let $A$ be a C*-algebra, let $a,b\in A_{+}$, and let $\varepsilon>0$. \begin{enumerate} \item\label{lembdd_it1} If $a=x(b-\varepsilon)_{+}$ for some $x\in A$, then $a=yb$ for some $y\in A$ with $\|y\|\leq \varepsilon^{-1}\|a\|$. \item\label{lembdd_it2} If $a\in \overline{A(b-\varepsilon)_{+}}$ then there is a sequence $(v_{n})$ in $A$ such that $\|a-v_{n}b\|\to 0$ and $\|v_{n}\|\leq \varepsilon^{-1}(\|a\|+\frac{1}{n})$ for all $n\in \mathbb{N}$. \end{enumerate} \end{lma} \begin{proof} We define continuous functions $f_{\varepsilon},g_{\varepsilon}\colon [0,\infty)\to [0,\infty)$ for $\varepsilon>0$ as in \cite[Lemma~2.4]{KR02}, that is, \[ f_{\varepsilon}(t)= \begin{cases} \sqrt{\dfrac{t-\varepsilon}{t}} &t\geq \varepsilon\\ 0 &t<\varepsilon \end{cases} \ \ \ \text{and}\ \ \ g_{\varepsilon}(t)= \begin{cases} \dfrac{1}{t} &t\geq \varepsilon\\ \varepsilon^{-2}t &t<\varepsilon. \end{cases} \] Then $tf_{\varepsilon}(t)^{2}=(t-\varepsilon)_{+}$ and $f_{\varepsilon}(t)^{2}=(t-\varepsilon)_{+}g_{\varepsilon}(t)$. Thus $bf_{\varepsilon}(b)^{2}=(b-\varepsilon)_{+}$ and $f_{\varepsilon}(b)^{2}=(b-\varepsilon)_{+}g_{\varepsilon}(b)$. Note that $\|g_{\varepsilon}(b)\|\leq \varepsilon^{-1}$. To prove \eqref{lembdd_it1}, put $y=xf_{\varepsilon}(b)^{2}$. Then $yb= xf_{\varepsilon}(b)^{2}b=x(b-\varepsilon)_{+}=a$. Also, \[ yy^{*}=xf_{\varepsilon}(b)^{4}x^{*}=x(b-\varepsilon)_{+}^{2}g_{\varepsilon}(b)^{2}x^{*} \leq \|g_{\varepsilon}(b)^{2}\|x(b-\varepsilon)_{+}^{2}x^{*}\leq \varepsilon^{-2}aa^{*}. \] Thus $\|y\|\leq \varepsilon^{-1}\|a\|$. For \eqref{lembdd_it2}, let $a\in \overline{A(b-\varepsilon)_{+}}$ and {f}ix $n\in \mathbb{N}$. Then there is $w_{n}\in A$ such that $\|a-w_{n}(b-\varepsilon)_{+}\|<\frac{1}{n}$. Put $a_{n}=w_{n}(b-\varepsilon)_{+}$. Thus $\|a_{n}\|\leq \|a\|+\frac{1}{n}$. By \eqref{lembdd_it1} there is $v_{n}\in A$ such that $a_{n}=v_{n}b$ and $\|v_{n}\|\leq \varepsilon^{-1}(\|a\|+\frac{1}{n})$. Then $\|a-v_{n}b\|=\|a-a_{n}\|<\frac{1}{n}$, and so $\|a-v_{n}b\|\to 0$. \end{proof} \section{The weak tracial Rokhlin property}\label{sec_wtRp} In this section, we give a definition for the weak tracial Rokhlin property for finite group actions on simple not necessarily unital C*-algebras. We show that the weak tracial Rokhlin property implies pointwise outerness and so the resulting crossed product is simple. (The proof of this is not as easy as the unital case.) Then we compare it with other notions of the weak tracial Rokhlin property for actions on simple unital C*-algebras. Moreover, we show that the Rokhlin property in the sense of \cite[Definition~3.2]{Sa15} implies the weak tracial Rokhlin property for actions on simple C*-algebras. \begin{df}\label{w.t.R.p} Let $\alpha \colon G \rightarrow {\mathrm{Aut}} (A)$ be an action of a finite group $G$ on a simple not necessarily unital C*-algebra $A$. We say that $\alpha$ has the \emph{weak tracial Rokhlin property} if for every $\varepsilon > 0$, every finite subset $F \subseteq A$, and all positive elements $x,y \in A$ with $\|x\|=1$, there exists a family of orthogonal positive contractions $(f_{g})_{g \in G}$ in $A$ such that, with $f=\sum_{g \in G} f_{g}$, the following hold: \begin{enumerate} \item\label{w.t.R.p_it1} $\|f_{g}a-af_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{w.t.R.p_it2} $\|\alpha_{g}(f_{h})-f_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{w.t.R.p_it3} $(y^{2}-yfy - \varepsilon)_{+} \precsim_{A} x$; \item\label{w.t.R.p_it4} $\|fxf\| > 1-\varepsilon$. \end{enumerate} \end{df} Pointwise outerness can be regarded as a form of freeness for actions on noncommutative C*-algebras (see \cite{Ph.free}). An action $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ is called \emph{pointwise outer} if for any $ g \in G\setminus \{1\}$, the automorphism $\alpha_{g}$ is outer, i.e., it is not of the form ${\mathrm{Ad}} (u)$ for any unitary $u$ in the multiplier algebra of $A$. \begin{prop} \label{outer} Let $\alpha$ be an action of a finite group $G$ on a simple not necessarily unital C*-algebra $A$. If $\alpha$ has the weak tracial Rokhlin property then $\alpha$ is pointwise outer. \end{prop} \begin{proof} The idea of the proof is similar to that of \cite[Proposition~3.2]{Sa15}. However, we need Condition~(4) in \autoref{w.t.R.p} instead of Condition~(iii) in \cite[Definition~3.2]{Sa15}, and so we need more estimates. Suppose to the contrary that there are $g_{0}\in G\setminus\{1\}$ and a unitary $u $ in the multiplier algebra of $A$ such that $\alpha_{g_{0}}={\mathrm{Ad}} (u)$. Choose $\varepsilon$ with $0<\varepsilon<1$ such that \[ \dfrac{\sqrt{1-\varepsilon}-n\varepsilon}{n}>0 \ \ \text{and}\ \ \left(\dfrac{\sqrt{1-\varepsilon}-n\varepsilon}{n}\right)^{2}-4\varepsilon>0, \] where $n={\mathrm{card}}(G)$. By \cite[Lemma~2.5.11]{Lin.book} (with $f(t)=t^{\frac{1}{2}}$ there), there is $\delta>0$ such that if $x,y\in A$ are positive contractions with $\|xy-yx\|<\delta$ then $\|x^{\frac{1}{2}}y-yx^{\frac{1}{2}}\|<\varepsilon$. We may assume that $\delta<\varepsilon$. Choose a positive element $b\in A^{\alpha}$ with $\|b\|=1$. Applying \autoref{w.t.R.p} with $F=\{b, bu^{*}\}$, with $\delta$ in place of $\varepsilon$, with $b$ in place if $x$, and with $y=0$, we obtain a family of orthogonal positive contractions $(f_{g})_{g\in G}$ in $A$ such that \begin{enumerate} \item\label{outer_it1} $\|f_{g}b-bf_{g}\|<\delta$ and $\|f_{g}bu^{*}-bu^{*}f_{g}\|<\delta$ for all $g\in G$; \item\label{outer_it2} $\|\alpha_{g}(f_{h})-f_{gh}\|<\delta$ for all $g,h\in G$; \item\label{outer_it3} $\|fbf\|>1-\delta$ where $ f= \sum_{g \in G} f_{g}$. \end{enumerate} Using \eqref{outer_it1} we have ($1$ denotes the neutral element of $G$): \begin{align*} \Big\|fb^{\frac{1}{2}}-\sum_{g\in G} \alpha_{g}(f_{1}b^{\frac{1}{2}})\Big\|&= \Big\| \sum_{g \in G} f_{g}b^{\frac{1}{2}}-\sum_{g\in G} \alpha_{g}(f_{1})b^{\frac{1}{2}}\Big\| \\ &\leq \sum_{g \in G} \|f_{g}-\alpha_{g}(f_{1})\| < n\delta. \end{align*} Thus, using \eqref{outer_it3} at the third step we get \[ n\delta>\|fb^{\frac{1}{2}}\|-\sum_{g\in G} \|\alpha_{g}(f_{1}b^{\frac{1}{2}})\| \geq \|fbf\|^{\frac{1}{2}}-n\|f_{1}b^{\frac{1}{2}}\|>\sqrt{1-\delta}-n\|f_{1}b^{\frac{1}{2}}\|. \] Hence, \begin{equation}\label{equout1} \|f_{1}b^{\frac{1}{2}}\|>\dfrac{\sqrt{1-\delta}-n\delta}{n}> \dfrac{\sqrt{1-\varepsilon}-n\varepsilon}{n}. \end{equation} By \eqref{outer_it1}, $\|f_{1}b-bf_{1}\|<\delta$ and so $\|f_{1}^{\frac{1}{2}}b-bf_{1}^{\frac{1}{2}}\|<\varepsilon$. Thus, \begin{equation}\label{equout2} \|f_{1}b-f_{1}^{\frac{1}{2}}bf_{1}^{\frac{1}{2}}\|\leq \|f_{1}^{\frac{1}{2}}b-bf_{1}^{\frac{1}{2}}\|<\varepsilon. \end{equation} Similarly, since $\|f_{g}b-bf_{g}\|<\delta$ we have \begin{equation}\label{equout3} \|f_{g_{0}}b-f_{g_{0}}^{\frac{1}{2}}bf_{g_{0}}^{\frac{1}{2}}\|<\varepsilon. \end{equation} Note that $f_{g_{0}}^{\frac{1}{2}}bf_{g_{0}}^{\frac{1}{2}}\perp f_{1}^{\frac{1}{2}}bf_{1}^{\frac{1}{2}}$ and thus by (\ref{equout1}) we have \begin{equation}\label{equout5} \|f_{g_{0}}^{\frac{1}{2}}bf_{g_{0}}^{\frac{1}{2}}-f_{1}^{\frac{1}{2}}bf_{1}^{\frac{1}{2}}\| \geq \|f_{1}^{\frac{1}{2}}bf_{1}^{\frac{1}{2}}\|= \|f_{1}^{\frac{1}{2}}b^{\frac{1}{2}}\|^{2}\geq \|f_{1}b^{\frac{1}{2}}\|^{2}> \left(\dfrac{\sqrt{1-\varepsilon}-n\varepsilon}{n}\right)^{2}. \end{equation} Moreover, using \eqref{outer_it1} we have \begin{align}\label{equout4} \|uf_{1}bu^{*}-bf_{1}\|&=\|uf_{1}bu^{*}-\alpha_{g_{0}}(b)f_{1}\| \\ &=\|uf_{1}bu^{*}-ubu^{*}f_{1}\| \notag\\ &\leq \|f_{1}bu^{*}-bu^{*}f_{1}\|<\delta \notag. \end{align} Finally, using \eqref{outer_it2}, (\ref{equout2}), (\ref{equout3}), (\ref{equout5}), and (\ref{equout4}) we obtain \begin{align*} \|\alpha_{g_{0}}(f_{1}b)-uf_{1}bu^{*}\|&\geq \|f_{g_{0}}^{\frac{1}{2}}bf_{g_{0}}^{\frac{1}{2}}-f_{1}^{\frac{1}{2}}bf_{1}^{\frac{1}{2}}\|- \|f_{g_{0}}^{\frac{1}{2}}bf_{g_{0}}^{\frac{1}{2}}-f_{g_{0}}b\| - \|f_{g_{0}}b- \alpha_{g_{0}}(f_{1}b)\|\\ &\quad - \|uf_{1}bu^{*}-bf_{1}\| - \|bf_{1}-f_{1}^{\frac{1}{2}}bf_{1}^{\frac{1}{2}}\|\\ &>\left(\dfrac{\sqrt{1-\varepsilon}-n\varepsilon}{n}\right)^{2}-\varepsilon-2\delta-\varepsilon\\ &>\left(\dfrac{\sqrt{1-\varepsilon}-n\varepsilon}{n}\right)^{2}-4\varepsilon>0, \end{align*} which is a contradiction. This shows that $\alpha$ is pointwise outer. \end{proof} The following result can be immediately deduced from \cite[Theorem~3.1]{Kishimoto} together with \autoref{outer}. \begin{cor}\label{simpleweak} Let $\alpha$ be an action of a finite group $G$ on a simple C*-algebra $A$. If $\alpha$ has the weak tracial Rokhlin property, then $A \rtimes _{\alpha} G$ is simple. \end{cor} \begin{cor} \label{Moritaweak} Let $\alpha$ be an action of a finite group $G$ on a simple C*-algebra $A$. If $\alpha$ has the weak tracial Rokhlin property, then the fixed point algebra $A^{\alpha}$ is Morita equivalent to $A \rtimes _{\alpha} G$. \end{cor} \begin{proof} By \cite{Rosenberg}, there exists a projection $p$ in the multiplier algebra of $A \rtimes_{\alpha} G$ such that $A^{\alpha} \cong p(A\rtimes _{\alpha} G)p$. Note that $A \rtimes _{\alpha} G$ is simple, by \autoref{simpleweak}, and so $A^{\alpha}$ is isomorphic to a full corner of $A \rtimes _{\alpha} G$. This shows that $A^{\alpha}$ is Morita equivalent to $A \rtimes_{\alpha} G$. \end{proof} The following lemma shows that if \autoref{w.t.R.p} holds for some $y\in A_{+}$ then it also holds for any $z\in A_{+}$ which is ``smaller" than $y$ (i.e., for any positive $z$ in $\overline{yAy}$). (Note that $\overline{Ay}\cap A_{+}= \overline{yA}\cap A_{+}=\overline{yAy}\cap A_{+} $.) \begin{lma}\label{lemdefwtrp} Let $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple C*-algebra $A$. Let $x \in A_{+}$ with $\|x\|=1$. Suppose that a positive element $y \in A$ has the following property: for every $\varepsilon > 0$ and every finite subset $F \subseteq A$ there exists a family of orthogonal positive contractions $(f_{g})_{g \in G}$ in $A$ such that, with $f=\sum_{g \in G} f_{g}$, the following hold: \begin{enumerate} \item\label{lemdefwtrp_it1} $\|f_{g}a-af_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{lemdefwtrp_it2} $\|\alpha_{g}(f_{h})-f_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{lemdefwtrp_it3} $(y^{2}-yfy - \varepsilon)_{+} \precsim_{A} x$; \item\label{lemdefwtrp_it4} $\|fxf\| > 1-\varepsilon$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} Then every positive element $z\in\overline{Ay}$ also has the same property. \end{lma} \begin{proof} The idea of the proof is similar to the argument given in the proof of \cite[Lemma~3.5]{AGJP17}. Let $z \in \overline{Ay}$ be a positive element, and let a finite subset $F \subseteq A$, $\varepsilon >0$, and an element $x \in A_{+}$ with $\|x\|=1$ be given. Let $ \delta$ be such that $0<\delta < \min\left\{1, \frac{\varepsilon}{4(2\|z\|+1)} \right\}$. Since $z \in \overline{Ay}$, there exists a nonzero element $w \in A$ such that $\|z-wy\| < \delta$. Choose $\eta>0$ such that $ \eta < \min \{\varepsilon, \frac{\varepsilon}{2\|w\|^{2}} \}$. By assumption, there exists a family of orthogonal positive contractions $(f_{g})_{g \in G}$ in $A$ such that, with $f=\sum_{g \in G} f_{g}$, the following hold: \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{lemdefwtrp_it5} $\|f_{g}a-af_{g}\| < \eta$ for all $a\in F$ and all $g\in G$; \item\label{lemdefwtrp_it6} $\|\alpha_{g}(f_{h})-f_{gh}\| < \eta$ for all $g,h\in G$; \item\label{lemdefwtrp_it7} $(y^{2}-yfy - \eta)_{+} \precsim_{A} x$; \item\label{lemdefwtrp_it8} $\|fxf\| > 1-\eta$. \end{enumerate} Since $\eta < \varepsilon$, \eqref{lemdefwtrp_it5}, \eqref{lemdefwtrp_it6}, and \eqref{lemdefwtrp_it8} also hold for $\varepsilon$ in place of $\eta$. It remains to show that $(z^{2}-zfz-\varepsilon)_{+} \precsim_{A} x$. To see this, first by \autoref{lemmkey} at the first step and by \eqref{lemdefwtrp_it7} at the last step, we have \begin{align*} \left(wy^{2}w^{*}-wyfyw^{*}- {\varepsilon}/{2}\right)_{+} &\precsim_{A} w\left(y^{2}-yfy- \frac{\varepsilon}{2\|w\|^{2}}\right)_{+}w^{*}\\ &\precsim_{A} \left(y^{2}-yfy- \frac{\varepsilon}{2\|w\|^{2}}\right)_{+}\\ &\precsim_{A} (y^{2}-yfy-\eta)_{+} \precsim_{A} x. \end{align*} On the other hand, we have \begin{align*} \big\|z^{2}-zfz -&\left(wy^{2}w^{*}-wyfyw^{*}- \tfrac{\varepsilon}{2}\right)_{+}\big\|\\ &\leq \|z^{2}-zfz -(wy^{2}w^{*}-wyfyw^{*})\|+ \tfrac{\varepsilon}{2}\\ &\leq \|z^{2}-wyz\|+\|wyz-wy^{2}w^{*}\|+\|zfz-wyfz\|\\ &\quad +\|wyfz-wyfyw^{*}\|+\tfrac{\varepsilon}{2} \\ &\leq \delta (2\|z\|+2\|wy\|)+ \tfrac{\varepsilon}{2}\\ &\leq \delta(4\|z\|+2\delta)+ \tfrac{\varepsilon}{2} \leq \delta(4 \|z\|+2)+ \tfrac{\varepsilon}{2} < \varepsilon. \end{align*} Therefore, by \autoref{lemkr}, $(z^{2}-zfz-\varepsilon)_{+} \precsim_{A} \left(wy^{2}w^{*}-wyfyw^{*}- \frac{\varepsilon}{2}\right)_{+} \precsim_{A} x$, as desired. This finishes the proof. \end{proof} \begin{rem}\label{rmkdefwtrp} Let $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple not necessarily unital C*-algebra $A$. \begin{enumerate} \item\label{rmkdefwtrp_it1} If the property stated in \autoref{w.t.R.p} holds for some $y\in A_{+}$ then it also holds for every positive element $z\in \overline{Ay}$ (by \autoref{lemdefwtrp}). \item\label{rmkdefwtrp_it2} If $A$ is $\sigma$-unital then $\alpha$ has the weak tracial Rokhlin property if \emph{some} strictly positive element $y$ in $A$ has the property stated in \autoref{w.t.R.p}. This follows from \eqref{rmkdefwtrp_it1} and $A=\overline{yAy}=\overline{Ay}$. \item\label{rmkdefwtrp_it3} in \autoref{w.t.R.p}, it is enough to take $y$ in a norm dense subset of $A_{+}$. Moreover, if $(e_{i})_{i\in I}$ is a approximate identity for $A$, it is enough to take $y$ from the set $\{e_{i} \mid i\in I \}$. This follows from \autoref{lemdefwtrp} and the fact that the set $\left\{y \in A_{+} \mid y\in \overline{Ae_{i}} \ \text{for some}\ i\in I \right\}$ is norm dense in $A_{+}$. \item\label{rmkdefwtrp_it4} In \autoref{w.t.R.p}, if moreover $A$ is purely infinite then Condition~\eqref{w.t.R.p_it3} is unnecessary. Also, if $A$ is finite then Condition~\eqref{w.t.R.p_it4} is redundant (this is proved in \cite{AGJP17}, however, we do not need it here). \end{enumerate} \end{rem} Nawata in \cite{Nawata} introduced a notion of Rokhlin property for finite group actions on $\sigma$-unital C*-algebras which coincides with Izumi's definition in the unital case. Later, Santiago in \cite{Sa15} defined the Rokhlin property for actions of finite groups on (not necessarily unital) C*-algebras. His definition coincides with Nawata's definition in the separable case. Hence, Santiago's definition and Izumi's definition coincide in the separable unital case. We show that an action of a finite group on a simple C*-algebra with the Rokhlin property has the weak tracial Rokhlin property. \begin{prop}\label{propsan} Let $A$ be a simple C*-algebra and let $\alpha \colon G\to \mathrm{Aut}(A)$ be an action of a finite group $G$ on $A$. If $\alpha$ has the Rokhlin property in the sense of Definition~3.2 of \cite{Sa15}, then it has the weak tracial Rokhlin property. \end{prop} \begin{proof} Let $\alpha$ have the Rokhlin property in the sense of Definition~3.2 of \cite{Sa15}. Let $x,y\in A_{+}$ with $\|x\|=1$, let $F\subseteq A$ be a finite subset, and let $\varepsilon>0$. We may assume that $y\neq 0$ and $x,y\in F$. Also, by \autoref{lemdefwtrp}, we may further assume that $\|y\| \leq 1$. Since $\alpha$ has the Rokhlin property, there exists a family of orthogonal positive contractions $(f_{g})_{g\in G}$ in $A$ such that, with $f=\sum_{g\in G}f_{g}$, we have: \begin{enumerate} \item\label{propsan_it1} $\|\alpha_{g}(f_{h})-f_{gh}\|<\frac{\varepsilon}{2}$ for all $g,h\in G$; \item\label{propsan_it2} $\|f_{g}a- af_{g}\|<\frac{\varepsilon}{2}$ for all $g\in G$ and all $a\in F$; \item\label{propsan_it3} $\| fa-a\|<\frac{\varepsilon}{2}$ for all $a\in F$. \end{enumerate} Then, Conditions~\eqref{w.t.R.p_it1} and \eqref{w.t.R.p_it2} in \autoref{w.t.R.p} are satisfied (by \eqref{propsan_it1} and \eqref{propsan_it2} above). Since $y\in F$, by \eqref{propsan_it3} we have \[ \|y^{2}-yfy\|\leq \|y-yf\|<{\varepsilon}/{2}<\varepsilon. \] Thus $(y^{2}-yfy-\varepsilon)_{+}=0\precsim_{A} x$. Hence, Condition~\eqref{w.t.R.p_it3} in \autoref{w.t.R.p} is also satisfied. To prove Condition~\eqref{w.t.R.p_it4}, by \eqref{propsan_it3} and that $x\in F$ we have \[ \|fxf-x\|\leq \|fxf-xf\|+\|xf-x\| \leq \|fx-x\|+\|xf-x\|<\varepsilon. \] Thus $\|fxf\|> \|x\|-\varepsilon=1-\varepsilon$, and so Condition~\eqref{w.t.R.p_it4} in \autoref{w.t.R.p} holds. Therefore, $\alpha$ has the weak tracial Rokhlin property. \end{proof} We remark here that the converse of~\autoref{propsan} does not hold in general. In \autoref{exa_blackadar} we give a finite group action on a simple nonunital C*-algebra with the weak tracial Rokhlin property but which does not have the Rokhlin property. There are several weaker versions of the tracial Rokhlin property for actions on simple unital C*-algebras. In the sequel, we compare them with our definition of the weak tracial Rokhlin property given in \autoref{w.t.R.p}. First we recall these definitions for the convenience of the reader. \begin{df}[\cite{GHS}, Definition~2.2]\label{unitalghs} Let $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple unital C*-algebra $A$. We say that $\alpha$ has the \emph{weak tracial Rokhlin property} if for every $\varepsilon > 0$, every finite set $F \subseteq A$, and every positive element $x \in A$ with $\|x\|=1$, there exists a family of positive contractions $(f_{g})_{g \in G}$ in $A$ such that, with $f=\sum_{g \in G} f_{g}$, the following hold: \begin{enumerate} \item\label{unitalghs_it1} $\|f_{g}a-af_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{unitalghs_it2} $\|\alpha_{g}(f_{h})-f_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{unitalghs_it3} $1-f \precsim_{A} x$; \item\label{unitalghs_it4} $\|fxf\| > 1-\varepsilon$; \item\label{unitalghs_it5} $\|f_{g}f_{h}\|<\varepsilon$ for all $g,h\in G$ with $g\neq h$. \end{enumerate} \end{df} We remark that the element $1-f$ in \eqref{unitalghs_it3} is not necessarily positive. The Cuntz subequivalence $1-f \precsim_{A} x$ means that for any $\delta >0$ there are $v,w\in A$ such that $\| (1-f) - vxw \|<\delta$, or equivalently, $(1-f)^{2} \precsim_{A} x$ as defined in \autoref{secpre} (see \cite{Cu78}). \begin{df}[\cite{Gar17}, Definition~3.11]\label{unitalgar} Let $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple unital C*-algebra $A$. Then $\alpha$ has the \emph{weak tracial Rokhlin property} if for every $\varepsilon > 0$, every finite subset $F \subseteq A$, and every positive element $x \in A$ with $\|x\|=1$, there exists a family of orthogonal positive contractions $(f_{g})_{g \in G}$ in $A$ such that, with $f=\sum_{g \in G} f_{g}$, the following hold: \begin{enumerate} \item\label{unitalgar_it1} $\|f_{g}a-af_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{unitalgar_it2} $\|\alpha_{g}(f_{h})-f_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{unitalgar_it3} $1-f \precsim_{A} x$; \item\label{unitalgar_it4} $\|fxf\| > 1-\varepsilon$. \end{enumerate} \end{df} \begin{prop}\label{propequiw} Let $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple unital C*-algebra $A$. The following statements are equivalent: \begin{itemize} \item[(a)] $\alpha$ has the weak tracial Rokhlin property in the sense of \autoref{unitalgar}; \item[(b)] $\alpha$ has the weak tracial Rokhlin property in the sense of \autoref{w.t.R.p}; \item[(c)] $\alpha$ has the weak tracial Rokhlin property in the sense of \autoref{unitalghs}. \end{itemize} \end{prop} \begin{proof} The implication (a)$\Rightarrow$(c) is clear. To prove (c)$\Rightarrow$(b), by \autoref{rmkdefwtrp}\eqref{rmkdefwtrp_it2}, it is enough to take $y=1$ in \autoref{w.t.R.p}. So, let $\varepsilon >0$, let $F \subseteq A$ be a finite subset, and let $x \in A_{+}$ with $\|x\|=1$. We may assume that $F$ is contained in the closed unit ball of $A$ and that $\varepsilon<1$. Set $n={\mathrm{card}}(G)$. By \cite[Lemma~2.5.12]{Lin.book}, there exists $\delta >0$ such that if $(f_{g})_{g \in G}$ is a family of positive contractions in $A$ satisfying $\|f_{g}f_{h}\| < \delta$ for all $g, h \in G$ with $g \neq h$, then there are orthogonal positive contractions $(a_{g})_{g \in G}$ in $A$ such that $\|f_{g}-a_{g}\| < \tfrac{\varepsilon^{2}}{4n}$ for all $g \in G$. We may assume that $\delta < \tfrac{\varepsilon}{3}$. Applying \autoref{unitalghs} with $F, x$ as given and with $\delta$ in place of $\varepsilon$, we obtain positive contractions $(f_{g})_{g \in G}$ in $A$ that satisfy Conditions~\eqref{unitalghs_it1}--\eqref{unitalghs_it5} of \autoref{unitalghs}. By the choice of $\delta$, there exist orthogonal positive contractions $(a_{g})_{g \in G}$ in $A$ such that $\|f_{g}-a_{g}\| < \tfrac{\varepsilon^{2}}{4n}<\tfrac{\varepsilon}{4n}$ for all $g \in G$. We show that the family $(a_{g})_{g \in G}$ satisfies the conditions of \autoref{w.t.R.p}. The verification of Conditions~\eqref{w.t.R.p_it1} and \eqref{w.t.R.p_it2} in \autoref{w.t.R.p} is easy. Put $a=\sum _{g \in G} a_{g}$. Then $\|f-a\|\leq \sum _{g \in G} \|f_{g} - a_{g}\| < \tfrac{\varepsilon^{2}}{4}<\tfrac{\varepsilon}{4}$. In particular, $\|f\|\leq 1+\tfrac{\varepsilon}{4}<\tfrac{5}{4}$. Using \eqref{unitalgar_it4} in \autoref{unitalghs}, we get \[ \|axa\|\geq \|fxf\| - \tfrac{\varepsilon}{4} - \tfrac{\varepsilon}{4} \|f\| >1-\delta - \tfrac{\varepsilon}{4} - \tfrac{5\varepsilon}{16} >1-\tfrac{\varepsilon}{3} - \tfrac{\varepsilon}{4} - \tfrac{5\varepsilon}{16} >1-\varepsilon \] which is \eqref{w.t.R.p_it4} in \autoref{w.t.R.p}. It remains to show that $(1-a-\varepsilon)_{+}\precsim_{A} x$. First we have \[ \big\| (1-a)^{2}- (1-f)^{2}\| \leq \|a-f\|+\|1-f\|\cdot\|a-f\| < \tfrac{\varepsilon^{2}}{4}+ \tfrac{3\varepsilon^{2}}{4}=\varepsilon^{2}. \] By this and \autoref{lemkr} at the second step and by the remark following \autoref{unitalghs} at the third step, we get \[ (1-a-\varepsilon)_{+} \sim_{A} ((1-a)^{2}-\varepsilon^{2})_{+} \precsim_{A} (1-f)^{2} \precsim_{A} x, \] as desired. This finishes the proof of (c)$\Rightarrow$(b). To show (b)$\Rightarrow$(a), let $F$, $x$, and $\varepsilon$ be as in \autoref{unitalgar}. We will find orthogonal positive contractions $(f_{g})_{g \in G}$ in $A$ satisfying \eqref{unitalgar_it1}--\eqref{unitalgar_it4} of \autoref{unitalgar}. We may assume that $F$ is contained in the closed unit ball of $A$. Choose $\delta$ with $ 0< \delta < \tfrac{\varepsilon}{2n+1}$ (where $n={\mathrm{card}}(G)$). Applying \autoref{w.t.R.p} with $\delta$ in place of $\varepsilon$, with $y=1$, and with $x, F$ as given, there are orthogonal positive contractions $(e_{g})_{g \in G}$ in $A$ such that, with $e= \sum _{g \in G} e_{g}$, the following hold: \begin{enumerate} \item\label{propequiw_pf_it1} $\|e_{g}a-ae_{g}\| < \delta$ for all $a\in F$ and all $g\in G$; \item\label{propequiw_pf_it2} $\|\alpha_{g}(e_{h})-e_{gh}\| < \delta$ for all $g,h\in G$; \item\label{propequiw_pf_it3} $(1-e- \delta)_{+} \precsim_{A} x$; \item\label{propequiw_pf_it4} $\|exe\| > 1-\delta$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} We write $G=\{g_{1},\ldots,g_{n}\}$. We define a c.p.c.~order zero map $\phi: \mathbb{C}^{n} \rightarrow A$ by $\phi(\lambda_{1},\ldots, \lambda_{n})= \sum _{i=1} ^{n} \lambda _{i} e_{g_{i}}$. Then $\phi(1)=e$. Let $\eta: [0,1] \rightarrow [0,1]$ be the continuous function defined by \[ \eta(\lambda)= \begin{cases} (1-\delta)^{-1} \lambda & 0 \leq \lambda \leq 1- \delta,\\ 1& 1-\delta< \lambda \leq 1. \end{cases} \] Using the functional calculus for c.p.c.~order zero maps (\cite[Corollary~4.2]{WZ08}), we define $\psi= \eta(\phi)$. Thus $\psi: \mathbb{C}^{n} \rightarrow A$ is a c.p.c.~order zero map. Similar to the argument given in the proof of \cite[Lemma~2.8]{ABP17}, we see that $\|\psi(z)- \phi(z)\| \leq \delta \|z\|$ for all $z \in \mathbb{C}^{n}$ with $\|z\|=1$, and that \[ 1-\psi(1) = \tfrac{1}{1-\delta} \left(1- \phi(1)-\delta \right)_{+} \sim_{A} \left(1 -\phi(1)- \delta \right)_{+} \precsim_{A} x. \] Put $f_{g_{i}} = \psi(E_{i})$, where $E_{i}$ is the element of $\mathbb{C}^{n}$ having 1 in the $i$-th coordinate and 1 elsewhere. Thus, $(f_{g})_{g \in G}$ is a family of orthogonal positive contractions in $A$ and we have \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{propequiw_pf_it5} $\|f_{g}-e_{g}\| \leq \delta$ for all $g \in G$. \end{enumerate} Moreover, with $ f= \sum _{g \in G} f_{g}$, we have $1-f=1- \psi(1) \precsim_{A} x$, which is \eqref{unitalgar_it3} in \autoref{unitalgar}. Using \eqref{propequiw_pf_it5}, it is easy to see that Conditions~\eqref{unitalgar_it1}, \eqref{unitalgar_it2}, and \eqref{unitalgar_it4} in \autoref{unitalgar} follow from \eqref{propequiw_pf_it1}, \eqref{propequiw_pf_it2}, and \eqref{propequiw_pf_it4} above, respectively. This completes the proof of (b)$\Rightarrow$(a). \end{proof} In~\cite{Hirshberg}, to study the behavior of $\mathcal{Z}$-absorption by taking crossed products, the authors defined another weak version of the tracial Rokhlin property called the generalized tracial Rokhlin property. To complete the comparison between different notions of weak tracial Rokhlin property in the unital case and \autoref{w.t.R.p}, we quote the following result from \cite{AGJP17}. \begin{prop}[\cite{AGJP17}]\label{prop_AGJP17} Suppose that $\alpha \colon G \to {\mathrm{Aut}}(A)$ is an action of a finite group $G$ on a simple unital C*-algebra $A$. If $\alpha$ has the weak tracial Rokhlin property, then $\alpha$ has the generalized tracial Rokhlin property in the sense of \cite[Definition~5.2]{Hirshberg}. Moreover, if $A$ is finite, then the converse also holds. \end{prop} In the following lemma, we give a (seemingly) stronger equivalent definition of the weak tracial Rokhlin property for finite group actions on simple C*-algebras. This lemma says that we can take two different unknowns $x,z$ in Conditions~\eqref{w.t.R.p_it3} and \eqref{w.t.R.p_it4} of \autoref{w.t.R.p} instead of $x$. The idea of the proof of this lemma will be used also in a number of places later. \begin{lma}\label{lemtrp2} Let $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple C*-algebra $A$. Then $\alpha$ has the weak tracial Rokhlin property if and only if the following holds. For every $\varepsilon > 0$, every finite subset $F \subseteq A$, and all positive elements $x,y,z \in A$ with $x\neq 0$ and $\|z\|=1$, there exists a family of orthogonal positive contractions $(f_{g})_{g \in G}$ in $A$ such that, with $f=\sum_{g \in G} f_{g}$, the following hold: \begin{enumerate} \item\label{lemtrp2_it1} $\|f_{g}a-af_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{lemtrp2_it2} $\|\alpha_{g}(f_{h})-f_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{lemtrp2_it3} $(y^{2}-yfy - \varepsilon)_{+} \precsim_{A} x$; \item\label{lemtrp2_it4} $\|fzf\| > 1-\varepsilon$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} \end{lma} \begin{proof} The backward implication is obvious. For the forward implication, let $\alpha$ have the weak tracial Rokhlin property and let $\varepsilon, F, x,y,z$ be as in the statement. We may assume that $F$ is contained in the closed unit ball of $A$. Let $n={\mathrm{card}}(G)$. Choose $\delta$ such that $0<\delta<1$ and \[ \left(\dfrac{\delta}{2-\delta}(1-\tfrac{\varepsilon}{2})\right)^{2}>1-\varepsilon. \] Put $z_{1}=(z^{{1}/{2}}-\delta)_{+}$. Since $A$ is simple, \cite[Lemma~2.6]{Ph14} implies that there is a positive element $d\in \overline{z_{1}Az_{1}}$ such that $d\precsim_{A} x$ and $\|d\|=1$. Applying \autoref{w.t.R.p} with $y$ and $F$ as given, with $\frac{\varepsilon}{2}$ in place of $\varepsilon$, and with $d$ in place of $x$, there exist orthogonal positive contractions $(f_{g})_{g \in G}$ in $A$ such that, with $f=\sum_{g \in G} f_{g}$, the following hold: \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{lemtrp2_pf_it5} $\|f_{g}a-af_{g}\| < \frac{\varepsilon}{2}$ for all $a\in F$ and all $g\in G$; \item\label{lemtrp2_pf_it6} $\|\alpha_{g}(f_{h})-f_{gh}\| < \frac{\varepsilon}{2}$ for all $g,h\in G$; \item\label{lemtrp2_pf_it7} $(y^{2}-yfy - \frac{\varepsilon}{2})_{+} \precsim_{A} d$; \item\label{lemtrp2_pf_it8} $\|fdf\| > 1-\frac{\varepsilon}{2}$. \end{enumerate} Clearly, \eqref{lemtrp2_it1}, \eqref{lemtrp2_it2}, and \eqref{lemtrp2_it3} follow from \eqref{lemtrp2_pf_it5}, \eqref{lemtrp2_pf_it6}, and \eqref{lemtrp2_pf_it7}, respectively. To see \eqref{lemtrp2_it4}, first note that have $d\in \overline{z_{1}Az_{1}}\subseteq \overline{Az_{1}}=\overline{A(z^{{1}/{2}}-\delta)_{+}}$. Thus by \autoref{lembdd} there exists a sequence $(v_{n})_{n\in \mathbb{N}}$ in $A$ such that $\|v_{n}z^{{1}/{2}}-d\|\to 0$ and $\|v_{n}\|\leq (\|d\|+\frac{1}{n})\delta^{-1}=(1+\frac{1}{n})\delta^{-1}$. Then $\|fv_{n}z^{\frac{1}{2}}f-fdf\|\to 0$. Since $\|fdf\|>1-\frac{\varepsilon}{2}$ and $\delta<1$, there is $n\in \mathbb{N}$ such that $\|fv_{n}x^{\frac{1}{2}}f\|>1-\frac{\varepsilon}{2}$ and $\frac{1}{n}<1-\delta$. Hence, \[ 1-\tfrac{\varepsilon}{2}<\|fv_{n}z^{\frac{1}{2}}f\|\leq \|z^{\frac{1}{2}}f \| \|v_{n}\| \leq \|z^{\frac{1}{2}}f \| (1+\tfrac{1}{n})\delta^{-1}\leq \|z^{\frac{1}{2}}f \| (2-\delta)\delta^{-1}. \] Thus, \[ \|fzf\|=\|z^{\frac{1}{2}}f\|^{2}> \left(\dfrac{\delta}{2-\delta}(1-\tfrac{\varepsilon}{2})\right)^{2}>1-\varepsilon. \] This completes the proof. \end{proof} Now we give an equivalent definition for the weak tracial Rokhlin property in terms of the central sequence algebra. For a C*-algebra $A$, we write \[ A_{\infty}= {\ell^{\infty}(\mathbb{N}, A)} /{c_{0}(\mathbb{N}, A)}. \] We consider the elements of $A$ in $A_{\infty}$ as the equivalence classes of constant sequences. We denote by $A_{\infty} \cap A'$ the relative commutant of $A$ in $A_{\infty}$. Also, $\pi_{\infty} \colon \ell^{\infty}(\mathbb{N}, A)\to A_{\infty}$ denotes the quotient map. If $\alpha \colon G \rightarrow \mathrm{Aut} (A)$ is an action, then we denote by $\alpha _{\infty}$ the induced action of $G$ on $A_{\infty}$. \begin{prop}\label{propcs} Let $\alpha \colon G \rightarrow \mathrm{Aut} (A)$ be an action of a finite group $G$ on a simple separable C*-algebra $A$. Then $\alpha$ has the weak tracial Rokhlin property if and only if for every $x, y \in A_{+}$ with $\|x\|=1$, there exists a family of orthogonal positive contractions $(f_{g})_{g \in G}$ in $ A _{\infty} \cap A'$ such that, with $f = \sum_{g \in G} f_{g}$, the following hold: \begin{enumerate} \item\label{propcs_it1} $(\alpha_{\infty})_{g}(f_{h})=f_{gh}$ for all $g,h\in G$; \item\label{propcs_it2} $y^{2}-yfy \precsim_{A_{\infty}} x$; \item\label{propcs_it3} $\|fxf\|=1$. \end{enumerate} If moreover $A$ is unital, then Condition~\eqref{propcs_it2} can replaced by $1-f\precsim_{A_{\infty}} x$. \end{prop} \begin{proof} Assume that $\alpha$ has the weak tracial Rokhlin property. Let $x, y \in A_{+}$ with $\|x\|=1$. Let $\{a_{1}, a_{2}, \ldots\}$ be a norm dense countable subset of the closed unit ball of $A$. Put $F_{n}=\{a_{1},\ldots,a_{n}\}$, $n\in \mathbb{N}$. Applying \autoref{lemtrp2} with $(x-\tfrac{1}{2})_{+}$ in place of $x$, with $x$ in place of $z$, with $F_{n}$ in place of $F$, and with $\tfrac{1}{n}$ in place of $\varepsilon$, we obtain mutually orthogonal positive contractions $(f_{(g,n)})_{g \in G}$ in $A$ satisfying \eqref{lemtrp2_it1}--\eqref{lemtrp2_it4} in \autoref{lemtrp2}. Let $f_{g}\in A_{\infty}$ denote the equivalence class of $(f_{(g,n)})_{n \in \mathbb{N}}$. Then $(f_{g})_{g \in G}$ is a family of orthogonal positive contractions in $A _{\infty} \cap A'$. It is easy to see that Conditions~\eqref{propcs_it1} and \eqref{propcs_it3} in the statement hold. To see \eqref{propcs_it2}, put $h_{n}=\sum_{g \in G} f_{(g,n)}$. Then $f:=\sum_{g \in G} f_{g}$ is the equivalence class of $(h_{n})_{n \in \mathbb{N}}$ in $A_{\infty}$ and $(y^{2}-yh_{n}y-\tfrac{1}{n})_{+} \precsim_{A} (x-\tfrac{1}{2})_{+}$. By \autoref{lembdd1}, there is $v_{n} \in A$ such that $\|(y^{2}-yh_{n}y-\tfrac{1}{n})_{+}- v_{n}xv_{n}^{*}\|< \tfrac{1}{n}$ and $\|v_{n}\| \leq 2\|y\|$. Hence $\|y^{2}-yh_{n}y-v_{n}xv_{n}^{*}\|< \tfrac{2}{n}$. Let $v$ be the equivalence class of $(v_{n})_{n \in \mathbb{N}}$ in $A_{\infty}$. (Note that $(v_{n})_{n \in \mathbb{N}}$ is a bounded sequence.) Then $y^{2}-yfy=vxv^{*} \precsim_{A_{\infty}} x$. Using \autoref{lemkr} and \cite[Lemma~2.5.12]{Lin.book}, the other implication follows. If moreover $A$ is unital and we replace Condition~\eqref{propcs_it2} by $1-f\precsim_{A_{\infty}} x$, then the forward implication follows by taking $y=1$ in the preceding argument. The backward implication follows from the previous case and the fact that $y^{2}-yfy =y(1-f)y \precsim_{A_{\infty}} 1-f$. \end{proof} Let $\alpha \colon G \rightarrow \mathrm{Aut} (A)$ be an action of a finite group $G$ on a simple separable C*-algebra $A$, and let $F(A)$ be the Kirchberg central sequence algebra. It is shown in \cite{G19} that $\alpha$ has the weak tracial Rokhlin property if and only if for every $x\in A_{+}$ with $\|x\|=1$ there are orthogonal positive contractions $(f_{g})_{g \in G}$ in $ F(A)$ such that $1_{F(A)}-f \precsim_{F(A)} x$, $(\alpha_{\omega})_{g}(f_{h})=f_{gh}$ for all $g,h\in G$, and $\|fxf\|=1$, where $f = \sum_{g \in G} f_{g}$ and $\alpha_{\omega}$ is the induced action on $F(A)$. \section{Permanence properties}\label{sec_per} The purpose of this section is to study the behavior of the weak tracial Rokhlin property for finite group actions on simple not necessarily unital C*-algebras under restriction to subgroups and invariant hereditary subalgebras, and direct limits and tensor products of actions. The following proposition is a nonunital version of \cite[Lemma~5.6]{Echterhoff-Phillips}. The proof is similar to the unital case. \begin{prop} \label{subgrp} Let $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple C*-algebra $A$ with the weak tracial Rokhlin property. If $H$ is a subgroup of $G$, then the restriction of $\alpha$ to $H$ also has the weak tracial Rokhlin property. \end{prop} Next, we show that the weak tracial Rokhlin property is preserved by passing to invariant hereditary subalgebras. \begin{prop} \label{propwtrpher} Let $A$ be a simple C*-algebra and let $\alpha \colon G\to \mathrm{Aut}(A)$ be an action of a finite group $G$ with the weak tracial Rokhlin property. Let $B$ be an $\alpha$-invariant hereditary C*-subalgebra of $A$ and let $\beta \colon G\to \mathrm{Aut}(B)$ be the restriction of $\alpha$ to $B$. Then $\beta$ has the weak tracial Rokhlin property. \end{prop} \begin{proof} Suppose we are give a finite set $F \subseteq B$, $\varepsilon >0$, and $x, y \in B_{+}$ with $\|x\|=1$. We can assume that $x, y \in F$ and that $F$ is contained in the closed unit ball of $B$. Without loss of generality $\varepsilon<1$. Set $n={\mathrm{card}}(G)$. By \cite[Lemma~2.5.12]{Lin.book}, there is $\delta >0$ such that if $(e_{g})_{g \in G}$ is a family of positive contractions in $B$ satisfying $\|e_{g}e_{h}\| < 2\delta$ for all $g, h \in G$ with $g \neq h$, then there are orthogonal positive contractions $(f_{g})_{g \in G}$ in $B$ such that $\|f_{g}-e_{g}\| < \tfrac{\varepsilon}{4n}$ for any $g \in G$. We may assume that $\delta < \tfrac{\varepsilon}{28n}$. Since any approximate identity for $B^{\beta}$ is also an approximate identity for $B$, we can choose a positive contraction $b\in B^{\beta}$ such that \begin{equation}\label{app.herd} \|ab-a\|<\delta \ \ \text{and} \ \ \|ba-a\| <\delta,\ \end{equation} for any $a \in F$. We set $z=bxb$ and $w=byb$. So by \eqref{app.herd} we have \begin{equation}\label{eq1_wtrp.herd} \|z-x\|<2\delta\ \ \text{and}\ \ \|w-y\|<2\delta . \end{equation} Applying \autoref{w.t.R.p} to $\alpha$ with $F \cup \{b\}$ in place of $F$, with $z/\|z\|$ in place of $x$, with $w$ in place of $y$, and with $\delta$ in place of $\varepsilon$, we obtain orthogonal positive contractions $(r_{g})_{g \in G}$ in $A$ such that, with $r=\sum_{g \in G} r_{g}$, the following hold: \begin{enumerate} \item \label{it1.wtrp.herd} $\|r_{g}a -ar_{g}\| < \delta$ for all $a \in F \cup \{b\}$ and all $g \in G$; \item \label{it2.wtrp.herd} $\|\alpha_{g}(r_{h})-r_{gh}\| < \delta$ for all $g, h \in G$; \item \label{it3.wtrp.herd} $(w^{2}-wrw-\delta)_{+} \precsim_{A} z$; \item \label{it4.wtrp.herd} $\|rzr\| > \|z\| (1-\delta)$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} We put $e_{g}=br_{g}b$ for any $g \in G$, and $e=\sum_{g \in G} e_{g}$. For any $g,h\in G$ with $g\neq h$, using \eqref{it1.wtrp.herd} and that $r_{g}r_{h}=0$ at the third step, we have \[ \|e_{g}e_{h}\|=\|br_{g}b^{2}r_{h}b\|\leq \| r_{g}b^{2}r_{h}\|< 2\delta. \] Hence by the choice of $\delta$ there are orthogonal positive contractions $(f_{g})_{g \in G}$ in $B$ such that $\|f_{g}-e_{g}\| < \tfrac{\varepsilon}{4n}$ for any $g \in G$. We set $f=\sum_{g \in G} f_{g}$. Then we have \begin{equation}\label{eq2_wtrp.herd} \|f_{g}-e_{g}\| < \tfrac{\varepsilon}{4}\ \ \text{and}\ \ \|f-e\|<\tfrac{\varepsilon}{4}. \end{equation} In particular, $\|e\|<\|f\|+\tfrac{\varepsilon}{4}<2$. We show that the family $(f_{g})_{g \in G}$ satisfies the conditions of \autoref{w.t.R.p} for the action $\beta \colon G\to \mathrm{Aut}(B)$, that is, we will show that the following hold: \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{it5.wtrp.herd} $\|f_{g}a-af_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{it6.wtrp.herd} $\|\beta_{g}(f_{h})-f_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{it7.wtrp.herd} $(y^{2}-yfy - \varepsilon)_{+} \precsim_{B} x$; \item\label{it8.wtrp.herd} $\|fxf\| > 1-\varepsilon$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} To see \eqref{it5.wtrp.herd}, using \eqref{eq2_wtrp.herd}, \eqref{app.herd}, and \eqref{it1.wtrp.herd}, for any $a\in F$ and any $g\in G$ we get \begin{align*} \|f_{g}a-af_{g}\| &\leq \|f_{g}a-e_{g}a\|+\|e_{g}a-br_{g}ab\|+\|br_{g}ab-bar_{g}b\|\\ &\ \ \ +\|bar_{g}b - ae_{g}\|+ \|ae_{g} -af_{g}\|\\ &<\tfrac{\varepsilon}{4} +2\delta +\delta+ 2\delta+ \tfrac{\varepsilon}{4}<\varepsilon. \end{align*} To prove \eqref{it6.wtrp.herd}, using \eqref{eq2_wtrp.herd} and \eqref{it2.wtrp.herd}, for any $g,h\in G$ we have \begin{align*} \|\beta_{g}(f_{h})-f_{gh}\| &\leq \|\alpha_{g}(f_{h})-\alpha_{g}(e_{h})\| + \|b\alpha_{g}(r_{h})b-br_{gh}b\| +\|e_{gh}- f_{gh}\|\\ & <\tfrac{\varepsilon}{4} + \delta + \tfrac{\varepsilon}{4}<\varepsilon. \end{align*} To see \eqref{it7.wtrp.herd}, first using \eqref{eq1_wtrp.herd} at the second step and using \eqref{app.herd} and \eqref{eq2_wtrp.herd} at the {f}ifth step, we obtain \begin{align*} \left\|(y^{2}-yfy)-(w^{2}-wrw-\delta)_{+} \right\|&\leq \delta+ \|y^{2}-w^{2}\|+\|yfy-wrw\|\\ &< \delta +4 \delta + \|yfy-byeyb\|\\ &\leq 5\delta + \|yfy-byfy\|+\|byfy-byey\|\\ &\ \ \ +\|byey- byeyb\|\\ &< 5\delta + \|y-by\| +\| f-e\| +\|e\|\cdot\|y-yb\|\\ &<5\delta + \delta +\tfrac{\varepsilon}{4} +2\delta=8\delta+ \tfrac{\varepsilon}{4} <\tfrac{\varepsilon}{3}+\tfrac{\varepsilon}{4}<\varepsilon. \end{align*} Now, using \autoref{lemkr} and \eqref{it3.wtrp.herd}, we get \[ (y^{2}-yfy - \varepsilon)_{+} \precsim_{A} (w^{2}-wrw-\delta)_{+} \precsim_{A} z \precsim_{A} x. \] Since $B$ is a hereditary subalgebra of $A$, we get $(y^{2}-yfy - \varepsilon)_{+}\precsim_{B} x$, which is \eqref{it7.wtrp.herd}. To show \eqref{it8.wtrp.herd}, first by \eqref{eq1_wtrp.herd} we have $\|z\|>1-2\delta$, and so by \eqref{it4.wtrp.herd} we get \begin{equation}\label{eq3_wtrp.herd} \|rzr\|>\|z\|(1-\delta)>(1-2\delta)(1-\delta)>1-3\delta. \end{equation} Second, using \eqref{app.herd} at the third step, using \eqref{it1.wtrp.herd} at the forth and fifth steps, and using \eqref{eq3_wtrp.herd} at the seventh step, we obtain \begin{align}\label{eq4_wtrp.herd} \|exe\| &= \|brbxbrb\|\\ \notag &\geq \|brxrb\| -\|br(bxb-x)rb\|\\ \notag &>\|brxrb\| -2\delta\\\notag &>\|rbxrb\|-n\delta -2\delta\\ \notag &>\|rbxbr\|-2n\delta -2\delta\\ \notag &=\|rzr\|-(2n+2)\delta\\ \notag &>1-3\delta-(2n+2)\delta \notag\\ &=1-(2n+5)\delta\geq 1-\tfrac{\varepsilon}{4} \notag. \end{align} Then using \eqref{eq2_wtrp.herd} at the third step and using \eqref{eq4_wtrp.herd} at the forth step, we get \begin{align*} \|fxf\| &\geq \|exe\| - \|exe-exf\|- \|exf-fxf\|\\ &\geq \|exe\| - \|e\|\cdot \|e-f\| -\|e-f\| \\ &>\|exe\| - \tfrac{3\varepsilon}{4} \\ &> 1- \tfrac{\varepsilon}{4} - \tfrac{3\varepsilon}{4}=1-\varepsilon, \end{align*} which is \eqref{it8.wtrp.herd}. This completes the proof of \eqref{it5.wtrp.herd}--\eqref{it8.wtrp.herd}, and shows that $\beta \colon G\to \mathrm{Aut}(B)$ has the weak tracial Rokhlin property. \end{proof} We can reformulate the weak tracial Rokhlin property as a local property as follows. This result is immediate from \autoref{w.t.R.p}. \begin{lma}\label{lemloc} Let $\alpha \colon G\to \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple C*-algebra $A$. Suppose that for every finite set $F \subseteq A$ and every $\varepsilon>0$ there is an $\alpha$-invariant simple C*-subalgebra $B$ of $A$ such that $F \subseteq_{\varepsilon} B$ and the restriction of $\alpha$ to $B$ has the weak tracial Rokhlin property. Then $\alpha \colon G\to \mathrm{Aut}(A)$ has the weak tracial Rokhlin property. \end{lma} The next corollary follows immediately from \autoref{lemloc}. (See Theorem~9.4.34 of \cite{Ph17} for the definition of the direct limit of actions.) \begin{cor}\label{corindlim} Let $G$ be a finite group. Let $\left( (G,A_{i},\alpha^{(i)})_{i\in I}, (\varphi_{j,i})_{i\leq j}\right)$ be a direct system of simple $G$-algebras. Let $A$ be the direct limit of the $A_{i}$ and let $\alpha \colon G\to \mathrm{Aut}(A)$ be the direct limit of the $\alpha^{(i)}$. If $\alpha^{(i)}$ has the weak tracial Rokhlin property for each $i$, then so does $\alpha$. \end{cor} In the next theorem, we consider the minimal tensor product of actions on simple not necessarily unital C*-algebras when one of the actions has the weak tracial Rokhlin property (cf.~\cite[Problem~3.18]{Ph.free}, \cite[Propositions~2.4.6]{Wang13} and \cite[Theorem~3.2(i)]{Sa15}). \begin{thm}\label{proptensor} Let $\alpha \colon G\to \mathrm{Aut}(A)$ and $\beta \colon G\to \mathrm{Aut}(B)$ be actions of a finite group $G$ on simple C*-algebras $A$ and $B$. If $\alpha$ has the weak tracial Rokhlin property then so does $\alpha \otimes \beta \colon G \to \mathrm{Aut}(A \otimes_{\min} B)$. \end{thm} \begin{proof} Suppose that we are given a finite set $F \subseteq A \otimes_{\min} B$, $\varepsilon >0$, and $x, y \in (A\otimes_{\min} B)_{+}$ with $\|x\|=1$. Set $D=A \otimes_{\min} B$. We shall find a family of orthogonal positive contractions $(f_{g})_{g \in G}$ in $A \otimes_{\min} B$ such that, with $f= \sum _{g \in G} f_{g}$, the following hold: \begin{enumerate} \item\label{proptensor_it1} $\|f_{g}a-af_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{proptensor_it2} $\|\alpha_{g}(f_{h})-f_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{proptensor_it3} $(y^{2}-yfy - \varepsilon)_{+} \precsim_{D} x$; \item\label{proptensor_it4} $\|fxf\| > 1-\varepsilon$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} We may assume that there exist $c_{1},\ldots,c_{m}$ in $A$ and $d_{1},\ldots,d_{m}$ in $B$ such that $F=\{ c_{i}\otimes d_{i} \mid 1\leq i \leq m\}$ and that $\|c_{i}\|, \|d_{i}\| \leq 1$ for all $1\leq i \leq m$. By Part~\eqref{rmkdefwtrp_it3} of \autoref{rmkdefwtrp}, we may assume that $y=y_{1} \otimes y_{2}$ for some $y_{1} \in A_{+}$ and $y_{2} \in B_{+}$ with $\| y_{1} \|, \| y_{2}\| \leq 1$. Choose $\delta$ such that $0 < \delta < \tfrac{\varepsilon}{3}$ and $\left((1-\delta)^{2}-4 \delta-\delta^{2}\right) > 1-\varepsilon$. There exists $\delta_{1}$ such that $\tfrac{1}{2} < \delta_{1} < 1$ and $\left((1-\delta)^{2}-4 \delta-\delta^{2}\right)\delta_{1}^{2} > 1-\varepsilon$. Put $z=(x-\delta _{1})_{+}$. It follows from Kirchberg's Slice Lemma (\cite[Lemma~4.1.9]{Ro02}) that there are elements $x_{1} \in A_{+}$ and $x_{2} \in B_{+}$ such that $x_{1} \otimes x_{2} \precsim_{D} z$ and that $\|x_{1}\|=\|x_{2}\|=1$. By \autoref{lembdd1}, there exists $ w \in A \otimes_{\min} B$ such that $\|wxw^{*}-x_{1} \otimes x_{2}\| <\delta^{2}$ and $\|w\| \leq \delta_{1} ^{-{1}/{2}} < \delta_{1}^{-1}$. Thus there is $ v \in A \otimes_{\min} B$ where $v= \sum _{i=1} ^{k} v_{i} \otimes w_{i}$ for some $ v_{i} \in A$ and $w_{i} \in B$, $i=1,\ldots,k$, such that \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{proptensor_it5} $\|vxv^{*}-x_{1}\otimes x_{2}\| < \delta^{2}$ and $\|v\| < \delta_{1}^{-1}$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} Put $E= \{c_{i} \mid 1 \leq i \leq m \} \cup \{v_{i} \mid 1 \leq i \leq k\} $. By \cite[Proposition~2.7(v)]{KR00}, there is $ n \in \mathbb{N}$ such that \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{proptensor_it6} $(y^{2}_{2}-\delta)_{+} \precsim_{B} x_{2} \otimes 1_{n}$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} By \autoref{outer}, $\alpha$ is pointwise outer, and so $A$ is not elementary. It follows from \cite[Corollary~IV.1.2.6]{Bl06} that $A$ is not of Type~I. Now, \cite[Lemma~2.4]{Ph14} implies that there is a nonzero element $x_{0} \in A_{+}$ such that \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{proptensor_it7} $x_{0} \otimes 1_{n} \precsim_{A} x_{1}$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} Put $M=1+ \sum _{i=1} ^{k} \|v_{i}\| + \sum_{i=1} ^{k} \|w_{i}\|$. Choose $\eta >0$ such that $\eta < \tfrac{\delta}{2Mk {\mathrm{card}}(G)}$. Applying \autoref{lemtrp2} to the action $\alpha$ with $E$ in place of $F$, with $\eta$ in place of $\varepsilon$, with $x_{0}$ in place of $x$, with $y_{1}$ in place of $y$, and with $x_{1}$ in place of $z$, we obtain a family of orthogonal positive contractions $(r_{g})_{g \in G}$ in $A$ such that, with $r=\sum _{g \in G} r_{g}$, the following hold: \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{proptensor_it8} $\|r_{g}c-cr_{g}\| < \eta$ for all $c\in E$ and all $g\in G$; \item\label{proptensor_it9} $\|\alpha_{g}(r_{h})-r_{gh}\| < \eta$ for all $g,h\in G$; \item\label{proptensor_it10} $(y_{1}^{2}-y_{1}ry_{1} - \eta)_{+} \precsim_{A} x_{0}$; \item\label{proptensor_it11} $\|rx_{1}r\| > 1-\eta$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} On the other hand, since $B$ has an approximate identity contained in $B^{\beta}$, we can choose a positive contraction $s \in B^{\beta}$ such that \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{proptensor_it12} $\|y_{2}sy_{2}-y_{2}^{2}\| < \eta$, $\|[s,d_{i}]\| < \eta$ for all $ 1 \leq i \leq m$, $\| [s, w_{j}] \| < \eta$ for all $ 1 \leq j \leq k$, and $ \|sx_{2}s\| > 1- \eta$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} Put $f_{g}= r_{g} \otimes s$ for all $g \in G$, and put $f =\sum _{g \in G} f_{g}$. Then $(f_{g})_{g \in G}$ is a family of mutually orthogonal positive contractions in $A \otimes_{\min} B$. We show that \eqref{proptensor_it1}--\eqref{proptensor_it4} hold. For \eqref{proptensor_it1}, let $1\leq i \leq m$. Then by \eqref{proptensor_it8} and \eqref{proptensor_it12} we have \begin{align*} \|[f_{g}, c_{i}\otimes d_{i}]\| & = \|[r_{g}\otimes s, c_{i}\otimes d_{i}]\|\\ &=\|(r_{g}c_{i}) \otimes (sd_{i})- (c_{i}r_{g}) \otimes (d_{i}s)\|\\ &\leq \|[r_{g},c_{i}] \otimes (sd_{i})\|+ \|(c_{i}r_{g}) [s,d_{i}]\|\\ &<\eta+\eta <2\delta <\varepsilon. \end{align*} Part~\eqref{proptensor_it2} follows from \eqref{proptensor_it9}. To prove \eqref{proptensor_it3}, first using \eqref{proptensor_it12} at the third step we have \begin{align*} \| \left(y^{2}-yfy\right)&-\left(y_{1}^{2}-y_{1}ry_{1}-\eta\right)_{+} \otimes \left(y_{2}^{2}-\delta\right)_{+}\|\\ &\leq \|\left(y_{1}^{2} \otimes y^{2}_{2}\right)-\left(y_{1}ry_{1}\right) \otimes \left(y_{2}sy_{2}\right)- \left(y_{1}^{2}-y_{1}ry_{1}\right) \otimes y_{2}^{2}\|+\eta+ \delta\\ &= \|\left(y_{1}ry_{1}\right) \otimes \left(y_{2}sy_{2}-y_{2}^{2}\right)\|+2\delta\\ &< \delta + 2 \delta < \varepsilon. \end{align*} Then by \autoref{lemkr} at the first step, by \eqref{proptensor_it6} and \eqref{proptensor_it10} at the second step, and by \eqref{proptensor_it7} at the forth step, we get \begin{align*} \left(y^{2}-yfy-\varepsilon\right)_{+} &\precsim_{D} \left(y_{1}^{2}-y_{1}ry_{1}- \eta\right)_{+} \otimes \left(y^{2}_{2}- \delta\right)_{+} \\ & \precsim_{D} x_{0} \otimes (x_{2} \otimes 1_{n}) \\ & \sim_{D} (x_{0} \otimes 1_{n}) \otimes x_{2} \\ & \precsim_{D} x_{1} \otimes x_{2} \precsim_{D} z \precsim_{D} x, \end{align*} which is \eqref{proptensor_it3}. To prove \eqref{proptensor_it4}, first by \eqref{proptensor_it8} and \eqref{proptensor_it12} we have \begin{align*} \|fv-vf\| &= \left\| \sum_{i=1}^{k}(rv_{i}) \otimes (sw_{i})- \sum_{i=1}^{k}(v_{i}r) \otimes (w_{i}s) \right\|\\ &\leq \sum_{i=1}^{k}\| (rv_{i}-v_{i}r) \otimes (sw_{i})\|+ \sum_{i=1}^{k}\|(v_{i}r) \otimes (sw_{i}-w_{i}s) \|\\ &\leq M \sum_{i=1}^{k} \sum_{g\in G}\|r_{g}v_{i}-v_{i}r_{g}\|+ M \sum_{i=1}^{k} \|sw_{i}-w_{i}s\|\\ &< Mk\, {\mathrm{card}}(G) \eta + Mk \eta \leq 2Mk\, {\mathrm{card}}(G) \eta < \delta. \end{align*} Also, by \eqref{proptensor_it11} and \eqref{proptensor_it12} we get \begin{align*} \|f (x_{1} \otimes x_{2}) f\| &= \|(rx_{1}r) \otimes (s x_{2} s)\|= \|rx_{1}r\| \cdot \|s x_{2} s\|\\ &> (1-\delta)(1-\eta) >(1-\delta)^{2}. \end{align*} Then by using these two latter inequalities and \eqref{proptensor_it5} we calculate \begin{align*} (1-\delta)^{2} &< \|f(x_{1} \otimes x_{2})f \| \\ & \leq \|fvxv^{*}f\|+\|f(vxv^{*}-x_{1} \otimes x_{2})f\| \\ & < \|vfxv^{*}f\|+\|(fv-vf)xv^{*}f\|+ \delta ^{2} \\ & \leq \|vfxfv^{*} \|+\|vfx(v^{*}f-fv^{*})\|+\delta \|v\|+ \delta ^{2}\\ & \leq \|v\|^{2} \|fxf\|+ \delta \|v\|+ \delta \|v\|+ \delta ^{2} \\ & \leq \delta _{1} ^{-2} \|fxf\|+2 \delta \delta_{1}^{-1}+ \delta^{2} \\ & < \delta_{1} ^{-2}\|fxf\|+4 \delta+ \delta^{2}. \end{align*} Therefore, by the choice of $\delta_{1}$ we obtain \[ \|fxf\| > \left( (1-\delta)^{2}-4 \delta-\delta^{2}\right)\delta_{1}^{2} > 1-\varepsilon \] which is \eqref{proptensor_it4}. This completes the proof. \end{proof} It follows from \autoref{proptensor} that if $\alpha \colon G \to \mathrm{Aut}(A)$ is an action of a finite group $G$ on a simple C*-algebra $A$ and $\alpha$ has the weak tracial Rokhlin property, then $\alpha \otimes \mathrm{id} \colon G \to \mathrm{Aut}(A \otimes \mathcal{K})$ has the weak tracial Rokhlin property. The following corollary follows from \autoref{proptensor} by taking $B=M_{n}$ and $\beta$ to be the trivial action. \begin{cor}\label{cortensor} Let $\alpha \colon G\to \mathrm{Aut}(A)$ be an action of a finite group $G$ with the weak tracial Rokhlin property on a simple not necessarily unital C*-algebra $A$. Then the induced action of $G$ on $M_{n}(A)$ has the weak tracial Rokhlin property for any $n\in \mathbb{N}$. \end{cor} \section{The tracial Rokhlin property}\label{sec_trp} We begin this section by proving that the weak tracial Rokhlin property for finite group actions on simple, separable, not necessarily unital C*-algebras with tracial rank zero is equivalent to a stronger condition, which we call the tracial Rokhlin property. Then we study this property independently and in particular, we show that it coincides with Phillips's notion of tracial Rokhlin property for finite group actions on simple unital C*-algberas. Similar results were proved in \cite[Lemma~1.8]{Ph15} and \cite[Theorem~2.7]{Wang18} for the unital case. One the main tools applied in the proof of these results is the tracial state space. However, in our nonunital setting, we use techniques from Cuntz subequivalence in the proof of the following theorem instead of using traces. First we recall the following known lemma. Indeed, it follows from \cite[Lemmas~2.5.5 and 2.5.12]{Lin.book}. \begin{lma}\label{lemperturb} For every $\varepsilon>0$ and every $n\in\mathbb{N}$ there exists $\delta=\delta(\varepsilon,n)>0$ with the following property. If $A$ is a C*-algebra and $a_{1},\ldots,a_{n}$ are positive contractions in $A$ such that $\|a_{i}^{2}-a_{i}\|<\delta$ and $\|a_{i}a_{j}\|<\delta$ for all $i,j=1,\ldots,n$ with $i\neq j$, then there are mutually orthogonal projections $p_{1},\ldots,p_{n}$ in $A$ such that $\|a_{i}-p_{i}\|<\varepsilon$ for all $i=1,\ldots,n$. \end{lma} \begin{thm}\label{wtr=tr} Let $A$ be a simple separable not necessarily unital C*-algebra with tracial rank zero, and let $\alpha \colon G \rightarrow {\mathrm{Aut}}(A)$ be an action of a finite group $G$ on $A$. Then $\alpha$ has the weak tracial Rokhlin property if and only if the following holds: for every finite subset $F \subseteq A$, every $\varepsilon > 0$, and all positive elements $x,y \in A$ with $\|x\|=1$, there exists a family of orthogonal projections $(p_{g})_{g \in G}$ in $A$ such that, with $p=\sum_{g \in G} p_{g}$, the following hold: \begin{enumerate} \item\label{it1.wtrp=trp} $\|p_{g}a-ap_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{it2.wtrp=trp} $\|\alpha_{g}(p_{h})-p_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{it3.wtrp=trp} $(y^{2}-ypy - \varepsilon)_{+} \precsim_{A} x$; \item\label{it4.wtrp=trp} $\|pxp\| > 1-\varepsilon$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} \end{thm} \begin{proof} It is clear that the condition of the statement implies that $\alpha$ has the weak tracial Rokhlin property. For the converse, suppose that $\alpha$ has the weak tracial Rokhlin property. Choose $\delta >0$ with $\delta < \min( \varepsilon/8, \varepsilon/2({\mathrm{card}}(G)+1))$. Without lose of generality, we can assume that $x, y \in F$, that $F$ is contained in the closed unit ball of $A$, and that $\alpha_{g}(F)=F$, for any $g \in G$. We may further assume that $y \in A^{\alpha}$ (by \autoref{rmkdefwtrp}\eqref{rmkdefwtrp_it3} and that $A$ has an approximate identity contained in $A^{\alpha}$). We claim that there exists $z \in A_{+}\setminus \{0\}$ such that \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item \label{it5.wtrp=trp} $z\oplus \bigoplus_{g \in G} \alpha_{g}(z) \precsim_{A} x$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} In fact, by~\cite[Lemma 2.1]{Ph14}, there is $ z_{1} \in A_{+} \setminus \{0\}$ such that $z_{1} \otimes 1_{k} \precsim_{A} x$, where $k={\mathrm{card}}(G)+1$. Then by \cite[Lemma 2.4]{Ph14}, there is $z \in A_{+} \setminus \{0\}$ such that $z \precsim_{A} \alpha_{g}(z_{1})$ for all $g \in G$. Hence, $z \oplus \bigoplus_{g \in G} \alpha_{g}(z) \precsim_{A} z_{1} \otimes 1_{k} \precsim_{A} x$. This proves \eqref{it5.wtrp=trp}. Since $A$ has tracial rank zero, we can apply \autoref{main.trk} to find a finite dimensional C*-subalgebra $B$ of $A$ such that, with $q=1_{B}$, the following hold: \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item \label{it6.wtrp=trp} $\|qa -aq\| < \delta$ for all $a \in F$; \item \label{it7.wtrp=trp} $qAq \subseteq _{\delta} B$; \item \label{it8.wtrp=trp} $(y^{2} - qyq-\delta)_{+} \precsim_{A} z$; \item \label{it9.wtrp=trp} $\|qxq\| > 1-\delta$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} Since $\alpha$ has the weak tracial Rokhlin property, \autoref{lemtrp2} together with an argument given in the proof of \autoref{propcs} imply that there are orthogonal positive contractions $(f_{g})_{g \in G}$ in $A_{\infty} \cap A'$ such that, with $f= \sum_{g \in G} f_{g}$, we have \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item \label{it10.wtrp=trp} $(\alpha_{\infty})_{g}(f_{h})=f_{gh}$ for all $g, h \in G$; \item \label{it11.wtrp=trp} $y^{2}-yfy \precsim_{A_{\infty}} z$; \item \label{it12.wtrp=trp} $\|fqxqf\|= \|qxq\| >1-\delta$ (by \eqref{it9.wtrp=trp}). \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} Note that $\|fqxqf\|= \max \{\|f_{g}qxqf_{g}\| \colon g \in G\|\}$. Hence, by \eqref{it12.wtrp=trp}, there exists $g_{0} \in G$ such that \begin{equation}\label{inq1.wtrp=trp} \|f_{g_{0}}qxqf_{g_{0}}\| > 1-\delta. \end{equation} Put $D = A_{\infty} \cap B'$. Note that $A_{\infty} \cap A' \subseteq D$ and that $D$ has real rank zero since the sequence algebra of a real rank zero C*-algebra is again of real rank zero and $B$ is finite dimensional. In particular, the hereditary subalgebra $\overline{qf_{g_{0}}D f_{g_{0}}q}$ of $D$ has real rank zero. Thus there is a projection $r_{1} \in \overline{qf_{g_{0}}Df_{g_{0}}q}$ such that \[ \|r_{1}qf_{g_{0}}-qf_{g_{0}}\|<\delta\ \ \text{and} \ \ \|r_{1}qf_{g_{0}}r_{1}-qf_{g_{0}}\|<\delta. \] Since $r_{1} \leq q$, we get \begin{equation} \label{inq2.wtrp=trp} \|r_{1}f_{g_{0}}-qf_{g_{0}}\|<\delta\ \ \text{and} \ \ \|r_{1}f_{g_{0}}r_{1}-qf_{g_{0}}\|<\delta. \end{equation} Put $r_{g}=\alpha_{g}(r_{1})$ for every $g \in G\setminus\{1\}$, and $r =\sum_{g \in G} r_{g}$. Thus $(r_{g})_{g \in G}$ is a family of projections in $A_{\infty}$. Since $r_{1} \in \overline{f_{g_{0}}Df_{g_{0}}} \subseteq \overline{f_{g_{0}}A_{\infty}f_{g_{0}}}$, we have $r_{g} \in \overline{f_{gg_{0}}A_{\infty}f_{gg_{0}}}$, and so $r_{g}r_{h}=0$ when $g \neq h$; that is, $(r_{g})_{g \in G}$ is a family of orthogonal projections in $A_{\infty}$. We show that \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item \label{it13.wtrp=trp} $\|r_{g}a-ar_{g}\| <\varepsilon/2$ for all $a \in F$ and all $g, h \in G$, \item \label{it14.wtrp=trp} $(\alpha_{\infty})_{g}(r_{g})=r_{gh}$ for all $g, h \in G$, \item \label{it15.wtrp=trp} $(y^{2}-ryr-\varepsilon/2)_{+} \precsim_{A_{\infty}} x$, and \item \label{it16.wtrp=trp} $\|rxr\| >1-\varepsilon/2$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} Observe that~\eqref{it14.wtrp=trp} follows from the definition of $r_{g}$. For \eqref{it13.wtrp=trp}, let $a \in F$. By \eqref{it7.wtrp=trp}, there is $b \in B$ such that $\|qaq-b\| <\delta$. Using this at third step, and \eqref{it6.wtrp=trp} at the forth step, we get \begin{align*} \|r_{1}a-ar_{1}\|&=\|qr_{1}qa-aqr_{1}q\| \\ &\leq \|qr_{1}qa-qr_{1}qaq\|+\|qr_{1}qaq-qaqr_{1}q\| +\|qaqr_{1}q-aqr_{1}q\| \\ & \leq \|qr_{1}\|\cdot\|qa-qaq\|+2\delta+\|qaq-aq\|\cdot\|r_{1}q\| \\ & \leq 4\delta <\varepsilon/2. \end{align*} Thus $\|r_{g}a-ar_{g}\| < \varepsilon/2$, for all $a \in F$ and all $g\in G$. Now, using~\eqref{it14.wtrp=trp} and that $\alpha_{g}(F)=F$, for all $g \in G$, we get~\eqref{it13.wtrp=trp}. To see~\eqref{it15.wtrp=trp}, first by \eqref{inq2.wtrp=trp} at the fifth step (also recall that $y \in A^{\alpha}$ and $f_{g} \in A_{\infty}\cap A'$), we get \begin{align}\label{inq3.wtrp=trp} \notag \Big\|(y^{2}-yrfry)&- \Big[(y^{2}-yfy)+ \sum_{g \in G}f_{gg_{0}}^{1/2}\alpha_{g}((y^{2}-yqy-\delta)_{+})f_{gg_{0}}^{1/2}\Big]\Big\| \\\notag & \leq \Big\|yfy-yrfry-\sum_{g \in G}f_{gg_{0}}^{1/2}\alpha_{g}(y^{2}-yqy)f_{gg_{0}}^{1/2}\Big\| +\delta \\ \notag &\leq \Big\|yfy-\sum_{g \in G}f_{gg_{0}}^{1/2} y^{2}f_{gg_{0}}^{1/2}\Big\| +\Big\|yrfry-\sum_{g \in G}yf_{gg_{0}}^{1/2}\alpha_{g}(q)f_{gg_{0}}^{1/2}y\Big\|+\delta \\ &=\Big\|\sum_{g \in G}yr_{g}f_{gg_{0}}r_{g}y -\sum_{g \in G}y \alpha_{g}(q)f_{gg_{0}}y\Big\| +\delta \\\notag &\leq \|y^{2}\|\cdot\Big \|\sum_{g \in G} \alpha_{g}(r_{1}f_{g_{0}}r_{1})-\alpha_{g}(qf_{g_{0}})\Big\| +\delta \\ &\leq {\mathrm{card}}(G) \delta + \delta=({\mathrm{card}}(G)+1)\delta <\varepsilon/2 \notag. \end{align} On the other hand, $yrfry \leq yry$ and so $y^{2}-yry \leq y^{2}-yrfry$. Then \cite[Lemma~1.7]{Ph14} implies that $(y^{2}-yry-\varepsilon/2)_{+} \precsim_{A_{\infty}} (y^{2}-yrfry-\varepsilon/2)_{+}$. By this at the first step, by \eqref{inq3.wtrp=trp} at the second step, by \eqref{it8.wtrp=trp} and \eqref{it11.wtrp=trp} at the third step, and by \eqref{it5.wtrp=trp} at the fifth step, we get \begin{align*} (y^{2}-yry-\varepsilon/2)_{+}&\precsim_{A_{\infty}} (y^{2}-yrfry-\varepsilon/2)_{+}\\ & \precsim_{A_{\infty}} (y^{2}-yfy)+\sum_{g \in G}f_{gg_{0}}^{1/2}\alpha_{g}((y^{2}-yqy-\delta)_{+})f_{gg_{0}}^{1/2} \\ &\precsim_{A_{\infty}} z+ \sum_{g \in G}f_{gg_{0}}^{1/2} \alpha_{g}(z) f_{gg_{0}}^{1/2} \\ &\precsim_{A_{\infty}} z\oplus \bigoplus_{g \in G} \alpha_{g}(z)\\ & \precsim_{A_{\infty}} x, \end{align*} which is \eqref{it15.wtrp=trp}. To prove \eqref{it16.wtrp=trp}, by \eqref{inq2.wtrp=trp} at the sixth step and by \eqref{inq1.wtrp=trp} at the seventh step, we calculate \begin{align*} \|rxr\|&=\|rx^{1/2}\|^{2}=\|x^{1/2}rx^{1/2}\|\\ & \geq \|x^{1/2}r_{1}x^{1/2}\|=\|r_{1}xr_{1}\| \\ &\geq \|r_{1}f_{g_{0}}r_{1}xr_{1}f_{g_{0}}r_{1}\| \\ & \geq \|f_{g_{0}}qxqf_{g_{0}}\|-2\delta \\ & >1-3\delta > 1-\varepsilon/2. \end{align*} This completes the proof of \eqref{it13.wtrp=trp}--\eqref{it16.wtrp=trp}. For each $g \in G$, let $(r_{g,n})_{n \in {\mathbb{Z}}_{> 0}} \in \ell^{\infty}(A)$ be a representing sequence for $r_{g}$, i.e., $\pi_{\infty}(r_{g,1}, r_{g,2}, \cdots)=r_{g}$. Since $r_{g}$ is a projection, we can assume that $r_{g,n}$ is also a projection for all $g \in G$ and all $n \in {\mathbb{Z}}_{> 0}$. Choose $\eta>0$ according to \autoref{lemperturb} for $\varepsilon/(4{\mathrm{card}}(G))$ in place of $\varepsilon$. It follows from \eqref{it15.wtrp=trp} that there is $v \in A_{\infty}$ such that $\|(y^{2}-yry-\varepsilon/2)_{+}-vxv^{*}\| <\varepsilon /4$, and so \begin{equation}\label{inq4.wtrp=trp} \|y^{2}-yry-vxv^{*}\| < 3\varepsilon/4. \end{equation} Let $(v_{n})_{n \in {\mathbb{Z}}_{> 0}}$ be a representing sequence for $v$. We can choose $n$ large enough such that $\|r_{g,n}r_{h,n}\| <\eta$ for all $g, h \in G$ with $g \neq h$ (since $r_{g}r_{h}=0$), and that the following hold (by \eqref{it13.wtrp=trp}, \eqref{it14.wtrp=trp}, \eqref{it16.wtrp=trp}, and \eqref{inq4.wtrp=trp}): \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item \label{it17.wtrp=trp} $\|ar_{g,n}-r_{g,n}a\| <\varepsilon/2$ for all $g \in G$ and all $a \in F$; \item \label{it18.wtrp=trp} $\|\alpha_{g}(r_{h,n})-r_{gh,n}\| <\varepsilon /2$ for all $g, h \in G$; \item \label{it19.wtrp=trp} $\|y^{2}-yr_{n}y-v_{n}xv_{n}^{*}\|<3\varepsilon/4$, where $r_{n}=\sum_{g \in G} r_{g,n}$; \item \label{it20.wtrp=trp} $\|r_{n}xr_{n}\| > 1-\varepsilon/2$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} By the choice of $\eta$, there are orthogonal projections $(p_{g})_{g \in G}$ in $A$ such that $\|r_{g,n}-p_{g}\| <\varepsilon/(4{\mathrm{card}}(G))$ for all $g \in G$. Since $\|r_{g,n}-p_{g}\| <\varepsilon/4$, \eqref{it17.wtrp=trp} implies \eqref{it1.wtrp=trp}, and \eqref{it18.wtrp=trp} implies \eqref{it2.wtrp=trp}. Put $p=\sum_{g \in G}p_{g}$. Then $\|r_{n}-p\| \leq \sum_{g \in G}\|r_{g,n}-p_{g}\| < \varepsilon/4$. Then by \eqref{it19.wtrp=trp}, we have $\|y^{2}-ypy-v_{n}xv_{n}^{*}\|<3\varepsilon/4+\varepsilon/4=\varepsilon$ and so $(y^{2}-ypy-\varepsilon)_{+} \precsim_{A} x$, which is \eqref{it3.wtrp=trp}. Finally, \eqref{it20.wtrp=trp} implies that $\|pxp\|>1-\varepsilon$, which is \eqref{it4.wtrp=trp}. This finishes proof. \end{proof} By the preceding theorem, in \autoref{w.t.R.p} we can replace orthogonal positive contractions by orthogonal projections if the C*-algebra $A$ is separable and has tracial rank zero. (This is also true if $A$ is a (not necessarily unital) Kirchberg algebra \cite{G19}.) However, this is not true in general, as there are examples of finite group actions with the weak tracial Rokhlin property on stably projectionless C*-algebras (see Section~\ref{subsec_examples}). Inspired by~\autoref{wtr=tr} and \cite[Definition~1.2]{Ph11}, we give the following definition. \begin{df} \label{t.R.p} Let $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple C*-algebra $A$. We say that $\alpha$ has the \emph{tracial Rokhlin property} if for every finite subset $F \subseteq A$, every $\varepsilon > 0$, and all positive elements $x,y \in A$ with $\|x\|=1$, there exists a family of orthogonal projections $(p_{g})_{g \in G}$ in $A$ such that, with $p=\sum_{g \in G} p_{g}$, the following hold: \begin{enumerate} \item\label{t.R.p_it1} $\|p_{g}a-ap_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{t.R.p_it2} $\|\alpha_{g}(p_{h})-p_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{t.R.p_it3} $(y^{2}-ypy - \varepsilon)_{+} \precsim_{A} x$; \item\label{t.R.p_it4} $\|pxp\| > 1-\varepsilon$. \end{enumerate} \end{df} In \cite{Ph11}, Phillips introduced the tracial Rokhlin property for finite group actions on simple separable unital C*-algebras. Note that the only difference between \autoref{t.R.p} and \cite[Definition~1.2]{Ph11} is that in \cite[Definition~1.2]{Ph11} one has $1-p\precsim_{A} x$ in place of Condition~\eqref{t.R.p_it3} of \autoref{t.R.p} (see also \autoref{lemmvn}). However, in the unital case, both definitions are equivalent. \begin{rem}\label{rmkdeftrp} A result similar to \autoref{lemdefwtrp} holds for the tracial Rokhlin property (\autoref{t.R.p}). More precisely, \autoref{lemdefwtrp} holds if we replace ``orthogonal positive contractions" by ``orthogonal projections." In particular, if \autoref{t.R.p} holds for some $y\in A_{+}$ then it also holds for every positive element $z\in\overline{Ay}$. Therefore, if moreover $A$ is unital then $\alpha$ has the tracial Rokhlin property (in the sense of \autoref{t.R.p}) if and only if the condition of \autoref{t.R.p} holds only for $y=1$ (and every $\varepsilon,F,x$ as in that definition). This implies that \autoref{t.R.p} and \cite[Definition~1.2]{Ph11} are equivalent in the unital case. \end{rem} \begin{rem}\label{rmkperm} Many results we proved for the weak tracial Rokhlin property have analogues for the tracial Rokhlin property in which one considers ``orthogonal projections" instead of ``orthogonal positive contractions." More precisely, we mean \autoref{lemdefwtrp}, \autoref{rmkdefwtrp}, \autoref{lemtrp2} (see also \autoref{lemtrp1} below in which one also has $p\in A^{\alpha}$), \autoref{subgrp}, \autoref{propwtrpher} (except that we require $B$ to have an approximate identity (not necessarily increasing) consisting of projections in the fixed point algebra), \autoref{lemloc}, \autoref{corindlim}, and \autoref{cortensor}. Some other results with a different statement are addressed in the sequel. \end{rem} Phillips in \cite[Problem~3.2]{Ph.free} asked whether there is a reasonable formulation of the tracial Rokhlin property for finite group actions on simple unital C*-algebras in terms of the central sequence algebra. We give an answer to this problem in the \emph{not necessarily unital} simple case as follows. (See also \autoref{propcs} which deals with the weak tracial Rokhlin property.) \begin{prop}\label{propcs.tr} Let $\alpha \colon G \rightarrow \mathrm{Aut} (A)$ be an action of a finite group $G$ on a simple separable C*-algebra $A$. Then $\alpha$ has the tracial Rokhlin property if and only if for every $x, y \in A_{+}$ with $\|x\|=1$, there exists a family of orthogonal projections $(p_{g})_{g \in G}$ in $ A _{\infty} \cap A'$ such that, with $p = \sum_{g \in G} p_{g}$, the following hold: \begin{enumerate} \item\label{propcs.tr_it1} $(\alpha_{\infty})_{g}(p_{h})=p_{gh}$ for all $g,h\in G$; \item\label{propcs.tr_it2} $y^{2}-ypy \precsim_{A_{\infty}} x$ in; \item\label{propcs.tr_it3} $\|pxp\|=1$. \end{enumerate} If moreover $A$ is unital, then Condition~\eqref{propcs.tr_it2} can replaced by $1-p\precsim_{A_{\infty}} x$. \end{prop} \begin{proof} The proof is essentially the same as that of \autoref{propcs} except that one needs \autoref{lemperturb} for the backward implication. \end{proof} Phillips asked the following problem for actions on simple unital C*-algebras. \begin{pbm}[\cite{Ph.free}, Problem~3.18]\label{prob_3.18} Let $A$ and $B$ be infinite dimensional simple unital C*-algebras, and let $\alpha \colon G \rightarrow {\mathrm{Aut}}(A)$ be an action with the tracial Rokhlin property and $\beta \colon G \rightarrow {\mathrm{Aut}}(B)$ be an arbitrary action. Does it follow that $\alpha \otimes \beta \colon G \rightarrow {\mathrm{Aut}}(A \otimes_{\min} B)$ has the tracial Rokhlin property? \end{pbm} There are some partial answers to this problem. Lemma~3.9 of \cite{Ph11} is the very special case $B = M_n$ and $\beta$ is inner. If $A$ has tracial rank zero and $B$ has tracial rank at most one, then by \cite[Proposition~3.19]{Ph.free}, $\alpha \otimes \beta$ has the tracial Rokhlin property. Moreover, it follows from~\cite[Proposition~2.4.6]{Wang13} that this problem has an affirmative answer provided that $A \otimes_{\min} B$ has Property~(SP). The following is a more general answer to this problem in the \emph{not necessarily unital} simple case. (See also \autoref{proptensor} which deals with the weak tracial Rokhlin property.) \begin{thm}\label{thmtensor} Let $\alpha \colon G\to \mathrm{Aut}(A)$ and $\beta \colon G\to \mathrm{Aut}(B)$ be actions of a finite group $G$ on simple C*-algebras $A$ and $B$. Let $\alpha$ have the tracial Rokhlin property and let $B^{\beta}$ have an approximate identity (not necessarily increasing) consisting of projections. Then the action $\alpha \otimes \beta \colon G \to \mathrm{Aut}(A \otimes_{\min} B)$ has the tracial Rokhlin property. In particular, \autoref{prob_3.18} has an affirmative answer. \end{thm} \begin{proof} The proof is similar to that of \autoref{proptensor}. \end{proof} The following result gives a criterion for the nonunital tracial Rokhlin property in terms of the unital tracial Rokhlin property in the case that the underlying algebra has ``enough" projections. \begin{prop}\label{proploctrp} Let $\alpha \colon G\to \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple C*-algebra $A$. Suppose that $A$ has an approximate identity (not necessarily increasing) $(p_{i})_{i\in I}$ consisting of projections such that each $p_{i}$ is in $A^{\alpha}$. Then $\alpha$ has the tracial Rokhlin property if and only if the restriction of $\alpha$ to $p_{i}Ap_{i}$ has the tracial Rokhlin property for every $i \in I$. \end{prop} \begin{proof} The ``{if}" part follows from the analogue of \autoref{lemloc} for the tracial Rokhlin property, and the ``{only if}" part follows from the analogue of \autoref{propwtrpher} for the tracial Rokhlin property. \end{proof} In \autoref{t.R.p}, if moreover $A$ has Property~(SP) then we may assume that $x$ is a nonzero projection, thanks to the following lemma. \begin{lma}\label{lemtrp} Let $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple C*-algebra $A$ with Property~(SP). Then $\alpha$ has the tracial Rokhlin property if and only if for every positive element $y \in A$, every $\varepsilon > 0$, every finite subset $F \subseteq A$, and every nonzero projection $q \in A$, there exists a family of orthogonal projections $(p_{g})_{g \in G}$ in $A$ such that, with $p=\sum_{g \in G} p_{g}$, the following hold: \begin{enumerate} \item\label{lemtrp_it1} $\|p_{g}a-ap_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{lemtrp_it2} $\|\alpha_{g}(p_{h})-p_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{lemtrp_it3} $(y^{2}-ypy - \varepsilon)_{+} \precsim_{A} q$; \item\label{lemtrp_it4} $\|pqp\| > 1-\varepsilon$. \end{enumerate} \end{lma} \begin{proof} The forward implication is obvious. For the converse, let $x, y \in A$, $F \subseteq A$, and $\varepsilon>0$ be as in \autoref{t.R.p}. Choose $\delta$ such that $0<\delta<1$ and \[ \left(\dfrac{\delta}{2-\delta}(1-\tfrac{\varepsilon}{2})\right)^{2}>1-\varepsilon. \] Note that this is possible since $\left(1-\frac{\varepsilon}{2}\right)^{2}>1-\varepsilon$. Put $z=(x^{ {1}/{2}}-\delta)_{+}.$ Then $(x^{{1}/{2}}-\delta)_{+}\neq 0$ since $\|x\|=1$ and $\delta<1$. Since $A$ has Property~(SP), there is a nonzero projection $q\in \overline{zAz}$. Applying the condition of the statement with $y$, $F$, and $q$ as given and with $ {\varepsilon}/{2}$ in place of $\varepsilon$, we obtain a family of orthogonal projections $(p_{g})_{g \in G}$ in $A$ such that \eqref{lemtrp_it1} and \eqref{lemtrp_it2} hold with $ {\varepsilon}/{2}$ in place of $\varepsilon$, and we have $(y^{2}-ypy - \varepsilon)_{+} \precsim_{A} q$ and $\|pqp\|>1-\frac{\varepsilon}{2}$. Thus \eqref{t.R.p_it1} and \eqref{t.R.p_it2} in \autoref{t.R.p} hold. As $q\in \overline{zAz}$, we have \[ (y^{2}-ypy - \varepsilon)_{+}\precsim_{A} q \precsim_{A} z =(x^{ {1}/{2}}-\delta)_{+}\precsim_{A} x. \] Therefore, \eqref{t.R.p_it3} in \autoref{t.R.p} holds. The proof of \eqref{t.R.p_it4} in that definition proceeds similar to the last part of the proof of \autoref{lemtrp2}. \end{proof} \section{Examples}\label{subsec_examples} In this section, we present examples of finite group actions with the weak tracial Rokhlin property on some simple nonunital C*-algebras. These examples are mainly based on Theorems~\ref{proptensor} and \ref{thmtensor} and a significant result proved in \cite{AGJP17} (\autoref{ex.t.Z.s} below). We remark that this result from \cite{AGJP17} is used only in the present section. In particular, we give examples of finite group actions with the weak tracial Rokhlin property on the Razak-Jacelon algebra ${\mathcal{W}}$. The Razak-Jacelon algebra ${\mathcal{W}}$ is a simple, separable, nuclear, stably projectionless C*-algebra, which can be regarded as a stably finite analogue of the Cuntz algebra $\mathcal{O} _{2}$ (see \cite{Jac13}). As a special case of \autoref{proptensor} we give the following example. This example uses the product type action of $\mathbb{Z}_{2}$ on UHF~algebras verified in \cite{Ph15}. \begin{prop}\label{example1} Let $A =\bigotimes _{k=1} ^{\infty} {M}_{3}$ be the UHF~algebra of type $3^{\infty}$ and let $B$ be a (not necessarily unital) simple C*-algebra. We define $\alpha \colon \mathbb{Z}_{2} \to \mathrm{Aut}(A)$ by \[ \alpha=\bigotimes_{k=1}^{\infty} \mathrm{Ad} \begin{pmatrix}1&0&0 \cr 0&1&0 \cr 0&0&-1 \end{pmatrix}. \] Then the action $ \alpha\otimes \mathrm{id} : \mathbb{Z}_{2} \rightarrow \mathrm{Aut}(A \otimes B)$ has the weak tracial Rokhlin property. \end{prop} \begin{proof} By \cite{Ph15} (see also \cite[Example~10.3.23 and Remark~10.4.9]{Ph17}), $\alpha$ has the tracial Rokhlin property (but it does not have the Rokhlin property). Hence $\alpha$ has the weak tracial Rokhlin property, and so does $ \alpha\otimes \mathrm{id} : \mathbb{Z}_{2} \rightarrow \mathrm{Aut}(A \otimes B)$, by \autoref{proptensor}. \end{proof} \begin{rm} \autoref{example1} suggests how to construct an example of an action with the weak tracial Rokhlin property on a stably projectionless C*-algebra. To see this, let $B=\mathcal{W}$ and let $\alpha$ be as \autoref{example1}. Then $ M_{3^{\infty}} \otimes {\mathcal{W}} \cong {\mathcal{W}}$ is stably projectionless and $\alpha \otimes {\mathrm{id}}_{{\mathcal{W}}} \colon {\mathbb{Z}}_{2} \to {\mathrm{Aut}}(M_{3^{\infty}} \otimes {\mathcal{W}})$ is an example of an action with the weak tracial Rokhlin property. This action does not have the tracial Rokhlin property since $ M_{3^{\infty}} \otimes {\mathcal{W}}$ does not have any nonzero projections. (We do not know whether this action has the Rokhlin property.) \end{rm} For $m \in {\mathbb{Z}}_{> 0}$, let $S_{m}$ denote the group of all permutations of $\{1,2,\ldots, m\}$. For a C*-algebra $A$, we let $A^{\otimes m}$ denote the minimal tensor product of $m$ copies of $A$. The permutation action $\beta$ of $S_{m}$ on $A^{\otimes m}$ is defined by \[ \beta_{\sigma}(a_{1},\ldots,a_{m})=(a_{\sigma^{-1}(1)}, \ldots , a_{\sigma^{-1}(m)}). \] In~\cite{AGJP17}, the permutation action on the Razak-Jacelon algebra $\mathcal{W}$ and the minimal tensor product of tracially $\mathcal{Z}$-absorbing C*-algebras are investigated. Also, in that paper, a notion of tracial $\mathcal{Z}$-absorption for simple (not necessarily unital) C*-algebras is defined. We quote the following result from \cite{AGJP17}. Note that, by \cite[Corollary~19.3]{GL17}, $\mathcal{W} \cong {\mathcal{W}}^{\otimes m}$. \begin{thm}[\cite{AGJP17}]\label{ex.t.Z.s} Let $A$ be a simple tracially $\mathcal{Z}$-absorbing C*-algebra and let $m \in {\mathbb{Z}}_{> 0}$. Suppose that $A^{\otimes m}$ is finite. Then the permutation action $\beta \colon S_{m} \rightarrow {\mathrm{Aut}} (A)$ has the weak tracial Rokhlin property. In particular, the permutation action of the symmetric group $S_{m}$ on $\mathcal{W} \cong {\mathcal{W}}^{\otimes m}$ has the weak tracial Rokhlin property. \end{thm} It is not clear whether the permutation action of $S_{m}$ on $\mathcal{W} \cong {\mathcal{W}}^{\otimes m}$ has the Rokhlin property. The corresponding action on ${\mathcal{Z}}$ does not have even any higher dimensional Rokhlin property with commuting towers (\cite[Corollary~4.8]{HrsPh1}). Note that for any nontrivial group $G$, there exists a unique (up to conjugacy) action of $G$ with the Rokhlin property on $\mathcal{W}$ (by \cite[Corllary~3.7]{Nawata}). To recall, this action is $\nu^{G}:= \mu^{G} \otimes {\mathrm{id}}$ on $ M_{n^{\infty}}\otimes {\mathcal{W}} \cong {\mathcal{W}}$ where $n={\mathrm{card}}(G)$ and $\mu^{G}$ is an action of $G$ on $M_{n^{\infty}}$ defined by \[ \mu^{G}_{g}= \bigotimes _{n=1}^{\infty} {\mathrm{Ad}} (\lambda_{g}), \] for every $g \in G$, which $\lambda$ is the left regular representation on $G$. Observe that the action $\nu^{G}$ on $\mathcal{W}$ has the weak tracial Rokhlin property, by \autoref{propsan}. In the following, we construct examples of finite group actions with the (weak) tracial Rokhlin property on simple C*-algebras with tracial rank zero. Note that, the weak tracial Rokhlin property is equivalent to the tracial Rokhlin property on simple separable C*-algebras with tracial rank zero (by \autoref{wtr=tr}). \begin{cor}\label{cor_per_tr0} Let $A$ be a nonelementary simple (not necessarily unital) C*-algebra with tracial rank zero and let $m \in {\mathbb{Z}}_{> 0} \setminus \{1\}$. Then the permutation action $\beta \colon S_{m} \rightarrow {\mathrm{Aut}} (A^{\otimes m})$ has the weak tracial Rokhlin property. If moreover $A$ is separable, then this action has the tracial Rokhlin property. \end{cor} \begin{proof} First we claim that $A^{\otimes m}$ is a simple C*-algebra with tracial rank zero. Since $A$ has tracial rank zero, it has real rank zero (by \autoref{corrrsr}). Choose a nonzero projection $p$ in $A$, and put $B=pAp$. By \autoref{corrrsr}, $B$ has also tracial rank zero, and so does $B^{\otimes m}$, by \cite[Theorem~4.8]{HLX04}. (Note that in \cite[Theorem~4.8]{HLX04} the algebras are assumed to be unital.) Since $B^{\otimes m}$ is a corner of $A^{\otimes m}$, \autoref{cormorita} implies that $A^{\otimes m}$ has tracial rank zero, and the claim is proved. By \autoref{corrrsr}, $A^{\otimes m}$ has stable rank one and so it is stably finite. Also, it is proved in \cite{AGJP17} that every nonelementary simple C*-algebra with tracial rank zero is tracially $\mathcal{Z}$-absorbing. Now, \autoref{ex.t.Z.s} yields that the permutation action $\beta \colon S_{m} \rightarrow {\mathrm{Aut}} (A^{\otimes m})$ has the weak tracial Rokhlin property. If moreover $A$ is separable, \autoref{wtr=tr} implies that this action has the tracial Rokhlin property. \end{proof} It is not clear that the action $\beta$ in the preceding corollary does not have the Rokhlin property. It may depend on $A$. For example, it follows from \cite[Corollary~3.10]{Nawata} that if either $K_{0}(A)=\mathbb{Z}$ or $K_{1}(A)=\mathbb{Z}$, then $\beta$ does not have the Rokhlin property. The preceding results can be used to construct examples of actions with the weak tracial Rokhlin property which do not have the Rokhlin property. \begin{eg}\label{example_z} We take $A=\mathcal{Z}$. Then the flip action on $A\otimes A\cong \mathcal{Z}$ has the weak tracial Rokhlin property (by \cite{Hirshberg} or \autoref{ex.t.Z.s}) but it does not have the Rokhlin property. In fact, it does not have even the tracial Rokhlin property in the sense of Phillips (\cite[Definition~1.2]{Ph11}), because $\mathcal{Z}$ has no nontrivial projections. For a nonunital example, we can take $A=\mathcal{Z}\otimes \mathcal{K}$ and consider the flip action on $A\otimes A \cong A$. Then \autoref{ex.t.Z.s} implies that the flip action on $A$ has the weak tracial Rokhlin property. However, $K_{0}(A)=\mathbb{Z}$ and so \cite[Corollary~3.10]{Nawata} yields that the flip action on A does not have the Rokhlin property. We do not know whether this action has the tracial Rokhlin property. \end{eg} \begin{prop}\label{prop_ex_z} Let $A$ be a simple (not necessarily unital) $\mathcal{Z}$-absorbing C*-algebra. Then for every finite nontrivial group $G$ there is an action $\alpha\colon G \to \mathrm{Aut}(A)$ with the weak tracial Rokhlin property. If moreover, $A$ is separable with either $K_{0}(A)=\mathbb{Z}$ or $K_{1}(A)=\mathbb{Z}$, then $\alpha$ does not have the Rokhlin property. \end{prop} \begin{proof} Let $n={\mathrm{card}}(G)$ and let $\beta$ denote the permutation action of $S_{n}$ on $\mathcal{Z} \cong\mathcal{Z}^{\otimes n} $. Then by \autoref{ex.t.Z.s}, $\beta \colon S_{n} \to \mathrm{Aut}(\mathcal{Z}) $ has the weak tracial Rokhlin property. Applying \autoref{proptensor}, we conclude that the action $ \beta \otimes \mathrm{id}$ of $S_{n}$ on $A\cong \mathcal{Z} \otimes A $ has the weak tracial Rokhlin property. Note that $G$ embeds in $S_{n}$ as a subgroup. Let $\alpha\colon G \to \mathrm{Aut}(A)$ be the restriction of the action $ \beta \otimes \mathrm{id}$ to $G$. Then \autoref{subgrp} implies that $\alpha$ has the weak tracial Rokhlin property. If moreover, $A$ is separable with either $K_{0}(A)=\mathbb{Z}$ or $K_{1}(A)=\mathbb{Z}$, then it follows from \cite[Corollary~3.10]{Nawata} that $\alpha$ does not have the Rokhlin property. \end{proof} We give another example of an action on a nonunital C*-algebra which has the tracial Rokhlin property but does not have the Rokhlin property. \begin{eg}\label{exa_blackadar} Let $\alpha$ be Blackadar's action of $\mathbb{Z}_{2}$ on $M_{2^{\infty}}$ (see \cite[Section~5]{Bla90} and \cite[Example~3.1]{Ph15}). Since the crossed product $M_{2^{\infty}} \rtimes _{\alpha} {\mathbb{Z}}_{2}$ is not an AF~algebra, it follows from \cite[Theorem~2.2]{Ph11} that $\alpha$ does not have the Rokhlin property. However, $\alpha$ has the tracial Rokhlin property, by \cite[Proposition~3.4]{Ph15}. Now consider the action $\alpha \otimes {\mathrm{id}} \colon {\mathbb{Z}}_{2} \to M_{2^{\infty}}\otimes \mathcal{K}$. By \autoref{proptensor}, $\alpha \otimes {\mathrm{id}}$ has the weak tracial Rokhlin property. Since $M_{2^{\infty}}\otimes \mathcal{K}$ has tracial rank zero, \autoref{wtr=tr} implies that $\alpha \otimes {\mathrm{id}}$ has the tracial Rokhlin property. We show that $\alpha \otimes {\mathrm{id}}$ does not have the Rokhlin property. Let $p$ be a rank one projection in ${\mathcal{K}}$ and set $B= (1_{A} \otimes p)(A \otimes {\mathcal{K}})(1_{A} \otimes p)\cong A$. Then $B$ is an invariant hereditary C*-subalgebra of $M_{2^{\infty}}\otimes \mathcal{K}$. Let $\beta$ denote the restriction of $\alpha \otimes {\mathrm{id}} $ to $B$. Suppose that $\alpha \otimes {\mathrm{id}}$ has the Rokhlin property. Then by \cite[Theorem~3.2(ii)]{Sa15}, $\beta$ has the Rokhlin property, and so does $\alpha$ (because $\alpha$ and $\beta$ are conjugate), which is a contradiction. Therefore, $\alpha \otimes {\mathrm{id}}$ does not have the Rokhlin property. \end{eg} Let $\alpha$ be as in the preceding example, and let $B$ be any simple stably projectionless C*-algebra. Then $\alpha \otimes {\mathrm{id}} \colon {\mathbb{Z}}_{2} \to M_{2^{\infty}}\otimes B$ has the weak tracial Rokhlin property (by \autoref{proptensor}) but not the tracial Rokhlin property, since $M_{2^{\infty}}\otimes B$ does not have any nonzero projections. \section{Crossed products}\label{sec_cross} The main goal of this section is to show that several classes of simple C*-algebras are closed under taking crossed products and fixed point algebras by actions of finite groups with the (weak) tracial Rokhlin property. In particular, this is true for the class of simple C*-algebras with tracial rank zero. This extends a result of Phillips (\cite[Theorem~2.6]{Ph11}) to the nonunital case and is evidence that our definition of (weak) tracial Rokhlin property on simple not necessarily unital C*-algebras is suitable. \autoref{propcross} below is essential for obtaining our main results in this section. To prove it, we need the following lemma, which is also of interest in its own right. \begin{lma}\label{lemtrp1} Let $\alpha \colon G \rightarrow \mathrm{Aut}(A)$ be an action of a finite group $G$ on a simple C*-algebra $A$. Then $\alpha$ has the tracial Rokhlin property if and only if the following holds. For every $\varepsilon > 0$, every finite subset $F \subseteq A$, and all positive elements $x,y,z \in A$ with $x\neq 0$ and $\|z\|=1$, there exist a family of orthogonal projections $(p_{g})_{g \in G}$ in $A$ such that, with $p=\sum_{g \in G} p_{g}$, we have $p\in A^{\alpha}$ and the following hold: \begin{enumerate} \item\label{lemtrp1_it1} $\|p_{g}a-ap_{g}\| < \varepsilon$ for all $a\in F$ and all $g\in G$; \item\label{lemtrp1_it2} $\|\alpha_{g}(p_{h})-p_{gh}\| < \varepsilon$ for all $g,h\in G$; \item\label{lemtrp1_it3} $(y^{2}-ypy - \varepsilon)_{+} \precsim_{A} x$; \item\label{lemtrp1_it4} $\|pzp\| > 1-\varepsilon$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} \end{lma} \begin{proof} The backward implication is obvious. For the forward implication, suppose that $\alpha$ has the tracial Rokhlin property, and let $\varepsilon, F, x,y,z$ be as in the statement. We may assume that $F$ is contained in the closed unit ball of $A$. Let $n={\mathrm{card}}(G)$ and put $\varepsilon_{0}=\min\{\frac{\varepsilon}{5}, \frac{1}{2}, \frac{\varepsilon}{1+2\|y\|^{2}}\}$. By \cite[Lemmas~2.5.1 and 2.5.4]{Lin.book}, there is $\delta_{0}>0$ satisfying the following property: if $b$ is a self-adjoint element in a C*-algebra $B$ and $e$ is a projection in $B$ such that $\|b-e\|<n\delta_{0}$, then there is a projection $p$ in $C^{*}(b)$ and a unitary $u\in B^{\sim}$ satisfying $ueu^{*}=p$ and $\|u-1\|<\varepsilon_{0}$, where $1$ denotes the unit of $B^{\sim}$. We may assume that $\delta_{0}<\varepsilon_{0}$. Choose $\delta$ with $0<\delta<1$ such that \[\left(\dfrac{\delta}{2-\delta}(1-\tfrac{\delta_{0}}{2})\right)^{2}>1-\delta_{0}. \] Put $z_{1}=(z^{ {1}/{2}}-\delta)_{+}$. Since $A$ is simple, \cite[Lemma~2.6]{Ph14} implies that there is a positive element $d\in \overline{z_{1}Az_{1}}$ such that $d\precsim_{A} x$ and $\|d\|=1$. Applying \autoref{t.R.p} with $y$ and $F$ as given, with $ {\delta_{0}}/{2}$ in place of $\varepsilon$, and with $d$ in place of $x$, we obtain a family of orthogonal projections $(e_{g})_{g \in G}$ in $A$ such that, with $e=\sum_{g \in G} e_{g}$, the following hold: \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{lemtrp1_it5} $\|e_{g}a-ae_{g}\| < \frac{\delta_{0}}{2}$ for all $a\in F$ and all $g\in G$; \item\label{lemtrp1_it6} $\|\alpha_{g}(e_{h})-e_{gh}\| < \frac{\delta_{0}}{2}$ for all $g,h\in G$; \item\label{lemtrp1_it7} $(y^{2}-yey - \frac{\delta_{0}}{2})_{+} \precsim_{A} d$; \item\label{lemtrp1_it8} $\|ede\| > 1-\frac{\delta_{0}}{2}$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} By \eqref{lemtrp1_it8}, similar to the proof of \autoref{lemtrp2} (with $\delta_{0}$ in place of $\varepsilon$, with $z$ and $z_{1}$ as given, and with $e$ in place of $f$) we get \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{lemtrp1_it9} $\|eze\| > 1-\delta_{0}$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} Now we give an argument analogous to the proof of \cite[Lemma~3.17]{Ph11} to obtain the desired $p$ in $A^{\alpha}$. Put $b=\frac{1}{n}\sum_{g\in G}\alpha_{g}(e)$. Then by \eqref{lemtrp1_it6}, $\|b-e\|<n\delta_{0}$. Note that $b\in A^{\alpha}$. By the choice of $\delta_{0}$, there is a projection $p$ in $A^{\alpha}$ and a unitary $u\in A^{\sim}$ such that $ueu^{*}=p$ and $\|u-1\|<\varepsilon_{0}$ where $1$ denotes the unit of $A^{\sim}$. Put $p_{g}=ue_{g}u^{*}$, for all $g\in G$. Thus $(p_{g})_{g \in G}$ is a family of orthogonal projections in $A$ and $p=\sum_{g \in G} p_{g}$. We have \begin{enumerate} \setcounter{enumi}{\value{TmpEnumi}} \item\label{lemtrp1_it10} $\|p_{g}-e_{g}\| <2\varepsilon_{0}$, for all $g\in G$, and $\|p-e\| <2\varepsilon_{0}$. \setcounter{TmpEnumi}{\value{enumi}} \end{enumerate} Now we prove \eqref{lemtrp1_it1}--\eqref{lemtrp1_it4}. For \eqref{lemtrp1_it1}, by \eqref{lemtrp1_it5} and \eqref{lemtrp1_it10} we have \[ \|p_{g}a-ap_{g}\|\leq \|p_{g}a-e_{g}a\|+\|e_{g}a-ae_{g}\|+\|ae_{g}-ap_{g}\|<5\varepsilon_{0}<\varepsilon. \] For \eqref{lemtrp1_it2}, using \eqref{lemtrp1_it6} and \eqref{lemtrp1_it10} we get \[ \|\alpha_{g}(p_{h})-p_{gh}\|\leq \|\alpha_{g}(p_{h})-\alpha_{g}(e_{h})\|+ \|\alpha_{g}(e_{h})-e_{gh}\|+\|e_{gh}-p_{gh}\|<5\varepsilon_{0}<\varepsilon. \] To see \eqref{lemtrp1_it3}, first by \eqref{lemtrp1_it10} we have \[ \big \|(y^{2}-ypy)-(y^{2}-yey - \tfrac{\delta_{0}}{2})_{+}\big\| \leq \tfrac{\delta_{0}}{2}+ \|ypy-yey\|<\varepsilon_{0}+2\varepsilon_{0}\|y\|^{2}\leq\varepsilon. \] Hence by \eqref{lemtrp1_it8} and \autoref{lemkr}, $(y^{2}-ypy - \varepsilon)_{+} \precsim_{A} (y^{2}-yey - \frac{\delta_{0}}{2})_{+} \precsim_{A} q \precsim_{A} x$ in $A$. To prove \eqref{lemtrp1_it4}, by \eqref{lemtrp1_it9} and \eqref{lemtrp1_it10} we can compute \begin{align*} \|pzp\|&=\|eze+(p-e)zp+ez(p-e)\|\\ &\geq \|eze\|-2\|p-e\|\\ &>1-\delta_{0}-4\varepsilon_{0} \geq 1-\varepsilon. \end{align*} This finishes the proof. \end{proof} \begin{prop}\label{propcross} Let $G$ be a finite group and let $\mathcal{C}$ be a class of simple (separable) C*-algebras with the following properties: \begin{enumerate} \item\label{propcross_1} if $A$ is a simple (separable) C*-algebra and $p\in A$ is a nonzero projection, then $A\in\mathcal{C}$ if and only if $pAp\in\mathcal{C}$ (in particular, this is the case if $\mathcal{C}$ is closed under Morita equivalence); \item\label{propcross_2} if $A\in\mathcal{C}$ is unital and $\alpha$ is an action of $G$ on $A$ with the tracial Rokhlin property then $A\rtimes_{\alpha} G\in \mathcal{C}$; \item\label{propcross_3} if $A\in \mathcal{C}$ and $B$ is a C*-algebra with $A \cong B$, then $B\in \mathcal{C}$. \end{enumerate} Then $\mathcal{C}$ is closed under taking crossed products by actions of $G$ with the tracial Rokhlin property (i.e., \eqref{propcross_2} above holds without the assumption that $A$ is unital). \end{prop} \begin{proof} Let $\mathcal{C}$ be a class of simple C*-algebras as in the statement. Let $A\in \mathcal{C}$ and let $\alpha \colon G \to \mathrm{Aut}(A)$ be an action with the tracial Rokhlin property. We show that $A\rtimes_{\alpha} G\in \mathcal{C}$. We may assume that $A$ is nonzero. By \autoref{lemtrp1}, there exists a nonzero projection $p$ in $A^{\alpha}$. Let $\alpha \colon G\to \mathrm{Aut}(pAp)$ be the restriction of $\alpha$ to $pAp$. Now, \autoref{rmkperm} implies that $\beta$ has the tracial Rokhlin property. By Condition~\eqref{propcross_1}, $pAp \in \mathcal{C}$. Thus, by Condition~\eqref{propcross_2}, $pAp \rtimes _{\beta} G \in \mathcal{C}$. Observe that $pAp \rtimes _{\beta} G\cong p(A \rtimes _{\alpha} G)p$. In fact, the map $\varphi: pAp \rtimes _{\beta} G\to p(A \rtimes _{\alpha} G)p$ defined by $\varphi(\sum_{g\in G}b_{g}u_{g})=\sum_{g\in G}b_{g}\delta_{g}$ where $b_{g}\in pAp$ for all $g\in G$, is easily seen to be a surjective $*$-isomorphism. Thus, by Condition~\eqref{propcross_3}, $p(A \rtimes _{\alpha} G)p \in \mathcal{C}$. Now Condition~\eqref{propcross_2} implies that $ A \rtimes _{\alpha} G \in \mathcal{C}$. If we assume that $\mathcal{C}$ is a class of \emph{separable} C*-algebras satisfying Conditions~\eqref{propcross_1} and \eqref{propcross_3}, and satisfying the separable version of Condition~\eqref{propcross_2}, then the argument above again works, since $ A \rtimes _{\alpha} G$ is separable if $A$ is separable and $G$ is finite. \end{proof} In the following theorem, we extend Phillips's result \cite[Theorem~2.6]{Ph11} to the nonunital case. We also consider the fixed point algebra. \begin{thm}\label{thmcross} Let $A$ be a simple separable (not necessarily unital) C*-algebra with tracial rank zero and let $\alpha$ be an action of a finite group $G$ on $A$ with the weak tracial Rokhlin property. Then the crossed product $A \rtimes _{\alpha} G$ and the fixed point algebra $A^{\alpha}$ are simple C*-algebras with tracial rank zero. \end{thm} \begin{proof} It follows from \autoref{wtr=tr} that $\alpha$ in fact has the tracial Rokhlin property. Let $\mathcal{C}$ denote the class of simple C*-algebras with tracial rank zero. By \autoref{cormorita}, $\mathcal{C}$ satisfies Condition~\eqref{propcross_1} in \autoref{propcross}. Also, by \cite[Theorem~2.6]{Ph11}, $\mathcal{C}$ satisfies Condition~\eqref{propcross_2} in \autoref{propcross}. (Note that the assumption of separability is unnecessary in \cite[Theorem~2.6]{Ph11}.) Clearly, $\mathcal{C}$ satisfies Condition~\eqref{propcross_3} in \autoref{propcross}. Thus \autoref{propcross} yields the first part of the statement about the crossed product. The second part of the statement about the fixed point algebra follows from \autoref{Moritaweak} which says that $A^{\alpha}$ is Morita equivalent to $A \rtimes _{\alpha} G$, and \autoref{cormorita} which implies that the tracial rank zero is preserved under Morita equivalence in the class of simple C*-algebras. \end{proof} We remark that if the action $\alpha$ in the preceding theorem has the tracial Rokhlin property (in the sense of \autoref{t.R.p}), then the assumption that $A$ is separable is unnecessary. The following corollary is immediate from \autoref{cor_per_tr0} and \autoref{thmcross}. \begin{cor}\label{cor_corss_tr0} Let $A$ be a simple separable (not necessarily unital) C*-algebra with tracial rank zero and let $\beta \colon S_{m} \rightarrow A^{\otimes m}$ be the permutation action defined before \autoref{ex.t.Z.s}, where $m\geq 2$. Then the crossed product $A^{\otimes m} \rtimes _{\beta} S_{m}$ is a simple C*-algebra with the tracial rank zero. \end{cor} We need the following two lemmas in the proof of \autoref{thm_cross_wtrp}. In the first lemma we extend a special case of \cite[Theorem~4.2]{JO98} to the nonunital case. \begin{lma} \label{proj.crossed} Let $A$ be a simple C*-algebra with Property~(SP) and let $\alpha$ be a pointwise outer action of a discrete group $G$ on $A$. Then every nonzero hereditary C*-subalgebra of the reduced crossed product $A \rtimes _{\alpha, \mathrm{r}} G$ has a nonzero projection which is Murray-von~Neumann equivalent (in $A \rtimes _{\alpha, \mathrm{r}} G$) to a projection in $A$. In particular, $A \rtimes _{\alpha, \mathrm{r}} G$ has Property~(SP). \end{lma} \begin{proof} Assume that $B$ is a hereditary C*-subalgebra of $A \rtimes _{\alpha, \mathrm{r}} G$. Since $A$ is simple, by Proposition~1.4 of \cite{Pas}, the action $\alpha$ is pointwise spectrally nontrivial. Thus, Lemma~3.2 and Proposition~3.9 of \cite{Pas} imply that there exist a nonzero hereditary C*-subalgebra $E$ of $A$ and an injective $*$-homomorphism $\phi \colon E \rightarrow B$ such that $\phi(x) \sim x$ in $A \rtimes _{\alpha, \mathrm{r}} G$, for all $x \in E_{+}$. Since $A$ has Property~(SP), we can choose a nonzero projection $p$ in $E$. Hence, $\phi(p)$ is a projection in $B$ and $ p \precsim_{A} \phi(p)$. Thus there is a projection $q\in A \rtimes _{\alpha, \mathrm{r}} G$ such that $q \sim_{\mathrm{MvN}} p$ and $ q \leq \phi(p)$. Since $ \phi(p) \in B$ and $B$ is a hereditary C*-subalgebra of $A \rtimes _{\alpha, \mathrm{r}} G$, we have $q \in B$. This completes the proof. \end{proof} Alternatively, the statement follows from \cite[Theorem~4.2]{JO98} by taking $N=\{1\}$. Observe that the assumption that $A$ is unital is not used in the proof of \cite[Theorem~4.2]{JO98}. Recall from \cite[Definition~4.1]{KR00} that a C*-algebra $A$ is called \emph{purely infinite} if there are no characters on $A$ and if for every pair of positive elements $a, b$ in $A$ such that $a\in \overline{AbA}$ we have $a\precsim_{A} b$. In particular, if $A$ is a nonzero simple C*-algebra then $A$ is purely infinite if and only if $A\ncong \mathbb{C}$ and $a\sim_{A} b$ for any $a,b\in A_{+}\setminus\{0\}$. The following lemma is known in the unital case \cite[Corollary~4.4]{JO98}. This nonunital version seems to be new. \begin{lma}\label{lem_pi} Let $\alpha \colon G \to \mathrm{Aut}(A)$ be an action of a discrete group $G$ on a (not necessarily unital) C*-algebra $A$. Let $x$ be a nonzero positive element in the reduced crossed product $A \rtimes _{\alpha, \mathrm{r}} G$. Set $D=A \rtimes _{\alpha, \mathrm{r}} G$. \begin{enumerate} \item\label{lem_pi_it1} For every $\varepsilon >0$ there is $b\in A_{+}\setminus \{0\}$ such that $(x-\varepsilon)_{+}\precsim_{D} b$. \item\label{lem_pi_it2} If $\alpha$ is pointwise outer and $A$ is simple, then there exists $a\in A_{+}\setminus \{0\}$ such that $a\precsim_{D} x$. \end{enumerate} In particular, if $\alpha$ is pointwise outer and $A$ is simple and purely infinite, then the reduced crossed product $A \rtimes _{\alpha, \mathrm{r}} G$ is also simple and purely infinite. \end{lma} It is not clear whether the assumption of pointwise outerness is really needed in the preceding lemma. When $A$ is a simple unital purely infinite C*-algebra and $G$ is finite, then---without any assumption on $\alpha$---the crossed product $A \rtimes _{\alpha} G$ is a finite direct sum of simple unital purely infinite C*-algebras, by \cite[Theorem~4.4]{JO98}. \begin{proof}[Proof of \autoref{lem_pi}] To prove \eqref{lem_pi_it1}, first we choose some nonzero positive element $c=\sum_{g\in G}c_{g}u_{g}$ in $C_{c}(G,A)$ such that $\|x-c\|<\varepsilon$. Since $c=c^{*}$, we can write $2c=\sum_{g\in G}\left(c_{g}u_{g}+(c_{g}u_{g})^{*}\right)$. For every $g\in G$, choose $a_{g},b_{g}\in A$ such that $c_{g}=a_{g}b_{g}$. Then we have \begin{align*} c_{g}u_{g}+(c_{g}u_{g})^{*} &=a_{g}(b_{g}u_{g})+(b_{g}u_{g})^{*}a_{g}^{*}\\ &\leq a_{g}a_{g}^{*} + (b_{g}u_{g})^{*}b_{g}u_{g}\\ &=a_{g}a_{g}^{*}+ \alpha_{g^{-1}}( b_{g}^{*}b_{g}), \end{align*} where the inequality follows from the fact that for all elements $y,z\in A$ we have $zy+y^{*}z^{*}\leq zz^{*}+y^{*}y$. Put $b=\tfrac{1}{2}\sum_{g\in G}\left(a_{g}a_{g}^{*}+ \alpha_{g^{-1}}( b_{g}^{*}b_{g})\right)$. Then the inequality above implies that $c\leq b$. Thus $b$ is a nonzero positive element in $A$, and applying \autoref{lemkr} we obtain $(x-\varepsilon)_{+}\precsim_{D} c \precsim_{D} b$. Part~\eqref{lem_pi_it2} follows from \cite[Lemma~5.1]{Hirshberg} which its proof also works in the nonunital case. To see the last part of the statement, let $\alpha$ be pointwise outer and let $A$ be simple and purely infinite. By \cite[Theorem~3.1]{Kishimoto}, $A \rtimes _{\alpha, \mathrm{r}} G$ is simple. Let $x, y$ be nonzero positive elements in $A \rtimes _{\alpha, \mathrm{r}} G$ and let $\varepsilon>0$. Then by \eqref{lem_pi_it1} and \eqref{lem_pi_it2} there are $a,b \in A_{+}\setminus\{0\}$ such that $(x-\varepsilon)_{+}\precsim_{D} b$ and $a\precsim_{D} y$. Since $A$ is simple and purely infinite, we have $a\sim_{A} b$, and so $(x-\varepsilon)_{+}\precsim_{D} b \precsim_{D} a \precsim_{D} y$. Letting $\varepsilon \to 0$, we get $x\precsim_{D} y$. Similarly, we get $y\precsim_{D} x$, and so $x\sim_{D} y$. This implies that $A \rtimes _{\alpha, \mathrm{r}} G$ is purely infinite. \end{proof} In the following theorem we list some properties preserved under taking crossed products by finite group actions with the weak tracial Rokhlin property. We do not know whether some other properties such as having stable rank one or real rank zero are also preserved (see also \autoref{thmclass1} below). In \cite{AGJP17}, a notion of tracial $\mathcal{Z}$-absorption for simple (not necessarily unital) C*-algebras is defined. Part~\eqref{thm_cross_wtrp_it4} in the following theorem is proved in \cite{AGJP17}. \begin{thm}\label{thm_cross_wtrp} The following classes of not necessarily unital C*-algebras are preserved under taking crossed products and fixed point algebras by finite group actions with the weak tracial Rokhlin property: \begin{enumerate} \item\label{thm_cross_wtrp_it1} simple C*-algebras; \item\label{thm_cross_wtrp_it2} simple separable C*-algebras of tracial rank zero; \item\label{thm_cross_wtrp_it3} simple purely infinite C*-algebras; \item\label{thm_cross_wtrp_it4} (\cite{AGJP17, Hirshberg}) simple tracially $\mathcal{Z}$-absorbing C*-algebras; \item\label{thm_cross_wtrp_it5} simple C*-algebras with Property~(SP). \end{enumerate} \end{thm} \begin{proof} Part~\eqref{thm_cross_wtrp_it1} follows from Corollaries~\ref{simpleweak} and \ref{Moritaweak}. Part~\eqref{thm_cross_wtrp_it2} is the conclusion of \autoref{thmcross}. By \autoref{outer}, the weak tracial Rokhlin property implies pointwise outerness. Also, the fixed point algebra is isomorphic to a hereditary subalgebra of the crossed product \cite{Rosenberg}. Then \autoref{lem_pi} implies Part~\eqref{thm_cross_wtrp_it3}, and \autoref{proj.crossed} implies Part~\eqref{thm_cross_wtrp_it5}. Finally, Part~\eqref{thm_cross_wtrp_it4} is proved in \cite{Hirshberg} for the unital case, and in \cite{AGJP17} for the nonunital case. \end{proof} The existence of projections in \autoref{t.R.p} enables us to prove that crossed products by actions with the tracial Rokhlin property preserve more structural properties of C*-algebras: \begin{thm}\label{thmclass1} The following classes of not necessarily unital C*-algebras are preserved under taking crossed products and fixed point algebras by finite group actions with the tracial Rokhlin property: \begin{enumerate} \item\label{thmclass1_it1} simple C*-algebras; \item\label{thmclass1_it2} simple C*-algebras with tracial rank zero; \item\label{thmclass1_it3} simple purely infinite C*-algebras; \item\label{thmclass1_it4} simple tracially $\mathcal{Z}$-absorbing C*-algebras; \item\label{thmclass1_it5} simple C*-algebras with Property~(SP); \item\label{thmclass1_it9} simple separable nuclear $\mathcal{Z}$-absorbing C*-algebras. \end{enumerate} \end{thm} \begin{proof} Parts~\eqref{thmclass1_it1}--\eqref{thmclass1_it5} follow from \autoref{thm_cross_wtrp}, since the tracial Rokhlin property implies the weak tracial Rokhlin property. Note that, by the remark following \autoref{thmcross}, in Part~\eqref{thmclass1_it2} we do not need to assume the separability. To prove \eqref{thmclass1_it9}, let $\mathcal{C}$ denote the class of simple separable nuclear $\mathcal{Z}$-absorbing C*-algebras. By \cite[Corollary~3.2]{TW07}, $\mathcal{Z}$-stability is preserved under Morita equivalence in the class of separable C*-algebras. Moreover, by \cite[Theorem~3.15]{HRW07}, nuclearity is preserved under Morita equivalence. Thus the class $\mathcal{C}$ satisfies \eqref{propcross_1} in \autoref{propcross}. On the other hand, \cite[Corollary~5.7]{Hirshberg} implies that the class $\mathcal{C}$ also satisfies \eqref{propcross_2} in \autoref{propcross}. Therefore, by \autoref{propcross} the class $\mathcal{C}$ is preserved under taking crossed products by finite group actions of with the tracial Rokhlin property. The corresponding result about $A^{\alpha}$ holds by noting that $A\rtimes _{\alpha} G$ and $A^{\alpha}$ are Morita equivalent (and separable). \end{proof}

\section{Introduction} In the Bayesian setting, when the full likelihood is too complex and difficult to specify or when robustness with respect to data or to model misspecifications is required, several authors have proposed the use of surrogate likelihoods in the Bayes formula, in place of the full likelihood. Although this approach is non-orthodox, it is widely used in the recent statistical literature and theoretically motivated in several papers. See, among others, Pauli {\em et al.} (2011) and Ribatet {\em et al.} (2012) for the use of composite likelihoods, Lazar (2003), Lin (2006), Schennach (2005), Greco {\em et al.} (2008), Chang and Mukerjee (2008), Ventura {\em et al.} (2010), and Yang and He (2012) for the use of empirical and quasi-likelihoods; see also the review by Ventura and Racugno (2016), which addresses the use of pseudo-likelihoods in the Bayesian framework. To deal with complex models or model misspecifications, useful surrogate likelihoods can be obtained trough proper scoring rules. A scoring rule (see, for instance, the recent overviews by Machete, 2013, and Dawid and Musio, 2014, and references therein) is a special kind of loss function designed to measure the quality of a probability distribution for a random variable, given its observed value. It is proper if it encourages honesty in the probability evaluation. Proper scoring rules supply unbiased estimating equations for any statistical model, which can be chosen to increase robustness or for ease of computation. The Brier score (Brier, 1950), the logarithmic score (Good, 1952), the Tsallis score (Tsallis, 1988), and the Hyv\"arinen score (Hyv\"arinen, 2005) are widely known. In particular, when using the logarithmic score, the full likelihood and the composite likelihood (Varin {\em et al.}, 2011) are obtained as special cases of proper scoring rules (see for instance Dawid and Musio, 2014). While frequentist scoring rule inference has been widely discussed (see Dawid {\em et al.}, 2016, and references therein), Bayesian inference is relatively unexplored. Few exceptions are Dawid and Musio (2015) who focus on Bayesian model selection, Ghosh and Basu (2016) who consider robust Bayes estimation using the density power divergence measure, and Musio {\em et al.} (2017) who give an illustration of Bayesian inference for directional data through the Hyv\"arinen score; see also Pauli {\em et al.} (2011) and Ribatet {\em et al.} (2012) for the use of composite likelihoods in the Bayes formula. To perform Bayesian inference, a suitable prior distribution on the parameter of interest must be elicited. In this paper we focus on default priors which are frequently used in Bayesian applications and which are still an active area of research (see, among others, Berger, 2006, Berger {\em et al.}, 2009, 2012, Ghosh, 2011, Walker, 2016, Leisen {\em et al.}, 2017). The two most common objective priors are the Jeffreys prior (Jeffreys, 1961), which uses the information about the parameter contained in the Fisher information, and the reference prior (Bernardo, 1979), which is based on the maximization of a distance in information between the prior and the posterior. The goal of this paper is twofold. The first aim is to discuss the use of scoring rules in order to compute a posterior distribution, called here the SR-posterior distribution. In particular, we suggest a SR-posterior distribution obtained by extending to the general scoring rule setting the curvature adjustment of the composite likelihood proposed by Chandler and Bate (2007) and Ribatet {\em et al.} (2012). The calibration is necessary in order to show that the SR-posterior distribution is, up to order $O_p(n^{-1/2})$, normally distributed, with the same asymptotic variance of the scoring rule estimator. Indeed, a calibrated scoring rule is needed to reach the right asymptotic variance in the normal approximation, as well as a correct shape of the posterior distribution. The second aim is to propose the elicitation of a default prior for the unknown parameter of interest. In particular, we focus on reference priors as pioneered by Bernardo (1979); for a review see Bernardo (2005) and Ghosh (2011). Our purpose is to construct reference priors obtained by maximizing $\alpha-$divergences from the SR-posterior distribution. The $\alpha-$divergences are a well known class of discrepancy functions which include as a special case the Kullback-Leibler divergence. We show that, for $0 \leq |\alpha| <1$, the maximizer is a Jeffreys-type prior that is proportional to the square root of the determinant of the Godambe information matrix. The paper unfolds as follows. Section 2 gives background on scoring rules. Section 3 discusses the properties of the SR-posterior distribution. Section 4 presents the construction of reference priors based on $\alpha-$divergences using SR-posteriors. Examples are presented in Section 5 both in the context of complex models and in the context of robustness theory. Further suggestions and comments can be found in the conclusions. \section{Scoring rules} A scoring rule is a loss function designed to measure the quality of a given probability distribution $Q$ for a random variable $X$, in view of the result $x$ of $X$; see Dawid (1986). The function $S(x;Q)$ takes values in $\mathbb{R}$ and its expected value under $P$ will be denoted by $S(P;Q)$. The scoring rule $S$ is called proper relative to the class of distributions $\mathcal{P}$ if the following inequality is satisfied for all $P,\,Q\in\mathcal{P}$: \begin{equation} \label{eq:1} S(P;Q)\geq S(P;P). \end{equation} It is strictly proper relative to $\mathcal{P}$ if equation (\ref{eq:1}) is satisfied with equality if and only if $Q = P$. Note that in the following we identify a distribution $Q$ by its probability density $q$ with respect some measure $\mu$; so the two notations $S(x;q)$ and $S(x;Q)$ are indistinguishable. An important example of proper scoring rules is the logarithmic score, which is defined as $S(x;Q)=-\log{q(x)}$ (Good, 1952) and corresponds to minus the log-likelihood function. Other widely used proper scoring rules are the Tsallis and the Hyv\"arinen scores. The Tsallis score (Tsallis, 1988, Basu {\em et al.}, 1998) is given by \begin{equation} \label{} S(x;Q) = (\gamma - 1) \int\! q(y)^\gamma \, d\mu(y) - \gamma q(x)^{\gamma-1}, \quad \gamma>1 . \end{equation} The Tsallis score gives in general robust procedures (Basu {\em et al.}, 1998, Dawid {\em et al.}, 2016), and $\gamma$ is a tuning constant that governs the trade-off between efficiency and robustness. The Hyv\"arinen score in its original formulation (Hyv\"arinen, 2005) with $x\in\mathbb{R}^{d}$ is defined as \begin{equation} \label{hiv} S(x,Q) = \varDelta \log{ q(x)} + \frac{1}{2}|| \nabla \log{ q(x)}||^2, \end{equation} where $||\cdot||$ is the standard norm, $\nabla$ denotes the gradient vector and $\varDelta$ the Laplacian operator. \subsection{Inference based on scoring rules} Let $x = (x_1,\ldots , x_n)$ be the realization of a random sample of size $n$ from a distribution with density function $p(x|\theta)$, with $\theta \in \Theta \subseteq \mathbb{R}^{d}$, ${d} \geq 1$. Moreover, let $L(\theta) = L(\theta;x)$ and $\ell(\theta)=\log L(\theta;x)$ be the likelihood and the log-likelihood functions, respectively. The validity of inference about $\theta$ using scoring rules can be justified by invoking the general theory of unbiased $M$-estimating functions. Indeed, inference based on proper scoring rules is a special kind of $M$-estimation (see, e.g., Dawid {\em et al.}, 2016, and references therein). The class of $M$-estimators is broad and includes a variety of well-known estimators. For example it includes the maximum likelihood estimator (MLE), the maximum composite likelihood estimator (see e.g.\ Varin {\em et al.}, 2011), and robust estimators (see e.g.\ Hampel {\em et al.}, 1986, Huber and Ronchetti, 2009), among many others. Given a proper scoring rule $S(x;P_\theta)=S(x;\theta)$, let us denote by $S(\theta)=\sum_{i=1}^n S(x_i;\theta)$ the total empirical score. Moreover, let $s(x;\theta)$ be the gradient vector of $S(x;\theta)$ with respect to $\theta$, i.e.\ $s(x;\theta)=\partial S(x;\theta)/\partial \theta$. Under broad regularity conditions (see Mameli and Ventura, 2015, and references therein), the scoring rule estimator $\tilde\theta$ is the solution of the unbiased estimating equation $$ S_{\theta}(\theta) = \sum_{i=1}^n s(x_i;\theta) = 0\,, $$ (see Dawid and Lautitzen, 2005, Dawid, 2007, Dawid {\em et al.}, 2016) and it is asymptotically normal, with mean $\theta$ and covariance matrix $V(\theta)/n$, with \begin{eqnarray} V (\theta) = K(\theta)^{-1} J(\theta) (K(\theta)^{-1}), \label{var} \end{eqnarray} where $K(\theta) = E_\theta (\partial s(X;\theta)/\partial \theta^{T})$ and $J(\theta) = E_\theta(s(X;\theta) s(X;\theta)^{T} )$ are the sensitivity and the variability matrices, respectively. The matrix $G(\theta)=V(\theta)^{-1}$ is known as the Godambe information matrix (see Godambe, 1960) and its form is due to the failure of the second Bartlett identity (also called information identity) since, in general, $K(\theta) \neq J(\theta)$. In the special case of the logarithmic score, we have that $G(\theta)=K(\theta)=J(\theta)$ is the Fisher information matrix. Asymptotic inference for scoring rule is usually based on first-order approximations, but several examples have shown the inaccuracy of these methods; see for instance Dawid et al. (2016) and Mameli and Ventura (2015). In order to improve the accuracy of first order methods, refinements have been obtained by resorting to higher-order asymptotic theory, which encompass the classical results for likelihood quantities while allowing for the failing of the information identity; see Mameli and Ventura (2015) and Mameli {\em et al.}\ (2017). \section{Posterior distributions based on scoring rules} The use of surrogate likelihoods in the Bayes formula has received great attention in the last decade (see the review by Ventura and Racugno, 2016, and references therein). In particular, the use of the composite likelihood, in the Bayes formula has been considered for instance by Smith and Stephenson (2009), Pauli {\em et al.}\ (2011) and Ribatet {\em et al.}\ (2012). Since the composite likelihood does not satisfy the information identity, it is necessary to calibrate it in order to reach the right asymptotic variance in the normal approximation, as well as a correct shape of the posterior distribution. The correct curvature of the posterior distribution based on composite likelihoods can also be reached by using the composite score function in the Approximate Bayesian Computation (ABC) procedure, as discussed in Ruli {\em et al.}\ (2016). While the use of surrogate likelihoods for Bayesian inference has been widely discussed in the recent literature to deal with complex models or robustness, the use of scoring rules in the Bayesian framework has not been deeply explored; exceptions are Dawid and Musio (2015), Ghosh and Basu (2016) and Musio {\em et al.}\ (2017). Paralleling the derivation of posterior distributions based on composite likelihoods, a SR-posterior distribution can be obtained by using a scoring rule instead of the full likelihood in Bayes formula. However, since the scoring rule does not satisfy the information identity, it must be suitably calibrated before deriving the SR-posterior distribution. In particular, here we suggest a SR-posterior distribution obtained by extending to the scoring rule setting the curvature adjustment of the composite likelihood proposed by Chandler and Bate (2007) and Ribatet {\em et al.}\ (2012). In the following we study the asymptotic normality of the SR-posterior distribution; this result validates the use of scoring rule in a Bayes formula (see Ventura and Racugno, 2016, Section 3.3). Let $\pi(\theta)$ be a prior distribution for the parameter $\theta$. The proposed SR-posterior distribution is defined as \begin{equation} \label{star} \pi_{SR}(\theta|x) \propto \pi(\theta) \exp{\{-S (\theta^*)\}}, \end{equation} with $\theta^*= \theta^*(\theta)=\tilde{\theta}+C(\theta-\tilde{\theta})$, where $C$ is a $d\times d$ fixed matrix such that $C^T K(\theta)C=G(\theta)$. A possible choice of the matrix $C$ is given by $C = M^{-1}M_A$, with $M_A^{T}M_A = G$ and $M^{T}M = K$; for details, see Ribatet {\em et al.}\ (2012) and references therein. Note that in formula \eqref{star}, when $S(\theta)=-\sum_{i=1}^n\log p(x_i|\theta)$ is the logarithmic scoring rule, the posterior $\pi_{SR}(\theta|x)$ reduces to the classical posterior distribution based on the likelihood function, i.e.\ $\pi(\theta|x)\propto \pi(\theta)\exp{\{\sum_{i=1}^n\log p(x_i|\theta)\}}$. The choice of a prior distribution $\pi(\theta)$ to be used with a scoring rule involves the same problems typical of the standard Bayesian perspective, namely the elicitation of a proper prior distribution and the fact that default prior distributions can often be improper. The study of the possible impropriety of the SR-posterior has to be considered case-by-case, similarly to what happens for the genuine posterior distribution based on a default prior. Under the same regularity conditions necessary for the asymptotic results for scoring rule inference (see for instance Mameli and Ventura, 2015) as $n \rightarrow \infty$, it can be shown that the scoring rule posterior distribution (\ref{star}) is, up to order $O_p(n^{-1/2})$, normally distributed with mean $\tilde{\theta}$ and variance $H(\tilde{\theta})^{-1}/n$, i.e.\ \begin{equation} \label{posterior} \pi_{SR}(\theta|x) \stackrel{\cdot}{\sim} N_{d} \left( \tilde{\theta},\frac{H(\tilde{\theta})^{-1}}{n} \right), \end{equation} where $H(\theta)=C^{T}(\partial^2 S(\theta)/\partial\theta^2)C/n$. Note that $H(\tilde{\theta})$ converges almost surely to $G(\theta)$ as $n \rightarrow \infty$. This result is stated in the next Theorem \ref{uno} that gives the expansion of the scoring rule posterior distribution (\ref{star}) up to third order. Before stating Theorem \ref{uno}, we introduce the notation which will be used in the rest of the paper. We use indices to denote derivatives of $S(x;\theta)$ and $\pi(\theta)$ with respect to the components of the parameter $\theta$. Therefore, for example, $S_{jkh}(\theta)=\partial^3 S(\theta)/\partial \theta^j\partial\theta^k\partial\theta^h$, $S_{ijkh}(\theta)=\partial^4 S(\theta)/\partial \theta^i\partial \theta^j\partial\theta^k\partial\theta^h$, for $1\leq i,j,k,h\leq d$, $\pi_{i}(\theta)=\partial \pi(\theta)/\partial \theta^i$ and $\pi_{ij}(\theta)=\partial^2 \pi(\theta)/\partial \theta^i\partial\theta^j$. In the sequel, a tilde over a quantity means evaluation of that quantity in $\tilde\theta$. For example, $\tilde{\pi}=\pi(\tilde\theta)$ and $\tilde{H}=H(\tilde\theta)$. Further, $\tilde{S}_{jkh}=S_{jkh}(\tilde{\theta})$, $\tilde{S}_{ijkh}=S_{ijkh}(\tilde{\theta})$, $1\leq i,j,k,h\leq d$, $\tilde\pi_i=\pi_{i}(\tilde\theta)$ and $\tilde\pi_{ij}=\pi_{ij}(\tilde\theta)$. Let $c_{ij}$ and $\tilde{h}^{ij}$ be the elements of the matrices $C$ and $\tilde{H}^{-1}=H(\tilde\theta)^{-1}$, respectively. Moreover, we use the Einstein summation convention so that when an index appears twice in an expression, summation on that index is intended. \begin{theorem} \label{uno} Let $w=(w^1,\dots,w^{d})^T=n^{1/2}(\theta-\tilde\theta)$. The SR-posterior distribution for $w$ can be written as \begin{eqnarray} \label{post-expansion} \pi_{SR} (w|x)=\phi_d(w;\tilde H^{-1}) \left[ 1+n^{-1/2}A_1(w) + n^{-1}A_2(w)\right]+O_p(n^{-3/2}), \end{eqnarray} where $\phi_d(w;\tilde H^{-1})$ is the density of a $d-$variate normal distribution with zero mean vector and variance matrix $ \tilde H^{-1}$, $$ A_1(w)= \frac{\tilde\pi_i}{\tilde\pi}w^i-\frac{1}{6}\frac{\tilde S_{ijk}}{n}c_{ir}c_{js}c_{kt}w^rw^sw^t $$ and \begin{eqnarray*} A_2(w)&=& \frac{1}{2}\frac{\tilde\pi_{ij}}{\tilde\pi}(w^iw^j-\tilde h^{ij})\\ &&-\frac{1}{6}\frac{\tilde\pi_i}{\tilde\pi} \frac{\tilde S_{jkh}}{n}c_{jr}c_{ks}c_{ht} (w^i w^r w^s w^t-3 \tilde h^{ir}\tilde h^{st})\\ &&-\frac{1}{24}\frac{\tilde S_{ijkh}}{n} c_{ir}c_{js}c_{kt}c_{hu} (w^r w^s w^t w^u -3\tilde h^{rs}\tilde h^{tu})\\ &&+\frac{1}{72}\frac{\tilde S_{ijk}\tilde S_{rst}}{n^2} \left[c_{ia}c_{jb}c_{kc}c_{rd}c_{se}c_{tf}(w^aw^bw^cw^dw^ew^f-9\tilde h^{ab}\tilde h^{cd}\tilde h^{ef}) \right.\\ &&\quad\quad\quad \left.+c_{ia}c_{rb}c_{jc}c_{sd}c_{ke}c_{tf}(w^aw^bw^cw^dw^ew^f-6\tilde h^{ab}\tilde h^{cd}\tilde h^{ef})\right]. \end{eqnarray*} \end{theorem} The proof is given in the Appendix. In practice, Theorem \ref{uno} shows that the SR-posterior distribution (\ref{star}) has the correct curvature, namely the variance of the posterior distribution (\ref{posterior}) is asymptotically equivalent to that of the minimum scoring rule estimator. When in particular $S(x;\theta)$ is the logarithmic score, it can be shown that (\ref{post-expansion}) reduces to the expansion given in Datta and Mukerjee (2004) for the classic posterior distribution $\pi(\theta|x) \propto \pi(\theta) L(\theta;x)$. \section{Reference priors obtained from $\alpha-$divergences} The information on $\pi(\theta|x)$ induced by $\pi(\theta)$ may be measured in terms of a divergence $D(\cdot,\cdot)$ between the prior and the posterior distribution: the higher the divergence, the lower the influence of the prior on the posterior. Let $D_{\pi}(x)=D(\pi(\theta),\pi(\theta|x))$, and let $p(x)$ be the marginal distribution of $X$. Minimizing the information in a prior is equivalent to maximize the expected divergence $D_\pi$ from the corresponding posterior, i.e.\ the functional \begin{equation}\label{functional} T(\pi)=\int_{\mathcal{X}}D_\pi (x)p(x)dx. \end{equation} Here, we focus on the well-known family of $\alpha-$divergences, defined as \[ D_{\pi}(x)=\frac{1}{\alpha(1-\alpha)}\int_{\Theta}\left\lbrace 1-\left(\frac{\pi(\theta)}{\pi(\theta|x)}\right)^{\alpha}\right\rbrace \pi(\theta|x)d\theta, \] which for $\alpha\rightarrow 0$ reduces to the Kullback-Liebler divergence, for $\alpha=1/2$ corresponds to twice the Hellinger distance and for $\alpha=-1$ is equivalent to the $\chi^2-$divergence. In this section, we extend the results of Ghosh {\em et al.} (2011) and Liu {\em et al.} (2014) to the context of scoring rules. In particular, in the following theorem we propose reference priors obtained by maximizing the expected $\alpha-$divergence from the SR-posterior distribution (\ref{star}). In the sequel, we denote by $g_{ij}$ and $g^{ij}$ the components of matrices $G(\theta)$ and $G(\theta)^{-1}$ respectively. Moreover, let $E_\theta(S_{ijk}/n)=B^S_{ijk}(\theta)+o(n^{-1/2})$ and let $a^S=(a^S_1,\dots,a^S_{d})^T$, with $a^S_i=B^S_{khl}c_{ki}c_{hu}c_{lt}g^{ut}$. Similarly, we define $E_\theta(\ell_{ijk}/n)=B^\ell_{ijk}(\theta)+o(n^{-1/2})$ and $a^\ell=(a^\ell_1,\dots,a^\ell_{d})^T$, with $a^\ell_i=B^\ell_{ihk}i^{hk}$. Here, $\bi^{ij}$ are the components of the inverse of the Fisher information matrix $I=E_\theta(-\partial^2\ell(\theta)/\partial\theta\partial\theta^T)$. Finally, $\Sigma=(\sigma^{ij})_{ij}$, with $\sigma^{ij}=nCov_\theta(\tilde\theta^i,\hat\theta^j)$, where $\hat\theta$ is the maximum likelihood estimator for $\theta$. Here, all indices range between $1$ and $d$. \begin{theorem} \label{due} When $0\leq |\alpha|<1$, the prior which asymptotically maximizes the expected $\alpha-$divergence to the SR-posterior distribution (\ref{star}) is \begin{eqnarray}\label{nostraprior} \pi_G (\theta)\propto |G(\theta)|^{1/2}. \end{eqnarray} When $\alpha=-1$, the prior which asymptotically maximizes the expected $\chi^2-$divergence to the SR-posterior distribution (\ref{star}) is such that \begin{equation} \label{log-solution} \frac{\partial\log\pi_{\chi}(\theta)}{\partial\theta}= \frac{1}{4}\left[ 6a^S G^{-1} + 4a^\ell I^{-1}+ |G|^{-1}\frac{\partial |G|}{\partial\theta}(5G^{-1}-4\Sigma) +2G \nabla\cdot (5G^{-1}-4\Sigma) \right] \Gamma, \end{equation} where $\nabla\cdot G^{-1}= (\partial g^{1j}/\partial\theta^j, \dots, \partial g^{{d} j}/\partial\theta^j)^T$, $\nabla\cdot \Sigma= (\partial \sigma^{1j}/\partial\theta^j, \dots, \partial \sigma^{{d} j}/\partial\theta^j)^T$ and $\Gamma=(G^{-1}-4I^{-1}+4\Sigma)^{-1}$. \end{theorem} The proof is given in the Appendix. It is important noticing that, if we use the full likelihood in the Bayes formula, (\ref{nostraprior}) and (\ref{log-solution}) reduce to the results obtained in Liu {\em et al.} (2014) and, for the case of a scalar parameter, in Ghosh {\em et al.} (2011). As shown in the Appendix, for $|\alpha|>1$ and $\alpha=1$ a maximizer for the expected $\alpha-$divergence does not exist. Note that, when dealing with the multidimensional parameter case, the use of the square root of the determinant of the Godambe information matrix may be questionable as in the case of the Jeffreys prior. Following the advices for default priors, one possibility is to assume that the components of $\theta$ are a priori independent and to use the one-dimensional reference prior for each of the parameters. An alternative is to consider a sequential scheme as in Berger and Bernardo (1992). On the other hand, it should be noted that the ordering of the parameters is relevant. Unless the practitioner has a specific ordering in mind, different orderings may lead to different reference priors. For $\alpha=-1$, the preceding expression (\ref{log-solution}) depends on quantities related to both the scoring rule and the likelihood. In practice this makes the proposed prior useful only when the likelihood function is available, as for instance in the context of robust inference. We also note that the required calculations remarkably simplify in the scalar parametric case. The proposed prior distribution (\ref{nostraprior}) shares some important properties with the Jeffreys prior. The most relevant is invariance with respect to one-to-one changes in the parametrization. As shown in Liu {\em et al.} (2014), when the functional (\ref{functional}) is invariant, the maximizer is also invariant. Thus, invariance of the proposed prior distributions follows from the well known invariant properties of $\alpha$-divergences. Anyway, for $0\leq|\alpha|<1$, it can be easily seen that $G(\theta)$ is a second order tensor so that, if $\psi(\theta)$ is a new parametrisation, $G_{rs}(\psi)=G_{ij}(\theta(\psi))\theta^i_r(\psi) \theta^j_s(\psi)$, where $\theta^i_r=\partial\theta^i/\partial\psi^r$. Thus it follows that $\pi_G(\psi)=\pi_G(\theta(\psi))|\partial\theta(\psi)/\partial\psi|$. We would like to stress that this is a general property that holds for every model and every scoring rule. Other properties of the proposed prior distribution (\ref{nostraprior}) in general depend on the scoring rule under consideration. In the following we present two special examples. \subsection{The Tsallis scoring rule} Let $X$ be distributed as a location model with parameter $\mu\in\mathbb{R}$, so that $p(x|\mu)=p_0(x-\mu)$. The Tsallis scoring rule in this case takes the following form \begin{eqnarray*} S(x;\mu)&=&(\gamma-1)\int p_0(x-\mu)^\gamma dx-\gamma p_0(x-\mu)^{\gamma-1}\\ &=& \mbox{const}-\gamma p_0(x-\mu)^{\gamma-1}. \end{eqnarray*} Its derivatives with respect to $\mu$ are all functions of $x-\mu$, so that the corresponding expected values, i.e. $K(\mu)$, $J(\mu)$ and $G(\mu)$ are independent of $\mu$. Thus, $\pi_G(\mu)\propto 1$, as the Jeffreys prior. Also for scale models, $\pi_G(\sigma)$ with the Tsallis score coincides with the Jeffreys prior. Indeed, let $X$ be distributed as a scale model with parameter $\sigma>0$, so that $p(x|\sigma)=p_0(x/\sigma)/\sigma$. The Tsallis scoring rule in this case reduces to \begin{eqnarray*} S(x;\sigma)&=&(\gamma-1)\int \frac{1}{\sigma^\gamma} p_0\left(\frac{x}{\sigma}\right)^\gamma dx-\gamma \frac{1}{\sigma^{\gamma-1}} p_0\left(\frac{x}{\sigma}\right)^{\gamma-1}\\ &=& \frac{1}{\sigma^{\gamma-1}}\left[ \mbox{const}-\gamma p_0\left(\frac{x}{\sigma}\right)^{\gamma-1}\right]. \end{eqnarray*} It is easy to show that the first three derivatives of $S(x;\sigma)$ can be written as \[ S_\sigma=\frac{1}{\sigma^\gamma} f_1\left(\frac{x}{\sigma} \right), \quad S_{\sigma\sigma}=\frac{1}{\sigma^{\gamma+1}} f_2\left(\frac{x}{\sigma} \right), \quad S_{\sigma\sigma\sigma}=\frac{1}{\sigma^{\gamma+2}} f_3\left(\frac{x}{\sigma} \right), \] where $f_1$, $f_2$ and $f_3$ are suitable functions. Notice that the expectation of any function of $X/\sigma$ with respect to $p(x;\sigma)$ does not depend on $\sigma$ itself. Thus, $J(\sigma)\propto 1/\sigma^{2\gamma}$ and $K(\sigma)\propto 1/\sigma^{\gamma+1}$, so that $G(\sigma) \propto 1/\sigma^{2}$ and $\pi_G(\sigma)\propto 1/\sigma$. \subsection{Hyv\"arinen scoring rule} Let $X$ be a non-degenerate random variable belonging to the one-parameter natural exponential family \begin{eqnarray} p(x|\theta)&=&\exp{\{\theta x-k(\theta)+a(x)\}},\; \quad x\in\mathbb{R} \ . \label{expfam} \end{eqnarray} Let denote by $a'(x_i)=\partial a(x_i)/ \partial x_i$ and $a''(x_i)=\partial^2 a(x_i)/ \partial x_i^2$. The Hyv\"arinen total empirical score in this case reduces to \begin{equation}\label{hyvexp} S(\theta)=-\left\lbrace 2\sum_{i=1}^na''(x_i)+\sum_{i=1}^n\left[ \theta +a'(x_i)\right] ^2\right\rbrace. \end{equation} The Hyv\"arinen score estimator for this family is \begin{equation*} \tilde\theta=-\frac{\sum_{i=1}^na'(X_i)}{n} \ , \end{equation*} which can be computed without knowledge of the normalizing constant $k(\theta)$; see Barndorff-Nielsen (1976), Hyv\"arinen (2007) and Parry {\em et al.} (2012). The Hyv\"arinen score estimator is an unbiased estimator for the parameter $\theta$ and its variance coincides with the inverse of the Godambe information $G(\theta)^{-1}\propto Var\left(a'(X)\right)$; see Mameli and Ventura (2015). Thus, $\pi_G(\theta)\propto Var(a'(X))^{-1/2}$, while the Jeffreys prior is $\pi_J(\theta)\propto (\partial ^2 k(\theta)/ \partial \theta^2)^{1/2}$ which requires the knowledge of $k(\theta)$. In models in which $k(\theta)$ is difficult to evaluate, the Jeffreys prior is not available. \section{Examples} In many applications, classical likelihood based methods may be infeasible, for example in models with complex dependency structure, or when robustness with respect to data or to model misspecification is required. In this section, we consider the analysis of these statistical issues from a scoring rule viewpoint. In particular, we illustrate the behaviour of the SR-posterior distribution and of the corresponding reference prior by means of three examples. In these examples different non-informative priors are also considered. MCMC methods based on data augmentation, such as the Gibbs sampler, seem difficult to adapt in this setting. The reason is that, in general, the proposed prior and the associated scoring rule may not be conjugate and, the full conditionals may not correspond to known densities. Therefore, except for Example 5.1 in which different posteriors are computed by numerical integration, the posterior distributions are computed via the classical Metropolis-Hastings algorithm using a multivariate normal proposal. The {\tt R} code for fully reproducing the examples of this section can be found in the Electronic Supplementary Material. \subsection{Hyv\"arinen scoring rule for the von Mises-Fisher distribution} The first example deals with the study of directional models (Mardia and Jupp, 2000). These models arise naturally in many areas of scientific research, such as meteorology for wind direction, biology for animal movement, in the investigation of biological processes and many others fields. Inference for directional models is difficult because typically the density function contains an intractable normalization constant, which cannot be explicitly computed in closed form. In this setting and to avoid the issue of the intractable normalising constant, Mardia {\em et al.} (2016) propose to use the Hyv\"arinen scoring rule. An important feature of this score is related to the homogeneity property for which the distribution needs to be known up to the normalization constant, thus avoiding the calculation of the density function. Let us consider the von Mises-Fisher density, which is a directional distribution defined on the unit sphere ${\cal S}_{q-1}\subset\mathbb{R}^q$ given by \begin{equation*} p(z|\kappa)\propto \exp(-\kappa \mu_0^Tz), \quad z\in {\cal S}_{q-1}, \end{equation*} with $\kappa\in \mathbb{R}^+$ scalar concentration parameter and $\mu_0=(\cos \theta_0, \sin \theta_0)$, with $\theta_0 \in \mathbb{R}$ known. When $q=2$ and the data are represented in polar coordinates $(z_{h1}, z_{h2}) = (\cos{\theta_h}, \sin{\theta_h})$, $h=1,\ldots,n$, the Hyv\"arinen estimator for $\kappa$ can be derived in closed form and it is given by \begin{equation} \tilde{\kappa} = \frac{2\sqrt{\bar{R}^2(1 + \bar{R}_2^2) + 2(\bar{C}^2 - \bar{S}^2)\bar{C}_2 + 4\bar{C}\bar{S}\bar{S}_2}}{(1 - \bar{R}_2^2)}, \end{equation} where $\bar{C} = \frac{1}{n}\sum_{h=1}^n\cos(\theta_h)\,,\bar{S} = \frac{1}{n}\sum_{h=1}^n\sin(\theta_h)$, $\bar{C}_2 = \frac{1}{n}\sum_{h=1}^n\cos(2\theta_h)$, $\bar{S}_2 = \frac{1}{n}\sum_{h=1}^n\sin(2\theta_h)$ and $\bar{R}_2 = \sqrt{\bar{C}^2_2+ \bar{S}^2_2}.$ For $\theta_0=0$, Mardia {\em et al.} (2016) show that the asymptotic variance of $\tilde{\kappa} $ is $$ \frac{V(\kappa)}{n} = \frac{\kappa\left[2\kappa-3A_1(\kappa)\right]}{nA_1^2(\kappa)}, $$ with $A_{1}(\kappa)=I_{1}(\kappa)/I_{0}(\kappa)$, where $I_1$ and $I_0$ are the modified Bessel functions of order 0 and 1, respectively. The reference prior (\ref{nostraprior}) is thus \[ \pi_G(\kappa)\propto \sqrt{\frac{A_1^2(\kappa)}{\kappa\left[2\kappa-3A_1(\kappa)\right]}}. \] Figure \ref{fig2} shows the plot of the reference prior $\pi_G(\kappa)$ and, for comparison, it also presents the classical non-informative prior \(\pi(\kappa)\propto 1/\kappa\) for the scale parameter. It can be seen from the plot that $\pi_G(\kappa)$ puts finite mass at 0, leading to more appropriate inference when the true parameter values are close to the boundary of the parametric space. Actually, using properties of first order Bessel functions, it can be proved that $\lim_{\kappa\rightarrow 0^+} \pi_G(\kappa)=2^{-1/2}$. \begin{figure}[htbp] \centering \includegraphics[scale=0.7]{g_prior-1.pdf} \caption{Priors $\pi_G(\kappa)\propto |G(\kappa)|^{1/2}$ and \(\pi(\kappa)\propto 1/\kappa\).} \label{fig2} \end{figure} In order to illustrate that the calibration of the Hyv\"arinen scoring rule is necessary to obtain a posterior distribution (\ref{star}) with the right curvature, consider a sample of size $n=50$ from the von Mises-Fisher distribution with $(\kappa, \mu_0) = (3,0)$. Hereafter, we take the location parameter $\mu_0$ to be fixed and equal to zero, and consider the problem of estimating $\kappa$. Figure \ref{fig1} compares the full posterior (black line) based on the likelihood function of the von Mises-Fisher model, with the calibrated (red) and non calibrated (green) SR-posteriors based on the Hyv\"arinen scoring rule. Here, the non-informative prior \(\pi(\kappa)\propto 1/\kappa\) is used. The vertical line corresponds to the Hyv\"arinen scoring rule estimate of \(\kappa\), i.e.\ $\tilde\kappa=2.36$. It can be noted that the non calibrated SR-posterior is too narrow compared with the genuine posterior, while the calibrated SR-posterior shows a curvature similar to the genuine posterior. \begin{figure}[htbp] \centering \includegraphics[scale=0.8]{unnamed-chunk-2-1.pdf} \caption{Comparison of calibrated SR-posterior vs non calibrated and proper posteriors.} \label{fig1} \end{figure} To compare the behaviour of the SR-posterior distributions, based on the two priors $\pi_G(k)\propto |G|^{1/2}$ and \(\pi(\kappa)\propto 1/\kappa\), two scenarios have been considered and the SR-posteriors with the two priors have been computed. The SR-posteriors are given in Figure \ref{fig3} for different values of the parameters $n$ and $\kappa$. In the first scenario, the true value of \(\kappa\) approaches the boundary and/or the sample size is small. In particular, we consider samples of size $n=10,30,50$ with \(\kappa=1\). It can be noted that for small samples sizes or for the true value of \(\kappa\) near zero, the SR-posterior with \(\pi(\kappa)\propto 1/\kappa\) may not be proper or puts too much mass at zero, i.e.\ it has a vertical asymptote at zero. On the other hand, the SR-posterior with the reference prior $\pi_G(\kappa)$ is proper and unimodal. \begin{figure}[htbp] \centering \includegraphics[scale=0.5]{unnamed-chunk-5-1.pdf} \includegraphics[scale=0.5]{unnamed-chunk-4-1.pdf} \caption{SR-posteriors with the $\pi_G(\kappa) \propto |G(\kappa)|^{1/2}$ prior and with the \(\pi(\kappa)\propto 1/\kappa\) prior.} \label{fig3} \end{figure} In the second scenario, the true value of \(\kappa\) is away from the boundary and the sample size is moderate or large. We consider samples of size \(n=30,50\), with \(\kappa=5\). The SR-posteriors in these cases look very similar, as it can be noted from Figure \ref{fig3}. Indeed, for high values of \(\kappa\) the two SR-posteriors are indistinguishable. \subsection{Pairwise likelihood for the multivariate equi-correlated normal model} Let $X$ be a $q-$dimensional random vector with mean $\mu$ and covariance matrix $\Sigma$, with $\Sigma_{rr} = \sigma^2$ and $\Sigma_{rs} = \rho\sigma^2$ for $r\neq s$, with $r,s = 1,\ldots, q$ and $\rho\in (-1/(q-1),1)$. The pairwise log-likelihood for $\theta = (\mu, \sigma^2,\rho)$ is (see Pace {\em et al.}, 2011) \begin{eqnarray} S_p (\theta) &=& -\frac{nq(q - 1)}{2}\log{\sigma^2} -\frac{nq(q - 1)}{4}\log{(1 - \rho^2)} - \frac{q - 1 + \rho}{2\sigma^2(1 - \rho^2)}SSW+\nonumber\\&-& \frac{q(q - 1)SSB + nq(q - 1)( \bar{y} - \mu)^2}{2\sigma^2(1 + \rho)},\nonumber \end{eqnarray} where $SSW =\sum_{i=1}^n\sum_{r=1}^q(x_{ir} - \bar{x}_i )^2$, $SSB =\sum_{i=1}^n( \bar{x}_i- \bar{x})^2$, $\bar{x}_i =\sum_{r=1}^q x_{ir} /q$ and $\bar{x} =\sum_{i=1}^n\sum_{r=1}^q x_{ir}/(nq)$. The reference prior (\ref{nostraprior}) is proportional to the square root of the determinant of the Godambe information matrix whose components are given in Pace {\em et al.}\ (2011). Consider, first, the simplest situation with $\mu$ and $\sigma^2$ known. The SR-posterior distribution based on the pairwise likelihood with the uniform prior and the proposed reference prior is computed for different values of \(\rho\) and \(q\) (see Figure \ref{fig4}). We notice that the two priors give very similar results when the value of the parameter $\rho$ is away from zero. \begin{figure}[htbp] \centering \includegraphics[scale=0.8]{unnamed-chunk-6-1.pdf} \caption{SR-posteriors with the $\pi_G(\rho)$ and with the uniform prior in $(0,1)$ for different values of $\rho$ and $q$ where $n=10$, $\sigma^2=1$ and $\mu=0$.} \label{fig4} \end{figure} Consider now the multi-parametric case. We compare the SR-posteriors based on the pairwise likelihood with the reference prior and with a non-informative prior derived by using two alternative parametrizations: \begin{itemize} \item[(a)] $\theta = (\mu, \sigma, \rho)$, for which we assume the non-informative prior $\pi(\theta) \propto 1/\sigma$. \item[(b)] $\xi = (\mu, \tau, \kappa)$, with $\tau = \log(\sigma)$ and $\kappa = \text{logit}(\rho)$, for which we assume the flat prior $\pi(\xi)\propto 1$. \end{itemize} The need for calibration of the composite likelihood has been studied by Ribatet {\em et al.} (2012); see also Pauli {\em et al.} (2011). The calibrated SR-posterior distribution has been computed with the parametrization \(\xi\) by using formula (\ref{posterior}), in order to avoid the constraints on $\theta \in \Theta$. As a first example, consider \(n=10\), \(q=10\) and a sample generated under the equi-correlated normal model with \(\theta=(0, 1, 0.5)\). The SR-posterior distributions with the reference prior and the two non-informative priors in cases $(a)$ and $(b)$ are shown in Figure \ref{fig5}. The SR-posterior distributions, described by means of histograms, show no appreciable differences. \begin{figure}[htbp] \centering \includegraphics[scale=0.7]{unnamed-chunk-7-1.pdf} \caption{SR-posterior with the reference prior $\pi_G(\theta)$ compared with the two alternative flat priors (a) and (b). The data are generated from the equi-correlated model with $n=10$, $q=10$, $\mu=0$, $\sigma^2=1$ and $\rho=0.5$} \label{fig5} \end{figure} As a second example, consider the more problematic scenario with \(n=10\), \(q=4\) and data generated under the equi-correlated normal model with \(\theta=(0, 0.5, 0.1)\). The SR-posteriors with the reference prior and the two non-informative priors are shown in Figure \ref{fig6}. \begin{figure}[htbp] \centering \includegraphics[scale=0.6]{unnamed-chunk-8-1.pdf} \caption{SR-posterior with the reference prior $\pi_G(\theta)$ (third row) compared with the two alternative flat priors (a) (first row) and (b) (second row). The data are generated from the equi-correlated model with $n=10$, $q=4$ $\mu=0$, $\sigma^2=0.5$ and $\rho=0.1$} \label{fig6} \end{figure} It can be noted that a flat prior on \(\theta\) may not necessarily be flat also on \(\xi\), at least not for \(\sigma\) and \(\rho\). Indeed, the uniform prior on \(\rho\) implies a standard logistic prior on \(\kappa\), which is quite informative. This explains the ``regularised" behaviour of the marginal posteriors of $\rho$ and $\sigma$ in case $(a)$. On the other hand, the proposed reference prior behaves quite similarly to the uniform improper prior \(\pi(\xi)\propto 1\). \subsection{Tsallis scoring rule for linear regression models} The third example is related to robustness procedures, which are more stable and reliable than their classical likelihood-based counterparts under misspecification. As it is well known, the ordinary Bayes estimator based on the classical based likelihood posterior density lacks robustness. Gosh and Basu (2016) studied the robustness properties of the posterior distribution based on the Tsallis scoring rule. There is a trade-off between robustness ad efficiency in the Tsallis scoring rule: small values of $\gamma$ in \eqref{} entails to a small loss in efficiency but a reduced robustness; while large values of $\gamma$ are associated to great robustness but low efficiency. Here, we consider a linear regression model to demonstrate the robustness performance of the Tsallis scoring rule and of the derived prior.\\ Consider the linear regression model \begin{equation*} y = X \beta+ \sigma \epsilon, \end{equation*} where $X$ is a fixed $n \times p$ matrix, $\beta \in \mathbb{R}^p$ ($p \geq 1$) is the vector of unknown regression coefficients, $\sigma > 0$ is a scale parameter, and $\epsilon$ is an $n-$dimensional vector of random errors from a standard normal distribution. Let $\theta=(\beta,\sigma^2)$. The Tsallis total empirical score in this case assumes the following form (Ghosh and Basu, 2013) \begin{equation*} S_T(\theta)= \frac{\gamma}{(2\pi\sigma^2)^{\frac{\gamma-1}{2}}}\sum_{i=1}^n e^{-\frac{(\gamma-1)}{2\sigma^2}(y_i-x_i^T\beta)^2}-\frac{n (\gamma-1)}{\sqrt{\gamma}(2\pi\sigma^2)^{(\gamma-1)/2}}. \end{equation*} The asymptotic variance of the Tsallis estimator of $\theta$ is given in Ghosh and Basu (2013). In particular, the asymptotic distribution of $(X^TX)^{1/2}(\tilde{\beta}-\beta)$ is a $N_p(0,v^{\beta}_{\gamma}I_p)$, with $v^{\beta}_{\gamma}=\sigma^2\left( 1+\frac{(\gamma-1)^2}{2\gamma-1}\right)^{3/2}$, while $\sqrt{n}(\tilde{\sigma}^2 - \sigma^2)$ follows a normal distribution with mean 0 and variance $v^e_{\gamma}$, where $$ v^{e}_{\gamma}= \frac{4\sigma^4}{(2+(\gamma-1)^2)^2}\left(2(1+2(\gamma-1)^2)\left(1+\frac{(\gamma-1)^2}{2\gamma-1}\right)^{5/2}-(\gamma-1)^2\gamma^2\right). $$ Since the asymptotic distributions of $\tilde{\beta}$ and $\tilde{\sigma}^2$ are independent, the proposed reference prior is proportional to the square root of the determinant of the Godambe information, i.e.\ \begin{eqnarray} \pi_G(\theta)\propto \left(v^{\beta}_{\gamma}v^{e}_{\gamma}\right)^{-1/2}. \label{prioreg} \end{eqnarray} Therefore, a prior for $\beta$ is $\pi_G(\beta)\propto \left(v^{\beta}_{\gamma}\right)^{-1/2}$, while a prior for $\sigma^2$ is $\pi_G(\sigma^2)\propto \left(v^{e}_{\gamma}\right)^{-1/2}.$ In order to illustrate the behaviour of the Tsallis scoring rule in the context of the linear regression model, we consider the GFR dataset (Heritier {\em et al.}, 2009), which contains measurements of the glomerular filtration rate (GFR) and serum creatinine (CR) on $n=30$ subjects. The GFR is the volume of fluid filtered from the renal glomerular capillaries into the Bowmans capsule per unit of time (typically in millilitres per minute) and, clinically, it is often used to determine renal function. Its estimation, when not measured, is of clinical importance and several techniques are used for that purpose. One of them is based on CR, an endogenous molecule, synthesized in the body, which is freely filtered by the glomerulus (but also secreted by the renal tubules in very small amounts). Several models have been proposed in the literature to explain the logarithm of GFR as a function of CR. Here, following Heritier {\em et al.} (2009), we consider a model for GFR based on CR$^{-1}$ and AGE, i.e. $$ \text{GFR} = \beta_0 + \beta_1 \, \frac{1}{\text{CR}} + \beta_2 \, \text{AGE} + \varepsilon. $$ Both covariates are scaled to have mean zero and unit variance, whereas the response is scaled to have unit variance. The data are illustrated in Figure \ref{scatter}: note that there are some observations which look like outliers. The need and consequence to resort to robust procedures to deal with these data have been highlighted in several papers; see, among others, Heritier {\em et al.} (2009) and Farcomeni and Ventura (2012). \begin{figure} \centering \includegraphics[scale=0.8]{gfreda-1.pdf} \caption{Scatterplot diagram of GFR data.} \label{scatter} \end{figure} For the parameter \(\theta = (\beta, \log\sigma)\) we assume both the usual non-informative flat prior \(\pi(\theta)\propto 1\) and the proposed reference prior (\ref{prioreg}) in the SR-posterior distribution, while the classical posterior distribution is based on the non-informative flat prior. Figure \ref{reg1} gives the violin plots of the marginal SR-posterior distributions based on the Tsallis score, with \(\gamma = 1.25\), and of the classical posterior distribution. From Figure \ref{reg1} it can be noted that the proposed reference prior behaves similarly to the non-informative prior. Moreover, the classical marginal posterior distribution shows in general heavier tails and, in particular, the robust and the classical posterior distributions give different inferences on $\beta_2$ and $\sigma$. \begin{figure} \centering \includegraphics[scale=0.7]{unnamed-chunk-18-1.pdf} \caption{SR-posteriors based on the Tsallis score with $\gamma=1.25$ (with the flat or the reference prior) vs the genuine posterior for the GFR dataset.} \label{reg1} \end{figure} \newpage Figure \ref{figa} gives useful monitoring plots in robustness studies. In particular, these plots illustrate the SR-posterior distributions based on the Tsallis score with reference prior as a function of $\gamma$. This approach (see, e.g., Riani {\em et al.}, 2014) provides tools for gaining knowledge and better understanding of the properties of robust procedures. The horizontal lines in Figure \ref{figa} correspond to the posterior mode of the genuine posterior based on the flat prior. \begin{figure} \centering \includegraphics[scale=0.5]{unnamed-chunk-23-1.pdf} \includegraphics[scale=0.5]{unnamed-chunk-24-1.pdf} \includegraphics[scale=0.5]{unnamed-chunk-25-1.pdf} \includegraphics[scale=0.5]{unnamed-chunk-26-1.pdf} \caption{Marginal SR-posteriors based on the Tsallis score, for $ \beta_1, \beta_2$ and $\sigma$ with the reference prior, and values of $R^2$ for varying values of $\gamma$.} \label{figa} \end{figure} The value $\gamma=1$ corresponds to complete overlap of the two methods. As far as $\gamma$ increases, robustness is achieved although at the price of losing efficiency at the central model; the value $\gamma=1.25$ gives approximately 0.95\% efficiency under the normal distribution. As a final remark, note that the Tsallis scoring rule provides a cogent framework for dealing with robust procedures in the Bayesian framework. \section{Conclusions} In this paper, we discuss the use of scoring rules in order to compute a posterior distribution, useful to deal with complex models or if robustness with respect to data or to model misspecification is required. Indeed, scoring rules provide a flexible and robust way of combining data-driven information with prior distributions, either subjective or non-informative. One should devise a scoring rule that properly captures the structure of the data, otherwise the resulting posterior inferences are not reliable. In this respect, the result of Section 3 shows that a suitably calibrated SR-posterior distribution is, up to first order, normally distributed with the same asymptotic variance of the scoring rule estimator. When dealing with default priors, in Section 4 reference priors for a vector parameter based on maximising $\alpha-$divergences are discussed in the framework of scoring rules. We show that, for $0 \leq |\alpha| < 1$, the result is a Jeffreys-type prior that is proportional to the square root of the determinant of the Godambe information matrix. A possible direction of further research is to extend the proposed methodology considering the class of monotone and regular divergences which is a broad family of divergences asymptotically equivalent to $\alpha-$divergences; see Corcuera and Giummol\`e (1998). \section*{Appendix} \subsection*{Proof of Theorem \ref{uno}} \begin{proof} The posterior (\ref{star}) can equivalently be written as \[ \pi_{SR} (\theta|x)=\frac{\pi(\theta) \exp{\{-S (\theta^*)+S(\tilde\theta)\}}} {\int_\Theta \pi(\theta) \exp{\{-S (\theta^*)+S(\tilde\theta)\}} d\theta}, \] with $\theta^*=\tilde\theta+C(\theta-\tilde\theta)$ and $C$ fixed such that $C^TK(\theta)C=G(\theta)$. Let $w=(w^1,\dots,w^d)^T=n^{1/2}(\theta-\tilde\theta)$. Then, $\theta=\tilde\theta+n^{-1/2}w$ and $\theta^*=\tilde\theta+n^{-1/2}Cw$. A posterior for $w$ is \[ \pi_{SR} (w|x)=\frac{b(w,x)}{\int b(w,x) dw}, \] with \[ b(w,x)= \pi(\tilde\theta+n^{-1/2}w) \exp{\{-S (\tilde\theta+n^{-1/2}Cw)+S(\tilde\theta)\}}. \] Now \[ \pi(\tilde\theta+n^{-1/2}w)= \tilde\pi\left(1+n^{-1/2}R_1(w)+\frac{1}{2}n^{-1}R_2(w)\right)+O_p(n^{-3/2}), \] with $$ R_1(w)=\frac{\tilde\pi_i}{\tilde\pi}w^i \quad\quad \mbox{ and }\quad\quad R_2(w)=\frac{\tilde\pi_{ij}}{\tilde\pi}w^{ij}, $$ where $w^{ij\dots}=w^iw^j\dots$ is a product of components of $w$. Moreover, \begin{eqnarray} -S (\tilde\theta+n^{-1/2}Cw)+S(\tilde\theta)&=& -n^{-1}\left(\frac{1}{2}(Cw)^{ij}\tilde S_{ij} +\frac{1}{6}n^{-1/2}(Cw)^{ijk}\tilde S_{ijk}\right. \nonumber\\ &&\left. +\frac{1}{24}n^{-1}(Cw)^{ijkh}\tilde S_{ijkh}\right)+O_p(n^{-3/2}) \nonumber\\ &=&-\frac{1}{2}n^{-1}w^T\left(C^T\frac{\partial^2\tilde S}{\partial\theta\partial\theta^T}C\right)w-\frac{1}{6}n^{-1/2}R_3(w)\nonumber\\ &&-\frac{1}{24}n^{-1}R_4(w) +O_p(n^{-3/2})\nonumber \\ &=&-\frac{1}{2}w^T\tilde H w-\frac{1}{6}n^{-1/2}R_3(w)-\frac{1}{24}n^{-1}R_4(w) +O_p(n^{-3/2}),\quad\quad \end{eqnarray} where $$R_3(w)=n^{-1}(Cw)^{ijk}\tilde S_{ijk} \quad\quad\mbox{ and }\quad\quad R_4(w)=n^{-1}(Cw)^{ijkh}\tilde S_{ijkh}.$$ The numerator in Bayes formula can thus be written as \begin{eqnarray} \label{num} \lefteqn{b(w,x)=\tilde\pi\left( 1+n^{-1/2}R_1(w) +\frac{1}{2}n^{-1} R_2(w)\right)}\nonumber\\ &&\exp{\left\{ -\frac{1}{2}w^T\tilde H w \right\}}\left(1-\frac{1}{6}n^{-1/2}R_3(w)-\frac{1}{24}n^{-1}R_4(w) +\frac{1}{72}n^{-1}R_3(w)^2\right) +O_p(n^{-3/2})\nonumber\\ &=& \tilde\pi \exp{\left\{ -\frac{1}{2}w^T\tilde H w \right\}}\left[ 1+n^{-1/2}\left( R_1(w)-\frac{1}{6}R_3(w)\right)\right.\nonumber\\ &&\left.+n^{-1}\left( \frac{1}{2}R_2(w) -\frac{1}{6}R_1(w)R_3(w)-\frac{1}{24}R_4(w)+ \frac{1}{72}R_3(w)^2 \right) \right] +O_p(n^{-3/2}). \end{eqnarray} The denominator can be approximated using the moments of the $d-$variate normal distribution $N_d(0,\tilde H^{-1})$. We have that \begin{eqnarray} \label{denom} \int b(w,x) dw&=&\tilde\pi (2\pi)^{d/2}|\tilde H|^{-1/2} \left[1+n^{-1}\left( \frac{1}{2}E(R_2(W)) -\frac{1}{6}E(R_1(W)R_3(W))\right.\right.\nonumber\\ &&\left.\left.-\frac{1}{24}E(R_4(W))+ \frac{1}{72}E(R_3(W)^2) \right) \right] + o(n^{-1}), \end{eqnarray} being $E(R_1(W))$ and $E(R_3(W))$ both related to odd moments and thus equal to zero. Now, $$ E(R_2(W))= \frac{\tilde\pi_{ij}}{\tilde\pi} \tilde h^{ij} , \quad\quad E(R_1(W)R_3(W))= 3\frac{\tilde\pi_{i}}{\tilde\pi} \frac{\tilde S_{jkh}}{n}c_{jr}c_{ks}c_{ht}\tilde h^{ir}\tilde h^{st} , $$ $$ E(R_4(W))= 3 \frac{\tilde S_{ijkh}}{n}c_{ir}c_{js}c_{kt}c_{hu}\tilde h^{rs}\tilde h^{tu} $$ and \begin{eqnarray*} E(R_3(W)^2)&=& \frac{\tilde S_{ijk}\tilde S_{rst}}{n^2}(9c_{ia}c_{jb}c_{kc}c_{rd}c_{se}c_{tf}\tilde h^{ab}\tilde h^{cd}\tilde h^{ef}\\ &&+6c_{ia}c_{rb}c_{jc}c_{sd}c_{ke}c_{tf}\tilde h^{ab}\tilde h^{cd}\tilde h^{ef}) . \end{eqnarray*} Putting together expressions (\ref{num}) and (\ref{denom}), we finally obtain the result with $$ A_1(w)=R_1(w)-\frac{1}{6}R_3(w)$$ and \begin{eqnarray*} A_2(w)&=&\frac{1}{2}\left(R_2(w)-E(R_2(W))\right) -\frac{1}{6}\left(R_1(w)R_3(w)-E(R_1(W)R_3(W))\right)\\ &&-\frac{1}{24}\left(R_4(w)-E(R_4(W))\right) +\frac{1}{72}\left(R_3(w)^2-E(R_3(W)^2)\right). \end{eqnarray*} \end{proof} \subsection*{Proof of Theorem \ref{due}} \begin{proof} The proof follows the same steps as in Liu {\em et al.} (2014), Section 3, generalized for the use of a SR-posterior (\ref{star}) instead of the classic posterior distribution. Let $\pi(\theta)$ be a prior distribution for $\theta$ and $p(x|\theta)$ be the conditional distribution of $X$ given $\theta$. Thus, the marginal distribution of $X$ can be written as $p(x)=p(x|\theta)\pi(\theta)/\pi(\theta|x)$, with $\pi(\theta|x)$ the conditional distribution of $\theta$ given $x$. The functional (\ref{functional}) associated to an $\alpha-$divergence between the prior $\pi(\theta)$ and the SR-posterior (\ref{star}) can be written as \begin{eqnarray} \label{functional2} T(\pi)&=& \frac{1}{\alpha(1-\alpha)}\left[ 1- \int_{\mathcal{X}} \left( \int_{\Theta} \pi(\theta)^\alpha \pi_{SR}(\theta|x)^{1-\alpha} d\theta \right) p(x) dx \right] \nonumber \\ &=&\frac{1}{\alpha(1-\alpha)}\left[ 1- \int_{\Theta} \left( \int_{\mathcal{X}} \pi(\theta)^\alpha \pi_{SR}(\theta|x)^{1-\alpha} \pi(\theta|x)^{-1}p(x|\theta)dx \right) \pi(\theta)d\theta \right] \nonumber \\ \label{func} &=&\frac{1}{\alpha(1-\alpha)}\left[ 1-\int_{\Theta} \pi(\theta)^{\alpha+1}E_\theta\left[\pi_{SR}(\theta|X)^{1-\alpha}\pi(\theta|X)^{-1}\right]d\theta\right] , \end{eqnarray} where $E_\theta(\cdot)$ denotes expectation with respect to the conditional distribution of $X$ given $\theta$. For $\alpha=0$ or $1$, we need to interpret $T(\pi)$ as its limiting value, when it exists. Now, we apply the shrinkage argument for evaluating the conditional expectation in (\ref{functional2}), $E_\theta\left[\pi_{SR}(\theta|X)^{1-\alpha}\pi(\theta|X)^{-1}\right]$. See Datta and Mukerjee (2004) for details on the shrinkage method. Let us first consider the case of $0<|\alpha|<1$. The first step of the shrinkage method involves fixing a suitable prior distribution $\bar\pi(\theta)$ and calculating the expected value of $\pi_{SR}(\theta|x)^{1-\alpha}\pi(\theta|x)^{-1}$ with respect to the corresponding posterior $\bar\pi(\theta|x)$, i.e.\ $E^{\bar\pi}\left[\pi_{SR}(\theta|x)^{1-\alpha}\pi(\theta|x)^{-1}|x\right]$. Notice that here we use the classic posterior based on the prior $\bar\pi(\theta)$, i.e. the conditional distribution of $\theta$ given $x$, $\bar\pi(\theta|x)$. Using up to first order the asymptotic expansion (\ref{post-expansion}) for $\pi_{SR}(\theta|x)$ and similar expansions for $\bar\pi(\theta|x)$ and $\pi(\theta|x)$, we obtain \[ \pi_{SR}(\theta|X)^{1-\alpha}=\left(\frac{n^d|\tilde H|}{(2\pi)^d}\right)^{(1-\alpha)/2} \exp{\left\{ \frac{n(1-\alpha)}{2}(\theta-\tilde\theta)^T\tilde H(\theta-\tilde\theta)\right\}}+O_p(n^{-1/2}) \] and \[ \pi(\theta|X)^{-1}\bar\pi(\theta|X)=1 +O_p(n^{-1/2}). \] Since $\tilde H$ tends to $G$ as $n\rightarrow \infty$, we have that \begin{eqnarray*} E^{\bar\pi}\left[\pi_{SR}(\theta|x)^{1-\alpha}\pi(\theta|x)^{-1}|x\right] &=&\left(\frac{n^d|\tilde H|}{(2\pi)^d}\right)^{(1-\alpha)/2} \int_{\Theta} \exp{\left\{ \frac{n(1-\alpha)}{2}(\theta-\tilde\theta)^T\tilde H(\theta-\tilde\theta)\right\}} d\theta\\ &&+O(n^{-1/2})\\ &=&\left(\frac{n^d|G(\theta)|}{(2\pi)^d}\right)^{-\alpha/2}(1-\alpha)^{-d/2}+O(n^{-1/2}). \end{eqnarray*} The second step of the shrinkage argument requires to integrate again with respect to the distribution of $X$ given $\theta$ and with respect to the prior $\bar\pi(\theta)$, thus obtaining \begin{eqnarray*} \lefteqn{ \int_\Theta E_\theta\left[ E^{\bar\pi}\left[\pi_{SR}(\theta|X)^{1-\alpha}\pi(\theta|X)^{-1}|X\right]\right] \bar\pi(\theta) d\theta }\\ &&=\int_\Theta \left(\frac{n^d|G(\theta)|}{(2\pi)^d}\right)^{-\alpha/2}(1-\alpha)^{-d/2} \, \bar\pi(\theta) \, d\theta +O(n^{-1/2}). \end{eqnarray*} Finally, by letting the prior $\bar\pi(\theta)$ go to $\theta$, we obtain \begin{eqnarray} \label{expectation} E_\theta\left[\pi_{SR}(\theta|X)^{1-\alpha}\pi(\theta|X)^{-1}\right]&=& \left(\frac{n}{2\pi}\right)^{-d\alpha/2}|G(\theta)|^{-\alpha/2}(1-\alpha)^{-d/2}+O(n^{-1/2}). \end{eqnarray} For $0<|\alpha|<1$, by substituting (\ref{expectation}) in (\ref{func}) we can see that the selection of a prior $\pi(\theta)$ corresponds to the minimization with respect to $\pi(\theta)$ of the functional \begin{equation}\label{alpha} \frac{1}{\alpha(1-\alpha)}\int_\Theta \pi(\theta)^{\alpha+1}|G(\theta)|^{-\alpha/2}d\theta. \end{equation} Notice that the preceding expression can be interpreted as an increasing transformation of the $(-\alpha)-$divergence between a density that is proportional to $|G(\theta)|^{1/2}$ and the prior $\pi(\theta)$. Thus, it is minimized if and only if the two densities coincide. This result cannot be extended to the case of $\alpha\geq 1$, as is evident from the right-hand-side of (\ref{expectation}). For $\alpha<-1$, (\ref{alpha}) turns out to be equivalent to a decreasing transformation of the $(-\alpha)-$divergence between a density that is proportional to $|G(\theta)|^{1/2}$ and the prior $\pi(\theta)$. Thus a maximizer for the expected $\alpha-$divergence in this case does not exist. For $\alpha\rightarrow 0$, following the same steps as for $0<|\alpha|<1$, it can be easily shown that maximization of the average Kullback-Leibler divergence is asymptotically equivalent to minimization of \[ \int_\Theta \pi(\theta) \log\left(\frac{\pi(\theta)}{|G(\theta)|^{1/2}}\right) d\theta, \] which is attained by choosing again a Jeffreys-type prior proportional to $ |G(\theta)|^{1/2}$. Indeed, the preceding expression is equal up to an additive constant to Kullback-Leibler divergence between a density that is proportional to $ |G(\theta)|^{1/2}$ and the prior $\pi(\theta)$. The case $\alpha=-1$ corresponds to the Chi-square divergence. For this case the proof requires higher order terms in the expansion of the scoring rule posterior distribution and also of both the conditional distributions of $\theta$ given $x$ calculated with respect to the priors $\pi(\theta)$ and $\bar\pi(\theta)$. Thus, we use (\ref{post-expansion}) up to order $O_p(n^{-3/2})$. In particular, let $w=n^{1/2}(\theta-\tilde\theta)$ and $w_1=n^{1/2}(\theta-\hat\theta)$, where $\tilde\theta$ is the minimum scoring rule estimator and $\hat\theta$ the maximum likelihood estimator for $\theta$. The different posterior distributions for $w$ and $w_1$ given $x$ can be written as \begin{eqnarray*} \pi_{SR} (w|x)=\phi_d(w;\tilde H^{-1}) \left[ 1+n^{-1/2}A_1(w) + n^{-1}A_2(w)\right]+O_p(n^{-3/2}), \end{eqnarray*} \begin{eqnarray*} \pi(w_1|x)=\phi_d(w_1;(\hat J^\ell)^{-1}) \left[ 1+n^{-1/2}A^\ell_1(w_1) + n^{-1}A^\ell_2(w_1)\right]+O_p(n^{-3/2}), \end{eqnarray*} \begin{eqnarray*} \bar\pi(w_1|x)=\phi_d(w_1;(\hat J^\ell)^{-1}) \left[ 1+n^{-1/2}\bar A^\ell_1(w_1) + n^{-1}\bar A^\ell_2(w_1)\right]+O_p(n^{-3/2}), \end{eqnarray*} where $A_1$ and $A_2$ are defined as in Theorem \ref{uno}, $J^\ell(\theta)=(-\partial^2\ell/\partial\theta\partial\theta^T)/n$ is the observed information and $A^\ell_1$, $\bar A^\ell_1$, $A^\ell_2$, $\bar A^\ell_2$ are obtained by calculating $A_1$ and $A_2$ with the logarithmic score as scoring rule and $\pi$ and $\bar\pi$ as priors; for instance, $$ A^\ell_1(w_1)= \frac{\hat\pi_i}{\hat\pi}w_1^i+\frac{1}{6}\frac{\hat \ell_{ijk}}{n}w_1^iw_1^jw_1^k, \quad\mbox{and}\quad \bar A^\ell_1(w_1)= \frac{\hat{\bar\pi}_i}{\hat{\bar\pi}}w_1^i+\frac{1}{6}\frac{\hat \ell_{ijk}}{n}w_1^iw_1^jw_1^k, $$ with $\ell$ the log-likelihood and $\ell_{ijk}$ its third order derivatives. As usual, a tilde or a hat over a quantity indicate that it is calculated at $\tilde\theta$ and $\hat\theta$ respectively. Following the shrinkage argument, we need first to evaluate \begin{eqnarray*} E^{\bar\pi}\left[\pi_{SR}(\theta|x)^2 \pi(\theta|x)^{-1}|x\right] &=& n^{d/2} \int \pi_{SR}(w|x)^2 \pi(w|x)^{-1} \bar\pi(w|x) dw . \end{eqnarray*} For doing this, we use the fact that \begin{eqnarray*} \pi_{SR} (w|x)^2 &=& (2\pi)^{-d/2} |\tilde H|^{1/2} 2^{-d/2} \phi_d(w;\tilde H^{-1}/2) \\ &&\left[1+n^{-1/2}2A_1(w) + n^{-1}(2A_2(w)+A_1(w)^2)\right]+O_p(n^{-3/2}) \end{eqnarray*} and \begin{eqnarray*} \pi(w_1|x)^{-1} \bar\pi(w_1|x) &=& 1+n^{-1/2}(\bar A^\ell_1(w_1)-A^\ell_1(w_1)) \\ &&+ n^{-1}(A^\ell_1(w_1)^2+\bar A^\ell_2(w_1)-A^\ell_2(w_1)-\bar A^\ell_1(w_1) A^\ell_1(w_1)) +O_p(n^{-3/2}) \end{eqnarray*} and we calculate the last expression in $w_1=w+n^{1/2}(\tilde\theta-\hat\theta)$. After tedious calculations, we obtain \begin{eqnarray*} \lefteqn{E^{\bar\pi}\left[\pi_{SR}(\theta|x)^2 \pi(\theta|x)^{-1}|x\right]=} \\ &=& (2\pi)^{-d/2}\left(\frac{n}{2}\right)^{d/2} |\tilde H|^{1/2} \left\{1 + \frac{1}{n^{1/2}} \left( \frac{\hat{\bar\pi}_{i}}{\hat{\bar\pi}} -\frac{\hat{\pi}_{i}}{\hat{\pi}} \right)n^{1/2}(\tilde\theta-\hat\theta)^i \right.\\ &&\left.+ \frac{1}{2n} \left[ \left(-\frac{\hat{\bar\pi}_{ij}}{\hat{\bar\pi}} +\frac{\hat\pi_{ij}}{\hat\pi} \right) \hat \bj^{ij} \right.\right.\\ &&\left.\left. +\left( -\frac{\tilde\pi_{ij}}{\tilde\pi}+\frac{\tilde\pi_{i}}{\tilde\pi}\frac{\tilde\pi_{j}}{\tilde\pi} +2\frac{\tilde\pi_{i}}{\tilde\pi}\frac{\hat{\bar\pi}_{j}}{\hat{\bar\pi}} -2\frac{\tilde\pi_{i}}{\tilde\pi}\frac{\hat{\pi}_{j}}{\hat{\pi}} +\frac{\hat\pi_{i}}{\hat\pi}\frac{\hat{\pi}_{j}}{\hat{\pi}} -\frac{\hat\pi_{i}}{\hat\pi}\frac{\hat{\bar\pi}_{j}}{\hat{\bar\pi}} +\frac{1}{2}\frac{\hat{\bar\pi}_{ij}}{\hat{\bar\pi}} -\frac{1}{2}\frac{\hat\pi_{ij}}{\hat\pi} \right) \tilde h^{ij} \right.\right. \\ && \left.\left. +\left( \frac{\tilde\pi_{i}}{\tilde\pi} -\frac{1}{2}\frac{\hat{\bar\pi}_{i}}{\hat{\bar\pi}} +\frac{1}{2}\frac{\hat{\pi}_{i}}{\hat{\pi}}\right) \frac{\tilde S_{khl}}{n} c_{kj}c_{hs}c_{lt}\tilde h^{st}\tilde h^{ij} +\left( -\frac{\hat{\bar\pi}_{i}}{\hat{\bar\pi}} +\frac{\hat{\pi}_{i}}{\hat{\pi}}\right) \frac{\tilde \ell_{jhk}}{n} \hat\bj^{hk}\hat\bj^{ij} \right.\right.\\ &&\left.\left.+ \left( -\frac{\hat\pi_{ij}}{\hat\pi}+\frac{\hat{\bar\pi}_{ij}}{\hat{\bar\pi}} +2\frac{\hat{\pi}_{i}}{\hat{\pi}} \frac{\hat{\pi}_{j}}{\hat{\pi}} -2\frac{\hat{\pi}_{i}}{\hat{\pi}} \frac{\hat{\bar\pi}_{j}}{\hat{\bar\pi}} \right) n(\tilde\theta-\hat\theta)^{ij} \right]+ R \right\} + O(n^{-1}), \end{eqnarray*} where $\hat\bj^{ij}$ are the components of $(\hat J^\ell)^{-1}$ and $R=R(\tilde\theta,\hat\theta)$ is a function that does not involve the prior nor its derivatives. Now, let $E_\theta(S_{ijk}/n)=B^S_{ijk}(\theta)+o(n^{-1/2})$, $E_\theta(\ell_{ijk}/n)=B^\ell_{ijk}(\theta)+o(n^{-1/2})$ and recall that $E_\theta(\tilde H)=G(\theta)+o(n^{-1/2})$ and $E_\theta(\hat J^\ell)=I(\theta)+o(n^{-1/2})$, where $I$ is the expected information matrix. Moreover, note that $E_\theta[(\tilde\theta-\hat\theta)^i ]=o(n^{-1/2})$ and $E_\theta[n(\tilde\theta-\hat\theta)^{ij}]=g^{ij}+\bi^{ij}-2\sigma^{ij}+o(n^{-1/2})$, where $g_{ij}$ and $g^{ij}$ are the components of $G$ and $G^{-1}$ respectively, $\bi^{ij}$ are the components of $I^{-1}$ and $\sigma^{ij}=nCov_\theta(\tilde\theta,\hat\theta)^{ij}$. Furthermore, let $a^S_j=B^S_{khl}c_{kj}c_{hs}c_{lt}g^{st}$, $a^\ell_j=B^\ell_{jhk}\bi^{hk}$. By integrating again the above expression with respect to the distribution of $X$ given $\theta$ we get \begin{eqnarray*} \lefteqn{E_\theta\left[E^{\bar\pi}\left[\pi_{SR}(\theta|X)^2 \pi(\theta|X)^{-1}|X\right]\right]=} \\ &=& (2\pi)^{-d/2}\left(\frac{n}{2}\right)^{d/2} |G|^{1/2} \left\{1 + \frac{1}{2n} \left[ \left(-\frac{\bar\pi_{ij}}{\bar\pi} +\frac{\pi_{ij}}{\pi} \right) \bi^{ij} \right.\right.\\ &&\left.\left. +\left( -\frac{3}{2}\frac{\pi_{ij}}{\pi}+\frac{\pi_{i}}{\pi}\frac{\bar\pi_{j}}{\bar\pi} +\frac{1}{2}\frac{\bar\pi_{ij}}{\bar\pi} \right) g^{ij} \right.\right. \\ && \left.\left. +\left(\frac{3}{2}\frac{\pi_{i}}{\pi} -\frac{1}{2}\frac{\bar\pi_{i}}{\bar\pi} \right) a^S_j g^{ij} -\left( \frac{\bar\pi_{i}}{\bar\pi} -\frac{\pi_{i}}{\pi}\right) a^\ell_j \bi^{ij} \right.\right.\\ &&\left.\left.+ \left( -\frac{\pi_{ij}}{\pi}+\frac{{\bar\pi}_{ij}}{{\bar\pi}} +2\frac{{\pi}_{i}}{{\pi}} \frac{{\pi}_{j}}{{\pi}} -2\frac{{\pi}_{i}}{{\pi}} \frac{{\bar\pi}_{j}}{{\bar\pi}} \right) (g^{ij}+\bi^{ij}-2\sigma^{ij}) \right]+ F \right\} + O(n^{-1}), \end{eqnarray*} where $F=F(\theta)$ is a function that does not involve the prior nor its derivatives. By integrating again with respect to the prior $\bar\pi(\theta)$ and by letting the prior go to $\theta$, we finally obtain the expected value needed in (\ref{func}) for $\alpha=-1$: \begin{eqnarray} \label{expectation-chi} \lefteqn{E_\theta\left[\pi_{SR}(\theta|X)^2\pi(\theta|X)^{-1}\right]=}\nonumber \\ &=&(2\pi)^{-d/2}\left(\frac{n}{2}\right)^{d/2} |G|^{1/2}\left\{1+\frac{1}{4n}\left[ \left(-3g^{ij} +4\bi^{ij}-4\sigma^{ij}\right)\frac{\pi_{ij}}{\pi}\right.\right.\nonumber\\ &&\left.\left. +\left( 3a^S_jg^{ij}+2a^\ell_j\bi^{ij}+|G|^{-1}\frac{\partial |G|}{\partial\theta^j}(g^{ij}+2\bi^{ij}-4\sigma^{ij})+2 \frac{\partial (g^{ij}+2\bi^{ij}-4\sigma^{ij})}{\partial\theta^j} \right)\frac{\pi_i}{\pi} \right.\right.\nonumber\\ &&\left.\left. +2g^{ij}\frac{\pi_{i}}{\pi}\frac{\pi_{j}}{\pi} +M \right]\right\}+O(n^{-1}), \end{eqnarray} where $M=M(\theta)$ is a function that does not involve the prior nor its derivatives. Substituting (\ref{expectation-chi}) in (\ref{func}) with $\alpha=-1$, we find that maximization of the average Chi-square divergence is equivalent to maximization of \begin{eqnarray*} \lefteqn{\int_\Theta |G|^{1/2} \left[\left(-3g^{ij} +4\bi^{ij}-4\sigma^{ij}\right)\frac{\pi_{ij}}{\pi} +2g^{ij}\frac{\pi_{i}}{\pi}\frac{\pi_{j}}{\pi} \right.}\\ && \left. +\left(3a^S_jg^{ij}+2a^\ell_j\bi^{ij}+|G|^{-1}\frac{\partial |G|}{\partial\theta^j} (g^{ij}+2\bi^{ij}-4\sigma^{ij} ) +2 \frac{\partial (g^{ij}+2\bi^{ij}-4\sigma^{ij})}{\partial\theta^j} \right)\frac{\pi_i}{\pi} \right] d\theta. \end{eqnarray*} Let $y=y(\theta)=(y_1(\theta),\dots,y_{d}(\theta))^T$, with $y_i(\theta)=\pi_i(\theta)/\pi(\theta)$, $i=1,\dots, {d}$, and let $y'$ be the matrix of partial derivatives of $y$ with respect to $\theta$, i.e.\ $y'=(y_{ij})_{ij}$, $y_{ij}=\partial y_i/\partial\theta^j$, $i,j=1,\dots { d}$. Then, $\pi_{ij}/\pi=y_{ij}+y_iy_j$ and the quantity to be maximized can be rewritten as \begin{eqnarray*} \lefteqn{\int_\Theta |G|^{1/2} \left[\frac{}{} \left(-3g^{ij} +4\bi^{ij}-4\sigma^{ij}\right)y_{ij}+\left(-g^{ij} +4\bi^{ij}-4\sigma^{ij}\right)y_iy_j\right.}\\ &&\left.+\left(3a^S_jg^{ij}+2a^\ell_j\bi^{ij}+|G|^{-1}\frac{\partial |G|}{\partial\theta^j}(g^{ij}+2\bi^{ij}-4\sigma^{ij} ) +2 \frac{\partial (g^{ij}+2\bi^{ij}-4\sigma^{ij})}{\partial\theta^j} \right)y_i \right] d\theta. \end{eqnarray*} Let us denote the integrand function by $U(\theta,y,y')$. The solution to the maximization problem is found by solving the system of Euler-Lagrange equations: \begin{equation*} \label{EL} \frac{\partial U}{\partial y_i}-\sum_{j=1}^d\frac{\partial}{\partial\theta^j}\left( \frac{\partial U}{\partial y_{ij}} \right)=0, \quad\quad i=1,\dots,{d}. \end{equation*} After some calculations, we obtain the solution to the variational problem as \begin{eqnarray*} y(\theta)=\frac{1}{4}\left[ 6a^S G^{-1} + 4a^\ell I^{-1}+ |G|^{-1}\frac{\partial |G|}{\partial\theta}(5G^{-1}-4\Sigma) +2G \nabla\cdot (5G^{-1}-4\Sigma) \right] \Gamma, \end{eqnarray*} where $a^S=(a^S_1,\dots,a^S_{d})^T$, $a^\ell=(a^\ell_1,\dots,a^\ell_{d})^T$, $\Sigma=(\sigma^{ij})_{ij}$, $\Gamma=(G^{-1}-4I^{-1}+4\Sigma)^{-1}$, $\nabla\cdot G^{-1}= (\partial g^{1j}/\partial\theta^j, \dots, \partial g^{{d} j}/\partial\theta^j)^T$ and $\nabla\cdot \Sigma= (\partial \sigma^{1j}/\partial\theta^j, \dots, \partial \sigma^{{d} j}/\partial\theta^j)^T$. \end{proof} \section*{References} \begin{description} \item Barndorff-Nielsen, O.E. (1976). Plausibility inference. {\em J.\ Roy.\ Statist.\ Soc.} B, {\bf 38}, 103--131. \item Basu, A., Harris, I.R., Hjort, N.L., Jones, M.C. (1998). Robust and efficient estimation by minimising a density power divergence. {\em Biometrika}, {\bf 85}, 549--559. \item Berger, J.O. (2006). The case for objective Bayesian analysis. {\em Bayesian Analysis}, {\bf 1}, 1--17. \item Berger, J.O., Bernardo, J.M. (1992). On the development of reference priors (with discussion). In {\em Bayesian Statistics} 4 (J.M.\ Bernardo, J.O.\ Berger, A.P.\ Dawid and A.F.M.\ Smith, eds.), 35--60. Oxford Univ.\ Press. \item Berger, J.O., Bernardo, J.M., Sun, D. (2009). The formal definition of reference priors. {\em Journal of the American Statistical Association}, {\bf 107}, 636--648. \item Berger, J.O., Bernardo, J.M., Sun, D. (2012). Overall objective priors. {\em Bayesian Analysis}, {\bf 10}, 189--221. \item Bernardo, J.M. (1979). Reference posterior distributions for Bayesian inference (with discussion). {\em J.\ Roy.\ Statist.\ Soc.} B, {\bf 41}, 113--147. \item Bernardo, J.M. (2005). Reference analysis. In: {\em Handbook of Bayesian Statistics}, {\bf 25} (D.K. Dey and C.R. Rao, eds.), North-Holland, Amsterdam: Elsevier, 17--90. \item Brier, G.W. (1950). Verification of forecasts expressed in terms of probability. {\em Mon. Weather Rev.}, {\bf 78}, 1--3. \item Chandler, R.E., Bate, S. (2007). Inference for clustered data using the independence loglikelihood. {\em Biometrika}, {\bf 94}, 167--183. \item Chang, I., Mukerjee, R. (2008). Bayesian and frequentist confidence intervals arising from empirical-type likelihoods. {\em Biometrika}, {\bf 95}, 139--147. \item Corcuera, J.M., Giummol\`e, F. (1998). A characterization of monotone and regular divergences. {\em Ann.\ Inst.\ Statist.\ Math.}, {\bf 50}, 433--450. \item Datta, G.S., Mukerjee R. (2004). {\em Probability Matching Priors: Higher Order Asymptotics}. Lecture Notes in Statistics 178. Springer. \item Dawid, A.P. (1986). Probability forecasting. In: {\em Encyclopedia of Statistical Sciences} (S. Kotz, N. L. Johnson, and C. B. Read eds.), 210--218. Wiley-Interscience. \item Dawid, A.P. (2007). The geometry of proper scoring rules. {\em Ann.\ Inst.\ Statist.\ Math.}, {\bf 59}, 77--93. \item Dawid, A.P., Lauritzen, S.L. (2005). The geometry of decision theory. In: {\em Proceedings of the Second International Symposium on Information Geometry and its Applications}, 22--28. University of Tokyo. \item Dawid, A.P., Musio, M. (2014). Theory and Applications of Proper Scoring Rules. {\em Metron}, {\bf 72}, 169--183. \item Dawid, A.P., Musio, M. (2015). Bayesian model selection based on proper scoring rules (with discussion). {\em Bayesian Analysis}, {\bf 10}, 479--521. \item Dawid, A.P., Musio, M., Ventura, L. (2016). Minimum scoring rule inference. {\em Scand.\ J.\ Statist.}, {\bf 43}, 123--138. \item Farcomeni, A., Ventura, L. (2012). An overview of robust methods in medical research. {\em Statistical Methods in Medical Research}, {\bf 21}, 111--133. \item Godambe, V.P. (1960). An optimum property of regular maximum likelihood estimation. {\em Ann.\ Math.\ Statist.}, {\bf 31}, 1208--1211. \item Good, I.J. (1952). Rational decisions. {\em J.\ Roy.\ Statist.\ Soc.} B, {\bf 14}, 107--114. \item Ghosh, M. (2011). Objective Priors: An Introduction for Frequentists. {\em Stat.\ Sci.}, {\bf 26}, 187--202. \item Ghosh, M., Basu, A. (2013). Robust estimation for independent non-homogeneous observations using density power divergence with applications to linear regression. {\em Electr.\ J.\ Statist.}, {\bf 7}, 2420--2456. \item Ghosh, M., Basu, A. (2016). Robust Bayes estimation using the density power divergence. {\em Ann.\ Inst.\ Stat.\ Math.}, {\bf 68}, 413--437. \item Ghosh, M., Mergel, V., Liu, R. (2011). A general divergence criterion for prior selection {\em Ann. Inst. Stat. Math.} 43--58. \item Ghosh, J.K., Mukerjee, R. (1991). Characterization of priors under which Bayesian and frequentist Bartlett corrections are equivalent in the multi-parameter case. {\em J.\ Mult.\ Anal.}, {\bf 38}, 385--393. \item Greco, L., Racugno, W., Ventura L. (2008). Robust likelihood functions in Bayesian inference. {\em J.\ Statist.\ Plann.\ Inf.}, {\bf 138}, 1258--1270. \item Hampel, F.R., Ronchetti, E.M., Rousseeuw, P.J., Stahel, W.A. (1986). {\em Robust Statistics: The Approach Based on Influence Functions}. Wiley, New York. \item Heritier, S., Cantoni, E., Copt, S., Victoria-Feser, M.P. (2009). {\em Robust methods in biostatistics}. Wiley, Chichester, UK. \item Huber, P.J., Ronchetti, E.M. (2009). {\em Robust Statistics}. Wiley, New York. \item Hyv\"arinen, A. (2005). Estimation of non-normalized statistical models by score matching. {\em Journal of Machine Learning Research}, {\bf 6}, 695--709. \item Hyv\"arinen, A. (2007). Some extensions of score matching. {\em Computational Statistics and Data Analysis}, {\bf 51}, 2499--2512. \item Jeffreys, H. (1961). {\em Theory of Probability}. Oxford University Press, New York. \item Lazar, N.A. (2003). Bayes empirical likelihood. {\em Biometrika}, {\bf 90}, 319--326. \item Leisen, F., Villa, C., Walker, S.G. (2017). On a global objective prior from score rules. arXiv: 1706.00599v1 \item Lin, L. (2006). Quasi Bayesian likelihood. {\em Statistical Methodology}, {\bf 3}, 444--455. \item Liu, R., Chakrabarti, A., Samanta, T., Ghosh, J.K., Ghosh, M. (2014). On divergence measures leading to Jeffreys and other reference priors. {\em Bayesian Analysis}, {\bf 9}, 331--370. \item Machete, R. (2013). Contrasting probabilistic scoring rules. {\em J.\ Statist.\ Plann.\ Inf.}, {\bf 143}, 1781--1790. \item Mameli, V., Musio, M., Ventura, L. (2017). Bootstrap adjustments of signed scoring rule root statistics. {\em Comm.\ Statist.\ - Simul.\ Comput.}, to appear. \item Mameli, V., Ventura, L. (2015). Higher-order asymptotics for scoring rules. {\em J.\ Statist.\ Plann.\ Inf.}, {\bf 165}, 13--26. \item Mardia, K.V., Jupp, P.E. (2000). Directional statistics. Wiley, London. \item Mardia, K.V., Kent, J.T., Laha, A.K. (2016). Score matching estimators for directional distributions. . \item Musio, M., Mameli, V., Ruli, E., Ventura, L. (2017). Bayesian Inference for directional data through ABC and homogeneous proper scoring rules. {\em Proceeding of 61$st$ ISI World Statistics Congress}, to appear. \item Pace, L., Salvan, A., Sartori, N. (2011). Adjusting composite likelihood ratio statistics. {\em Statist.\ Sin.}, {\bf 21}, 129--148. \item Parry, M., Dawid, A.P., Lauritzen, S.L. (2012). Proper local scoring rules. {\em Ann.\ Statist.}, {\bf 40}, 561--592. \item Pauli, F., Racugno, W., Ventura, L. (2011). Bayesian composite marginal likelihoods. {\em Statist.\ Sin.}, {\bf 21}, 149--164. \item Riani, M., Atkinson, A.C., Perrotta, D. (2014). A parametric framework for the comparison of methods of very robust regression. {\em Statistical Science}, {\bf 29}, 128--143. \item Ribatet, M., Cooley, D. and Davison, A.C. (2012). Bayesian inference from composite likelihoods, with an application to spatial extremes. {\em Statist.\ Sin.}, {\bf 22}, 813--845. \item Ruli, E., Sartori, N., Ventura, L. (2016). Approximate Bayesian Computation with composite score functions. {\em Statist.\ Comp.}, {\bf 26}, 679--692. \item Schennach, S. (2005). Bayesian exponentially tilted empirical likelihood. {\em The Annals of Statistics}, {\bf 35}, 634--672. \item Smith, E.L., Stephenson, A.G. (2009). An extended Gaussian max-stable process model for spatial extremes. {\em J.\ Statist.\ Plann.\ Inf.}, {\bf 139}, 1266--1275. \item Tsallis, C. (1988). Possible generalization of Boltzmann-Gibbs statistics. {\em J.\ Statist.\ Physics}, {\bf 52}, 479--487. \item Varin, C., Reid, N. and Firth, D. (2011). An overview of composite likelihood methods. {\em Statist.\ Sin.}, {\bf 21}, 5--42. \item Ventura, L., Cabras, S., Racugno, W. (2010). Default prior distributions from quasi- and quasi-profile likelihoods. {\em J.\ Statist.\ Plann.\ Inf.}, {\bf 140}, 2937--2942. \item Ventura, L., Racugno, W. (2016). Pseudo-likelihoods for Bayesian inference. In: {\em Topics on Methodological and Applied Statistical Inference, Series Studies in Theoretical and Applied Statistics}, Springer-Verlag, 205--220. \item Walker, S.G. (2016). Bayesian information in an experiment and the Fisher information distance. {\em Statistics and Probability Letters}, {\bf 112}, 5-9. \item Yang, Y., He, X. (2012). Bayesian empirical likelihood for quantile regression. {\em The Annals of Statistics}, {\bf 40}, 1102--1131. \end{description} \end{document}

\section{Introduction}\label{intro} A {\em universal cycle}, or {\em u-cycle}, for a class of combinatorial objects is a cyclic word containing each object, encoded by a word, exactly once as a cyclic interval (called a {\em factor}). Universal cycles were introduced by Chung, Diaconis and Graham in~\cite{CDG1992}. However, the origin of the notion of a u-cycle is in the study of the celebrated de Bruijn cycles, which are cyclic words containing each word of length $n$ over some alphabet, for a given $n$, exactly once (see~\cite{deBruijn}). There are many ways to construct a de Bruijn cycle. A ``classical'' way to construct a de Bruijn cycle is via the notion of a {\em de Bruijn graph of order}~$n$, which is a directed graph representing overlaps between words of length~$n$. It can be easily shown that any de Bruijn graph is {\em balanced} and {\em strongly connected}, and thus it contains an {\em Eulerian cycle}. Moreover, one can show that a de Bruijn graph of order~$n$ is the {\em line graph} of the de Bruijn graph of order $n-1$, and thus any de Bruijn graph contains a {\em Hamiltonian cycle}, which can easily be translated to a de Bruijn cycle. \subsection{Martin's algorithm to generate de Bruijn sequences} An alternative construction of a de Bruijn sequence involves concatenating together, in lexicographic order, all the {\em Lyndon words} whose length divides $n$. An interesting, and most relevant to our paper fact is that this sequence was first generated by Martin~\cite{Martin1934} in 1934 using the following greedy algorithm. \medskip \begin{mdframed} \centerline{\bf Martin's greedy algorithm to generate a de Bruijn sequence} \medskip Start with the word $(k-1)^{n-1}$ and then repeatedly apply the following rule: Append the {\em smallest} letter in $\{0,1,\ldots,k-1\}$ so that factors of length $n$ in the resulting word are distinct. Once no more extension is possible, remove the $n-1$ rightmost letters. \end{mdframed} For example, for $k=3$ and $n=2$, the steps of the algorithm are: $2\rightarrow 20 \rightarrow 200 \rightarrow 2001 \rightarrow 20010 \rightarrow 200102 \rightarrow 2001021 \rightarrow 20010211 \rightarrow 200102112\rightarrow 20010211$. \subsection{Universal cycles for permutations} In this paper, permutations of length $n$ are called {\em $n$-permutations}. Universal cycles for permutations were one of the main objects considered in~\cite{CDG1992}. A~universal cycle for $n$-permutations is a cyclic word, each length $n$ factor of which is order-isomorphic to a unique $n$-permutation, and every $n$-permutation is order-isomorphic to some factor in the word. There is a long line of research on universal cycles for permutations, e.g.\ see \cite{AW2009,HRW2010,HRW2012,HI1996,I2006,J1993,Johnson2009,RW2010}. Even though the basic original idea to construct universal cycles for permutations is similar to that of constructing de Bruijn cycles via de Bruijn graphs, the situation with permutations is much more complicated. Indeed, just proving the existence of such a universal cycle uses a not so trivial trick of {\em clustering} the {\em graph of non-overlapping permutations}, an analogue of the de Bruijn graph in the case of permutations. The actual construction of universal cycles for permutations in~\cite{CDG1992} involved dealing with partially ordered sets and resulted in a nice conjecture on the minimal alphabet that was settled 17 years later by Johnson in~\cite{Johnson2009}. Namely, it turns out that it is possible to construct a universal cycle for $n$-permutations using just $n+1$ letters; using $n$ letters is impossible (unless using the {\em shorthand encoding of $n$-permutations by $(n-1)$-permutations, which is deleting the rightmost element in each $n$-permutation; see~\cite{HRW2012}}) as can be easily seen, e.g.\ in the case of $n=3$. We found the existent constructions of the universal cycles for permutations in the literature to be rather involved, and we asked ourselves whether the simple algorithm by Martin for de Bruijn cycles could be extended to the case of the permutations. We were pleased to discover that the answer to the question was positive with surprisingly small changes that were required to Martin's algorithm: instead of the initial word $(k-1)^{n-1}$ we use the increasing permutation of length $n-1$, and instead of the smallest available letter on a given iteration of the process we use the smallest available extension of the respective permutation, to be defined in Section~\ref{gener-sec}. Interestingly, beginning with a monotone permutation, that is, the increasing or decreasing permutation of appropriate length, is a necessary condition in our greedy algorithm, as we show in Section~\ref{unique-sub}. \subsection{Basic definitions} For a permutation, or word, $\pi$, the {\em reduced form} of $\pi$, denoted $\red(\pi)$, is obtained by replacing the $i$-th smallest element in $\pi$ by $i$. For example, $\red(3275)=2143$. Any $i$ consecutive letters of a word or permutation $w$ form a {\em factor} of $w$. If $w$ is a cyclic word then a factor can begin at the end of $w$ and end at the beginning of $w$. If $u$ is a factor of $w$, we also say that $w$ {\em covers} $\red(u)$. \begin{deff} A word $\Pi'_n$ is a {\em universal word}, or {\em u-word}, for $n$-permutations if $\Pi'_n$ covers {\em every} $n$-permutation {\em exactly once}. \end{deff} \begin{deff} A cyclic word $\Pi_n$ is a {\em universal cycle}, or {\em u-cycle}, for $n$-permutations if $\Pi_n$ covers {\em every} $n$-permutation {\em exactly once}. \end{deff} \subsection{Organization of the paper} This paper is organized as follows. In Section~\ref{gener-sec} we show our simple way to generate the words $\Pi'_n$ and $\Pi_n$, and in Sections~\ref{just-1} and~\ref{just-2} we justify that the constructions of $\Pi'_n$ and $\Pi_n$ work. Also, in Section~\ref{unique-sub} we show that our greedy algorithm to generate $\Pi'_n$ and $\Pi_n$ can only be applied to the increasing permutation of length $n-1$. Further, Section~\ref{prop-sec} discusses some properties of $\Pi'_n$ and $\Pi_n$, and Section~\ref{further-res-sec} gives several directions of future research. This includes a discussion on reducing the alphabets of $\Pi'_n$ and $\Pi_n$ in Section~\ref{reducing-sec} and a possibility of extending our approach to relevant object in the literature in Section~\ref{sec-shortened}. \section{Simple generation of u-words and u-cycles for permutations} \subsection{Generating $\Pi'_n$ and $\Pi_n$}\label{gener-sec} Let $\pi = \pi_1\pi_2\cdots \pi_m$ be a permutation of $m$ distinct integers. Then the {\em $i$-th extension} of $\pi$ to the right, $1\leq i\leq m$, is the permutation $$c_b(\pi_1)c_b(\pi_2)\cdots c_b(\pi_m)b,$$ where $b$ is the $i$-th smallest element in $\{\pi_1,\pi_2,\ldots,\pi_m\}$, and $$c_b(x)= \begin{cases} x & \mbox{if }x<b, \\ x+1 & \mbox{if }x\geq b. \end{cases} $$ The {\em $(m+1)$-st extension} is the permutation $\pi b$, where $b$ is the largest element in $\{\pi_1+1,\pi_2+1,\ldots,\pi_m+1\}$. We call the first extension the {\em smallest extension}, and the $(m+1)$-st extension the {\em largest extension} of~$w$. The following simple algorithm produces recursively a universal word $\Pi'_n$ of length $n!+n-1$ for $n$-permutations. \medskip \begin{mdframed} \centerline{\bf The greedy algorithm to construct $\Pi'_n$} \medskip Begin with the increasing permutation $$\Pi'_{n,0}:=12\cdots(n-1).$$ Suppose that a permutation $$\Pi'_{n,k}=a_{1}a_{2}\cdots a_{k+n-1}$$ has been constructed for $0 \leq k < n!$, and no two factors in $\Pi'_{n,k}$ of length $n$ are order-isomorphic. Let $i$ be minimal such that no factor of length~$n$ in $\Pi'_{n,k}$ is order-isomorphic to the $i$-th extension of $a_{k+1}a_{k+2}\cdots a_{k+n-1}$, and denote the last element of this extension by~$b$. Then $$\Pi'_{n,k+1}:=c_b(a_1)c_b(a_2)\cdots c_b(a_{k+n-1})b.$$ For some $k^*$, no extension of $\Pi'_{n,k^*}$ will be possible without creating a factor order-isomorphic to a factor in $\Pi'_{n,k}$. The greedy algorithm then terminates and outputs $\Pi'_n:=\Pi'_{n,k^*}$. \end{mdframed} \medskip\noindent For example, the steps of the algorithm for $n=3$ are as follows: $$12 \rightarrow 231 \rightarrow 3421 \rightarrow 45312 \rightarrow 564132 \rightarrow 6751324 \rightarrow 78613245 = \Pi'_3.$$ We note that for each $k$, $\Pi'_{n,k}$ is a permutation of $\{1,2,\ldots,k+n-1\}$. The following simple extension of the greedy algorithm turns the universal word for permutations $\Pi'_n$ into a universal cycle for permutations $\Pi_n$. \medskip \begin{mdframed} \centerline{\bf Generating the u-cycle $\Pi_n$ from $\Pi'_n$} \medskip Remove the last $n-1$ elements in $\Pi'_n$ and take the reduced form of the resulting sequence to obtain $\Pi_n$. \end{mdframed} \medskip\noindent For example, $\Pi_3$ is given by $$\Pi'_3=78613245 \rightarrow 786132\rightarrow \red(786132)= 564132=\Pi_3.$$ For another example, $\Pi_4$ is given by $$22\ 23\ 24\ 21\ 20\ 18\ 19\ 3\ 17\ 4\ 2\ 16\ 1\ 6\ 7\ 5\ 11\ 10\ 8\ 13\ 9\ 12\ 15\ 14.$$ \subsection{Justification of $\Pi'_n$ being a u-word for permutations}\label{just-1} In the following, let $n$ be arbitrary but fixed. For $\Pi'_{n,k} = a_{1}a_{2}\cdots a_{k+n-1}$, set $$\sigma_{k} := \red(a_{k}a_{k+1}\cdots a_{k+n-1}), \qquad \sigma'_{k} := \red(a_{k+1}a_{k+2}\cdots a_{k+n-1}),$$ and $$J_{k} = |\{j\le k:\, \sigma'_{j} = \sigma'_{k}\}|,$$ i.e.\ $J_{k}$ is the number of occurrences of the $(n-1)$-permutation $\sigma'_{k}$ in~$\Pi'_{n,k}$. By the definition of the greedy algorithm, all $i$-th extensions of~$\sigma'_{k}$ with $i < J_k$ occur in $\Pi'_{n,k}$, and the $J_k$-th extension of $\sigma'_{k}$ does not occur in~$\Pi'_{n,k}$. Therefore, the greedy algorithm terminates at~$k$ if and only if $J_k = n+1$. If $J_k \le n$, then $\sigma_{k+1}$ is the $J_k$-th extension of $\sigma'_{k}$ (and ends with~$J_k$). \begin{lem}\label{lem2} The greedy algorithm terminates at~$k$ if and only if $\sigma_{k} {\,=\,}12\cdots n$. \end{lem} \begin{proof} If $\sigma_{k} = 12\cdots n$, then $\sigma'_{k-1} = \sigma'_{k} = 12\cdots(n-1)$ and $J_{k-1} = n$. Thus, $J_k = n+1$, and the greedy algorithm terminates at~$k$. For the converse, assume that $\sigma_{k} \ne 12\cdots n$. By the preceding paragraph, this implies that $\sigma_{j} \ne 12\cdots n$ for all $j \le k$. Since each word $a_{j+1} \cdots a_{j+n-1}$, $1 \le j \le k$, is preceded by the letter $a_{j}$ and the permutation $\sigma_j$ occurs only at the position~$j$, we have $\sigma'_j = \sigma'_k$ for at most $n$ different indices $j \ge 1$. If $\sigma'_{k} \ne 12\cdots (n-1)$, then we have $\sigma'_{0} \ne \sigma'_{k}$ and thus $J_k \le n$. If $\sigma'_{k} = 12\cdots (n-1)$, then we have $\sigma'_j = \sigma'_k$ for at most $n-1$ different indices $j \ge 1$ (because $\sigma_{j} = 12\cdots n$ is not possible), which also gives that $J_k \le n$. Therefore, the greedy algorithm does not terminate at~$k$ if $\sigma_{k} \ne 12\cdots n$. \end{proof} \begin{lem}\label{lem3} $\Pi'_n$ covers all $n$-permutations. \end{lem} \begin{proof} We shall prove this lemma by contradiction. Suppose that a permutation $\pi_1 \pi_2 \cdots \pi_{n}$ is not covered by $\Pi'_n$. Then $\pi_2 \cdots \pi_{n}$ is covered at most $n-1$ times, hence $\red(\pi_2 \cdots \pi_{n})n$ is not covered. More generally, we have for $1 \le k \le n$ that, if $\red(\pi_k \cdots \pi_{n}) (n-k+2) \cdots n$ is not covered, then $$\red(\red(\pi_{k+1} \cdots \pi_{n}) (n-k+2) \cdots n)n = \red(\pi_{k+1} \cdots \pi_{n}) (n-k+1) \cdots (n-1)n$$ is not covered. We obtain that $\Pi'_n$ does not cover the permutation $1 2\cdots n$, contradicting Lemma~\ref{lem2}. \end{proof} \begin{thm}\label{u-word-thm} $\Pi'_n$ is a u-word for $n$-permutations. \end{thm} \begin{proof} By the nature of the greedy algorithm, $\Pi'_n$ cannot cover a permutation more than once. By Lemma~\ref{lem3}, $\Pi'_n$ covers all $n$-permutations. \end{proof} \subsection{Justification of \texorpdfstring{$\Pi_n$}{} being a u-cycle for permutations}\label{just-2} \begin{thm} $\Pi_n$ is a u-cycle for $n$-permutations. \end{thm} \begin{proof} By Theorem~\ref{u-word-thm}, it suffices to prove that $$\red(a_{k}a_{k+1} \cdots a_{k+n-1}) = \red(a_{k}\cdots a_{n!} a_1 \cdots a_{k+n-n!-1})$$ for all $n!-n+2 \le k \le n!$. First, note that $\sigma_i$ ends with $1$ for all $i < n$. Hence, we have that $$\Pi'_{n,n-1} = n(n+1)\cdots(2n-2) (n-1)(n-2)\cdots1$$ and thus $$a_k < a_1 < a_2 < \cdots < a_{n-1}$$ {for all} $k \ge n$. Next, we show that $\sigma_i$ ends with $n$ for all $i \ge n!-n+2$. Suppose that this is not true for some $i \ge n!-n+2$. Since $\Pi'_n$ is a u-word, we must have $\sigma'_{i-1} = \sigma'_{j-1}$ for some $j > i$. It follows from $\sigma_{n!} = 12\cdots n$ that $$a_{n!} < a_{n!+1} < \cdots < a_{n!+n-1}.$$ In particular, $$a_{n!} < a_{n!+1} < \cdots < a_{j+n-2}.$$ Then, $\sigma'_{i-1} = \sigma'_{j-1}$ implies that $$a_{n!+i-j} < a_{n!+i-j+1} < \cdots < a_{i+n-2}< \cdots < a_{n!+n-1}.$$ Iterating this argument gives that $a_{j} < a_{j+1} < \cdots < a_{n!+n-1}$ and thus $a_{i} < a_{i+1} < \cdots < a_{n!+n-1}$, contradicting the assumption $\sigma_i$ does not end with~$n$. Therefore, $\sigma_i$ ends with $n$ for all $i \ge n!-n+2$, which implies that $$a_k < a_{n!+1} < a_{n!+2} < \cdots < a_{n!+n-1}$$ for all $n!-n+2 \le k \le n!$. It follows that $$\sigma_{k} = \red(a_{k}a_{k+1} \cdots a_{n!}) (n!-k+2) \cdots n = \red(a_{k}\cdots a_{n!} a_1 \cdots a_{k+n-n!-1})$$ for $n!-n+2 \le k \le n!$, which proves the theorem. \end{proof} \subsection{Uniqueness of $\Pi_n$}\label{unique-sub} The construction of $\Pi_n$ begins with the permutation $12\cdots (n-1)$ and is followed by consecutive applications of the greedy algorithm choosing the smallest possible element to extend the already constructed sequence. A~natural question is what happens if we start with a different permutation of length $n-1$ and follow the greedy steps of our algorithm? Can we obtain other (non-equivalent) constructions of u-cycles for $n$-permutations? The following theorem answers this question. \begin{thm}\label{uniqueness-thm} Beginning the construction of $\Pi_n$ {\rm (}and thus of $\Pi'_n${\rm )} with an $(n-1)$-permutation $\Pi'_{n,0}\neq 12\cdots (n-1)$ and following the steps of the greedy algorithm will {\em not} produce a $u$-cycle for $n$-permutations. \end{thm} \begin{proof} Assume that $\sigma'_0 \neq 12\cdots(n-1)$. As in the proof of Lemma~\ref{lem2}, the greedy algorithm terminates at~$k$ if $\sigma_k = 12\cdots n$. Now, the second part of the proof of Lemma~\ref{lem2} shows that $J_k < n$ when $\sigma'_k = 12\cdots(n-1)$. Therefore, the permutation $12\cdots n$ is not covered by $\Pi'_n$, hence $\Pi'_n$ is not a u-word and $\Pi_n$ is not a u-cycle. \end{proof} \section{Properties of $\Pi_n$ and $\Pi'_n$}\label{prop-sec} In this section, we shall give some properties of $\Pi_n$ and $\Pi'_n$. All our results discussed are stated in terms of $\Pi'_n$, since the same results for $\Pi_n$ will trivially follow from the definition of $\Pi_n$. Our main result is that all permutations with $1$ before $n$ occur in the first half of $\Pi'_n$, and those with $n$ before $1$ occur in the second half. More precisely, consider the partition of the symmetric group $S_n = S_{1,n} \cup S_{n,1}$, where \begin{align*} S_{1,n} & = \{\pi_1\pi_2\cdots \pi_n \in S_n:\ i < j\ \mbox{for}\ \pi_i = 1,\ \pi_j = n\}, \\[6pt] S_{n,1} & = \{\pi_1\pi_2\cdots \pi_n \in S_n:\ i < j\ \mbox{for}\ \pi_i = n,\ \pi_j = 1\}. \end{align*} \begin{thm} \label{properties-thm} We have $a_{\frac{n!}{2}+1} = 1$, \begin{gather*} \sigma_{\frac{n!}{2}} = n12\cdots(n-1), \quad \sigma_{\frac{n!}{2}+1} = 134\cdots n2, \quad \sigma_{\frac{n!}{2}+2} = 23\cdots (n-1)1n, \\ \left\{\sigma_k:\, k \le \frac{n!}{2}\right\} = S_{n,1} \quad \mbox{and} \quad \left\{\sigma_k:\, k > \frac{n!}{2}\right\} = S_{1,n}. \end{gather*} \end{thm} In particular, $\sigma_{\frac{n!}{2}} = n12\cdots(n-1)$ is the last occurrence of an $n$-permutation starting with~$n$, $\sigma_{\frac{n!}{2}+1} = 134\cdots n2$ is the first occurrence of an $n$-permutation starting with~$1$, and $\sigma_{\frac{n!}{2}+2} = 23\cdots (n-1)1n$ is the first occurrence of an $n$-permutation ending with~$n$. For the proof of Theorem~\ref{properties-thm}, we use the following lemmas. \begin{lem} \label{l:am} Let $m \ge 1$ be minimal such that $\sigma_m$ starts with $1$. Then $a_m = 1$, and $\sigma_k$ does not end with $n$ for all $k \le m$. \end{lem} \begin{proof} The fact that $a_m = 1$ is a direct consequence of the greedy algorithm. Suppose that $\sigma_k$ ends with $n$ for some $k \le m$. Then we have $\sigma'_j = \sigma'_{k-1}$ for $n$ different indices $j < k$. Since $\sigma_k \ne 12\cdots n$, we have that $j \ge 1$ for all these~$j$. Hence, there are $n$ different permutations $\sigma_j$ ending with~$\sigma'_{k-1}$. Therefore, one of these permutations starts with $1$, contradicting the definition of~$m$. \end{proof} \begin{lem} \label{l:Sn1} Let $m$ be as in Lemma~\ref{l:am}. We have $\sigma_k \in S_{n,1}$ for all $k < m$, $\sigma_{m-1} = n12\cdots(n-1)$ and $\sigma_m = 134\cdots n2$. \end{lem} \begin{proof} We use induction on~$k$. We have $\sigma_1 = 23\cdots n1 \in S_{n,1}$. Assume that $\sigma_j \in S_{n,1}$ for all $j \le k$, where $1 \le k < m$. If $\sigma'_k \in S_{n-1,1}$, then we have $\sigma_{k+1} \in S_{n,1}$ because $\sigma_{k+1}$ does not end with~$n$ by Lemma~\ref{l:am}. Now, assume that $\sigma'_k \in S_{1,n-1}$. Since $\sigma_k \in S_{n,1}$, $\sigma_k$ starts with~$n$. Moreover, we cannot have $\sigma'_j = \sigma'_k$ for $1 \le j < k$, since otherwise it implies that $\sigma'_j \in S_{1,n-1}$. Together with $\sigma_j \in S_{n,1}$, it follows that $\sigma_j$ would also have to start with $n$, contradicting that $\sigma_j \ne \sigma_k$. If $\sigma'_k \ne \sigma'_0$, then this implies that $\sigma_{k+1}$ ends with~$1$ and hence $\sigma_{k+1} \in S_{n,1}$. Finally, if $\sigma'_k = \sigma'_0$, then $\sigma_{k+1}$ is an $i$-th extension of $12\cdots(n-1)$ with $i \ge 2$, thus $\sigma_{k+1}$ starts with~$1$, which is possible only for $k=m-1$. Therefore, we have $\sigma_k \in S_{n,1}$ for all $k < m$. We have $\sigma_m \in S_{1,n}$ because it starts with~$1$, hence we obtain from the previous paragraph that $\sigma'_{m-1} = \sigma'_0$, $\sigma_{m-1} = n12\cdots(n-1)$ and $\sigma_m = 134\cdots n2$. \end{proof} \begin{lem} \label{l:S1n} Let $m$ be as in Lemma~\ref{l:am}. We have $\sigma_k \in S_{1,n}$ for all $k \ge m$, with $\sigma_{m+1} = 23\cdots (n-1)1n$. \end{lem} \begin{proof} Suppose that $\sigma_k \in S_{n,1}$ for some $k \ge m$. Then, we will show that, for some $i \ge m$, $\sigma_i$ starts with $n$ and ends with $n-1$. We obtain that, for some $i \ge m$, $\sigma_i$ ends with $n1$, and finally this yields that $n12\cdots (n-1)$ occurs after~$m$, contradicting that $\sigma_{m-1} = n12\cdots(n-1)$. First, note that $\sigma'_k = \sigma'_0$ and $\sigma_k \in S_{n,1}$ are possible only for $k = m-1$. Therefore, $\sigma_k \in S_{n,1}$ for $k \ge m$ implies that $\sigma'_j = \sigma'_k$ for at most $n$ different indices $j \le k$. Moreover, since $\sigma_j$ cannot start with~$1$ if $j < m$ or $j = k$, and since $k \ge m$, there are at most $n-2$ different indices $j < m$ such that $\sigma'_j = \sigma'_k$. Hence the $(n-1)$-st extension of $\sigma'_k$ occurs after~$m$. If $\sigma_k \in S_{n,1}$ does not start with~$n$, then its $(n-1)$-st extension is also in~$S_{n,1}$, with the position of $n$ shifted to the left; we obtain iteratively that there is some $i \ge m$ such that $\sigma_i$ starts with $n$ and ends with $n-1$. On the other hand, if $\sigma_k$ starts with~$n$, then the $(n-1)$-st extension of $\sigma'_{k-1}$ starts with $n$ and occurs at or after the position~$k$. Therefore, we can assume that, for some $k \ge m$, $\sigma_k$ starts with $n$ and ends with $n-1$. Now, we cannot have $\sigma'_j = \sigma'_k$ for some $1 \le j < m$, because $\sigma_j \ne \sigma_k$ would imply that $\sigma_j$ ends with $n$, which is not possible for $j < m$. Since $\sigma'_k \ne \sigma'_0$, we obtain that $\sigma'_k$ does not occur before~$m$, hence the first extension of $\sigma'_k$, which ends with $n1$, occurs after~$m$. Finally, assume that $\sigma_k$ ends with $n1$ for some $k \ge m$. Then, iterating $(n-1)$-st extensions gives that $n12\cdots(n-1)$ occurs after~$m$, contradicting that $\sigma_{m-1} = n12\cdots(n-1)$. Hence, we cannot have $\sigma_k \in S_{n,1}$ for $k \ge m$. Since $\sigma'_m = 23\cdots (n-1)1$ by Lemma~\ref{l:Sn1} and $\sigma_{m+1} \in S_{1,n}$, we have $\sigma_{m+1} = 23\cdots (n-1)1n$. \end{proof} \begin{proof}[Proof of Theorem~\ref{properties-thm}] Since $|S_{1,n}| = |S_{n,1}| = n!/2$, the theorem is a direct consequence of Lemmas~\ref{l:Sn1} and~\ref{l:S1n}, with $m = n!/2+1$. \end{proof} Note that, since $a_{\frac{n!}{2}+1} = 1$, we have, for all $k \le n!/2-n+2$, some $0 \le j \le n-2$ such that $\sigma_{k+j}$ ends with~$1$. This also follows from the fact that $\sigma_k \in S_{n,1}$ for all $k \le n!/2$. Similarly, since $\sigma_k \in S_{1,n}$ for all $k > n!/2$, we have, for all $k > n!/2$, some $0 \le j \le n-2$ such that $\sigma_{k+j}$ ends with~$n$. Therefore, $a_{n!+n-1}$ is the largest element to the right of $a_{\frac{n!}{2}+1} = 1$. \section{Directions of further research}\label{further-res-sec} \subsection{Reducing the alphabets of $\Pi'_n$ and $\Pi_n$}\label{reducing-sec} Our main goal in this paper was to come up with a {\em simple} construction of a u-cycle for permutations, {\em not} a construction requiring the {\em minimum} number of letters, which was a concern in, say~\cite{CDG1992,HRW2012,Johnson2009}. However, it is an interesting and challenging question to ask what is the minimum number of distinct letters that is required for our greedy construction. These considerations will be similar in nature to dealing with partially ordered sets in \cite{CDG1992}. For example, constructing $\Pi'_3$ in a greedy way with respect to the number of letters used, we end up with a word over a 6-letter alphabet, not over an 8-letter alphabet: \begin{align*} 12 & \rightarrow 231 \rightarrow 3421 \rightarrow 45312 \rightarrow 564132 \rightarrow 5641324 \rightarrow 56413245. \end{align*} Such an optimization is irrelevant for the construction of $\Pi_3$, but it would be relevant for the construction of $\Pi_n$ for $n\geq 4$. So, the idea is that in the original construction of $\Pi'_n$ some of the letters can be used more than once. This defines a partial order on the letters of the originally constructed $\Pi'_n$ indicating which letters have a choice to be used more than once. More precisely, the poset is defined by $a_j \prec a_k$ if $a_j < a_k$ and $|k-j|<n$, where $\prec$ denotes the poset relation. Then the minimal number of letters is the length of this partial order. Such partially ordered sets, in the cases of $n=3,4$, are shown in Figure~\ref{posets-n=3-4}, and we can see that $\Pi'_3$ (resp., $\Pi'_4$) actually requires 6 (resp., 13) distinct letters. It would be interesting to study properties of such partially ordered sets. For example, from the case of $n=4$, we see that such partially ordered sets are not necessarily graded. \begin{figure}[ht] \centering \begin{tikzpicture}[scale=.9] \node (1) at (0,0) [draw] {$1$}; \node (2) at (0,2) [draw] {$2$}; \node (3) at (0,4) [draw] {$3$}; \node (4) at (1,6) [draw] {$4$}; \node (5) at (1,8) [draw] {$5$}; \node (6) at (-1,6) [draw] {$6$}; \node (7) at (-1,8) [draw] {$7$}; \node (8) at (-1,10) [draw] {$8$}; \draw (1)--(2)--(3)--(4)--(5) (3)--(6)--(7)--(8); \begin{scope}[shift={(7,0)}] \node (1) at (0,0) [draw] {$1$}; \node (2) at (-1,1) [draw] {$2$}; \node (3) at (-2,2) [draw] {$3$}; \node (4) at (-2,3) [draw] {$4$}; \node (5) at (1,1) [draw] {$5$}; \node (6) at (0,2) [draw] {$6$}; \node (7) at (0,3) [draw] {$7$}; \node (8) at (2,2) [draw] {$8$}; \node (9) at (2,3) [draw] {$9$}; \node (10) at (0,4) [draw] {$10$}; \node (11) at (0,5) [draw] {$11$}; \node (12) at (2,4) [draw] {$12$}; \node (13) at (0,6) [draw] {$13$}; \node (14) at (2,5) [draw] {$14$}; \node (15) at (0,7) [draw] {$15$}; \node (16) at (0,8) [draw] {$16$}; \node (17) at (0,9) [draw] {$17$}; \node (18) at (0,10) [draw] {$18$}; \node (19) at (-2,4) [draw] {$19$}; \node (20) at (-2,5) [draw] {$20$}; \node (21) at (-2,6) [draw] {$21$}; \node (22) at (-2,7) [draw] {$22$}; \node (23) at (-2,8) [draw] {$23$}; \node (24) at (-2,9) [draw] {$24$}; \node (25) at (-2,10) [draw] {$25$}; \node (26) at (-2,11) [draw] {$26$}; \node (27) at (-2,12) [draw] {$27$}; \draw (1)--(2)--(3)--(4)--(19)--(20)--(21)--(22)--(23)--(24)--(25)--(26)--(27) (1)--(5)--(6)--(7)--(19) (2)--(6) (5)--(8)--(9)--(10)--(11)--(13)--(15)--(16)--(17)--(18) (7)--(10) (9)--(12)--(13) (12)--(14)--(15); \end{scope} \end{tikzpicture} \caption{Partially ordered sets for $\Pi'_3$ and $\Pi'_4$}\label{posets-n=3-4} \end{figure} We note that our greedy algorithm will never result in the optimum size of the alphabet, which is $n+1$ for $n$-permutations, because of the first $n$ steps of the algorithm when we are forced to use new letters, so the total number of letters will be $\geq 2n-2>n+1$ for $n\geq 4$. However, to find the exact optimal size for our construction is an interesting problem requiring understanding the partially ordered sets in question. If constructing a universal cycle for $n$-permutations over an $(n+1)$-letter alphabet is someone's main goal, it would be interesting to see if there is a (natural) modification of our algorithm that would help to achieve this goal (thus reproving the result of Johnson~\cite{Johnson2009}). Such a modification would be in describing (in a ``nice'' way) a sequence of extensions instead of applying the minimal possible extension every time. For example, we might be interested in alternation of applications of minimal and maximal extensions, or applying some sort of closest to average extension, etc. Of course, many, if not (almost) all such modifications will not result in a universal cycle. This effectively opens up the question on classification of sequences of extensions that result in universal cycles (leaving aside the question on the number of letters used by a particular universal cycle obtained this way). \subsection{Shortened universal words/cycles for permutations}\label{sec-shortened} It would be interesting to extend our approach to generate universal cycles for permutations in the context of shortened universal words/cycles for permutations that were introduced in \cite{KPV2017}. To illustrate the idea of one of the two ways suggested in \cite{KPV2017} to shorten universal cycles for permutations, consider the word $112$, which is claimed to be a universal cycle for all permutations of length 3, thus shortening a ``classical'' universal cycle for these permutations, say, $145243$. Indeed, we can treat equal elements as {\em incomparable elements}, while the relative order of these incomparable elements to the other elements must be respected. Thus, $112$ encodes all permutations whose last element is the largest one, namely, $123$ and $213$; starting at the second position (and reading the word cyclically), we obtain the word $121$ encoding the permutations $132$ and $231$, and finally, starting at the third position, we (cyclically) read the word $211$ encoding the permutations $312$ and $321$. For another example, the word $1232$ is also a universal cycle for permutations of length $3$. The main goal of \cite{KPV2017} is to study compression possibilities for (classical) universal cycles for permutations. In particular, \cite{KPV2017} shows that such universal cycles exist of lengths $n!-kn$ for $k=0,1,\ldots,(n-1)!$. The approach in \cite{KPV2017} to obtain these results is graph theoretical, and it actually does not provide any explicit constructions of the objects in the general case. Thus, (a~modification of) our approach to generate universal cycles could potentially be useful in generating shortened universal words/cycles, and this is an interesting direction of further research. An idea here would be to introduce a rule, or rules, overriding the rules of the algorithm, while still beginning with the increasing sequence of length $(n-1)$ for $n$-permutations. For example, we could require to use the minimum letter equal to some letter among the $(n-1)$ preceding letters, if possible, while following our original algorithm otherwise. So that the steps of such a modified algorithm in the case of $n=3$ could be $$12 \rightarrow 121 \rightarrow 1211 \rightarrow 12112$$ covering all 6 permutations non-cyclicly, or for the cyclic version, we can stop at 121 (132 and 231 are covered by 121, 321 and 312 by 211, and 123 and 213 by 112). However, correctness of such an algorithm, or its variations, must be addressed, which is outside of the scope in this paper. \section*{Acknowledgments} The first author was supported by the Fundamental Research Funds for the Central Universities (31020170QD101). The second author is grateful to the administration of the Center for Combinatorics at Nankai University for their hospitality during the author's stay in May 2017. The last author was partially supported by the National Science Foundation of China (Nos. 11626172, 11701424).

\section{Introduction} \label{} In the last few years, the semileptonic decays $\bar{B}^0 \to D^{(\ast)} \tau^- \bar{\nu}_{\tau}$ have been widely discussed in the literature as candidates for testing the Standard Model (SM) and searching for possible new physics (NP) in charged-current interactions. At $B$ factories, the Belle and \textit{BABAR} Collaborations have been continuously updating their measurements with better precision based on electron-positron colliders. Recently, the LHCb Collaboration has also entered the game with data taken at the LHC hadron collider. The three groups have reported measurements of the ratios in Refs.~\cite{, }. These measurements provide the average ratios \begin{equation} R(D)|_{\rm expt} = 0.397 \pm 0.049 , \qquad R(D^\ast)|_{\rm expt} = 0.308 \pm 0.017 , \label{} \end{equation} which exceed the SM expectations given in Refs.~\cite{} \begin{equation} R(D)|_{\rm SM} = 0.300 \pm 0.008 , \qquad R(D^\ast)|_{\rm SM} = 0.252 \pm 0.003 , \label{} \end{equation} by 1.9~$\sigma$ and 3.3~$\sigma$, respectively. The excess of $R(D^{(\ast)})$ over SM predictions has attracted a great deal of attention in the particle physics community and has led to many theoretical studies looking for NP explanations. In the paper \cite{} we included NP operators in the effective Hamiltonian and investigated their effects on physical observables of the decays $\bar{B}^0 \to D^{(\ast)} \ell^-\bar\nu_{\ell}$. We defined a full set of form factors corresponding to SM+NP operators and calculated them by employing the covariant confined quark model (CCQM). In the CCQM the transition form factors can be determined in the full range of momentum transfer, making the calculations straightforward without any extrapolation. This provides an opportunity to investigate NP operators in a self-consistent manner, and independently from the HQET. We first constrain the NP operators using experimental data, then analyze their effects on various observables including the ratios of branching fractions, the forward-backward asymmetries, and a set of polarization observables. We also derive the fourfold angular distribution for the cascade decay $\bar {B}^0\to D^{\ast\,+}(\to D^0\pi^+)\tau^-\bar\nu_\tau$ to analyze the polarization of the $D^{\ast}$ meson in the presence of NP by using the traditional helicity amplitudes. Recently the LHCb Collaboration has performed an angular analysis of the decay $\Lambda_b \to \Lambda^{(*)} + J/\psi$, where the $\Lambda_{b}$'s are produced in $pp$ collisions at $\sqrt{s} = 7$ TeV at the LHC (CERN)~\cite{}. They reported on the measurement of the relative magnitude of the helicity amplitudes in the decay $\Lambda_b \to \Lambda^{(*)} + J/\psi$ by a fit to several asymmetry parameters in the cascade decay distribution $\Lambda_b \to \Lambda(\to p\pi^-) + J/\psi (\to \ell^{+}\ell^{-})$ and $\Lambda_b \to \Lambda^\ast(\to pK^-) + J/\psi (\to \ell^{+}\ell^{-})$. In the paper~\cite{} we have performed a detailed analysis of the decay process $\Lambda_b \to \Lambda + J/\psi$ within the CCQM. We have worked out two variants of the threefold joint angular decay distributions in the cascade decay $\Lambda_b\to \Lambda(\to p\pi^-)\,+\,J/\psi(\to\ell^+\ell^-)$ for polarized and unpolarized $\Lambda_{b}$ decays. We have further listed results on helicity amplitudes which determine the rate and the asymmetry parameters in the decay processes $\Lambda_b \to \Lambda(\to p\pi^-)\,+\,J/\psi$ and $\Lambda_b \to \Lambda(\to p\pi^-)\,+\,\psi(2S)$. In the paper~\cite{} we have calculated the corresponding invariant and helicity amplitudes in the transitions $\Lambda_b~\to~\Lambda^{(\ast)}(J^P)~+~J/\psi$ where the $\Lambda^{(\ast)}(J^P)$ are $\Lambda$-type $(sud)$ ground and excited states with $J^P$ quantum numbers $J^P=\frac12^{\pm},\frac32^{\pm}$. We found that the values of the helicity amplitudes for the $\Lambda_b \to \Lambda^\ast(1520,\,\frac32^-),\Lambda^\ast(1890,\,\frac32^+)$ transitions are suppressed compared with those for the transitions to the ground state $\Lambda(1116,\,\frac12^+)$ also calculated in~\cite{} and the excited state $\Lambda^\ast(1405,\,\frac12^-)$. This analysis is important for the identification of the hidden charm pentaquark states $P_c^+(4380)$ and $P_c^+(4450)$ since the cascade decay $\Lambda_b~\to~\Lambda^\ast(\frac12^-,\frac32^\pm)(~\to~pK^-)~+~J/\psi$ involves the same final states as the decay $\Lambda_b^0~\to~P_c^+(~\to~p~J/\psi)~+~K^- $. \section{Extended Effective Hamiltonian for $B\to D^{(\ast)}\tau\bar\nu_\tau$ decay} We extend the SM effective Hamiltonian for the quark-level transition $b \to c \tau^- \bar{\nu}_{\tau}$ by including new operators: \begin{eqnarray} {\mathcal H}_{eff} &=& 2\sqrt {2}G_F V_{cb}[(1+V_L)\mathcal{O}_{V_L}+V_R\mathcal{O}_{V_R} +S_L\mathcal{O}_{S_L}+S_R\mathcal{O}_{S_R} +T_L\mathcal{O}_{T_L}], \label{] \mathcal{O}_{V_L} &=& \left(\bar{c}\gamma^{\mu}P_Lb\right) \left(\bar{\tau}\gamma_{\mu}P_L\nu_{\tau}\right), \qquad \mathcal{O}_{V_R} = \left(\bar{c}\gamma^{\mu}P_Rb\right) \left(\bar{\tau}\gamma_{\mu}P_L\nu_{\tau}\right), \nonumber\\ \mathcal{O}_{S_L} &=& \left(\bar{c}P_Lb\right)\left(\bar{\tau}P_L\nu_{\tau}\right),\qquad \mathcal{O}_{S_R} = \left(\bar{c}P_Rb\right)\left(\bar{\tau}P_L\nu_{\tau}\right), \qquad \mathcal{O}_{T_L} = \left(\bar{c}\sigma^{\mu\nu}P_Lb\right) \left(\bar{\tau}\sigma_{\mu\nu}P_L\nu_{\tau}\right). \nonumber \end{eqnarray} Here, $\sigma_{\mu\nu}=i\left[\gamma_{\mu},\gamma_{\nu}\right]/2$, $P_{L,R}=(1\mp\gamma_5)/2$ are the left and right projection operators, and $V_{L,R}$, $S_{L,R}$, and $T_L$ are the complex Wilson coefficients governing the NP contributions. In the SM one has $V_{L,R}=S_{L,R}=T_L=0$. We assume that NP only affects leptons of the third generation. The form factors which appear in the matrix elements including the NP operators are calculated in the framework of the CCQM. The model parameters, namely, the hadron size parameter $\Lambda$, the constituent quark masses $m_{q_i}$, and the universal infrared cutoff parameter $\lambda$, are determined by fitting calculated quantities of a multitude of basic processes to available experimental data or lattice simulations. It is important to note that within the SM (without any NP operators) our model calculation yields $R(D)=0.267$ and $R(D^\ast)=0.238$~\cite{}, which are consistent with other SM predictions given in Refs.~\cite{, } within $10\%$. In order to acquire the allowed regions for the NP Wilson coefficients, we assume that besides the SM contribution, only one of the NP operators in Eq.~(\ref{}) is switched on at a time. We then compare the calculated ratios $R(D^{(\ast)})$ with the recent experimental data. The experimental constraints are shown in Fig.~\ref{}. \begin{figure}[htbp] \begin{tabular}{lr} \includegraphics[scale=0.45]{constraint-VL.eps} & \includegraphics[scale=0.45]{constraint-VR.eps}\\ \includegraphics[scale=0.45]{constraint-SL.eps} & \includegraphics[scale=0.45]{constraint-TL.eps} \end{tabular} \caption{ The allowed regions of the Wilson coefficients $V_{L,R}$, $S_L$, and $T_L$ within $1\sigma$ (green, dark) and $2\sigma$ (yellow, light). The best-fit value in each case is denoted with the symbol $\ast$. The coefficient $S_R$ is disfavored at $2\sigma$ and therefore is not shown here.} \label{} \end{figure} The vector operators $\mathcal{O}_{V_{L,R}}$ and the left scalar operator $\mathcal{O}_{S_L}$ are favored while there is no allowed region for the right scalar operator $\mathcal{O}_{S_R}$ within $2\sigma$. Therefore we will not consider $\mathcal{O}_{S_R}$ in what follows. The tensor operator $\mathcal{O}_{T_L}$ is less favored, but it can still well explain the current experimental results. The stringent constraint on the tensor coupling mainly comes from the experimental data of $R(D^\ast)$. In each allowed region at $2\sigma$ we find the best-fit value for each NP coupling. The best-fit couplings read \begin{eqnarray} V_L &=&-0.23-i0.85, \quad V_R =0.03+i0.60,\quad S_L = -1.80-i0.07, \quad T_L =0.38+i0.06, \label{} \end{eqnarray} and are marked with an asterisk. The allowed regions of the coupling coefficients are then used to analyze the effect of the NP operators on different physical observables~\cite{}. \section{The decays $\Lambda_b \to \Lambda^{(\ast)}(\frac12^\pm,\frac32^\pm) + J/\psi$: matrix element and helicity amplitudes} \label{} The matrix element of the exclusive decay $\Lambda_1(p_1,\lambda_1)\to \Lambda_2(p_2,\lambda_2)\,+\,V(q,\lambda_V)$ is defined by (in the present application the vector meson label $V$ stands for the $J/\Psi$) \begin{equation} M(\Lambda_1\to \Lambda_2 + V) = \frac{G_F}{\sqrt{2}} \, V_{cb} \, V^\ast_{cs} \, C_{\rm eff} \, f_V \, M_V \, \langle \Lambda_2 | \bar s O_\mu b | \Lambda_1 \rangle \, \epsilon^{\dagger\,\mu}(\lambda_V) \,, \label{} \end{equation} where $M_V$ and $f_V$ are the mass and the leptonic decay constant of the vector meson $V$. The coefficient $C_{\rm eff}$ stands for the combination of Wilson coefficients $ C_{\rm eff}= C_1 + C_3 + C_5+ \xi \Big(C_2 + C_4 + C_6\Big)$. The color factor $\xi=1/N_c$ will be set to zero such that we only keep the leading term in the $1/N_c-$expansion. The hadronic matrix element $\langle \Lambda_2 | \bar s O^\mu b | \Lambda_1 \rangle$ is expressed in terms of six and eight, respectively, dimensionless invariant form factors $F^{V/A}_i(q^2)$. The three-quark currents with the appropriate quantum numbers of the the $\Lambda_Q (\frac12^\pm,\frac32^\pm)$ states are given by \begin{eqnarray} \Lambda_Q^{1/2^+} &\Longrightarrow& \epsilon_{a_1a_2a_3} \, Q_{a_1}\, \left( u_{a_2}C\gamma_5 d_{a_3} \right), \hspace*{1.5cm} \Lambda_Q^{1/2^-} \Longrightarrow \epsilon_{a_1a_2a_3} \,\gamma_5 Q_{a_1}\, \left( u_{a_2}C\gamma_5 d_{a_3} \right) \,, \nonumber\\ \Lambda_Q^{3/2^+} &\Longrightarrow& \epsilon_{a_1a_2a_3} \,\gamma_5 Q_{a_1}\, \left( u_{a_2}C\gamma_5\gamma_\mu d_{a_3} \right) \,, \qquad \Lambda_Q^{3/2^-} \Longrightarrow \epsilon_{a_1a_2a_3} \,Q_{a_1}\, \left( u_{a_2}C\gamma_5\gamma_\mu d_{a_3} \right) \,. \label{} \end{eqnarray} The nonlocal generalizations of the above currents are used in the CCQM to evaluate the appropriate form factors and helicity amplitudes. The numerical values of the normalized helicity amplitudes are listed in Table~\ref{}. \begin{table}[htb] \begin{center} \caption{Moduli squared of normalized helicity amplitudes.} \label{} \vspace*{.1cm} \def1.3{1.3} \begin{tabular}{|c|cccc|} \hline $\Lambda^\ast$ & 1116 & 1405 & 1890 & 1520 \\ \hline $J^P$ & $\frac12^+$ & $\frac12^-$ & $\frac32^+$ & $\frac32^-$ \\ \hline \qquad $|\hat H_{+\tfrac32 +1}|^2$ \qquad &\qquad 0 \qquad &\qquad 0 \qquad &\qquad $3.50\times 10^{-4}$\qquad &\qquad $0.84\times 10^{-4}$ \qquad \\ \qquad $|\hat H_{+\tfrac12 +1}|^2$\qquad &\qquad $2.34\times 10^{-3}$\qquad &\qquad $1.27\times 10^{-2}$\qquad &\qquad $3.19\times 10^{-2}$\qquad &\qquad $2.26\times 10^{-2}$\qquad \\ \qquad $|\hat H_{+\tfrac12 0}|^2$\qquad &\qquad $3.24\times 10^{-4}$\qquad &\qquad $5.19\times 10^{-3}$\qquad &\qquad $1.61\times 10^{-3}$\qquad &\qquad $1.82\times 10^{-3}$\qquad \\ \qquad $|\hat H_{-\tfrac12 0}|^2$\qquad &\qquad 0.53 \qquad &\qquad 0.51\qquad &\qquad $0.51$\qquad &\qquad 0.54\qquad \\ \qquad $|\hat H_{-\tfrac12 -1}|^2$\qquad &\qquad 0.47 \qquad &\qquad 0.47 \qquad &\qquad $ 0.45$ \qquad & \qquad 0.44 \qquad \\ \qquad $|\hat H_{-\tfrac32 -1}|^2$\qquad &\qquad 0 \qquad &\qquad 0 \qquad &\qquad $3.34\times 10^{-3}$\qquad &\qquad $1.06\times 10^{-3}$ \qquad \\[2mm] \hline \end{tabular} \end{center} \end{table} The helicity amplitudes $H_{\lambda_2,\lambda_V}$ of the produced $\Lambda^{(*)}$ states are clearly dominated by the helicity configuration $\lambda_2=-1/2$ as in the quark level transition $b \to s$. For the spin $1/2$ states in the transition $1/2^+ \to 1/2^\pm$ this implies that the two $\Lambda^{(*)}(1/2)$ states are almost purely left-handed.

\section{Introduction} Consider the multivariate $k$-sample problem expressed in the following canonical form: $p$-variate random vectors $\X_1, \ldots, \X_k $ and a positive scalar random variable $S$ are mutually independent and distributed as \begin{equation} \X_1\sim \Nc_p(\bmu_1, \si^2\V_1), \quad \ldots, \quad \X_k\sim \Nc_p(\bmu_k, \si^2\V_k), \quad S/\si^2\sim \chi_{(n)}^2, \label{} \end{equation} where the $p \times 1$ means $\bmu_1, \ldots, \bmu_k$ and the scale parameter $\si^2$ are unknown, and $\V_1, \ldots, \V_k$ are $p\times p$ known and positive definite symmetric matrices. In this model, we want to estimate $\bmu_1, \ldots, \bmu_k$ simultaneously relative to the quadratic loss function \begin{equation} L(\bde, \bom) = \frac{1}{\si^2} \sum_{i=1}^k \Vert\bde_i-\bmu_i\Vert_{\Q_i}^2 = \frac{1}{\si^2} \sum_{i=1}^k (\bde_i-\bmu_i)^\top\Q_i(\bde_i-\bmu_i) , \label{} \end{equation} where $\Q_1, \ldots, \Q_k$ are $p\times p$ known and positive definite symmetric matrices, $\Vert \a \Vert_\A^2=\a^\top \A \a$ with $\a^\top$ standing for the transpose of $\a$, $\bom=(\bmu_1, \ldots, \bmu_k, \si^2)$ is a vector of unknown parameters and $\bde=(\bde_1, \ldots, \bde_k)$ esimates $(\bmu_1, \ldots, \bmu_k)$. A typical example of Model~(\ref{}) is a $k$-sample problem. For $i \in \{1, \ldots, k\}$, a random sample of size $n_i$ is drawn from the $i$th population, say $\X_{i1}, \ldots, \X_{i n_i} \sim \Nc_p(\bmu_i, \si^2 \V_{i,0})$, where $\V_{i,0}$ is a known matrix. In this case, $\X_i$, $\V_i$, $S$, and $n$ in (\ref{}) correspond to $$ \Xb_i = \sum_{j=1}^{n_i}\X_{ij}/n_i, \quad {\V_{i,0}}/{n_i}, \quad \tr \left\{\sum_{i=1}^k\V_{i,0}^{-1}\sum_{j=1}^{n_i}(\X_{ij}-\Xb_i)(\X_{ij}-\Xb_i)^\top \right\}, \quad \sum_{i=1}^k(n_i-1)p, $$ respectively. Another example of (\ref{}) is $k$ linear regression models such that $\y_1 = \Z_1\bbe_1 + \bep_1, \ldots, \y_k = \Z_k\bbe_k + \bep_k$, where $\y_i$ is an $n_i \times 1$ vector of observations, $\bbe_i$ is a $p \times 1$ vector of regression coefficients, $\Z_i$ is an $n_i\times p$ matrix of explanatory variables and $\bep_i$ is an $n_i \times 1$ vector having $\Nc_{n_i}(\zero, \si^2\I_{n_i})$. In this case, $\X_i$, $\V_i$, $S$ and $n$ in (\ref{}) respectively correspond to $$ \bbeh_i=(\Z_i^\top\Z_i)^{-1}\Z_i^\top\y_i, \quad (\Z_i^\top\Z_i)^{-1}, \quad \sum_{i=1}^k \Vert\y_i-\Z_i\bbeh_i\Vert_{\I_{n_i}}^2, \quad \sum_{i=1}^k (n_i-p). $$ In this paper, we consider the case where the means are equal or nearly so. For example, suppose that Laboratories $L_1, \ldots, L_k$ use similar instruments to measure several common characteristics. One may then suspect that the $k$ population mean vectors are nearly equal. In such a situation, rather than estimating each mean by the corresponding sample mean separately, it might be preferable to use a pooled mean estimator. A classical way to address this problem is to compute an estimator which uses the pooled mean estimator upon acceptance of the null hypothesis $\mathcal{H}_0 : \bmu_1=\cdots = \bmu_k$, and to use separate sample means upon rejection of $\mathcal{H}_0$. However, the preliminary test estimators are non-smooth and do not necessarily improve on sample means. As an alternative method, we consider Bayesian methods including hierarchical Bayes and empirical Bayes estimators. A Bayesian approach to this problem uses the prior distribution of $\bmu_i$ assuming a multivariate normal distribution with common mean $\bnu$ and common variance $\tau^2$, where $\tau^2$ is assumed to have a hierarchical prior distribution. The mean $\bnu$ corresponds to $\mathcal{H}_0 : \bmu_1=\cdots = \bmu_k=\bnu$. Ghosh and Sinha \cite{gs1988} considered this prior distribution in the framework of estimating the single mean $\bmu_1$ in the two-sample problem and showed that the resulting empirical and hierarchical Bayes estimators have reasonable forms of shrinking $\X_1$ towards the pooled estimator of $\bnu$. They also derived conditions for those shrinkage estimators to be minimax. Shrinkage estimators with minimaxity have been studied since Stein \cite{stein1956} established the inadmissibility of the standard estimator $\X_1$. For papers related to the context of this paper, see \cite{em1976, js1961, perron1993, stein1981, straw1971, straw1973, s1996}. The Bayesian approach suggested here consists of two kinds of shrinkage. One is to shrink the $k$ sample means toward a pooled mean estimator, and the other is to shrink the $p$-dimensional pooled estimator toward a constant like zero, which was proposed in~\cite{perron1993}. The former shrinkage is treated in Section~\ref{}, and a class of minimax estimators is derived. Out of the class, we develop hierarchical and empirical Bayes estimators with minimaxity. The latter shrinkage is discussed in Section~\ref{}, and we show that all the estimators derived by the former shrinkage can be further improved by using the latter shrinkage. In Section~\ref{sec:BM}, we find a hierarchical Bayes and minimax estimator, thus an admissible and minimax estimator using both kinds of shrinkage when $\Q_1=\V_1^{-1}, \ldots, \Q_k=\V_k^{-1}$. The Bayes estimator, which consists of the two kinds of shrinkage, has a new form which we cannot handle with conventional techniques considered in Section~\ref{sec:2}. Thus, we need to extend the classes of minimax estimators in order to treat such a new type of shrinkage estimator. The minimaxity of the hierarchical Bayes estimator is successfully established. In Section~\ref{}, we investigate through simulations the performance of several hierarchical Bayes and empirical Bayes estimators and the preliminary test estimators. The results show that the hierarchical Bayes and empirical Bayes estimators for the two kinds of prior distributions perform well. All proofs are relegated to the Appendix. \section{Improvement by two kinds of shrinkage\label{sec:2}} \subsection{Shrinkage toward the pooled estimator\label{}} Let $\eta=1/\si^{2}$, $\bMu=(\bmu_1,\dots,\bmu_k)$ and $q=\min \{\Ch_{{\rm min}}(\V_1\Q_1), \ldots, \Ch_{{\rm min}}(\V_k\Q_k) \}$. For $\la \in (0, q)$, we begin by assuming the prior distribution \begin{equation} \vec(\bMu)\mid\bnu,\eta,\la \sim \Nc_{pk}\left[\J\bnu,\;\;\blockdiag \left\{\frac{(\V_1\Q_1-\la \I_p)\V_1}{\la\eta},\dots,\frac{(\V_k\Q_k-\la \I_p)\V_k}{\la\eta}\right\} \right], \quad \bnu \sim \mathcal{U}(\Re^p), \label{} \end{equation} where $\J$ is the $pk\times p$ matrix $\J=(\I_p,\dots,\I_p)^\top$ and $\mathcal{U}(\Re^p)$ denotes the improper uniform distribution over $\Re^p$. The posterior distribution of $\bMu$ given $\X=(\X_1, \ldots, \X_k)$ and $\bnu$ is $$ \vec(\bMu) \mid \X, \bnu, \eta, \la \sim \Nc_{pk} \left[ \vec\big\{\bMuh^*(\eta,\la,\bnu)\big\}, \blockdiag \left( {\V_1-\la\Q_1^{-1} \over \eta},\dots,{\V_k-\la\Q_k^{-1} \over \eta} \right) \right] $$ and the marginal distribution of $\X$ given $\bnu$ is \begin{equation} \vec(\X) \mid \bnu, \eta,\la \sim \Nc_{pk} \left[ \J\bnu, \blockdiag \left( {\V_1\Q_1\V_1\over\la\eta},\dots,{\V_k\Q_k\V_k\over\la\eta} \right) \right], \label{} \end{equation} where $\bMuh^*(\eta,\la,\bnu)=\big(\bmuh_1^*(\eta,\la,\bnu),\dots,\bmuh_k^*(\eta,\la,\bnu)\big)$ and \begin{equation} \bmuh_i^*(\eta,\la,\bnu)= \X_i - \la\Q_i^{-1}\V_i^{-1}(\X_i-\bnu)=\X_i - \la\eta \Q_i^{-1}\V_i^{-1}(\X_i-\bnu)/\eta. \label{} \end{equation} Also, the posterior distribution of $\bnu$ given $\X_1, \ldots, \X_k$ and the marginal density of $\X_1, \ldots, \X_k$ are \begin{equation} \begin{split} \bnu \mid \X_1, \ldots, \X_k, \eta, \la \sim& \Nc_p[ \bnuh,\A / (\la\eta)],\\ f_\pi (\x_1, \ldots, \x_k \mid \eta, \la) \propto& (\la\eta)^{p(k-1)/2}\exp \left (- \la\eta \sum_{i=1}^k\Vert\x_i-\bnuh\Vert_{\V_i^{-1}\Q_i^{-1}\V_i^{-1}}^2/2 \right), \end{split} \label{} \end{equation} where $\A = (\V_1^{-1}\Q_1^{-1}\V_1^{-1}+ \cdots + \V_k^{-1}\Q_k^{-1}\V_k^{-1})^{-1}$ and $\bnuh = \V_i^{-1}\Q_i^{-1}\V_1^{-1}\X_1 + \cdots + \V_k^{-1}\Q_i^{-1}\V_k^{-1}\X_i$. Then, the Bayes estimator of $\bmu_i$ with respect to the quadratic loss (\ref{}) is derived from (\ref{}) as \begin{equation} \bmuh_i^B(\eta, \la)=\bmuh_i^*(\eta, \la,\bnuh)=\X_i - \la\Q_i^{-1}\V_i^{-1}(\X_i-\bnuh). \label{} \end{equation} Because $\la$ and $\eta$ are unknown, we set $$ F = \frac{1}{S} \sum_{i=1}^k \Vert \X_i-\bnuh\Vert_{\V_i^{-1}\Q_i^{-1}\V_i^{-1}}^2 $$ and estimate $\la\eta$ by $\{p(k-1)-2\}/(FS)$ from the marginal likelihood in (\ref{}) and $\si^2$ by $\sih^2=S/(n+2)$. Then, the resulting empirical Bayes estimator is \begin{equation} \bmuh_i^{EB1} = \X_i - \min \left[ {\{p(k-1)-2\}/\{(n+2)F\}}, 1 \right]\Q_i^{-1}\V_i^{-1}(\X_i-\bnuh). \label{} \end{equation} Motivated from the empirical Bayes estimator, we consider the class of estimators \begin{equation} \bmuh_i^S(\phi_0) = \X_i - \phi_0(F) \Q_i^{-1}\V_i^{-1}(\X_i-\bnuh)/F, \label{} \end{equation} where $\phi_0(F)$ is an absolutely continuous function. We can provide conditions on $\phi_0(F)$ for minimaxity of $\bmuh_i^S(\phi_0)$. Theorem~\ref{thm:1} is a corollary of Proposition~\ref{prp:1} in Section~\ref{sec:BM}. \begin{thm} \label{thm:1} The estimator $(\bmuh_1^S(\phi_0), \ldots, \bmuh_k^S(\phi_0))$ is minimax relative to the quadratic loss $(\ref{})$ if {\rm (a)} $\phi_0(F)$ is non-decreasing in $F$ and {\rm (b)} $0<\phi_0(F)\leq 2 \{p(k-1)-2\}/(n+2)$. \end{thm} It follows from Theorem~\ref{thm:1} that the empirical Bayes estimator $(\bmuh_1^{EB1}, \ldots, \bmuh_k^{EB1})$ is minimax. Another minimax estimator is the hierarchical Bayes estimator. In addition to the prior distribution (\ref{}), we assume that $$ \pi(\la) \propto \;\la^{a -1} \mathbf{1}_{(0<\la<q)}, \quad \pi(\eta) \propto \;\eta^{-c } \mathbf{1}_{(\eta>0)}, $$ where $a $ and $c $ are constants, $q=\min \{ \Ch_{{\rm min}}(\V_1 \Q_1), \ldots, \Ch_{{\rm min}}(\V_k \Q_k)\}$ and $\mathbf{1}$ denotes the indicator function. It is noted that $\pi(\la)$ is proper for $a>0$, and $\pi(\eta)$ is improper for any $c$. From (\ref{}), the posterior distribution of $(\la, \eta)$ given $\X_1, \ldots, \X_k, S$ is $$ \pi(\la, \eta \mid \x_1, \ldots, \x_k,S) \propto \la^{p(k-1)/2+a -1}\eta^{\{n+p(k-1)\}/2-c}\exp \left( - {\la\eta }\sum_{i=1}^k\Vert\x_i-\bnuh\Vert_{\V_i^{-1}\Q_i^{-1}\V_i^{-1}}^2/2 -\eta S/2 \right), $$ where $\la\in (0,q)$ and $\eta>0 $. Then from (\ref{}), the hierarchical Bayes estimator of $\bmu_i$ relative to the quadratic loss (\ref{}) is written as \begin{align} \bmuh_i^{HB1} =& \X_i - {\rm E} (\la \mid \X_1, \ldots, \X_k, S)\Q_i^{-1}\V_i^{-1}(\X_i - \bnuh), \label{} \end{align} where \begin{align*} {\rm E} (\la \mid \X_1, \ldots, \X_k, S)={\int_0^q \int_0^\infty \la^{p(k-1)/2+a }\eta^{\{n+p(k-1)\}/2-c }\exp\{ - {(\eta S/2)}(\la F+1)\}d\eta d\la \over \int_0^q \int_0^\infty \la^{p(k-1)/2+a -1}\eta^{\{n+p(k-1)\}/2-c }\exp\{ - {(\eta S/2)}(\la F+1)\}d\eta d\la}. \end{align*} Making the transformations $x=F\la$ and $v=S\eta$ gives the expression ${\rm E} (\la \mid \X_1, \ldots, \X_k, S)= {\phi^{HB1}(F) /F}$, where $$ \phi^{HB1}(F)= {\int_0^{qF} x^{p(k-1)/2+a }/ (1+x)^{\{n+p(k-1)\}/2+1-c } dx \over \int_0^{qF} x^{p(k-1)/2+a -1}/(1+x)^{\{n+p(k-1)\}/2+1-c }dx}\\ $$ We can derive a condition on $a $ and $c $ for minimaxity of the hierarchical Bayes estimator. \begin{thm} \label{} The hierarchical Bayes estimator $(\ref{})$ is minimax if $p(k-1)\geq 3$, $a +c < (n+2)/2$ and \begin{equation} \{2p(k-1)+n-2\}a +2\{p(k-1)-2\}c \leq (n-2)p(k-1)/2 - 2n. \label{} \end{equation} \end{thm} \noindent \textit{Proof.} It suffices to show that $\phi^{HB1}(F)$ satisfies conditions (a) and (b) in Theorem~\ref{thm:1}. The numerator of the derivative of $\phi^{HB1}$ with respect to $F$ is proportional to \begin{align*} \frac{(qF)^{p(k-1)/2+a-1}}{(1+qF)^{\{n+p(k-1)\}/2+1-c }}\int_{0}^{qF}\frac{x^{p(k-1)/2+a-1}(qF-x)}{(1+x)^{\{n+p(k-1)\}/2+1-c}}\, dx, \end{align*} which establishes condition (a). To check condition (b), we note that $$ \int_0^\infty {x^\ell (1+x)^{-m}}dx = \int_0^1 w^\ell (1-w)^{m-\ell-2}dw =B(\ell+1, m-\ell-1), $$ for the beta function $B$. Using condition (a), we see that \begin{align*} \phi^{HB1}(F)\leq \lim_{F\to\infty} \phi^{HB1}(F) & = {\int_0^\infty x^{p(k-1)/2+a }/ (1+x)^{\{n+p(k-1)\}/2+1-c } dx \over \int_0^\infty x^{p(k-1)/2+a -1}/(1+x)^{\{n+p(k-1)\}/2+1-c }dx} = {p(k-1)+2a \over n- 2(a +c )}. \end{align*} Thus, condition (b) is satisfied if $a +c < n/2$ and if $$ \{p(k-1)+2a\}/\{n- 2(a +c )\} \leq 2 \{p(k-1)-2\}/(n+2), $$ which is equivalently rewritten by (\ref{}). Hence, the minimaxity of the hierarchical Bayes estimator is established. \hfill$\Box$ \begin{remark}{\rm When $\bnu=\zero$, the prior distribution (\ref{}) was treated in~\cite{berger1976,zrnan2013}. There exists a nonsingular matrix $\B_i$ such that $\B_i\V_i\B_i^\top=\I_p$ and $\B_i\Q_i^{-1}\B_i^\top=\bDe_i$ for a diagonal matrix $\bDe_i$. This reduces the problem to the simple model with $\V_i=\I_p$ and $\Q_i=\bDe_i^{-1}$. When we assume the uniform prior for $\bnu$ in (\ref{}), however, we cannot reduce the problem to such a simple situation. } \end{remark} \begin{remark}{\rm When $\eta$ and $\la$ are known, assuming the uniform distribution for $\bnu$ in (\ref{}) yields the Bayes estimator $\bmuh_i^*(\eta, \la,\bnuh)$. It is interesting to note that this Bayes estimator is also interpreted as an empirical Bayes estimator, because when $\bnu$ is an unknown parameter, $\bnuh$ is the maximum likelihood estimator (MLE) of $\bnu$ in the marginal likelihood $$ \vec(\X)\sim \Nc_{pk} [\J\bnu, \;\;(\la\eta)^{-1}\blockdiag ({\V_1\Q_1\V_1},\dots,{\V_k\Q_k\V_k} ) ]. $$ Thus, the estimator $\bmuh_i^*(\eta, \la,\bnuh)$ is a Bayes empirical Bayes estimator. } \end{remark} \subsection{Shrinkage of the pooled estimator} \label{} The empirical Bayes estimator $\bmuh_i^{EB1}$ and the hierarchical Bayes estimator $\bmuh_i^{HB1}$ shrink $k$ individual estimators $\X_i$ toward the common mean $\bnuh$. Because $\bnuh$ is a $p$-dimensional estimator against the uniform prior over $\Re^p$, we can consider to shrink $\bnuh$ for further improvement. Such a further improvement was studied in~\cite{perron1993}. We here assume the prior distribution \begin{align} \vec(\bMu)\mid\bnu,\eta,\la \sim& \Nc_{pk} \left[\J\bnu,\;\;\blockdiag \{ {(\V_1\Q_1-\la I_p)\V_1}/{(\la\eta)},\dots, {(\V_k\Q_k-\la I_p)\V_k}/{(\la\eta)} \} \right],\non\\ \bnu \mid \eta,\la,\xi \sim& \Nc_p [\zero, (\la-\xi)\A/(\la\xi\eta) ], \label{} \end{align} where $\xi$ and $\la$ satisfy the constraint $0<\xi<\la<q$ for $q=\min\{\Ch_{{\rm min}}(\V_1 \Q_1), \ldots, \Ch_{{\rm min}}(\V_k \Q_k)\}$. It is noted that the normal prior is considered for $\bnu$ instead of the uniform prior in (\ref{}). Then the posterior distribution of $\bMu$ given $\X=(\X_1, \ldots, \X_k)$ and $\bnu$, and the marginal distribution of $\X$ given $\bnu$ are given by (\ref{}) and (\ref{}), and the posterior distribution of $\bnu$ given $\X_1, \ldots, \X_k$ and the marginal density of $\X_1, \ldots, \X_k$ are \begin{equation} \begin{split} \bnu \mid \X_1, \ldots, \X_k, \eta, \la,\xi \sim& \Nc_p [ \left(1- {\xi}/{\la}\, \right)\bnuh, {(\la-\xi)}\A/{(\la^2\eta)} ],\\ f_\pi(\x_1,\ldots, \x_k \mid \la, \xi, \si^2) \propto& ({\la\eta})^{p(k-1)/2}\exp \left( -{\la\eta } \sum_{i=1}^k\Vert\x_i-\bnuh\Vert_{\V_i^{-1}\Q_i^{-1}\V_i^{-1}}^2/2 \right) \times \left({\xi\eta}\right)^{p/2} \exp ( -{\xi\eta } \Vert\bnuh\Vert_{\A^{-1}}^2/2 ). \end{split} \label{} \end{equation} Thus the Bayes estimator of $\bmu_i$ with respect to the prior in (\ref{}) is given by \begin{equation} \bmuh_i^B(\eta, \la,\xi)= \X_i - \la\Q_i^{-1}\V_i^{-1}(\X_i-\bnuh)-\xi\Q_i^{-1}\V_i^{-1}\bnuh. \label{} \end{equation} Since $\la$, $\xi$ and $\eta$ are unknown, we now set $$ F = \frac{1}{S} \sum_{i=1}^k\Vert\X_i-\bnuh \Vert^2_{\V_i^{-1}\Q_i^{-1}\V_i^{-1}}, $$ and estimate $\la\eta$ and $\xi\eta$ by $\{p(k-1)-2\}(FS)$ and $(p-2)/\Vert\bnuh\Vert_{\A^{-1}}^2$, respectively. When $\si^2$ $(=1/\eta)$ is estimated by $\sih^2=S/(n+2)$, the resulting empirical Bayes estimator is \begin{equation} \bmuh_i^{EB2} = \X_i - \min\left[ {\{p(k-1)-2\}/\{(n+2)F\}}, 1\right] \Q_i^{-1}\V_i^{-1}(\X_i-\bnuh) -\min\left[ {(p-2)/\{(n+2)G\}}, 1\right] \Q_i^{-1}\V_i^{-1}\bnuh, \label{} \end{equation} where $G=\Vert\bnuh\Vert_{\A^{-1}}^2/S$. The estimator $\bmuh_i^{EB2}$ enjoys shrinking $\X_i$ toward $\bnuh$ and shrinking $\bnuh$ toward zero. Motivated from the empirical Bayes estimator, we consider the class of double shrinkage estimators \begin{equation} \bmuh_i^S(\phi_0, \psi_0) = \X_i - {\phi_0(F)} \Q_i^{-1}\V_i^{-1}(\X_i-\bnuh)/F - {\psi_0(G)} \Q_i^{-1}\V_i^{-1}\bnuh/G, \label{} \end{equation} where $\phi_0(F)$ and $\psi_0(G)$ are absolutely continuous functions. We can provide conditions on $\psi_0(G)$ for improving on $\bmuh_i^S(\phi_0)$ in (\ref{}), which is stated in the following theorem. Theorem~\ref{thm:2} is a corollary of Proposition~\ref{prp:2} in the subsequent section. \begin{thm} \label{thm:2} The estimator $(\bmuh_1^S(\phi_0), \ldots, \bmuh_k^S(\phi_0))$ in $(\ref{})$ is improved upon by the double shrinkage estimator $(\bmuh_1^S(\phi_0, \psi_0)$, $\ldots$, $\bmuh_k^S(\phi_0,\psi_0))$ in $(\ref{})$ relative to the quadratic loss $(\ref{})$ if {\rm (a)} $\psi_0(G)$ is non-decreasing in $G$, and {\rm (b)} $0<\psi_0(G)\leq 2 (p-2)/(n+2)$. \end{thm} From Theorem~\ref{thm:2}, the empirical Bayes estimator $(\bmuh_1^{EB1}, \ldots, \bmuh_k^{EB1})$ in (\ref{}) is improved upon by $(\bmuh_1^{EB2}, \ldots, \bmuh_k^{EB2})$ in (\ref{}). Using Theorem~\ref{thm:2}, we provide a shrinkage estimator improving on the preliminary-test (PT) estimator. We here consider the case where $\Q_i=\V_i^{-1}$ for all $i\in\{1, \ldots, k\}$. With the new definition of $F$, the statistic $nF/\{p(k-1) \}$ has the Fisher--Snedecor distribution with $(p(k-1),n)$ degrees of freedom under the hypothesis $\mathcal{H}_0$. Then the PT estimator based on the statistic $F$ for testing the hypothesis is \begin{equation} \bmuh_i^{PT}= \X_i - \mathbf{1}_{[F\leq \{p(k-1)/n\}F_{p(k-1), n, \al}]}(\X_i -\bnuh) = \left\{\begin{array}{ll} \X_i & \ \text{if}\ F>\{p(k-1)/n\}F_{p(k-1), n, \al}, \\ \bnuh & \ \text{otherwise,} \end{array}\right. \label{eqn:PT} \end{equation} where $F_{p(k-1), n, \al}$ is the upper $\al$ point of the Fisher--Snedecor distribution with $(p(k-1), n)$ degrees of freedom. Because the PT estimator belongs to the class (\ref{}), Theorem~\ref{thm:2} implies that the estimator $(\bmuh_1^{PT}, \ldots, \bmuh_k^{PT})$ can be improved upon by $(\bmuh_1^{PT*}, \ldots, \bmuh_k^{PT*})$ with \begin{equation} \bmuh_i^{PT*} = \bmuh_i^{PT} - \min \left[ (p-2)/\{( n+2) G\}, 1 \right] \bnuh. \label{} \end{equation} \section{Bayes and minimax estimation\label{sec:BM}} \subsection{Extension of the class of minimax shrinkage estimators} In this section, we derive the hierarchical Bayes and minimax estimators. As shown in (\ref{}), the hierarchical Bayes estimator with double shrinkage does not belong to the class (\ref{}). To show minimaxity of the hierarchical Bayes estimators, we begin by extending the class of estimators (\ref{}) to the following class with more general form \begin{equation} \bmuh_i(\phi, \psi) = \X_i - \phi(F,G,S) \Q_i^{-1}\V_i^{-1}(\X_i - \bnuh)/F - \psi(F,G, S) \Q_i^{-1}\V_i^{-1}\bnuh/G, \label{eqn:GE} \end{equation} where $\phi(F,G,S)$ and $\psi(F,G,S)$ are absolutely continuous functions. The class (\ref{}) is a special case of (\ref{eqn:GE}), and Theorems \ref{thm:1} and \ref{thm:2} can be extended in the propositions given below. \begin{thm} \label{} An unbiased estimator of the risk function of $(\ref{eqn:GE})$ relative to the loss $(\ref{})$ is \begin{align} \text{UER}(\phi,\psi)=& \sum_{i=1}^k\tr(\V_i\Q_i) - {[2\{p(k-1)-2\}-(n+2)\phi]\phi/ F} - 4 \phi_F - {4\phi }(F\phi_F+G\phi_G)/F+{4S\phi\phi_S/F}\non\\ & - {\{2(p-2)-(n+2)\psi\}\psi/ G} - 4 \phi_G - {4\psi}(F\psi_F+G\psi_G)/G +{4S\psi\psi_S/ G}, \label{} \end{align} where $\phi_F=\partial \phi(F,G,S)/\partial F$, $\phi_G=\partial \phi(F,G,S)/\partial G$, $\phi_S=\partial \phi(F,G,S)/\partial S$, and $\psi_F$, $\psi_G$ and $\psi_S$ are defined similarly. \end{thm} The proof is given in the Appendix. Considering the case with $\psi(F,G,S)=0$ in (\ref{}), we can get a generalization of Theorem~\ref{thm:1}. \begin{prp} \label{prp:1} The estimator $(\bmuh_1(\phi), \ldots, \bmuh_k(\phi))$ with the form $\bmuh_i(\phi) = \X_i - \{\phi(F,G,S)/F\}(\X_i - \bnuh)$ is minimax relative to the quadratic loss $(\ref{})$ if {\rm (a)} $\phi(F,G,S)$ is non-decreasing in $F$ and $G$, and non-increasing in $S$; and {\rm (b)} $0<\phi(F,G,S)\leq 2 \{p(k-1)-2\}/(n+2)$. \end{prp} The unbiased risk estimator (\ref{}) also enables us to get a generalization of Theorem~\ref{thm:2}. \begin{prp} \label{prp:2} The estimator $(\bmuh_1(\phi), \ldots, \bmuh_k(\phi))$ given in Proposition~\ref{prp:1} is improved upon by the double shrinkage estimator $(\bmuh_1(\phi, \psi), \ldots, \bmuh_k(\phi,\psi))$ in $(\ref{eqn:GE})$ relative to the quadratic loss $(\ref{})$ if {\rm (a)} $\psi(F,G,S)$ is non-decreasing in $F$ and $G$, and non-increasing in $S$; and {\rm (b)} $0<\psi(F,G,S)\leq 2 (p-2)/(n+2)$. \end{prp} \subsection{Hierarchical Bayes minimax estimators} In this section, we concentrate attention on the case of $\Q_1=\V_1^{-1}, \ldots, \Q_k=\V_k^{-1}$. Although a natural prior is a hierarchical prior based on (\ref{}), the minimaxity of the resulting Bayes estimator is difficult to show, because the integrals with respect to $\xi$ and $\la$ are taken over the constraint $\xi<\la$. Instead of (\ref{}), we here consider the following prior distribution for $\bMu=(\bmu_1,\dots,\bmu_k)$: \begin{equation} \vec(\bMu)\mid \eta,\la,\xi\sim \Nc_{pk} \left[0,\;\;\eta^{-1} \left\{ {\la\over 1-\la} \, \big (\bSi^{-1}-\bSi^{-1}\J\A\J^\top\bSi^{-1}\big)+{\xi\over1-\xi }\, \bSi^{-1}\J\A\J^\top\bSi^{-1} \right\}^{-1} \right], \label{ prior} \end{equation} where $0<\la<1$, $0<\xi<1$, $\bSi=\blockdiag(\V_1,\dots,\V_k)$ and $\A=\big(\V_1^{-1} + \cdots + \V_k^{-1}\big)^{-1}$. The covariance matrix in the prior (\ref{ prior}) is positive definite for $0<\la, \xi<1$. It is noted that when $\Q_i=\V_i^{-1}$, integrating out the prior (\ref{}) with respect to $\bnu$ leads to the same density function as in (\ref{ prior}) with the constraint $\xi<\la$. The prior distribution (\ref{ prior}) does not have to assume this constraint. For $\la$, $\xi$ and $\eta$, we assume the following second-stage priors: \begin{equation} \pi(\la) \propto \;\la^{a -1}\mathbf{1}_{(0<\la< 1)},\quad \pi(\xi) \propto \;\xi^{b -1}\mathbf{1}_{(0<\xi< 1)},\quad \pi(\eta) \propto \;\eta^{-c}\mathbf{1}_{(\eta\geq L)}. \label{} \end{equation} Here $a $, $b$ and $c $ are constants and $L $ is a positive constant. When $a>0$, $b>0$ and $c>1$, the prior distribution (\ref{}) is proper. From (\ref{}), the posterior distribution of $(\la, \xi, \eta)$ given $\X_1, \ldots, \X_k, S$ is $$ \pi(\xi, \la, \eta\mid \x_1, \ldots, \x_k, S) \propto \la^{p(k-1)/2+a-1} \xi^{p/2+b-1} \eta^{(n+pk)/2-c}\exp\left\{ -{S \eta} (\la F+ \xi G + 1)/2\right\}, $$ where $0<\xi<q$, $0<\la<q$ and $\eta\geq L $. From (\ref{}), the hierarchical Bayes estimator of $\bmu_i$ is written as \begin{align} \bmuh_i^{HB2} = \X_i - {\rm E} (\la \mid \X, S)(\X_i - \bnuh) - {\rm E} (\xi \mid \X, S)\bnuh, \label{} \end{align} where $\X=(\X_1, \ldots, \X_k)$, \begin{align*} {\rm E} (\la \mid \X, S)={\int_0^1\int_0^1 \int_L^\infty \la^{p(k-1)/2+a }\xi^{p/2+b-1}\eta^{(n+pk)/2-c }\exp\{ - {\eta Sr 2}(\la F+\xi G +1)/2\}d\eta d\xi d\la \over \int_0^1 \int_0^1 \int_L ^\infty \la^{p(k-1)/2+a-1 }\xi^{p/2+b-1}\eta^{(n+pk)/2-c }\exp\{ - {\eta S }(\la F+\xi G + 1)/2\}d\eta d\xi d\la}. \end{align*} Making the transformations $x=F\la$, $y=G\xi$ and $v=S\eta$ gives the expression $ {\rm E} (\la \mid \X, S)= {\phi^{HB2}(F,G,S)/ F}$, where $$ \phi^{HB2}(F,G,S)= {\int_0^{F} \int_0^{G} \int_{L S}^\infty x^{p(k-1)/2+a }y^{p/2+b-1}v^{(n+pk)/2-c }\exp\{ - v(x+y+1)/2\}dv dy dx \over \int_0^{F} \int_0^{G} \int_{L S}^\infty x^{p(k-1)/2+a-1 }y^{p/2+b-1}v^{(n+pk)/2-c }\exp\{ - v(x+y+1)/2\}dv dy dx}. $$ Similarly, $ {\rm E} (\xi \mid \X, S)= {\psi^{HB2}(F,G, S) /G}$, where $$ \psi^{HB2}(F,G,S)= {\int_0^{F} \int_0^{G} \int_{L S}^\infty x^{p(k-1)/2+a-1 }y^{p/2+b}v^{(n+pk)/2-c }\exp\{ - v(x+y+1)/2\}dv dy dx \over \int_0^{F} \int_0^{G} \int_{L S}^\infty x^{p(k-1)/2+a-1 }y^{p/2+b-1}v^{(n+pk)/2-c }\exp\{ - v(x+y+1)/2\}dv dy dx}. $$ Thus, the hierarchical Bayes estimator in (\ref{}) is expressed as \begin{equation} \bmuh_i^{HB2} = \X_i - {\phi^{HB2}(F,G,S) }(\X_i - \bnuh)/F - {\psi^{HB2}(F,G, S) }\bnuh/G. \label{} \end{equation} This estimator does not belong to the class (\ref{}), but belongs to the class (\ref{eqn:GE}). For minimaxity of the hierarchical Bayes estimators, we check the conditions given in Propositions~\ref{}. Combining the condition for the minimaxity and the condition for the proper prior, we can get a condition for the Bayes estimator to be admissible and minimax. \begin{thm} \label{} The hierarchical Bayes estimator $(\ref{})$ is minimax if $p(k-1)\geq 3$, $p\geq 3$ and if $a +b+c < n/2$ and \begin{equation} \{2p(k-1)+n-2\}a +2\{p(k-1)-2\}(b+c) \leq p(k-1)(n-2)/2-2n, \label{} \end{equation} \begin{equation} (2p+n-2)b +2(p-2)(a+c) \leq p(n-2)/2-2n. \label{} \end{equation} Further, $(\ref{})$ is a Bayes minimax estimator, namely admissible and minimax, provided that the constants $a$, $b$ and $c$ satisfy the above conditions for $a>0$, $b>0$ and $c>1$. \end{thm} The latter part of Theorem~\ref{} follows from the fact that the prior distribution (\ref{}) is proper when $a >0$, $b >0$, $c>1$ and $L >0$. Then, the Bayes estimator (\ref{}) is admissible and minimax. Since $k\geq 2$, we can find such constants if $p\geq 5$ and $n>2(3p-4)/(p-4)$. This shows that $p$ is at least $5$. \begin{remark}{\rm Although the condition $L>0$ is imposed technically in order to guarantee the admissibility, we have no information on $L$. Furthermore, the Bayes estimator based on triple integrals is computationally hard to derive. From a practical point of view, we set $L=0$. In this case, the hierarchical prior (\ref{}) is improper and $\bmuh_i^{HB2}$ is no longer guaranteed to be admissible. However, the minimaxity still holds, and the resulting generalized Bayes estimator is \begin{equation} \bmuh_i^{HB2}=\X_i - {\phi^{HB2}(F,G)}(\X_i - \bnuh)/F - {\psi^{HB2}(F,G)} \bnuh/G, \label{} \end{equation} where $\phi^{HB2}(F,G,S)$ and $\psi^{HB2}(F,G,S)$ are expressed based on double integrals as \begin{align*} &\phi^{HB2}(F,G)={\int_0^{F} \int_0^{G} x^{p(k-1)/2+a }y^{p/2+b-1} (x+y+1)^{-\{(n+pk)/2-c+1 \}} dy dx \over \int_0^{F} \int_0^{G} x^{p(k-1)/2+a-1 }y^{p/2+b-1}(x+y+1)^{-\{(n+pk)/2-c+1 \}} dy dx},\\ &\psi^{HB2}(F,G)={\int_0^{F} \int_0^{G} x^{p(k-1)/2+a-1 }y^{p/2+b}(x+y+1)^{-\{(n+pk)/2-c+1 \}} dy dx \over \int_0^{F} \int_0^{G} x^{p(k-1)/2+a-1 }y^{p/2+b-1}(x+y+1)^{-\{(n+pk)/2-c+1 \}} dy dx}. \end{align*} In our simulation experiments given in the next section, we use this generalized Bayes estimator. }\end{remark} \section{Simulation study} \label{} We investigate the numerical performances of the risk functions of the preliminary-test estimator and several empirical and hierarchical Bayes estimators through simulation. To compare the preliminary-test and Bayes estimators fairly, we employ the quadratic loss function $L(\bde,\bom)$ in (\ref{}) for $ \Q_1= \V_1^{-1}, \ldots, \Q_k= \V_k^{-1}$. The estimators which we compare are the following eight: \begin{itemize} \item [] JS1: the sample-wise James--Stein estimator $$ \bmuh_i^{JS1}=\X_i - {p-2 \over n+2}{S\over \Vert\X_i\Vert^2_{\V_i^{-1}}}\X_i, $$ \item [] JS2: the overall James--Stein estimator for $\bMu=(\bmu_1,\dots,\bmu_k)$ $$ \widehat{\boldmath{\text{$\mathcal{M}$}}}^{JS2}=\X - {pk-2 \over n+2}{S\over \sum_{i=1}^k\Vert\X_i\Vert^2_{\V_i^{-1}}}\X, $$ \item [] PT: the preliminary-test estimator $\bmuh_i^{PT}$ given in (\ref{eqn:PT}), \item [] PT*: the preliminary-test and shrinkage estimator $\bmuh_i^{PT*}$ in (\ref{}), \item [] EB: the empirical Bayes estimator $\bmuh_i^{EB1}$ in (\ref{}), \item [] EB*: the improved empirical Bayes estimator $\bmuh_{EB2}$ in (\ref{}) \item [] HB1: the hierarchical Bayes estimator $\bmuh_i^{HB1}$ in (\ref{}), \item [] HB2: the hierarchical Bayes estimator $\bmuh_i^{HB2}$ in (\ref{}). \end{itemize} \begin{table}[b!] \caption{Values of PRIAL of estimators PT, PT*, EB1, EB2, HB1 and HB2} \begin{center} $ {\renewcommand\arraystretch{1.1}\small \begin{array}{l@{\hspace{5mm}} r@{\hspace{2mm}} r@{\hspace{2mm}} r@{\hspace{2mm}} r@{\hspace{2mm}} r@{\hspace{2mm}} r@{\hspace{2mm}} r@{\hspace{2mm}} r } \text{$(\bmu_1, \ldots, \bmu_5)$} &\text{JS1}&\text{JS2}&\text{PT}&\text{PT*}&\text{EB}&\text{EB*}&\text{HB1}&\text{HB2}\\ \hline \text{$(\zero, \zero, \zero, \zero, \zero)$} &54.1 &82.8 &74.1 &87.2 &70.3 &83.4 &69.7 &82.6 \\ \text{$(2, 2, 2, 2, 2)\otimes \j_5$} &9.0 &14.2 &74.1 &74.4 &70.3 &70.6 &69.7 &70.0 \\ \hline \text{$(-0.4, -0.2, 0.0, 0.2, 0.4)\otimes \j_5$} &48.3 &74.4 &64.1 &75.3 &63.6 &74.9 &64.7 &76.5 \\ \text{$(-1, -0.5, 0, 0.5, 1)\otimes \j_5$} &37.4 &48.2 &17.4 &23.0 &40.6 &46.2 &44.6 &52.2 \\ \text{$(-2, -1, 0, 1, 2)\otimes \j_5$} &25.0 &21.1 &-11.6 &-10.0 &17.5 &19.1 &17.5 &18.9 \\ \hline \text{$(1.2, 1.4, 1.6, 1.8, 2.0)\otimes \j_5$} &12.7 &23.3 &64.1 &64.6 &63.6 &64.2 &64.7 &65.2 \\ \text{$(1, 1.5, 2, 2.5, 3)\otimes \j_5$} &9.5 &18.1 &17.4 &17.9 &40.6 &41.1 &44.6 &44.9 \\ \text{$(0, 1, 2, 3, 4)\otimes \j_5$} &18.9 &17.1 &-11.6 &-10.8 &17.5 &18.3 &17.5 &17.6 \\ \end{array} } $ \end{center} \label{} \end{table} The significance size in the preliminary-test and related estimators $\bmuh_i^{PT}$ and $\bmuh_i^{PT*}$ is $\al=0.05$. For the hierarchical Bayes estimators $\bmuh_i^{HB1}$ and $\bmuh_i^{HB2}$, we used the constants $a=b=c=0.1$ and $L=0$. In this simulation, we generate random numbers of $\X_1, \ldots, \X_k$ and $S$ based on Model~(\ref{}) for $p=k=5$, $n=20$, $\si^2=2$ and $\V_i=(0.1\times i)\, \I_p$ for all $i\in\{1, \ldots, k\}$. For the mean vectors $\bmu_i$, we treat the eight cases: \begin{multline*}(\bmu_1, \ldots, \bmu_5) = (\zero, \zero, \zero, \zero, \zero), (2\j_5, 2\j_5, 2\j_5, 2\j_5, 2\j_5),\\ (-0.4\j_5, -0.2\j_5, \zero, 0.2\j_5, 0.4\j_5), (-\j_5, -0.5\j_5, \zero, 0.5\j_5, \j_5), (-2\j_5, -\j_5, \zero, \j_5, 2\j_5), \\ (1.2\j_5, 1.4\j_5, 1.6\j_5, 1.8\j_5, 2.0\j_5), (\j_5, 1.5\j_5, 2\j_5, 2.5\j_5, 3\j_5),(\zero, \j_5, 2\j_5, 3\j_5, 4\j_5), \end{multline*} where $\j_p=(1, \ldots, 1)^\top\in \Re^p$. The first two are cases of equal means, the next three are cases where $\bmu_1 + \cdots + \bmu_5 =\zero$, and the last three are unbalanced cases. For each estimator $\bde=(\bmuh_1, \ldots, \bmuh_5)$, based on 5000 replicates, we obtain an approximated value of the risk function $R(\bom, \bde)={\rm E} \{ L(\bde, \bom) \}$ for the loss function given in (\ref{}). Table~\ref{} reports the percentage relative improvement in average loss (PRIAL) of each estimator $\bde$ over $\bde^U=(\X_1, \ldots, \X_5)$, defined by $$ {\rm PRIAL} = 100\{ R(\bom, \bde^U) - R(\bom, \bde)\}/R(\bom, \bde^U). $$ Note from Theorem~\ref{thm:2} that PT and EB1 can be improved upon by PT* and EB2. These results are confirmed by the simulation results in Table~\ref{}. Concerning the hierarchical Bayes estimators HB1 and HB2, Theorem~\ref{} shows that the estimator HB2 is minimax, while we could not demonstrate that HB2 dominates HB1. However, Table~\ref{} illustrates that HB2 is better than HB1. Concerning PT, EB1 and HB1, their values of PRIAL are high under the hypothesis $\mathcal{H}_0 : \bmu_1=\cdots=\bmu_5$ but decrease when the means are far away from $\mathcal{H}_0$. The empirical Bayes estimator EB1 and the hierarchical Bayes estimator HB1 are comparable, and EB2 and HB2 are also comparable. The simulation results in these setups illustrate that the hierarchical Bayes estimator HB2 and the empirical Bayes estimator EB2 have good performances. \section{Proofs} We here provide the proofs of Theorems \ref{} and \ref{}. Theorems \ref{thm:1} and \ref{thm:2} can be derived from Theorem~\ref{} through Propositions \ref{prp:1} and \ref{prp:2}. \subsection{Proof of Theorem~\ref{}} For the proof, Stein's identity \cite{stein1981} and the chi-square identity due to Efron and Morris \cite{em1976} are useful. See also Bilodeau and Kariya \cite{bk1989} for a multivariate version of Stein's identity. \begin{lem} \label{} {\rm (i)} Assume that $\Y=(Y_1, \ldots, Y_p)^\top$ is a $p$-variate random vector having $\Nc_p(\bmu, \bSi)$ and that $\h$ is an absolutely continuous function from $\Re^p$ to $\Re^p$. Then, we have Stein's identity, viz. $$ {\rm E} \{(\Y-\bmu)^\top \h(\Y)\}= {\rm E} \big[ \tr\big\{ \bSi \bnabla_\Y \h(\Y)^\top\big\}\big], $$ provided the expectations in both sides exist, where $\bnabla_\Y=(\partial /\partial Y_1, \ldots, \partial /\partial Y_p)^\top$. {\rm (ii)} Assume that $S$ is a random variable such that $S/\si^2\sim \chi_{(n)}^2$ and that $g$ is an absolutely continuous function from $\Re$ to $\Re$. Then, we have the chi-square identity, viz. $$ {\rm E} \{S g(S)\} = \si^2 {\rm E} \{n g(S) + 2Sg'(S)\}, $$ provided the expectations exist on both sides. \end{lem} The risk function is decomposed, say, as $$ R\{ \bom, \bmuh(\phi, \psi)\} = \frac{1}{\si^2} \sum_{i=1}^k {\rm E} \Big \{\Vert\bmuh_i(\phi,\psi)-\bmu_i\Vert^2_{\Q_i^{-1}} \Big\} = I_1 - 2 I_2 - 2I_3 + I_4, $$ where $$ I_1 = \frac{1}{\si^2} \sum_{i=1}^k {\rm E} \Big\{\Vert\X_i-\bmu_i\Vert^2_{\Q_i^{-1}} \Big\} , \quad I_2 = {\phi\over \si^2 F} \sum_{i=1}^k {\rm E} \{(\X_i-\bmu_i)^\top\V_i^{-1}(\X_i-\bnuh) \}, $$ $$ I_3 = {\psi\over \si^2 G} \sum_{i=1}^k {\rm E} \{(\X_i-\bmu_i)^\top\V_i^{-1}\bnuh \}, \quad I_4 = {1\over \si^2}\sum_{i=1}^k {\rm E} \left\{ \Vert{\phi }(\X_i-\bnuh)/F+{\psi}\bnuh/G \Vert^2_{\V_i^{-1}\Q_i^{-1}\V_i^{-1}} \right\}. $$ with $\phi=\phi(F,G,S)$ and $\psi=\psi(F,G,S)$. It is easy to see $I_1 = \tr(\V_1\Q_1) + \cdots + \tr(\V_k\Q_k)$. Note that \begin{align*} S\times F&=\sum_{i=1}^{k}\X_i^\top\V_i^{-1}\Q_i^{-1}\V_i^{-1}\X_i- \left(\sum_{i=1}^{k}\V_i^{-1}\Q_i^{-1}\V_i^{-1}\X_i \right)^\top \A \left(\sum_{i=1}^{k}\V_i^{-1}\Q_i^{-1}\V_i^{-1}\X_i \right),\\ S\times G&= \left(\sum_{i=1}^{k}\V_i^{-1}\Q_i^{-1}\V_i^{-1}\X_i \right)^\top \A \left(\sum_{i=1}^{k}\V_i^{-1}\Q_i^{-1}\V_i^{-1}\X_i \right). \end{align*} Letting $\bnabla_i=\partial/\partial \X_i=(\partial /\partial X_{i1}, \ldots, \partial /\partial X_{ip})^\top$ for $\X_i=(X_{i1}, \ldots, X_{ip})^\top$, we have $\bnabla_i F=2\V_i^{-1}\Q_i^{-1}\V_i^{-1}(\X_i-\bnuh)/S$ and $\bnabla_i G=2\V_i^{-1}\Q_i^{-1}\V_i^{-1}\bnuh/S$. For $I_2$, it is observed that, by Stein's identity, \begin{align*} I_2& = \sum_{i=1}^k {\rm E} \Big[\tr [\bnabla_i \{(\X_i-\bnuh)^\top {\phi(F,G,S)/F} \} ]\Big] \\ & = \sum_{i=1}^k {\rm E} \Big[ \tr\Big[ \{\bnabla_i (\X_i-\bnuh)^\top \} {\phi(F,G,S)/ F} + \{\bnabla_i {\phi(F,G,S)/ F} \}(\X_i-\bnuh)^\top\Big] \\ & = \sum_{i=1}^k {\rm E} \Big[ \tr\Big[(\I_p-\V_i^{-1}\Q_i^{-1}\V_i^{-1}\A){\phi/F}\\ & \quad \quad + (\X_i-\bnuh)^\top \{ {2\V_i^{-1}\Q_i^{-1}\V_i^{-1}(\X_i-\bnuh) } ( -{\phi/ F^2}+{\phi_F/ F} )/S+ {2 \V_i^{-1}\Q_i^{-1}\V_i^{-1}\bnuh } {\phi_G/(F}S) \}\Big]\Big]\\ & = {\rm E} \left\{ p(k-1){\phi/ F} + 2 F ( -{\phi/ F^2}+{\phi_F/ F} ) + \sum_{i=1}^k {(\X_i-\bnuh)^\top\V_i^{-1}\Q_i^{-1}\V_i^{-1}\bnuh } {\phi_G /(FS)}\right \}. \end{align*} Noting that $\sum_{i=1}^k (\X_i-\bnuh)^\top\V_i^{-1}\Q_i^{-1}\V_i^{-1}=\zero$, we get \begin{equation} I_2 = {\rm E} [ { \{p(k-1) -2\} \phi / F} + 2 \phi_F]. \label{eqn:I2} \end{equation} Similarly, we have \begin{align*} I_3 & = \sum_{i=1}^k {\rm E} \Big[\tr [\bnabla_i \{\bnuh^\top {\psi(F,G,S)/ G} \} ]\Big] \\ & = \sum_{i=1}^k {\rm E} \Big[ \tr [(\V_i^{-1}\Q_i^{-1}\V_i^{-1}\A){\psi/ G} + \bnuh^\top \{ {2\V_i^{-1}\Q_i^{-1}\V_i^{-1}\bnuh } ( -{\psi/ G^2}+{\psi_G/G} )/S + {2 \V_i^{-1}\Q_i^{-1}\V_i^{-1}(\X_i-\bnuh) }{\phi_F /(GS)}\} ]\Big]. \end{align*} Since $\sum_{i=1}^k \bnuh^\top \V_i^{-1}\Q_i^{-1}\V_i^{-1}(\X_i-\bnuh)=\zero$, we get \begin{equation} I_3 = {\rm E} \{ (p-2)\psi / G + 2 \psi_G\}. \label{eqn:I3} \end{equation} Concerning $I_4$, note that \begin{align*} I_4=& {1\over \si^2} \, {\rm E} \Big\{{S }\phi^2(F,G,S)/F\Big\} + {1\over \si^2} \, {\rm E} \Big\{{S}\psi^2(F,G,S)/G\Big\} + {2\over \si^2} \, {\rm E} \left\{\sum_{i=1}^k(\X_i-\bnuh)^\top\V_i^{-1}\Q_i^{-1}\V_i^{-1}\bnuh {\phi\psi/(FG)}\right\} \\ =& {1\over \si^2} \, {\rm E} \{{S}\phi^2(F,G,S)/F \} + {1\over \si^2} \, {\rm E} \{{S }\psi^2(F,G,S)/G \}. \end{align*} Using the chi-square identity, we have \begin{align*} {\rm E} \{(S\phi^2)/( F\si^2)\} & = {\rm E} \{ {n\phi^2/ F} + 2S {\phi^2/(FS)} + 2S {2\phi } ( - {F}\phi_F/S - {G }\phi_G/S+\phi_S )/F \} \\ & = {\rm E} \{ (n+2) {\phi^2/F} - 4{\phi }\big(F\phi_F+G\phi_G\big)/F+ 4{S}\phi\phi_S/F \}. \end{align*} Similarly, $$ {\rm E} \{(S\psi^2)/( G\si^2)\} = {\rm E} \{ (n+2) {\psi^2/ G} - 4{\psi}\big(F\psi_F+G\psi_G\big)/G+ 4{S}\psi\psi_S/G \}. $$ Thus, one gets \begin{equation} I_4 = {\rm E} \{ (n+2) {\phi^2/F} - 4{\phi}\big(F\phi_F+G\phi_G\big)/F+ 4{S}\phi\phi_S/F + (n+2) {\psi^2/ G} - 4{\psi}\big(F\psi_F+G\psi_G\big)/G+ 4{S}\psi\psi_S/G \}. \label{eqn:I4} \end{equation} Combining (\ref{eqn:I2}), (\ref{eqn:I3}) and (\ref{eqn:I4}) gives the expression in Theorem~\ref{ $\Box$ \subsection{Proof of Theorem~\ref{}} It suffices to show that $\phi^{HB2}(F,G,S)$ satisfies conditions (a) and (b) in Proposition~\ref{prp:1}, and that $\psi^{HB2}(F,G,S)$ satisfies conditions (a) and (b) in Proposition~\ref{prp:2}. For simplicity, let $h(x,y,v)=x^{\al }y^{\be}v^{\ga }\exp\{ - v(x+y+1)/2\}$ for $\al=p(k-1)+a-1$, $\be=p/2+b-1$ and $\ga=(n+pk)/2-c$. Then, $\phi^{HB2}(F,G,S)$ and $\psi^{HB2}(F,G,S)$ are written as \begin{align*} \phi^{HB2}(F,G,S)=& {\int_0^F \int_0^G \int_{L S}^\infty x h(x,y,v) dv dy dx \Big/ \int_0^F \int_0^G \int_{L S}^\infty h(x,y,v) dv dy dx}, \\ \psi^{HB2}(F,G,S)=& {\int_0^F \int_0^G \int_{L S}^\infty y h(x,y,v) dv dy dx \Big/ \int_0^F \int_0^G \int_{L S}^\infty h(x,y,v) dv dy dx}. \end{align*} We check conditions (a) and (b) in Proposition~\ref{prp:1} for $\phi^{HB2}(F,G,S)$. The proof for $\psi^{HB2}(F,G,S)$ is omitted, because it can be shown similarly. We begin by showing that $\phi^{HB2}(F,G,S)$ is increasing in $F$. The derivative of $\phi^{HB2}(F,G,S)$ with respect to $F$ is proportional to \begin{multline*} \int_0^G \int_{L S}^\infty F h(F,y,v)dvdy\times \int_0^F\int_0^G\int_{L S}^\infty h(x,y,v)dvdydx\\ -\int_0^F\int_0^G\int_{L S}^\infty x h(x,y,v)dvdydx\times \int_0^G\int_{L S}^\infty h(F,y,v)dvdy\\ = \int_0^F(F-x) \left\{\int_0^G\int_{L S}^\infty h(F,y,v)dvdy\times \int_0^G\int_{L S}^\infty h(x,y,v)dvdy\right\} dx, \end{multline*} which is non-negative. We next show that $\phi^{HB2}(F,G,S)$ is increasing in $G$. The derivative of $\phi^{HB2}(F,G,S)$ with respect to $G$ is proportional to \begin{multline*} \int_0^F \int_{L S}^\infty x h(x,G,v)dvdx\times \int_0^F\int_0^G\int_{L S}^\infty h(x,y,v)dvdydx\\ -\int_0^F\int_0^G\int_{L S}^\infty x h(x,y,v)dvdydx\times \int_0^F\int_{L S}^\infty h(x,G,v)dvdx\\ = \left\{\int_0^F\int_{L S}^\infty h(x,G,v)dvdx \right\}^2 \\ \times \left[ {\rm E} ^*(X) {\rm E} \left\{ {\int_0^G\int_{L S}^\infty h(X,y,v)dvdy \over \int_{L S}^\infty h(X,G,v)dv} \right\} - {\rm E} ^*\left\{ X {\int_0^G\int_{L S}^\infty h(X,y,v)dvdy \over \int_{L S}^\infty h(X,G,v)dv}\right\} \right], \end{multline*} where $ {\rm E} ^*$ denotes expectation with respect to the probability $$ {\rm Pr} (X\in A)=\int_{A\cap[0,F]} \int_{L S}^\infty h(x,G,v)dvdx\Big/ \int_0^F\int_{L S}^\infty h(x,G,v)dvdx. $$ Let $g(X)=\int_0^G\int_{L S}^\infty h(X,y,v)dvdy / \int_{L S}^\infty h(X,G,v)dv$. If $g(X)$ is decreasing in $X$, it is seen that $g(X)$ and $X$ are monotone in opposite directions, which implies that $ {\rm E} ^*\{X g(X)\} - {\rm E} ^*(X) {\rm E} ^*\{g(X)\}\leq 0$ from the covariance inequality. Thus, we need to show that $g(X)$ is decreasing in $X$. Note that $g(x)$ is written as $$ g(x) = \int_0^G\int_{LS}^\infty y^\be v^\ga e^{-v(x+y+1)/2}dvdy\Big/ \int_{LS}^\infty G^\be v^\ga e^{-v(x+G+1)/2}dv. $$ The derivative $g'(x)$ is proportional to \begin{align*} - \int_0^G \int_{LS}^\infty y^\be & v^{\ga+1} e^{-v(x+y+1)/2}dvdy \int_{LS}^\infty v^\ga e^{-v(x+G+1)/2}dv\\ & + \,\int_0^G\int_{LS}^\infty y^\be v^\ga e^{-v(x+y+1)/2}dvdy \int_{LS}^\infty v^{\ga+1} e^{-v(x+G+1)/2}dv\\ \\ &= \left\{\int_{LS}^\infty v^\ga e^{-v(x+G+1)/2}dv\right\}^2 \left[ - {\rm E} ^\dagger \left\{ V {\int_0^G y^\be e^{-V(x+y+1)/2}dy\over e^{-V(x+G+1)/2}}\right\} + {\rm E} ^\dagger(V) {\rm E} ^\dagger \left\{ {\int_0^G y^\be e^{-V(x+y+1)/2}dy\over e^{-V(x+G+1)/2}}\right\} \right]\\ & = \left\{\int_{LS}^\infty v^\ga e^{-v(x+G+1)/2}dv\right\}^2 \left[ - {\rm E} ^\dagger \left\{ V \int_0^G y^\be e^{(G-y)V}dy\right\} + {\rm E} ^\dagger(V) {\rm E} ^\dagger \left\{ \int_0^G y^\be e^{(G-y)V}dy \right\} \right], \end{align*} where $ {\rm E} ^\dagger$ denotes the expectation with respect to the probability $$ {\rm Pr} (V\in A)=\int_{A\cap[LS,\infty)} v^\ga e^{-v(x+G+1)/2}dv\Big/ \int_{LS}^\infty v^\ga e^{-v(x+G+1)/2}dv $$ for fixed $x$. Because $\int_0^G y^\be e^{(G-y)V}dy$ is increasing in $V$ for $y<G$, the covariance inequality implies that $$ - {\rm E} ^\dagger \left\{ V \int_0^G y^\be e^{(G-y)V}dy\right\} + {\rm E} ^\dagger(V) {\rm E} ^\dagger \left\{ \int_0^G y^\be e^{(G-y)V}dy\right\}\leq 0, $$ which means that $g'(x)\leq 0$. Hence, $\phi^{HB2}(F,G,S)$ is increasing in $G$. We now show that $\phi^{HB2}(F,G,S)$ is decreasing in $S$. The derivative of $\phi^{HB2}(F,G,S)$ with respect to $S$ is proportional to \begin{align*} -L \int_0^F \int_0^G x h(x,y,L S)dydx & \int_0^F\int_0^G\int_{L S}^\infty h(x,y,v)dvdydx \\ +\, L\int_0^F\int_0^G\int_{L S}^\infty & x h(x,y,v)dvdydx \int_0^F\int_0^G h(x,y,L S)dydx \\ & \hspace{-5mm} = L \left\{\int_0^F\int_0^G h(x,y,L S)dydx\right\}^2 \times \\ & \quad \quad \left[ - {\rm E} ^{**}(X)\times {\rm E} ^{**} \left\{ { \int_0^G\int_{L S}^\infty h(X,y,v)dvdy \over \int_0^G h(X,y,L S)dy} \right\} + {\rm E} ^{**} \left\{ X { \int_0^G\int_{L S}^\infty h(X,y,v)dvdy \over \int_0^G h(X,y,L S)dy}\right\} \right], \end{align*} where $ {\rm E} ^{**}$ is the expectation with respect to the probability $$ {\rm Pr} (X\in A) = \int_{A\cap[0,F]} \int_0^G h(x,y,L S)dydx\Big/ \int_0^F\int_0^G h(x,y,L S)dydx. $$ Note that $$ { \int_0^G\int_{L S}^\infty h(X,y,v)dvdy} {\Big /} { \int_0^G h(X,y,L S)dy} = {\int_0^G\int_{LS}^\infty y^\be v^\ga e^{(LS-v)X - v(y+1)/2}dvdy} {\Big /} { \int_0^G y^\be(LS)^\ga e^{-LS(y+1)/2}dy}, $$ which is decreasing in $X$. Thus, $X$ and $\int_0^G\int_{L S}^\infty h(X,y,v)dvdy /\int_0^G h(X,y,L S)dy$ are monotone in opposite directions, so that the derivative $\phi^{HB2}(F,G,S)$ with respect to $S$ is negative due to the covariance inequality. Hence, condition (a) is satisfied for $\phi^{HB2}(F,G,S)$. Finally, from condition (a), we see that \begin{align*} \phi^{HB2}(F,G,S)\leq& \lim_{F\to\infty}\lim_{G\to\infty}\lim_{S\to 0} \phi^{HB2}(F,G,S) = {\int_0^\infty\int_0^\infty \int_{0}^\infty x^{\al+1 }y^{\be }v^\ga\exp\{ - y(x+y+1)/2\}dvdy dx \over \int_0^\infty \int_{0}^\infty x^{\al}y^{\be }v^\ga\exp\{ - y(x+y+1)/2\}dvdy dx}\\ =& {\Ga(\al+2)\Ga(\be+1)\Ga(\ga-\al-\be-2)\over \Ga(\al+1)\Ga(\be+1)\Ga(\ga-\al-\be-1)} ={\al+1 \over \ga-\al-\be-2}={p(k-1)+2a \over n- 2(a+b+c)}. \end{align*} Similarly, \begin{align*} \psi^{HB2}(F,G,S)\leq& \lim_{F\to\infty}\lim_{G\to\infty}\lim_{S\to 0} \psi^{HB2}(F,G,S) = {\int_0^\infty\int_0^\infty \int_{0}^\infty x^{\al }y^{\be+1 }v^\ga\exp\{ - y(x+y+1)/2\}dvdy dx \over \int_0^\infty \int_{0}^\infty x^{\al}y^{\be }v^\ga\exp\{ - y(x+y+1)/2\}dvdy dx}\\ =& {\Ga(\al+1)\Ga(\be+2)\Ga(\ga-\al-\be-2)\over \Ga(\al+1)\Ga(\be+1)\Ga(\ga-\al-\be-1)} ={\be+1 \over \ga-\al-\be-2}={p+2b \over n- 2(a+b+c)}. \end{align*} Thus, condition (b) in Proposition~\ref{prp:1} is satisfied if $a +b +c < n/2$ and if $$ \{p(k-1)+2a\}/\{n- 2(a+b+c)\} \leq 2 \{p(k-1)-2 \}( n+2), $$ which is equivalently rewritten by (\ref{}). Also, condition (b) in Proposition~\ref{prp:2} is satisfied if $a +b +c < n/2$ and if $$ (p+2b)/\{n- 2(a+b+c)\} \leq 2(p-2)/(n+2), $$ which leads to (\ref{}). Hence, the minimaxity of the hierarchical Bayes estimator is established. \section*{Acknowledgments} We thank the Editor-in-Chief, the Associate Editor and the reviewers for valuable comments and helpful suggestions which led to an improved version of this paper. Research of the first author was supported by Grant-in-Aid for JSPS Research Fellows (18J21162). Research of the second author was supported in part by Grant-in-Aid for Scientific Research (18K11188, 15H01943, 26330036) from Japan Society for the Promotion of Science.

\section{Introduction} A linear map $\Phi : M_d(\mathbb{C}) \rightarrow M_d(\mathbb{C})$ is covariant with respect to the unitary $d$-dimensional representation $U$ of the group $G$ if and only if \begin{equation}\label{cov_def} \bigforall_{g\in G}\bigforall_{X\in M_d(\mathbb{C})}\quad \Phi\big[U(g)X U^\dagger(g)\big]=U(g)\Phi[X]U^\dagger(g). \end{equation} If $\Phi$ is also completely positive and trace-preserving, then it is called the {\it covariant quantum channel}. This notion first appeared in \cite{Scutaru}, where Scutaru proved the Stinespring type theorem for covariant completely positive maps. However, the covariant quantum channels were first analyzed together with the covariant Markovian generators by Holevo \cite{Holevo1993,CQME}. Examples of covariant channels include depolarizing channels \cite{Keyl} and transpose depolarizing channels \cite{additiv,DHS}. The channels that are covariant with respect to the irreducible unitary representation are called irreducibly covariant and posess an interesting property. Namely, if the channel is irreducibly covariant with respect to a compact group, then its Holevo capacity is additive and linearly proportional to the minimum output entropy \cite{Holevo2000}. This discovery started a series of articles dealing with additivity and multiplicity of minimum output entropy for irreducibly covariant quantum channels \cite{additiv,WernerHolevo,DHS2,Holevo2005,DFH,Fukuda,King}. There are the channels known as Weyl-covariant which are covariant with respect to the Weyl operators \cite{CQME,Amosov,Holevo2005,DFH}. It turns out that they satisfy condition (\ref{cov_def}) with $G$ being the group generated by the Weyl operators. The properties of unitarily covariant and Weyl-covariant channels in dimension $d\geq 2$ were studied by Datta, Fukuda, and Holevo in \cite{DFH}, whereas Mendl and Wolf \cite{20} dealt with the channels covariant with respect to O($d$). Nuwairan \cite{23} introduced the so-called EPOSIC channels and showed that they form a set of extreme points of the SU(2)-irreducibly covariant channels. Finally, Jen\u{c}ov\'{a} and Pl\'{a}vala \cite{17} provided the optimality conditions in discriminating covariant quantum channels. In this paper, we construct the linear maps $\Phi$ that are covariant with respect to the unitary representation $U$ of a finite group $G$. In particular, we construct the Weyl maps by requiring that the group $G$ is the generalization of the quaternion group generated by the Weyl operators. It turns out that such $\Phi$ belong to the class of irreducibly covariant linear maps with simply reducible $U\otimes U^c$, which has been recently analyzed by Mozrzymas et. al. \cite{MSD}. We provide the necessary and sufficient conditions for the Weyl maps to be completely positive and trace-preserving. For the $d$-dimensional Weyl channels, there are $d-1$ choices of inequivalent representations $U$. The covariant maps possessing the additional symmetry $S \Phi[X]S^\dagger =\Phi[SXS^\dagger]$, with $S$ being a suitable permutation matrix, satisfy $\Phi=\Phi^\ast$, where $\Phi^\ast$ is the map dual to $\Phi$. Applying the additional conditions that follow from the group representation theory allows us to obtain the generalized Pauli channels for prime dimensions $d$. Finally, we provide two methods of constructing the Weyl-covariant positive, trace-preserving maps. \section{Generalization of the quaternion group} Consider a set of unitary Weyl operators in $\mathbb{C}^d$ \begin{equation} W_{kl}=\sum_{m=0}^{d-1}\omega^{km}|m+l\>\<m|,\qquad \omega=e^{2\pi\imag/d}, \end{equation} satisfying well known properties \begin{equation}\label{group_action} W_{kl}W_{rs}=\omega^{ks}W_{k+r,l+s},\qquad W_{kl}^\dagger=\omega^{kl}W_{-k,-l}. \end{equation} Moreover \begin{equation}\label{} \mathrm{Tr}(W^\dagger_{kl} W_{mn}) = \delta_{km}\delta_{ln} , \end{equation} that is, `renormalized' operators $W_{kl}/\sqrt{d}$ define an orthonormal basis in ${M}_d(\mathbb{C})$ (with respect to the Hilbert-Schmidt inner product). Let $G$ denote a finite group generated by $W_{kl}$. It consists in the following elements \begin{equation} G= \Big\{\, \omega^mW_{kl}\, |\, k,l,m=0,1,\dots,d-1\, \Big\}, \end{equation} and hence $|G|=d^3$. One easily find the conjugacy classes of $G$, \begin{equation}\label{conj} \mathcal{C}_{0}^l=\{\omega^lW_{00}\},\quad \mathcal{C}_{kl}=\{\omega^mW_{kl}|m=0,1,\dots,d-1\}, \end{equation} for $k,l=0,1,\dots,d-1$ and $(k,l)\neq (0,0)$. \begin{Proposition} There are `$d(d+1)-1$' irreducible representations of $G$ \begin{enumerate} \item $d^2$ one-dimensional (denoted by $\phi_0=\mathrm{id},\phi_1,\dots,\phi_{d^2-1}$), and \item $d-1$ $d$-dimensional unitary representations (denoted by $U_\alpha$, $\alpha=1,\dots,d-1$). \end{enumerate} \end{Proposition} \begin{proof} The dimensions of all the non-equivalent irreducible representations are related to the number of elements $|G|=d^3$ by \begin{equation}\label{AAA} |G|=\sum_{k=1}^Nn_kd_k^2, \end{equation} where $n_k$ is the number of $d_k$-dimensional representations. The number of nonequivalent irreducible representations is equal to the number of equivalence classes (\ref{conj}) of the group, of which there are \begin{equation}\label{BBB} d+d^2-1=\sum_{k=1}^Nn_k. \end{equation} Moreover, we know that every group has the identity representation, which is one-dimensional, so that $d_1=1$. The following property of the characters, \begin{equation} |G|=\sum_{g\in G}\left|\chi^\alpha(g)\right|^2, \end{equation} and the fact that one-dimensional representations satisfy the group action imply that $\chi^\alpha(g)\in\{1,\omega,\dots,\omega^{d-1}\}$. Moreover, there are at most $n_1=d^2$ orthogonal rows of the character table. The group $G$ is generated by the $d$-dimensional Weyl operators, and therefore $W_{kl}$ define an irreducible unitary representation of $G$. Denote its dimension by $d_2=d$. From (\ref{AAA}) and (\ref{BBB}), it follows that \begin{equation} d-1-n_2=\sum_{k=3}^Nn_k,\qquad\sum_{k=3}^Nn_k\left(1-\frac{d_k^2}{d^2}\right)=0. \end{equation} As $d_k\neq d$ for $k\geq 3$, the only solution is $n_k=0$, $k\geq 3$, and $n_2=d-1$. \end{proof} The character table has the structure presented below. \FloatBarrier \begin{figure}[ht!] \centering \tiny \includegraphics[width=0.4\textwidth]{1.jpg} \end{figure} \FloatBarrier In the above table, we introduced the matrices with the following entries: $\mathcal{I}_{ij}=1$, the Hadamard matrix $(\mathbb{F}_d)_{kl}=\omega^{kl}$ ($k,l=0,\dots,d-1$), $\mathcal{D}_m=d$, $\mathbb{O}_{mn}=0$, and $\Omega_{ij}=d\omega^{ij}$ ($i,j=1,\dots,d-1$). The characters of the irreducible representations \begin{equation} \chi^{\mathrm{id}},\chi^{\phi_1},\dots,\chi^{\phi_{d^2-1}},\chi^{U_1},\dots,\chi^{U_{d-1}} \end{equation} number the rows of the table, and the conjugacy classes \begin{equation}\label{klasy} \mathcal{C}_{0}^1,\dots,\mathcal{C}_{0}^{d-1}, \mathcal{C}_{0}^0\equiv\mathcal{C}_{00},\dots,\mathcal{C}_{0,d-1}, \dots, \mathcal{C}_{d-1,0},\dots,\mathcal{C}_{d-1,d-1} \end{equation} number the columns. From definition, the columns and rows of the matrix $\mathbb{F}_d\otimes\overline{\mathbb{F}}_d$ are mutually orthogonal. We check that the one-dimensional representations satisfy the group action (\ref{group_action}), \begin{equation} \chi^{md+n}(\mathcal{C}_{kl})\chi^{md+n}(\mathcal{C}_{rs}) =\chi^{md+n}(\mathcal{C}_{k+r,l+s}), \end{equation} $m,n=0,\dots,d-1$, as \begin{equation} \chi^{md+n}(\mathcal{C}_{kl})=\left(\mathbb{F}_d\otimes\overline{\mathbb{F}}_d \right)_{mn,kl}=\omega^{mk-nl}. \end{equation} Let us denote the known unitary representation (with the Weyl operators) by $U_1$. For $d=2$, this is the only irreducible representation of dimension higher than one. Observe that for prime $d>2$, the other unitary representations follow from the group action written in terms of the group generators, \begin{equation}\label{action} W_{kl}W_{rs}=(\omega W_{00})^{ks}W_{k+r,l+s}. \end{equation} \begin{Proposition}\label{prop} For prime $d$, the set of $d$-dimensional irreducible unitary representations of $G$ consist in $U_1$ and the representations constructed from $U_1$ in the following way, \begin{align} U_1\to U_\alpha:\quad &\omega^mW_{kl}\to \omega^mW_{\alpha k,\alpha l},\nonumber\\ &\omega^m W_{00}\to \omega^{m\alpha^2}W_{00},\label{U}\\ U_1\to \overline{U}_\alpha:\quad &\omega^mW_{kl}\to \omega^m\overline{W}_{\alpha k,\alpha l}=\omega^mW_{-\alpha k,\alpha l},\nonumber\\ &\omega^m W_{00}\to \omega^{-m\alpha^2}W_{00},\label{Ubar} \end{align} for $\alpha=1,\dots,\frac{d-1}{2}$. To number the rows of the character table, they are ordered as follows, \begin{equation} U_1,\dots,U_{\frac{d-1}{2}},\overline{U}_{\frac{d-1}{2}},\dots,\overline{U}_1. \end{equation} \end{Proposition} \begin{proof} The fact that $U_1$ is an irreducible representation of $G$ is obvious, as $U_1(g)$ have the right dimensions, characters, and they satisfy the group action. After making substitutions (\ref{U}) and (\ref{Ubar}) in eq. (\ref{action}), we get \begin{equation} W_{\alpha k,\alpha l}W_{\alpha r,\alpha s}=(\omega^{\alpha^2} W_{00})^{ks}W_{\alpha(k+r),\alpha(l+s)}, \end{equation} \begin{equation} \overline{W}_{\alpha k,\alpha l}\overline{W}_{\alpha r,\alpha s}=(\omega^{-\alpha^2} W_{00})^{ks}\overline{W}_{\alpha(k+r),\alpha(l+s)}, \end{equation} respectively, which agree with (\ref{action}). Observe that $\overline{U}_\alpha$ is the contragredient (dual) representation of $U_\alpha$. It is worth noting that for $\alpha=\frac{d+1}{2},\dots,d-1$, (\ref{U}) and (\ref{Ubar}) also produce the irreducible unitary representations, but they are equivalent to the ones already obtained (they correspond to the same rows of the character table). \end{proof} \section{Covariant quantum channels} {Consider an arbitrary finite group} $G$ and its unitary representation $U$ in $\mathbb{C}^d$. Let us construct the linear map \begin{equation}\label{kanal} \Phi=\sum_{g\in G}\mu(g)\mathrm{Ad}_{U(g)}, \end{equation} where the adjoint representation is defined by \begin{equation} \mathrm{Ad}_{U(g)}(X)=U(g)XU^\dagger(g). \end{equation} This class of maps defines a $\star$-algebra: if \begin{equation} \Psi=\sum_{g\in G}\nu(g)\mathrm{Ad}_{U(g)}, \end{equation} then \begin{equation}\label{} \Phi \Psi = \sum_{g,h\in G}\mu(g)\nu(h)\mathrm{Ad}_{U(gh)} = \sum_{g\in G}(\mu\ast \nu)(g) \mathrm{Ad}_{U(g)}, \end{equation} where \begin{equation}\label{} (\mu\ast \nu)(g) = \sum_{h \in G} \mu(h) \nu(h^{-1}g) , \end{equation} is the convolution on $G$. Moreover, the dual {map} $\Phi^\ast$ defined by \begin{equation}\label{} {\rm Tr}(\Phi[X]^\dagger Y) =: {\rm Tr}(X^\dagger \Phi^*[Y]) , \end{equation} reads \begin{equation} \Phi^* =\sum_{g\in G}\overline{\mu(g)}\mathrm{Ad}_{U(g)}. \end{equation} Now, formula (\ref{cov_def}) implies that $\Phi$ is covariant with respect to $U$ if and only if \begin{equation} \sum_{h\in G}\mu(h)\mathrm{Ad}_{U(g)U(h)} =\sum_{h\in G}\mu(h)\mathrm{Ad}_{U(h)U(g)} . \end{equation} Using $U(g)U(h)=U(gh)$, one has \begin{equation}\label{1} \sum_{h\in G}\mu(h)\mathrm{Ad}_{U(gh)}=\sum_{h\in G}\mu(h)\mathrm{Ad}_{U(hg)}. \end{equation} Note that \begin{equation} \sum_{h\in G}\mu(h)\mathrm{Ad}_{U(hg)}=\sum_{ghg^{-1}\in G}\mu(ghg^{-1})\mathrm{Ad}_{U(gh)}, \end{equation} which, together with (\ref{1}), results in the following condition for $\mu(g)$, \begin{equation} \bigforall_{g,h\in G}\quad \mu(g)=\mu(hgh^{-1}). \end{equation} Therefore, $\mu(g)$ depends on the classes $\mathcal{C}(g)$ of the group elements $g$ instead of the elements themselves. {Denoting the conjugacy classes by $C_k$ $(k=1,2,\ldots,r)$,} one has \begin{equation}\label{} \Phi=\sum_{k=1}^r \mu_k \Phi_k , \end{equation} where $\mu_k=\mu(g)$ for $g \in C_k$, and \begin{equation}\label{} \Phi_k=\sum_{g\in C_k}\mathrm{Ad}_{U(g)}. \end{equation} For the linear map $\Phi$ to describe the quantum channel, additional constraints have to be imposed upon $\mu(g)$. \begin{Theorem}\label{tw} The map $\Phi$ in (\ref{kanal}) is the quantum channel if its coefficients $\mu(g)$ satisfy \begin{equation}\label{warunki} \mu(g)\geq 0,\qquad \sum_{g\in G}\mu(g)=1. \end{equation} \end{Theorem} Interestingly, as we shall see, conditions (\ref{warunki}) are only sufficient and not necessary. {Now, consider} the group $G$ generated by the $d$-dimensional Weyl operators. Construct the linear map $\Phi$ that is covariant with respect to the representation $U_1$. Let us assume that $\mu(g)$ is the linear combination of the characters, \begin{equation}\label{mu} \mu(\mathcal{C}(g))=\frac{1}{|G|}\left(\sum_{\alpha=0}^{d^2-1}\nu_\alpha\chi^\alpha(\mathcal{C}(g)) +\sum_{\alpha=1}^{d-1}\tau_\alpha\chi^{U_\alpha}(\mathcal{C}(g))\right), \end{equation} where $\chi^\alpha$ is the character of the one-dimensional representation $\phi_\alpha$, and $\chi^{U_\alpha}$ denotes the character of the $d$-dimensional representation $U_\alpha$. Using the character table of $G$, one arrives at \begin{equation}\label{mu1} \mu(\mathcal{C}_{kl})=\frac{1}{|G|}\sum_{m,n=0}^{d-1}\nu_{mn}\omega^{mk-nl}, \qquad (k,l)\neq (0,0), \end{equation} \begin{equation}\label{mu2} \mu(\mathcal{C}_0^l)=\frac{1}{|G|}\sum_{m,n=0}^{d-1}\nu_{mn}+ \frac{d}{|G|}\sum_{\alpha=1}^{d-1}\tau_\alpha \omega^{\alpha l}, \qquad l=0,\dots,d-1. \end{equation} Therefore, the resulting map is the Weyl map \begin{equation}\label{inv} \Phi[X]=\sum_{k,l=0}^{d-1}\mu_{kl}W_{kl}XW_{kl}^\dagger, \end{equation} where $\mu_{kl}=d\mu(\mathcal{C}_{kl})$ and \begin{equation} \mu_{00}=\sum_{l=0}^{d-1}\mu(\mathcal{C}_0^l)=\frac{d}{|G|}\sum_{m,n=0}^{d-1}\nu_{mn}. \end{equation} Note that parameters $\tau_\alpha$ are completely redundant, {as} they do not enter the final formula for $\mu_{kl}$. \begin{example} For $d=2$, $G=Q_8$ is the quaternion group, for which $U(g)$ are given by the Pauli matrices $\{\pm\sigma_0,\pm\sigma_1,\pm\imag\sigma_2,\pm\sigma_3\}$. Now, we construct the map \begin{equation} \Phi[X]=\sum_{k=1}^2\mu(\mathcal{C}_0^k)X+\sum_{k=1}^3\mu(\mathcal{C}_k)\sigma_kX\sigma_k, \end{equation} where $\mathcal{C}_1=\{\pm\sigma_1\}$, $\mathcal{C}_2=\{\pm\imag\sigma_2\}$, $\mathcal{C}_3=\{\pm\sigma_3\}$, and $\mathcal{C}_0^k=\{(-1)^k\sigma_0\}$. The coefficients $\mu(\mathcal{C}_0^k)$ and $\mu(\mathcal{C}_k)$ read \begin{equation} \begin{split} \mu(\mathcal{C}_1)&=\frac 18 \left(\nu_{00}-\nu_{01}+\nu_{10}-\nu_{11}\right),\\ \mu(\mathcal{C}_2)&=\frac 18 \left(\nu_{00}-\nu_{01}-\nu_{10}+\nu_{11}\right),\\ \mu(\mathcal{C}_3)&=\frac 18 \left(\nu_{00}+\nu_{01}-\nu_{10}-\nu_{11}\right),\\ \mu(\mathcal{C}_0^0)&=\frac 18 \left(\nu_{00}+\nu_{01}+\nu_{10}+\nu_{11}+2\tau_1\right),\\ \mu(\mathcal{C}_0^1)&=\frac 18 \left(\nu_{00}+\nu_{01}+\nu_{10}+\nu_{11}-2\tau_1\right) , \end{split} \end{equation} and hence $\Phi$ can be rewritten as \begin{equation}\label{Pauli} \Phi[X]=\sum_{k=0}^3\mu_k\sigma_kX\sigma_k \end{equation} with $\mu(\mathcal{C}_k)=2\mu_k$, $k=1,\ 2,\ 3$, and \begin{equation} \mu_0=\frac 14 \left(\nu_{00}+\nu_{01}+\nu_{10}+\nu_{11}\right). \end{equation} Note that $\mu(\mathcal{C}_0^k)$ can be negative, provided that $\mu(\mathcal{C}_0^0) + \mu(\mathcal{C}_0^1) \geq 0$. \end{example} \begin{Theorem}\label{quantum_channel} The Weyl map $\Phi$ given in (\ref{inv}) is a quantum channel if and only if \begin{equation}\label{CP_war} \mu_{kl}\geq 0,\qquad k,l=0,\dots,d-1,\qquad \sum_{k,l=0}^{d-1}\mu_{kl}=1. \end{equation} \end{Theorem} \begin{proof} Recall that $\Phi$ is a quantum channel if and only if it is completely positive and trace-preserving. Complete positivity is guaranteed by $J(\Phi)\geq 0$, where $J(\Phi)$ is the associated Choi matrix \begin{equation} J(\Phi)=\sum_{i,j=0}^{d-1}e_{ij}\otimes\Phi[e_{ij}], \end{equation} and $e_{ij}=|i\>\<j|$. It is easy to see that the Choi matrix satisfies the eigenvalue equation \begin{equation} J(\Phi)|v_{kl}\>=d\mu_{kl}|v_{kl}\> \end{equation} with \begin{equation} |v_{kl}\>=\sum_{i=0}^{d-1}|i\>\otimes W_{kl}|i\>. \end{equation} Therefore, $J(\Phi)\geq 0$ is equivalent to $\mu_{kl}\geq 0$. Showing that $\Phi$ preserves the trace if and only if $\sum_{k,l=0}^{d-1}\mu_{kl}=1$ is elementary. \end{proof} \section{Construction of the irreducibly covariant quantum channels} \label{IC} In this section, we follow the method of constructing the irreducibly covariant linear maps presented in \cite{MSD}. Take a finite group $G$ and its irreducible unitary representation $U$. With $U$, we associate the contragredient representation \begin{equation}\label{contragredient} U^c(g)=U^T(g^{-1})\equiv\overline{U}(g). \end{equation} It turns out that the representation $U\otimes U^c$ is reducible, and therefore it can be expressed as the sum \begin{equation}\label{UotimesUc} U\otimes U^c=\bigoplus_{\alpha\in\Theta} m_\alpha \phi_\alpha, \end{equation} of irreducible unitary representations $\phi_\alpha$ with multiplicity \begin{equation}\label{m} m_\alpha=\frac{1}{|G|}\sum_{g\in G}\chi^\alpha(g^{-1})|\chi^U(g)|^2, \end{equation} where $\chi^\alpha$ is the character of the representation $\phi_\alpha$. The procedure in \cite{MSD} is applicable only to the multiplicity-free case, in which $m_\alpha=1$ for $\alpha\in\Theta$ and $m_\alpha=0$ for $\alpha\notin\Theta$. Then, $U\otimes U^c$ is called simply reducible. The linear map $\Phi$ that is irreducibly covariant with respect to the group $G$ is constructed in terms of its spectral decomposition, \begin{equation}\label{Phi} \Phi=\sum_{\alpha\in\Theta}\ell_\alpha\Pi_\alpha. \end{equation} In the above formula, $\ell_\alpha$ are the eigenvalues of $\Phi$, and \begin{equation}\label{Pi} \Pi_\alpha[X]=\frac{d_\alpha}{|G|}\sum_{g\in G}\chi^\alpha(g^{-1}) U(g)XU^\dagger(g), \end{equation} where $|G|$ denotes the order of $G$, $d_\alpha=\dim\phi_\alpha$. Additionally, $\Phi$ is a quantum channel if and only if it is completely positive and trace-preserving. \begin{Proposition}\label{speccase} The irreducibly covariant quantum channel (\ref{Phi}) is a special case of the covariant quantum channel in Theorem \ref{tw}, where $U$ is the irreducible unitary representation, $U\otimes U^c$ is simply reducible, and \begin{equation} \mu(g)=\frac{1}{|G|}\sum_{\alpha\in\Theta}d_\alpha\ell_\alpha\overline{\chi}^\alpha(g). \end{equation} \end{Proposition} The following theorem shows that the Weyl channels belong to the class presented in Proposition \ref{speccase}. \begin{Theorem} For every $U_\alpha$ and $\overline{U}_\alpha$, decomposition (\ref{UotimesUc}) is multiplicity-free -- that is, the only non-vanishing multiplicities correspond to $\phi_\alpha$ and equal $m_\alpha=1$. \end{Theorem} \begin{proof} From the character table, we know that \begin{equation} \left|\chi_\beta^{U_\alpha}(\mathcal{C}_{kl})\right|^2= \left|\chi_\beta^{\overline{U}_\alpha}(\mathcal{C}_{kl})\right|^2= 0 \end{equation} for $(k,l)\neq(0,0)$ and \begin{equation} \left|\chi_\beta^{U_\alpha}(\mathcal{C}_{0}^m)\right|^2= \left|\chi_\beta^{\overline{U}_\alpha}(\mathcal{C}_{0}^m)\right|^2= d^2 \end{equation} for $m=0,\dots,d-1$. This allows us to simplify eq. (\ref{m}) to \begin{equation} m_\alpha=\frac 1d \sum_{g\in\bigcup_{m=0}^{d-1}\mathcal{C}_0^m}\chi^\alpha(g^{-1})=1 \end{equation} for the one-dimensional representations $\phi_\alpha$ and \begin{equation} m_{U_\alpha}=\frac 1d \sum_{g\in\bigcup_{m=0}^{d-1}\mathcal{C}_0^m}\chi_\beta^{U_\alpha}(g^{-1})= \sum_{k=0}^{d-1}\omega^k=0 \end{equation} for the $d$-dimensional representations $U_\alpha$. Analogical calculations show that $m_{\overline{U}_\alpha}=0$. \end{proof} Now, using the unitary representation $U_1$ of the group $G$, we construct the irreducibly covariant linear map \begin{equation}\label{phi} \Phi=\sum_{k,l=0}^{d-1}\ell_{kl}\Pi_{kl}, \end{equation} where, from definition (\ref{Pi}), \begin{equation}\label{proj} \Pi_{kl}[X]=\frac{1}{d^2}\sum_{m,n=0}^{d-1}\omega^{-mk+nl}W_{mn}XW_{mn}^\dagger. \end{equation} \begin{Proposition} {The operators given in (\ref{proj}) are the rank-1 projectors onto the Weyl operators,} \begin{equation}\label{proj2} \Pi_{kl}[X]=\frac{1}{d}W_{kl}\mathrm{Tr}(W_{kl}^\dagger X). \end{equation} \end{Proposition} \begin{proof} {To prove this,} it is enough to see that, when acting on the unitary basis of the Weyl operators, eqs. (\ref{proj}-\ref{proj2}) both produce \begin{equation} \Pi_{kl}[W_{mn}]=\delta_{km}\delta_{nl}W_{mn}. \end{equation} \end{proof} \begin{Theorem} The irreducibly covariant linear map defined in (\ref{phi}) with $\Pi_{kl}$ given by (\ref{proj}) is the Weyl map -- that is, \begin{equation}\label{WeylCh} \Phi[X]=\sum_{k,l=0}^{d-1}p_{kl}W_{kl}XW_{kl}^\dagger. \end{equation} Moreover, it is the Weyl channel if and only if $p_{kl}\geq 0$ and $\sum_{k,l=0}^{d-1}p_{kl}=1$. \end{Theorem} \begin{proof} If we substitute (\ref{proj}) into (\ref{phi}), we see that \begin{equation} \Phi[X]=\frac{1}{d^2}\sum_{k,l,m,n=0}^{d-1}\omega^{-mk+nl}\ell_{kl}W_{mn}XW_{mn}^\dagger, \end{equation} which for \begin{equation}\label{p} p_{mn}:=\frac{1}{d^2}\sum_{k,l=0}^{d-1}\omega^{-mk+nl}\ell_{kl} \end{equation} is exactly (\ref{WeylCh}). Complete positivity and preserving the trace follow from Theorem \ref{quantum_channel} for $\ell_{kl}=\overline{\nu}_{kl}$ and $p_{kl}=\mu_{kl}$. \end{proof} \section{Generalized Pauli channels} For a prime $d$, there exists a subclass of the Weyl channels with additional symmetries, known as the generalized Pauli channels \cite{Ruskai,DCKS,DCKS2}. {The} Weyl channel $\Phi$ in (\ref{WeylCh}) is the generalized Pauli channel $\Phi_{GPC}$ if and only if \begin{equation}\label{GPC_group} \ell_{\alpha k,\alpha l}=\ell_{kl},\qquad p_{\alpha k,\alpha l}=p_{kl}. \end{equation} One finds the following Kraus representation, \begin{equation}\label{GPC_form} \Phi_{GPC}[X]=\pi_0X+\frac{1}{d-1}\left[\sum_{k=1}^{d}\pi_k\sum_{\alpha=1}^{d-1} W_{\alpha k,\alpha}XW_{\alpha k,\alpha}^\dagger+\pi_{d+1}\sum_{\alpha=1}^{d-1}W_{\alpha 0}XW_{\alpha 0}^\dagger\right], \end{equation} with $\pi_k\geq 0$, $\sum_{k=0}^{d+1}\pi_k=1$. Therefore, to obtain the generalized Pauli channels, we need to impose additional conditions on the quantum channels that are irreducibly covariant with respect to the group $G$ generated by the Weyl operators. Let us recall that two unitary representations $V_\alpha$ and $V_\beta$ of the group $G$ are equivalent if and only if there exists the transformation matrix $S$ such that \begin{equation}\label{equiv} \bigforall_{g\in G}\quad V_\beta(g)=SV_\alpha(g)S^\dagger. \end{equation} \begin{Proposition}\label{prop2} For the group generated by the Weyl operators, there are the following pairs of equivalent unitary representations, \begin{equation} \{U_\alpha,U_{d-\alpha}\},\quad\{\overline{U}_\alpha,\overline{U}_{d-\alpha}\}, \quad \alpha=1,\dots,d-1, \end{equation} where $S$ is the following permutation matrix \begin{equation}\label{S} S=\sum_{m=0}^{d-1}|m\>\<-m|. \end{equation} \end{Proposition} \begin{proof} The fact that the representations $U_\alpha$ and $U_{d-\alpha}$ (and also $\overline{U}_\alpha$ and $\overline{U}_{d-\alpha}$) are equivalent has been discussed in the proof to Proposition \ref{prop}. To find the transformation matrix $S$, we assume that \begin{equation} \label{} S=\sum_{m,n=0}^{d-1}s_{mn}W_{mn}. \end{equation} For the unitary representations $U_\alpha$ and $\overline{U}_\alpha$ of the group $G$, eq. (\ref{equiv}) implies that \begin{equation} \bigforall_{k,l=0,\dots,d-1}\quad W_{kl}=SW_{-k,-l}S^\dagger. \end{equation} Using the group action of $G$ generated by the Weyl operators $W_{kl}$, one gets (\ref{S}). \end{proof} \begin{Remark} The transformation matrix $S$ from Proposition \ref{prop2} recovers $A_{00}$ in the definition of the discrete Wigner function \cite{Wootters}, \begin{equation} \mathcal{W}_{kl}=\frac 1d \mathrm{Tr}(\rho A_{kl}), \end{equation} where \begin{equation}\label{A} A_{kl}=\sum_{m=0}^{d-1}\omega^{2(m-k)l}|m\>\<-m+2k|. \end{equation} \end{Remark} \begin{Proposition}\label{prop3} Let $S$ be the transformation matrix given in Proposition \ref{prop2}. If the Weyl channel $\Phi$ is covariant with respect to $S$, that is, \begin{equation}\label{S_cov} \Phi[SXS^\dagger]=S\Phi[X]S^\dagger, \end{equation} then it possesses real eigenvalues. \end{Proposition} \begin{proof} Calculate the inverse relation between $p_{kl}$ and $\ell_{kl}$ from formula (\ref{p}), \begin{equation} \ell_{mn}=\sum_{k,l=0}^{d-1}\omega^{-mk+nl}p_{kl}. \end{equation} The eigenvalues $\ell_{mn}$ are real if and only if \begin{equation}\label{real_EV} \ell_{d-m,d-n}={\ell}_{mn}. \end{equation} For the Weyl operators and $S$ from Proposition \ref{prop2}, we have \begin{equation} SW_{mn}S^\dagger=W_{-m,-n}. \end{equation} Hence, we see that $\Phi$ satisfies eq. (\ref{S_cov}) if and only if (\ref{real_EV}) holds. \end{proof} \begin{example}\label{ex} For $d=3$, the Weyl channel $\Phi$ is given by its eigenvalues $\ell_{kl}$. Condition (\ref{real_EV}) implies that \begin{equation} \ell_{01}=\ell_{02},\quad \ell_{10}=\ell_{20},\quad \ell_{11}=\ell_{22},\quad \ell_{12}=\ell_{21}. \end{equation} Observe that the eigenvalues of $\Phi$ are grouped as in (\ref{GPC_group}). Therefore, eq. (\ref{S_cov}) provides the necessary and sufficient conditions for the three-dimensional Weyl channel $\Phi$ to be the generalized Pauli channel. However, for a prime $d>3$, conditions (\ref{S_cov}) are no longer sufficient. \end{example} \begin{Theorem}\label{Th} Suppose that the Weyl channel $\Phi$ satisfies (\ref{S_cov}); that is, $\ell_{kl}$ are real. For every $\beta=1,\dots,\frac{d-1}{2}$, use the irreducible representation $U_\beta$ to construct the Weyl channel via \begin{equation}\label{QCmu} \Phi_\beta := \sum_{k,l=0}^{d-1}\ell_{kl}\Pi_{kl}^{(\beta)}, \end{equation} where $\Pi_{kl}^{(\beta)}$ are the rank-1 projectors onto $W_{\beta k,\beta l}$. Now, $\Phi$ is the generalized Pauli channel if and only if \begin{equation}\label{th} \Phi=\Phi_\beta , \end{equation} for any ${\beta=1,\dots,\frac{d-1}{2}}$. \end{Theorem} \begin{proof} Observe that $\Phi$ and $\Phi_\beta$ have the same spectrum but to different eigenvalues, as eq. (\ref{proj2}) produces \begin{equation}\label{projectors} \Pi_{kl}^{(\beta)}[X]=\frac{1}{d}W_{\beta k,\beta l}\mathrm{Tr}(W_{\beta k,\beta l}^\dagger X)=\Pi_{\beta m,\beta n}[X]. \end{equation} Therefore, if $\Phi$ is the quantum channel, then so are $\Phi_\beta$. Moreover, formula (\ref{th}), together with (\ref{projectors}), imposes the constraints $\ell_{\beta k,\beta l}=\ell_{kl}$, which are consistent with condition (\ref{GPC_group}) for the generalized Pauli channels. Note that the assumptions of Theorem \ref{Th} can be modified. If we drop the requirement that the Weyl channel $\Phi$ has real eigenvalues but check condition (\ref{th}) for $\beta=1,\dots,d-1$, then we arrive at the same results. \end{proof} \section{Class of covariant positive maps} Consider the positive covariant linear map \begin{equation} \Phi[X] = \sum_{\alpha=0}^{d^2-1} q_\alpha W_\alpha X W_\alpha^\dagger, \end{equation} where $W_\alpha$ are the Weyl operators. Recall that a map is positive if for any $X \geq 0$, one has $\Phi[X] \geq 0$ \cite{Paulsen,Bhatia}. Positive maps have received considerable attention recently due to their close relation to the entanglement theory \cite{HHHH}. Recall that a quantum state represented by a density operator $\rho \in \mathcal{B}(\mathcal{H}_1 {\,\otimes\,} \mathcal{H}_2)$ is separable if and only if $({\mathchoice{\rm 1\mskip-4mu l}{\rm 1\mskip-4mu l_1 {\,\otimes\,} \Phi)[\rho] \geq 0$ for all positive maps $\Phi : \mathcal{L}(\mathcal{H}_2) \longrightarrow \mathcal{L}(\mathcal{H}_1)$. Unfortunately, the structure of positive maps is still not fully understood, and the general construction of such maps is not known (see the recent review \cite{KYE,TOPICAL}). In what follows, we provide two constructions of the positive maps that are covariant with respect to the finite group generated by the Weyl operators. Let $\Delta \subset \{0,1,\ldots,d^2-1\}$ with $|\Delta|=N$. Consider the linear map $\Phi : M_d(\mathbb{C}) \longrightarrow M_d(\mathbb{C})$ given by \begin{equation}\label{PI} \Phi[X] = \sum_{\alpha \notin \Delta} \lambda_\alpha^+ F_\alpha X F_\alpha^\dagger + \sum_{\alpha \in \Delta} \lambda_\alpha^- F_\alpha X F_\alpha^\dagger , \end{equation} with $\lambda_\alpha^+ >0$ and $\lambda_\alpha^- < 0$. The operators $F_\alpha$ define an orthonormal basis in $ M_d(\mathbb{C})$, \begin{equation}\label{} \mathrm{Tr}( F_\alpha^\dagger F_\beta) = \delta_{\alpha\beta} \ ;\quad\alpha,\beta=0,1,\ldots,d^2-1\ . \end{equation} Clearly, if there are no negative eigenvalues, the map is completely positive. Now, the problem is to find the `balance' between $\lambda_\alpha^+$ and $\lambda_\alpha^- $ that guarantees the positivity of $\Phi$. Denote the operator norm (the largest singular value) of $A$ by $||A||$. \begin{Theorem}[\cite{CMP}] If $\sum_{\alpha \in \Delta} || F_\alpha||^2 < 1$ and \begin{equation}\label{+-} \lambda_\alpha^+ \geq \frac{ \sum_{\beta \in \Delta} | \lambda_\beta^-| || F_\beta||^2 }{1- \sum_{\beta \in \Delta} || F_\beta||^2 } \ ; \ \ \alpha \notin \Delta \ , \end{equation} then $\Phi$ defined by (\ref{PI}) is positive. \end{Theorem} Clearly, conditions (\ref{+-}) are only sufficient but not necessary. Now, taking $F_\alpha = W_\alpha/\sqrt{d}$, one has $|| F_\alpha||^2 = 1/d$, and hence one arrives at the following corollary. \begin{Corollary} If \begin{equation}\label{} \lambda_\alpha^+ \geq \frac{1}{d-N} \sum_{\beta \in \Delta} | \lambda_\beta^-| \ ; \ \ \alpha \notin \Delta \ , \end{equation} for any $N \leq d-1$ and $F_\alpha = W_\alpha/\sqrt{d}$, then formula (\ref{PI}) defines a covariant positive map. \end{Corollary} The above construction of covariant positive maps allows at most `$d-1$' negative $\lambda_\alpha$. \begin{example} For $d=2$, let us take $\lambda^-_0=-1$. Now, if $\lambda_\alpha^+ \geq 1$, then the map \begin{equation}\label{} \Phi[X] = \frac 12 \left( \sum_{\alpha=1}^3 \lambda_\alpha^+ \sigma_\alpha X \sigma_\alpha - X \right) \end{equation} is positive. In particular, taking $ \lambda_\alpha^+ = 1$ and using the well-known property \begin{equation}\label{} \sum_{\alpha=1}^3 \sigma_\alpha X \sigma_\alpha = 2 \mathbb{I}_2 \, \mathrm{Tr} X - X, \end{equation} one recovers the celebrated reduction map \begin{equation}\label{} \Phi[X] = \mathbb{I}_2 \mathrm{Tr} X - X. \end{equation} \end{example} \begin{example} The above example can be generalized to an arbitrary dimension in the following way: take $\lambda_\alpha^-=-1$ for $\alpha=0,1,\ldots,d-2$, and $\lambda_\alpha^+=d-1$ for $\alpha=d-1,\ldots,d^2-1$. Then, the map \begin{equation}\label{} \Phi[X]= \frac{1}{d(d-1)^2} \left\{ (d-1)\sum_{\alpha=d-1}^{d^2-1} W_\alpha X W_\alpha^\dagger - \sum_{\alpha=0}^{d-2} W_\alpha X W_\alpha^\dagger \right\} \end{equation} is positive and trace-preserving. \end{example} Let us turn our interest to prime dimensions. By $\{ |\psi^{(\alpha)}_1\>,\ldots,|\psi^{(\alpha)}_{d}\> \}$ with $\alpha=1,\ldots,d+1$, denote $d+1$ mutually unbiased bases (MUBs) in $\mathbb{C}^d$; that is, the orthonormal bases for which \begin{equation}\label{} | \< \psi^{(\alpha)}_k| \psi^{(\beta)}_l\>|^2 = \frac 1d \end{equation} when $\alpha \neq \beta$. In \cite{MUBs}, the authors considered the following family of positive, trace-preserving maps, \begin{equation}\label{Phi-MUB} \Phi[X]=\frac{1}{d-1}\left[2\mathbb{I}\mathrm{Tr} X-\sum_{\alpha=1}^{d+1} \sum_{k,l=0}^{d-1} \mathcal{O}_{kl}^{(\alpha)} \mathrm{Tr}(XP_l^{(\alpha)})P_k^{(\alpha)}\right]. \end{equation} In the above formula, $\mathcal{O}^{(\alpha)}$ are rotations in $\mathbb{R}^d$ around the axis determined by $\mathbf{n}=(1,\ldots,1)$; i.e., $\mathcal{O}^{(\alpha)} \mathbf{n}=\mathbf{n}$. \begin{Proposition} The map (\ref{Phi-MUB}) is Weyl-covariant if and only if $\mathcal{O}^{(\alpha)} = \mathbb{I}_d$ for $\alpha=1,\ldots,d+1$. \end{Proposition} \begin{proof} One can construct unitary operators from MUBs in the following way, \begin{equation} U_\alpha=\sum_{l=1}^d\omega^lP_l^{(\alpha)}. \end{equation} Recall that, for prime $d$, $U_\alpha^k$ are the rescalled Weyl operators \cite{DCKS}. {From the properties of the Weyl channels,} it follows that $\Phi$ is Weyl-covariant if and only if $\mathbb{I}$ and $U_\alpha^k$ are its eigenvectors. We check that \begin{equation}\label{ev} \begin{split} \Phi[U_\beta^m] =-\frac{1}{d-1}\sum_{\alpha=1}^{d+1}\sum_{k,l,n=1}^d\mathcal{O}_{kl}^{(\alpha)}\omega^{nm}P_k^{(\alpha)} \left(\delta_{nl}\delta_{\alpha\beta}+\frac 1d (1-\delta_{\alpha\beta})\right) =-\frac{1}{d-1}\sum_{k,l=1}^d\mathcal{O}_{kl}^{(\beta)}\omega^{lm}P_k^{(\beta)}, \end{split} \end{equation} and hence $U_\alpha^k$ are the eigenvectors of $\Phi$ if and only if \begin{equation}\label{O} \mathcal{O}^{(\alpha)}=\mathbb{I}_d. \end{equation} \end{proof} \begin{Remark} If $\mathcal{O}^{(\alpha)} = \mathbb{I}_d$ for $\alpha=1,\ldots,d+1$, then map (\ref{Phi-MUB}) reduces to \begin{equation}\label{red} \Phi[X]= \frac{1}{d-1}\left( 2 \mathbb{I}\, \mathrm{Tr} X - \sum_{\alpha=1}^{d+1} \Phi_\alpha[X] \right), \end{equation} where \begin{equation}\label{} \Phi_\alpha[X] = \sum_{k=0}^{d-1} P_k^{(\alpha)} X P_k^{(\alpha)} . \end{equation} Knowing that \begin{equation}\label{} \sum_{\alpha=1}^{d+1} \Phi_\alpha = {\mathchoice{\rm 1\mskip-4mu l}{\rm 1\mskip-4mu l + d\Phi_0 , \end{equation} with $\Phi_0[X]= \frac{1}{d}\mathbb{I}\mathrm{Tr} X$ being the completely depolarizing channel, we recover the reduction map in $M_d(\mathbb{C})$, \begin{equation}\label{} \Phi[X]= \frac{1}{d-1}\left( \mathbb{I}\, \mathrm{Tr} X - X \right). \end{equation} \end{Remark} Now, let us generalize the map in (\ref{red}) as follows. \begin{Proposition} Take a subset $\Gamma \subset \{1,2,\ldots,d+1\}$ with $|\Gamma|=k$. The map defined by \begin{equation}\label{gen} \Phi_\Gamma := \frac{1}{d-1}\left( 2(k-1) \Phi_0 + \sum_{\alpha \notin \Gamma} \Phi_\alpha - \sum_{\alpha \in \Gamma} \Phi_\alpha \right) \end{equation} is the generalized Pauli trace-preserving positive map. \end{Proposition} \begin{proof} Let $P$ be any rank-1 projector. We show that \begin{equation}\label{} \mathrm{Tr}(\Phi_\Gamma[P])^2 = \frac{1}{d-1}, \end{equation} which implies the positivity of $\Phi_\Gamma$ (cf. \cite{MUBs}). We consider two separate cases: \begin{enumerate} \item $k=1$, \item $k=2,\ldots,d+1$. \end{enumerate} If $k=1$, then $\Gamma=\{ \alpha_*\}$, and $\Phi_\Gamma$ simplifies to \begin{equation}\label{gen-1} \Phi_\Gamma := \frac{1}{d-1}\left( {\sum_{\alpha}}^\prime \Phi_\alpha - \Phi_{\alpha_*} \right) , \end{equation} where ${\sum}^\prime$ denotes the sum over $\alpha \neq \alpha_*$. We need to find \begin{eqnarray} \mathrm{Tr}(\Phi_\Gamma[P])^2 = \frac{1}{(d-1)^2} \left\{ \sum_{\alpha=1}^{d+1} \mathrm{Tr}(\Phi_\alpha[P])^2 +{\sum_{\alpha \neq \beta}}^\prime \mathrm{Tr}(\Phi_\alpha[P] \Phi_\beta[P] ) - 2 {\sum_{\alpha}}^\prime \mathrm{Tr}( \Phi_\alpha[P] \Phi_{\alpha_*}[P]) \right\}. \end{eqnarray} Observe that \begin{equation}\label{} \mathrm{Tr}( \Phi_\alpha[P] \Phi_\alpha[P] ) = \sum_{k=0}^{d-1} \sum_{l=0}^{d-1} \mathrm{Tr}\left( P_k^{(\alpha)} P P_k^{(\alpha)} P_l^{(\alpha)} P P_l^{(\alpha)} \right) = \sum_{k=0}^{d-1} \mathrm{Tr}( P P_k^{(\alpha)} )^2, \end{equation} and hence \begin{equation}\label{a1} \sum_{\alpha=1}^{d+1} \mathrm{Tr}(\Phi_\alpha[P])^2 = \sum_{\alpha=1}^{d+1} \sum_{k=0}^{d-1} \mathrm{Tr}( P P_k^{(\alpha)} )^2 = 2 , \end{equation} where we have used the well-known property of the MUBs \cite{Wu,Beatrix}. Moreover, using the fact that all the maps $\Phi_\alpha$ are self-dual and $\Phi_\alpha \Phi_\beta = \Phi_0$ for $\alpha \neq \beta$, one finds \begin{equation}\label{a2} \mathrm{Tr}(\Phi_\alpha[P] \Phi_\beta[P] ) = \mathrm{Tr}(P \Phi_\alpha[\Phi_\beta[P]] ) = \mathrm{Tr}(P \Phi_0[P] ) = \frac 1d . \end{equation} Finally, we arrive at \begin{eqnarray} \mathrm{Tr}(\Phi_\Gamma[P])^2 = \frac{1}{(d-1)^2} \left[ 2 + (d-1) - 2 \right] = \frac{1}{d-1} . \end{eqnarray} For $k>1$, {we arrive at} \begin{eqnarray} \mathrm{Tr}(\Phi_\Gamma[P])^2 &=& \frac{1}{(d-1)^2} \mathrm{Tr} \left\{ 4(k-1)^2 (\Phi_0[P])^2 + 4(k-1) \Phi_0[P] \left( \sum_{\alpha \notin \Gamma} \Phi_\alpha[X] - \sum_{\alpha \in \Gamma} \Phi_\alpha[X] \right) \right. \\ &+& \left. \sum_{\alpha \notin \Gamma} \Phi_\alpha[X] \sum_{\beta \notin \Gamma} \Phi_\beta[X] + \sum_{\alpha \in \Gamma} \Phi_\alpha[X] \sum_{\beta \in \Gamma} \Phi_\beta[X] - \sum_{\alpha \notin \Gamma} \Phi_\alpha[X] \sum_{\beta \in \Gamma} \Phi_\beta[X] - \sum_{\beta \in \Gamma} \Phi_\beta[X] \sum_{\alpha \notin \Gamma} \Phi_\alpha[X] \right\} , \nonumber \end{eqnarray} which simplifies to \begin{equation} \begin{split} \mathrm{Tr}(\Phi_\Gamma[P])^2\ =&\ \ \frac{1}{(d-1)^2} \left\{ \frac{4(k-1)^2}{d} + \frac{4(k-1)}{d} \left( d+1-2k \right) + \sum_{\alpha=1}^{d+1} \mathrm{Tr}(\Phi_\alpha[P])^2 \right. \\ +& \left. \sum_{\alpha\neq \beta \notin \Gamma} \mathrm{Tr}(\Phi_\alpha[P] \Phi_\beta[P] )+ \sum_{\alpha \neq \beta \in \Gamma} \mathrm{Tr}(\Phi_\alpha[P] \Phi_\beta[P] ) - 2\sum_{\alpha \notin \Gamma} \sum_{\beta \in \Gamma} \mathrm{Tr}(\Phi_\alpha[P] \Phi_\beta[P] ) \right\} , \end{split} \end{equation} Finally, using eqs. (\ref{a1}-\ref{a2}), one gets \begin{equation} \begin{split} \mathrm{Tr}(\Phi_\Gamma[P])^2\ &=\ \ \frac{1}{d(d-1)^2} \left\{ {4(k-1)^2} + {4(k-1)} \left( d+1-2k \right) + 2d +2 \left( \begin{array}{c} d+1-k \\ 2 \end{array} \right) \right. \\ &+ \left. 2 \left( \begin{array}{c} k \\ 2 \end{array} \right) - 2 k (d+1-k) \right\} = \frac{1}{d-1} , \end{split} \end{equation} which ends the proof. \end{proof} \section{Conclusions} { We analyzed a class of irreducibly covariant quantum channels with respect to the unitary representations of the finite group $G$ generated by the Weyl operators. Interestingly, all the $d$-dimensional representations of $G$ are multiplicity-free, and hence one can apply to them the analysis developed recently by Mozrzymas et. al. \cite{MSD}. The channel that is irreducibly covariant with respect to the $d$-dimensional representations of $G$ turns out to be the Weyl channel. For prime dimensions, one can consider the subclass of the Weyl channels known as the generalized Pauli channels \cite{Ruskai,DCKS,DCKS2}. Finally, we analyzed the class of irreducibly covariant maps which are positive but not necessarily completely positive. Such maps provide an important tool in the entanglement theory. We presented two classes of positive maps: one belonging to the general family of the Weyl maps and the other to the reduced family of the generalized Pauli maps. It would be interesting to further study the properties of positive maps -- for example, decomposability, optimality, or extremality. Also, the generalized Pauli maps were defined only for prime dimensions. One could consider a more general case with $d=p^r$ and prime $p$, where the maximal set of `$d+1$' MUBs can be found \cite{Wootters}.} \section*{Acknowledgements} This paper was partially supported by the National Science Centre project 2015/19/B/ST1/03095.

\section{Introduction} Many shape analysis tasks describe shapes as smooth manifolds and analyze them with respect to their geometry in terms of normals, curvatures, and geodesics that require accurate estimates of fist- and second-order derivatives over the shape to compute tangent spaces, the Riemannian metrics, or Laplace-Beltrami operators~\cite{reuter06,Levy2006,Windheuser2014,Litman2014}. When the surface is represented as a mesh or point cloud, these differential operators can only be computed approximately at large memory and computational costs. Once these differential operators have been used to compute normals or metrics, most of the high resolution information is discarded by operating only with the leading eigenfunctions of the Laplace-Beltrami operator necessary to estimate geodesics~\cite{Crane2013}, compute shape descriptors~\cite{aubry-et-al-11,reuter06,Levy2006,Litman2014} or shape matches~\cite{Ovsjanikov2012} with tractable problem sizes. This creates a paradox, as shapes are first discretized with fine meshes or large pointclouds to estimate high-dimensional shape operators that are mostly discarded to analyze the shape. This paradox appears because we only need large meshes or dense pointclouds to discretize accurate differential operators over the shape, not for shape analysis. We propose a paradigm shift that avoids this paradox by representing a shape with a subdivision surface~\cite{Notes2000} that parametrizes the surface with a small set of smooth base functions. Differentiability is then intrinsic to the shape representation and does not require large bases to compute principal curvatures, wave-kernel signatures, or Laplace-Beltrami eigenfunctions. Subdivision surfaces are a generalization of splines to surfaces of arbitrary topology that provide set of base functions to compactly parametrize the surface as well as smooth functions over it. As a result, subdivision surfaces provide a compressed shape representation well suited to shape analysis. Our first contribution (Section~\ref{}) is a method to estimate a subdivision surface from the triangular meshes common in shape analysis\cite{bronstein08}. This kind of manifold meshes, however, are far from the meshes or pointclouds obtained from range measurements or computer-vision models. As these low-level representations proliferate, they create a challenge for shape analysis to adapt to the characteristics of raw data (irregular surface sampling, noise, and outliers) that cause havoc in the estimation of differential operators. Up until now, raw measurements have been largely ignored in shape analysis by estimating a mesh representation from a pointcloud and performing analysis on this representation. This avoids dealing with noise and outliers but ignores the direct measurements and makes shape analysis dependent on the model used to estimate the surface.\footnote{Least-squares models \cite{Kazhdan2013,Calakli2011}, for instance, oversmooth surface edges and require less eigenfunctions to match than the $\ell_1$ models~\cite{Ummenhofer2015,Estellers2016} that emphasize corners.} For this reasons, our second contribution is a robust method that fits a subdivision surface to a noisy pointcloud and estimates the geometric operators required to do shape analysis on it. Our model is robust to both noise and outliers in the pointcloud and to erroneous correspondences between the input points and the subdivision surface. This is a critical aspect of fitting a parametric surface to raw data without developing the non-manifold artifacts that prevent shape analysis Our third contribution is a fast optimization algorithm (Section~\ref{}) that combines a sequential quadratic approximation with a good initialization strategy. This is important because fitting a subdivision surface is not a convex problem and vanilla initializations~\cite{Jaimez2017} oversmooth the surface to avoid small-scale local minima and prevent self-intersections. Our fourth contribution (Section~\ref{}) describes how to accurately compute many shape operators with a subdivision surface. We focus on Laplace-Beltrami operator, wave kernel signatures~\cite{aubry-et-al-11}, approximate geodesics \cite{Crane2013} and prove its efficacy in the full-pipeline task of shape matching~\cite{Ovsjanikov2012}. Our experiments show state-of-the-art performance at a fraction of the memory requirements of shape representations with triangular meshes. \section{Smooth Shapes}\label{} In this paper, a \emph{3D shape} is a smooth\footnote{ We call a $C^1$ object smooth if it is $C^2$ except for a finite subset. This differs from the mathematical notion, which identifies smoothness with $C^\infty$.}, compact oriented surface $S\subset{\mathbb{R}}^3$ without boundaries. For computational purposes, shapes are usually discretized by a mesh ${\mathbf{M}}=(V,F)$ in terms of two matrices: the $i$-th row of $V\in{\mathbb{R}}^{n\times3}$ stores the location of the $i$-th vertex of the mesh, while each row of $F\in{\mathbb{N}}^{m\times k}$ stores the indices of the vertices of a face of ${\mathbf{M}}$. We focus on quad meshes with $k=4$. In shape analysis, quad meshes usually discretize a smooth surfaces as a piecewise $C^1$ surface that is not smooth. Subdivision surfaces, on the other hand, only use meshes to discretize the domain of a smooth surface parametrization. To formally introduce this parametrization, we first revisit the topological space associated to a quad mesh and show how it can be used to define a smooth subdivision parametrization. \subsection{Meshes -- Piecewise Bilinear Surfaces}\label{} The regular quad mesh parametrization of a surface $S_{\mathbf{M}}$ consists of quadrilaterals ``glued'' together along their edges. Each quadrilateral in $S_{\mathbf{M}}$ is modeled by a bilinear patch that fits the fours vertices of a facet in ${\mathbf{M}}$ and is glued to its neighboring facets. This topological gluing extends the independent parametrization of each quadrilateral to the whole mesh and is formalized by the topological space, ${\mathcal{T}}_{\mathbf{M}}$, that defines the domain of the parametrization~\cite{Juettler-et-al-16}. Formally, given a quad mesh ${\mathbf{M}}=(V,F)$, the topological space ${\mathcal{T}}_{\mathbf{M}}$ is defined by a set of quads that is indexed by the mesh facets ${\mathcal{I}}_F=\{1,\ldots,m\}$ and glued together by the equivalence relation $\sim$, that is ${\mathcal{T}}_{\mathbf{M}}:=\square\times{\mathcal{I}}_F/\sim$. The \emph{reference quad} $\square=[0,1]\times[0,1]$ parametrizes each quadrilateral patch while the equivalence relation $\sim$ connects neighboring faces along their common edge. A parametrization of the surface is then obtained by defining a piecewise bilinear function $$B_{\mathbf{M}}^i\colon{\mathcal{T}}_{\mathbf{M}}\to{\mathbb{R}}$$ for each vertex in $x_i \in {\mathcal{T}}_{\mathbf{M}}$ satisfying $B_{\mathbf{M}}^i(x_j)=\delta_{ij}$. The resulting parametrization $$\Phi_{\mathbf{M}}\colon{\mathcal{T}}_{\mathbf{M}}\to S_{\mathbf{M}}\subset{\mathbb{R}}^3$$ defines the piecewise $C^1$ surface $S_{\mathbf{M}}:=\Phi_{\mathbf{M}}({\mathcal{T}}_{\mathbf{M}})$ by bilinear interpolation $$\Phi(u) = \sum_{i=1}^n B_{\mathbf{M}}^i(u)\cdot v_i.$$ With such a piecewise $C^1$ paremetrization, Riemannian quantities like the first or second fundamental form can only be approximated and need large meshes with small quadrilaterals for accurate approximations. This limits their use in real-time applications and calls for other shape representations.% \subsection{Smooth Subdivision Surfaces}\label{} A subdivision surface represents a smooth surface $S$ by a coarse control mesh ${\mathbf{M}}^0=(V^0,F^0)$ and a subdivision scheme. The subdivision scheme transforms a mesh ${\mathbf{M}}^k=(V^k,F^k)$ into a finer mesh ${\mathbf{M}}^{k+1}$ such that if we start with ${\mathbf{M}}^0:=(V,F)$, we obtain the smooth surface as the limit of iterating subdivision ad infinitum:$$S = \lim_{k\to\infty} S_{{\mathbf{M}}^k}.$$ In this paper we follow the Catmull-Clark subdivision scheme ${\mathbf{M}}^k\mapsto {\mathbf{M}}^{k+1}$ consisting of two steps: First, the topology of the new mesh ${\mathbf{M}}^{k+1}$ is defined by subdividing each quad in ${\mathcal{T}}_{\mathbf{M}}^{k}$ into its four quadrants that define four adjacent facets in ${\mathcal{T}}_{\mathbf{M}}^{k+1}$. Second, the location of each vertex of the new mesh ${\mathbf{M}}^{k+1}$ is computed as a linear combination of the positions of the nearby vertices. Figure \ref{} illustrates this subdivision process. This produces a compressed surface representation because the topology of ${\mathbf{M}}^{k+1}$ is completely determined by the topology of ${\mathbf{M}}^{k}$. Similarly, the space of bilinear functions over ${\mathcal{T}}_{{\mathbf{M}}^{k+1}}$ created by subdivision is determined by the space of bilinear functions over ${\mathcal{T}}_{{\mathbf{M}}^{k}}$. This means that we can parametrize the $i$-th function created by subdivision $B_{{\mathbf{M}}^{k+1}}^i$ over the domain ${\mathcal{T}}_{\mathbf{M}}^{k}$, and by induction over ${\mathcal{T}}_{\mathbf{M}}$, instead than ${\mathcal{T}}_{\mathbf{M}}^{k+1}$. We denote these functions $B_{\mathbf{M}}^{k,i}~\colon~{\mathcal{T}}_{\mathbf{M}}~\to~{\mathbb{R}}$ and use them to parametrize the surface $S_{{\mathbf{M}}^{k}}=\Phi_{\mathbf{M}}^k({\mathcal{T}}_{\mathbf{M}})$ as a function of the vertices of ${\mathbf{M}}^{k}$ In the second step, the position of the vertices of ${\mathbf{M}}^{k+1}$ are computed as linear combination of the positions of the vertices of ${\mathbf{M}}^{k}$. The weights of this linear combination are organized into a subdivision matrix $\hat{A}_k\in{\mathbb{R}}^{n_k\times~n_{k-1}}$ that computes the position of the vertices of ${\mathbf{M}}^k$ as $$V_k=\underbrace{\hat{A}_k\ldots\hat{A}_1}_{A_k \in \mathbb{R}^{n_k \times n}}~V=A_k~V.$$ Using this linear dependency, we can parametrize a surface approximation $S_{{\mathbf{M}}^k}$ resulting from $k$ subdivisions directly over ${\mathcal{T}}_{\mathbf{M}}$ as follows: \begin{align*} \Phi_{\mathbf{M}}^k(u) = \sum_{i=1}^{n_k}B_{\mathbf{M}}^{k,i}(u) \left(A_{k} V\right)_i = % \sum_{j=1}^{n}\sum_{i=1}^{n_k}B_{\mathbf{M}}^{k,i}(u) A_{k,ij} v_j \end{align*} By iterating this process, the subdivision surface of the limit $$S = \lim_{k\to\infty} S_{{\mathbf{M}}^k}$$ can be formulated as a limit of functions defined over the topological space ${\mathcal{T}}_{\mathbf{M}}$. Evaluating the limit surface at the location of the vertex thus amounts to computing the base functions $$ \Phi_j:=\lim_{k\to\infty}\sum_{i=1}^{n_k}B_{\mathbf{M}}^{k,i}\cdot A_{k,ij}.$$ The existence of this limit and the smoothness of the limit functions is guaranteed by the subdivision scheme. All the \emph{limit base functions} $\Phi_j$ are compactly supported because the subdivision rules act locally and converge to spline basis functions. This equivalence provides analytic derivatives and fast evaluation techniques \cite{Stam1998} for smooth functions or tangent vectors defined over the surface and makes subdivision surfaces well-suited for shape analysis. We focuses on Catmull-Clark subdivision~\cite{catmull-clark-78} for simplicity. This scheme ensures the existence and $C^2$ smoothness of the basis functions everywhere except at extraordinary vertices (vertices with valence different than four). This scheme is designed for control grids with quadrilateral connectivity, but our compression generalizes to any subdivision schemes that parametrizes $C^2$ surfaces. \begin{figure*} \centering \begin{tabular}{cccc} \includegraphics[trim=0cm 0cm 0cm 0cm, clip=true, width = 0.18\textwidth]{fig3-horse-sds-1k718ref0}& \includegraphics[trim=0cm 0cm 0cm 0cm, clip=true, width = 0.18\textwidth]{fig3-horse-sds-1k718ref1}& \includegraphics[trim=0cm 0cm 0cm 0cm, clip=true, width = 0.18\textwidth]{fig3-horse-sds-1k718ref2}& \includegraphics[trim=0cm 0cm 0cm 0cm, clip=true, width = 0.18\textwidth]{fig3-horse-sds-1k718ref3} \\ ${\mathbf{M}}^0$ & ${\mathbf{M}}^1$ & ${\mathbf{M}}^2$ & ${\mathbf{M}}^3$ \\ 1701 Control Points & % 6798 Vertices 27186 Vertices 108738 Vertices \end{tabular} \caption{Different subdivision levels ${\mathbf{M}}^k$ of Catmull-Clark subdivision scheme applied to a control mesh and a smooth function defined over the mesh (its 18-th $\Delta$ eigenfunction). ${\mathbf{M}}^3$ is already a good approximation of the smooth surface $S={\mathbf{M}}^\infty$.} \label{} \end{figure*} \section{Robust Fitting Model}\label{} We formulate the problem of fitting a subdivision surface $S$ to a set of surface samples ${\mathbf{P}}= \{p_1,\ldots, p_N \}$ and a set of normals $\mathbf{T}=\{t_1,\ldots,t_N\}$ as the minimization problem: \begin{align}\label{} \min_{S} \ E(S) :=& \underbrace{\dist(S,{\mathbf{P}})}_\text{Point Fit} + % \underbrace{\alpha\cdot{\mathcal{T}}(S, \mathbf{T})}_\text{Tangent Fit}+% \underbrace{\beta\cdot R(S)}_\text{Regularization}, \end{align} where $\alpha,\beta\geq0$ are model parameters. % The different terms take different information into account: $\dist(S,{\mathbf{P}})$ fits the surface $S$ to the points ${\mathbf{P}}$, ${\mathcal{T}}(S,\mathbf{T})$ matches the tangent plane of $S$ to the normal information $\mathbf{T}$, and $R(S)$ regularizes the discrete representation of $S$ to avoid degenerate faces in ${\mathbf{M}}$. If no normal information is provided, $\alpha = 0$. % \subsubsection*{Point Fit} Let ${\mathcal{D}}_{S}(p)=\min_{s\in S}\norm{s-p}$ denote the Euclidean distance between a point $p\in{\mathbb{R}}^3$ and the compact surface $S$, we define the \emph{point fit energy} as \begin{align} \dist(S,{\mathbf{P}}) := \sum_{j=1}^N{\mathcal{D}}_{S}(p_j)^q = \sum_{j=1}^N\min_{u_j\in{\mathcal{T}}_{\mathbf{M}}}\norm{\Phi(u_j)-p_j}^q \nonumber. \end{align} Using the $q$-th power of the distance with $q\in(1,2)$, instead of the common squared distance, makes our model robust to outliers and scanning artifacts in the input samples. The minimization in $u_1,\ldots,u_N$ results from the parametric representation of the surface $S=\Phi({\mathcal{T}}_{\mathbf{M}})$ and introduces a large number of additional variables over which the objective function is not convex. We avoid the optimization of the correspondence parameters $u_1,\ldots,u_N$ by using a second-order approximation of the squared distance function to a surface in Section~\ref{}. \subsubsection*{Tangent Fit} To obtain a surface that also matches the normals, we include the \emph{tangent energy} \begin{align} {\mathcal{T}}(S, \mathbf{T}) = \sum_{j=1}^{N} % \abs{\vecprod{t_j}{\partial_1 \Phi(u_j)}}^q + % \abs{\vecprod{t_j}{\partial_2 \Phi(u_j)}}^q, \end{align} % where $t_j$ is the input surface normal at point $p_j$ and $\partial_i\Phi(u_j)$ the $i$-th base vector of the tangent space $T_{\Phi(u_j)}S$. Each term in ${\mathcal{T}}$ aligns the tangent space of the subdivision surface in the direction orthogonal to the sample normal. This penalty is designed to be independent of the orientation of the normals to account for noisy point clouds where estimating consistent normal orientations is prone to fail. The $\ell_q$-norm makes this term robust to noise and outliers in the normals and the correspondence parameters $u_1,\ldots, u_N$. \subsubsection*{Regularization} The regularizer $R(S)$ penalizes the squared distance between the vertices of the mesh $M$ incident to the same quad. This keeps the size and shape of the quads regular and avoids skewed elements that cause instabilities in finite-element computations. The regularizer is thus a simple quadratic penalty that can be described by a sparse matrix $R$ that has a row for each edge in the mesh with entries $\pm 1$ at the columns of the vertices incident to the edge. We choose this regularizer for its simplicity. \subsection{Initialization and Surface Topology}\label{} A major challenge of fitting any explicit parametrization to a set of points is the lack of convex formulations for the optimization problem. Non-convexity introduces the difficulty of finding a suitable initialization that leads to a \textit{good} local minimum. It is an important step of the algorithm that conditions the topology of the surface. By parameterizing $S$ with a subdivision surface, our optimization problem is formulated in terms of the control mesh ${\mathbf{M}}=(V,F)$. To avoid solving a combinatorial problem, we fix the topology of the control mesh, and thus the topology of the surface, and only optimize for the location of the vertices $V$. We design the control mesh to satisfy two properties: it determines a surface with the same topology as the input data, and it has a small number of vertices to represent the surface compactly. Depending on the input data, we apply different methods to obtain a good initialization. If the point samples ${\mathbf{P}}$ are the vertices of a mesh $\overline{\mathbf{M}}$, the topology of the surface is already encoded by the mesh. Without loss of generality, we assume that $\overline{\mathbf{M}}$ is a triangular mesh without boundaries and use quadratic edge-collapse~\cite{Garland1997} to reduce the number of vertices in the mesh and preserve the surface topology. We then transform this triangular mesh into a quad mesh by solving a perfect matching problem that pairs triangles of the collapsed mesh to create quads~\cite{Remacle2012a}. For each edge $e$, let $\alpha_1(e),\ldots,\alpha_4(e)\in{\mathbb{S}}^1$ be the angles of the quad that results from removing $e$ and $\eta(e)\in{\mathbb{S}}^1$ the angle between the normals of the triangles incident to $e$. We set the cost of removing an edge $e$ to \begin{align*} \small c(e) = \begin{cases} \frac14\sum_{i=1}^4 d\left(\alpha_i(e),\frac\pi2\right)^2 + \tan(\eta(e))^2 & \text{if $\abs{\eta(e)}<\frac\pi2$}\\ \infty & \text{if $\abs{\eta(e)}\geq\frac\pi2$}. \end{cases} \end{align*} to assign infinity costs to edges that connect strongly bent triangles ($\eta(e)>\frac\pi2$) or edges that would create quads with straight or reflex angles. The cost favors rectangular quads to obtain a subdivision surface where finite-element computations are accurate and numerically stable. We solve the perfect matching and find the quad mesh with minimum cost with the \emph{Blossom V} method~\cite{Kolmogorov-MPC09}. When the input is a point cloud, we use an implicit representation to estimate the topology of the surface from the point samples and only then extract a control mesh. Estimating first the topology of the surface with an implicit representation avoids explicitly handling topology changes during the compression process. We use the popular Poisson reconstruction~\cite{Kazhdan2013} for simplicity but other techniques like voxel hashing work well. From this implicit representation, we extract a closed triangular mesh with marching cubes~\cite{Lorensen1987} and create a compact quad mesh with the collapse and blossom techniques described for meshes. \section{Efficient Optimization}\label{} Our optimization algorithm exploits the properties of subdivision surfaces, namely, the compact support of the basis functions and the ability to analytically evaluate its geometry, with a sequential quadratic program that is more efficient than the gradient-based algorithms proposed by~\cite{Cashman2013,Ili2006,Jaimez2017}. We solve \eqref{} as a sequence of convex problems \begin{align} v^{m+1} \leftarrow \min_v \, v^\top\left(Q^m v - b\right) \end{align} % that approximate the original energy around the current surface estimate by a least-squares problem. The approximation is derived from the surface geometry, while the sparsity of $Q^m$ results from compactness of the subdivision basis. We follow a \emph{Majorize-Minimize} (MM) principle to minimize $E(S)$ by iterating two steps until convergence. The first step finds an upper envelop of the objective function $E(S|S_0)\geq E(S)$ that coincides with $E$ at $S_0$. We derive an upper envelop of the data and tangent fits from the inequality \begin{align} \abs{d}^q \leq \frac{q}{2} \abs{d_0}^{q-2}d^2 + (1-\frac{q}{2}) \abs{d_0}^q && \forall d_0 \neq 0, ~q\in[1,2]\nonumber. \end{align}% The second step of the MM algorithm minimizes the upper envelop the upper envelop and drives the value of the original function downwards by solving $$S^{m+1}\leftarrow\min_{S}E(S|S^m).$$ In our case, the envelop \begin{footnotesize} \begin{align} E(S|S^m) = \sum_{j=1}^{n} w^m_j {\mathcal{D}}_{S}(p_j)^2 + \sum_{i=1}^2 \alpha^m_{ij}\vecprod{t_j}{\partial_i\Phi^m(u_j)}^2 + \beta R(S) \nonumber \end{align} \end{footnotesize} defines a weighted least-squares problem where the weights $w^m_j$ and $\alpha^m_{ij}$ are determined by the MM principle \begin{align} w^m_j &= \frac{q}2{\mathcal{D}}_{S^m}(p_j)^{q-2} & % \alpha^m_{i j} &= \frac{\alpha q}{2} % \abs{\vecprod{t_j}{\partial_i\Phi^m(u_j)}}_{\epsilon}^{q-2}. \nonumber \end{align} % To simplify the minimization step of MM, we approximate ${\mathcal{D}}_{S^m}(\cdot)^2$ with a quadratic function that parametrizes the squared distance to a surface in terms of its geometry as follows: \begin{align} {\mathcal{D}}^2_{S}(x) \approx (x-p)^\top\left[ \frac{d\cdot \tau_1\tau_1^\top}{d+\rho_1} + \frac{d\cdot \tau_2\tau_2^\top}{d+\rho_2} + \nu\nu^\top \right](x-p), \nonumber \end{align} where $p$ is a point on the surface $S$ close to $x$ (but not necessarily the projection of $x$ onto $S$), $d:=\norm{x-p}$ is the signed distance from $x$ to this point, and $\rho_i, \tau_i, \nu$ are the principal curvature radii, the principal curvature directions and the normal to the surface at $p$. This approximation is due to \cite{Pottmann2003a} and coincides with the second order Taylor approximation within the radius of curvature of the surface at $p$, where it can be relaxed into a positive definite quadratic form by taking the absolute values of $d,~\rho_1,~\rho_2$. Incorporating this approximation into the upper envelop corresponds to using a quasi-Newton algorithm for optimizing jointly the control vertices of the subdivision surface and the parametric coordinates of the surface point closest to $x$~\cite{Wang2006,Liu2008}. The approximation let us thus work with approximate correspondences between surface samples and input points, instead of explicitly finding the correspondences by solving a larger optimization problem. It also makes our model robust to erroneous correspondences by taking into account their approximate nature in the objective function. The approximation, however, can affect the majorizing property and puts our algorithm into the category of sequential quadratic programs instead of MM algorithms. The combination of the MM upper envelope with the quadratic approximation of the squared distance to $S$ reduces to weighted least-squares problems in the vertices $V$ (c.f. Appendix) that we solve efficiently with conjugate gradient solver warm-started with the solution of the previous MM minimization step. After optimizing $V$ at iteration $m$, we update the parameter values $U$ where we approximate the squared distance to the surface by sampling the surface $S^m=\Phi^m({\mathcal{T}}_{\mathbf{M}})$ uniformly and creating a kd-tree from these samples to find the parameter of the sample closest to each input point This results in an efficient algorithm that avoids the slow convergence rates of coordinate descent algorithms and the large optimization problems of the Levenberg-Marquard solvers advocated by \cite{Cashman2013,Ili2006,Jaimez2017}. \section{Smooth Shape Analysis}\label{} We have showed how to obtain smooth shapes from non-smooth input data like point clouds or meshes. We now show how to do \emph{smooth shape analysis} with a subdivision surface $S$. To this purpose, we focus on the computation of the Laplace-Beltrami operator and its eigenfunctions and use them to match shapes, compute shape descriptors, and approximate geodesics \subsection{Laplace-Beltrami-Operator}\label{} Given a smooth shape $S$, the Laplace-Beltrami-Operator $\Delta$ maps a twice continuously differentiable function $g\in~C^2(S)$ to a continuous function $$\Delta(g)\colon S\to{\mathbb{R}},$$ If $g$ satisfies Neumann boundary conditions, then Stokes' theorem $$\int_S~\Delta~g~\cdot~h~\,\mathrm{dx}~=~\int_S~\vecprod{\nabla~g}{\nabla~h}~\,\mathrm{dx}$$ defines a weak formulation of $\Delta$ for Sobolev functions $g\in~H^1(S)$. The Galerkin method uses this identity to define a discretization of $\Delta$ in the ansatz space $H^1_n(S)$ spanned by Sobolev functions $\Phi_1,\ldots,\Phi_n\in H^1(S)$. Using $H^1_n(S)$ as test space, $g$, $h$ and $\Delta g$ are represented by vectors $\alpha,\beta,\gamma\in{\mathbb{R}}^n$ as follows \begin{align*} g =& \sum_{i=1}^n\alpha_i\Phi_i& % h =& \sum_{i=1}^n\beta_i\Phi_i& % \Delta g =& \sum_{i=1}^n\gamma_i\Phi_i&% \end{align*} and the weak formulation of $\Delta$ in $H^1_n(S)$ reads \begin{align} \vecprod{\gamma}{\beta}_{D_0} = \vecprod{\alpha}{\beta}_{D_1} ~\forall\beta \in {\mathbb{R}}^n.\label{} \end{align} The Laplace-Beltrami operator in $H^1_n(S)$ thus depends on the mass and stiffness matrices $D_0,D_1\in{\mathbb{R}}^{n\times n}$ whose entries \begin{align*}\left(D_0\right)_{ij} = \int_S \Phi_i\cdot\Phi_j\,\mathrm{dx} ~~ \left(D_1\right)_{ij} = \int_S \vecprod{\nabla\Phi_i}{\nabla\Phi_j}\,\mathrm{dx}. \end{align*} compute the scalar product of functions and tangent vectors in the surface $S$. From \eqref{}, the Laplace-Beltrami operator in $H^1_n(S)$ can be parametrized by the matrix $D_2:=D_0^{-1}D_1$. This matrix has a non-negative spectrum and its eigenvector discretizes the eigenfunctions of the operator $\Delta$ in $H^1_n(S)$. Indeed, if $v\in{\mathbb{R}}^n$ is an eigenvector of $D_2$ with eigenvalue $\lambda$, then $g_v=\sum_{i=1}^n v_i\Phi_i$ is an eigenfunction of $\Delta$ with eigenvalue $\lambda$ in $H^1_n(S)$. \begin{figure*} \centering \begin{tabular}{ccc c} \includegraphics[width = 0.24\textwidth]{fig3-cat0-coarse12}& \includegraphics[width = 0.24\textwidth]{fig3-cat0-fine12}& \includegraphics[width = 0.24\textwidth]{fig3-cat0-sds-3ks12ref3}& \includegraphics[width = 0.2\textwidth]{wksall}\\ Mesh 3,400 vertices & Mesh 27,984 vertices & Subdiv. Surf. 1700 vertices & Wave-kernel signatures \end{tabular} \caption{$12^\text{th}$ $\Delta$-eigenfunction and wave-kernel signature of the a surface point in red computed with fine (blue line) and coarse triangular meshes (red line) and a subdivision surface (yellow). The subdivision surface obtains a wake signature comparable to the fine mesh while the signature of the coarse mesh lacks detail.} \label{} \end{figure*} To particularize this construction to functions defined on a subdivision surface $S$, we consider the mapping $$X(u,f) = \sum^{n}_{i=1} \Phi_i(u,f) v_i$$ that parametrizes the surface as a linear combination of $n$ limit basis functions $\Phi_1,\ldots,\Phi_n$ and $n$ mesh vertices $v_1,\ldots,v_n$. Similarly, we can parametrize Sobolev funcctions over the surface $g\in H^1(S)$ by $n$ scalar values $g_1,\ldots,g_n\in{\mathbb{R}}$ with a mapping \begin{small} \begin{align} g\colon S\to&{\mathbb{R}}\\ \sum_{i=1}^{n}\Phi_i(u,f)\cdot v_i \mapsto& % \sum_{i=1}^{n}\Phi_i(u,f)\cdot g_i \end{align} \end{small} that depends on each basis function $\Phi_i$ twice, once to describe the function $g$ and once to describe the function domain $S$. As a result, the gradient of a surface function is \begin{small} \begin{align} \nabla g(X(u,f))=JX(u,f) (G^{-1}(u,f) \sum_{i=1}^{n}\nabla\Phi_i(u,f)\cdot g_i)\nonumber, \end{align} \end{small} where $JX$ is the Jacobian of the surface mapping $X$ with respect to $u$. When we use this parametrization to compute the mass and stiffness matrix, we need to evaluate the first fundamental form of $S$ at a point $x\in S$ explicitly. Let $f\in{\mathcal{I}}_F$ be an arbitrary face index, then the first fundamental form at surface point $X(u,f)$ is \begin{align} G(u,f)=JX(u,f)^\top JX(u,f). \end{align} We compute the surface integrals in the mass and stiffness matrices by pulling-back the integrand to the reference quad as follows: \begin{align*} \int_S g(s)\,\mathrm{ds} = \sum_{f\in{\mathcal{I}}_F}\int_\square g\circ X(u,f)\sqrt{\det G(u,f)}\,\mathrm{du} \end{align*} We can compute the integral over the quad with arbitrary accuracy with high-order quadrature approximations. Experimentally, we obtain reliable mass and stiffness matrices with simple $3\times 3$ Gaussian quadrature. Note that when the subdivision surface $S$ is derived from a control mesh with $n$ vertices, $D_0, D_1, D_2\in{\mathbb{R}}^{n\times n}$. An eigenfunction $g_v$ of $\Delta$ is then obtained by solving an $n\times n$ generalized eigen problem and by parameterizing $g_v=\sum_{i=1}^n v_i\Phi_i$. In our figures, we use the refined ${\mathbf{M}}^k$ mesh and the refined eigenfunction $g_v^k$ for visualization. That is, $g_v^k$ is not derived from the huge mass and stiffness matrices of ${\mathbf{M}}^k$ but from the projection of $g_v\in C^1(S)$ onto the mesh ${\mathbf{M}}^k$. \subsection{Geodesic Distances} The classic approach to computing the geodesic distance $\chi$ to a surface point $S(u_0)$ is to solve the eikonal equation \begin{align} \Vert \nabla \chi \Vert = 1 \label{} \end{align} subject to boundary conditions $\chi(u_0) = 0$. This results in a nonlinear hyperbolic PDE that is computationally expensive to solve. To alleviate the computational burden, \cite{Crane2013} proposes a method to approximately solve \eqref{} in terms of the Heat equation. This method first computes a function $f\in H^1(S)$ whose gradient is parallel to $\nabla \chi$ by integrating the Heat flow $\dot{f} = \Delta f$. It then approximates $\nabla \chi$ by the vector field $X = -\frac{\nabla f}{\Vert \nabla f \Vert}$ by observing that $\nabla \chi$ is parallel to $\nabla f$ and has unit norm as a result of \eqref{}. The method then finds an approximate $\chi$ by minimizing $\int_{S} \Vert \nabla \chi - X \Vert^2$ through its Euler-Lagrange equation $ \Delta \chi = \nabla\cdot X$. As all the computations are formulated in terms of the Heat kernel $\Delta$, the method can be directly use to compute approximate geodesics in our subdivision surface framework. We refer to \cite{Crane2013} for the details of the method. \section{Related Work} \label{} Subdivision surfaces are a standard representation of shapes in computer graphics and animation, but they have largely been neglected in vision due to the challenges of working with noisy measurements that favor convex formulations and implicit representations. Computer graphics techniques fit a subdivision surface by minimize the sum of squared distances between the subdivision surface and the vertices of a clean and regular mesh~\cite{Hoppe1994,Ma2002}. Research has focused into efficient ways of minimizing this non-convex energy by gradient-based techniques \cite{Zheng2012}, quasi-Newton methods \cite{Marinov2005,Wang2006,Liu2008}, and sequential convex programing \cite{Marinov2005,Lavoue2005}. A very effective approach iteratively approximates the squared distance to the surface with a quadratic penalty \cite{Pottmann2003a} and minimizes the resulting convex energy. Our method method adapts this strategy to the robust energies necessary to fit a surface to noisy data and outperform the squared distance penalty of \cite{Marinov2005} in three points: 1) We use robust data terms to account for noise and outliers in the input data. A robust energy also reduces the effects of errors in the correspondences (between input samples and the surface point closest them) and the quadratic distance approximation. This is improves the accuracy of the surface independently of the noise present in the data, as our experiments show. 2) We use a second-order approximation to the squared distance function, instead of the first order of \cite{Marinov2005}, to take into account curvature information and fit better the small-scale structures of the surface. 3) We penalize deviations of the tangent space to improve the appearance of sharp creases. Computer Vision, on the other hand, has focused on developing energy models to fit subdivision surfaces to noisy data~\cite{Taylor2016a,Cashman2013,Ili2006,Jaimez2017}, but has ignored the impact of outliers and the optimization on the accuracy of the surface. \cite{Ili2006} fits a sub vision surface to range data by solving a weighted least-squares problem that assumes optimal correspondences between input point and domain parameters. This assumption is removed in \cite{Taylor2016a,Cashman2013,Jaimez2017} by explicitly optimizing the correspondences with least-squares models that preprocess the data to eliminate outliers and can only handle gross outliers. These methods also assume the topology of the surface to be know and initialize it with a manually-designed control mesh \cite{Taylor2016a,Cashman2013} or a sphere~\cite{Jaimez2017}. The sphere initialization is flexible but oversmooths the surface because to prevents self intersections as surface evolves. In terms of optimization, all these techniques use a Levenberg-Marquardt (LM) algorithm that estimates jointly the vertices of the control mesh and the correspondences between the input points and surface samples. This increases the size of the problem and slows the optimization because gradient-based techniques, like LM, only update the surface correspondences to adjacent faces of the control mesh. Our majorize-minimize (MM) algorithm places no limits on these updates and can jump to different valleys of the energy landscape. Our application of subdivision surfaces to shape analysis builds on the work of \cite{Juettler-et-al-16,Burkhart2010,Wawrzinek2016,Goes2016} that investigate the accuracy of computing mass and stiffness matrices on a subdivision surface, but do not consider the estimation of wave-kernel signatures, geodesics, or shape matches in the subdivision framework \section{Experiments} \label{} \subsection*{Experiments on Surface Compression} \begin{figure*} \centering \begin{tabular}{cccccc} \includegraphics[trim=17cm 0cm 17cm 0cm, clip=true, width = 0.09\textwidth]{thinker_PC0} & \includegraphics[trim=17cm 0cm 17cm 0cm, clip=true, width = 0.09\textwidth]{thinker_Poisson0} & \includegraphics[trim=17cm 0cm 17cm 0cm, clip=true, width = 0.09\textwidth]{thinker_n250_collapsedPoisson} & \includegraphics[trim=17cm 0cm 17cm 0cm, clip=true, width = 0.09\textwidth]{thinker0_n250_1storder} & \includegraphics[trim=17cm 0cm 17cm 0cm, clip=true, width = 0.09\textwidth]{thinker_n250_our} & \includegraphics[trim=2cm 0cm 1.75cm 0cm, clip=true, width = 0.09\textwidth]{statue_sa} \\ Point cloud & % Poisson surface collapsed Poisson \cite{Marinov2005} subdiv. surf.& our subdiv. surf. & recognizable \\ 139553 points & % 20932 vertices 250 vertices 250 vertices & 250 vertices & picture \end{tabular} \begin{tabular}{ccccc} \includegraphics[trim=18cm 1cm 18cm 0cm, clip=true, width = 0.15\textwidth]{lotion_PC} & \includegraphics[trim=18cm 0cm 18cm 0cm, clip=true, width = 0.15\textwidth]{lotion_Poisson0} & \includegraphics[trim=18cm 0cm 18cm 0cm, clip=true, width = 0.15\textwidth]{lotion_n250_collapsedPoisson} & \includegraphics[trim=18cm 0cm 18cm 0cm, clip=true, width = 0.15\textwidth]{lotion_n250_1storder} & \includegraphics[trim=18cm 0cm 18cm 0cm, clip=true, width = 0.15\textwidth]{lotion_n250_our} \\ Point cloud & % Poisson surface collapsed Poisson \cite{Marinov2005} subdiv. surf.& our subdiv. surf. \\ 107534 points & % 12928 vertices 250 vertices 250 vertices & 250 vertices \\ \end{tabular} \caption{Surfaces reconstructed from a Kinect point cloud (columns 1) with a state of the art Poisson surface reconstruction method \cite{Kazhdan2013} (column 2) and with the proposed subdivision surface (columns 5). Our technique reconstructs and represents the surface with an accuracy comparable to a high-resolution mesh while compressing the representation to less than 2\%. We reconstruct the nozzle of the bottle with higher accuracy thanks to our robust emery model. The subdivision surface of \cite{Marinov2005} (columns 4) shrinks the statue's V-shaped torso while it oversmooths its hands and produces artifacts at the base of the bottle where samples are irregularly spaced and overlap. } \label{} \end{figure*} Our experiments compare our subdivision representation to triangular meshes with two different settings: reconstructing a surface from noisy Kinect pointclouds, and approximating a fine artifact-free mesh with small-scale structures. For all our experiments we set $q=1.3$ and manually choose $\alpha,\beta$ to fit each shape and model ( either our model or the state-of-the-art methods we compare to ). When the input is a mesh, we set these parameters to minimize the Haussdorff distance between the subdivsision surface and the input mesh. Figures \ref{}, \ref{} show that our subdivision surface reconstructs a surface from a noisy pointcloud as accurately as the mesh obtained by the state-of-the-art reconstruction technique~\cite{Kazhdan2013} but reduces the number of representation variables to $2\%$ of the mesh size. Our subdivision surfaces preserve the smooth regions and sharp corners (torso and hands of Figure \ref{}), and texture details (human hair in Figure \ref{}) of the fine mesh that are lost in a triangular mesh of the same size obtained by quadratic edge collapse. For all our experiments with meshes, we use the high-resolution versions of the TOSCA dataset and fitted a subdivision surface with 1700 and 1000 vertices. Tables \ref{} report the execution time and the Hausdorff distance between the input mesh and our compressed subdivision representation for 72 high-resolution TOSCA meshes. Figures \ref{} show the subdivision surfaces fitted to the input meshes with the different models for a qualitative comparison. \textbf{Comparison to State-of-the-Art Models} When the input is a noisy pointcloud, our surfaces avoid the artifacts of \cite{Marinov2005} because our energy model is robust to outliers (Figure \ref{}). Robustness also guards us against erroneous correspondences in thin structures of clean meshes. We also fit better small curved structures because we use a second-order distance approximation, instead of the first-order of \cite{Marinov2005}, that takes into account curvature information. This is highlighted in the representation of mouths, eyes, and ears of Figures \ref{}. \textbf{Comparison to State-of-the-Art Optimization} We compare the subdivision surfaces obtained by optimizing our energy model with the proposed algorithm and by employing Levenberg-Marquardt to minimize the MM majorizer with respect to the mesh vertices and the correspondence parameters. Our optimization is an order of magnitude faster and less sensitive to local minima than the LM approach proposed by \cite{Taylor2016a,Cashman2013,Ili2006,Jaimez2017}. As a result, our surfaces evolve further away from their initialization to reproduce the small-scale details of the input meshes (ears and mouths of Figure \ref{}). Our fitting technique is moderately slower than the quadratic model of \cite{Marinov2005} as a result of our robust energy, as expected. This speed loss is compensated by our improved surface quality, see Table \ref{}. \textbf{Effects of the different terms in our energy model:} Our model always benefits from the robust term while the tangent-fit is slightly detrimental in rare cases, see columns 1--3 in Table \ref{}. \begin{figure*} \centering \begin{tabular}{cccc} \includegraphics[width = 0.20\textwidth]{david13_tosca}& \includegraphics[width = 0.20\textwidth]{david13_1storder}& \includegraphics[width = 0.20\textwidth]{david13_LM} & \includegraphics[width = 0.20\textwidth]{david13_our}\\ TOSCA mesh, 52565 vertices & \cite{Marinov2005} subdiv. surface & LM solver & our subdiv. surface\\ \includegraphics[trim=0cm 16cm 8cm 0cm, clip=true, width = 0.20\textwidth]{david13_tosca}& \includegraphics[trim=0cm 16cm 8cm 0cm, clip=true, width = 0.20\textwidth]{david13_1storder}& \includegraphics[trim=0cm 16cm 8cm 0cm, clip=true, width = 0.20\textwidth]{david13_LM} & \includegraphics[trim=0cm 16cm 8cm 0cm, clip=true, width = 0.20\textwidth]{david13_our}\\ zoom on TOSCA mesh & zoom on \cite{Marinov2005} surface & zoom on LM solver & zoom on our surface\\ \includegraphics[width = 0.20\textwidth]{centaur3_tosca}& \includegraphics[width = 0.20\textwidth]{centaur3_1storder}& \includegraphics[width = 0.20\textwidth]{centaur3_LM}& \includegraphics[width = 0.20\textwidth]{centaur3_our}\\ TOSCA mesh, 15768 vertices& \cite{Marinov2005} subdiv. surface & LM solver & our subdiv. surface\\ \includegraphics[trim=9cm 14cm 2cm 4.5cm, clip=true, width = 0.20\textwidth]{centaur3_tosca}& \includegraphics[trim=9cm 14cm 2cm 4.5cm, clip=true, width = 0.20\textwidth]{centaur3_1storder}& \includegraphics[trim=9cm 14cm 2cm 4.5cm, clip=true, width = 0.20\textwidth]{centaur3_LM}& \includegraphics[trim=9cm 14cm 2cm 4.5cm, clip=true, width = 0.20\textwidth]{centaur3_our}\\% trims left, bottom, right, top zoom on TOSCA mesh & zoom on \cite{Marinov2005} surface & zoom on LM solver & zoom on our surface\\ \end{tabular} \caption{Comparison to state-of-the-art: high-resolution input mesh (column 1) mesh and subdivision surfaces fitted with \cite{Marinov2005} (column 2) and our proposed model optimized with Levenberg-Marquard (column 3) and our proposed optimization algorithm (column 4). Comparing columns 2 and 4, shows the benefits of using a robust $q$-power distance function and a second-order, instead of a first-order, approximation of the squared distance to the surface to fit thin surface structures like the ears or lips. Comparing columns 3 and 4 show the benefits of quadratic sequential program to avoid small-scale local minima close to the initialization that do not capture the shape details of the nose, lips, or mesh ears. All the subdivision surfaces have 1700 control vertices and are visualized by refining ${\mathbf{M}}_0$ 3 times.} \label{} \end{figure*} \begin{figure*} \centering \begin{tabular}{cccc} \includegraphics[width=0.20\textwidth]{horse6_tosca}& \includegraphics[width=0.20\textwidth]{horse6_1storder}& \includegraphics[width=0.20\textwidth]{horse6_LM}& \includegraphics[width=0.20\textwidth]{horse6_our}\\ TOSCA mesh, 19248 vertices & \cite{Marinov2005} subdiv. surface & LM solver & our subdiv. surface \\ \includegraphics[trim=0cm 9.5cm 14cm 0cm, clip=true, width=0.20\textwidth]{horse6_tosca}& \includegraphics[trim=0cm 9.5cm 14cm 0cm, clip=true, width=0.20\textwidth]{horse6_1storder}& \includegraphics[trim=0cm 9.5cm 14cm 0cm, clip=true, width=0.20\textwidth]{horse6_LM}& \includegraphics[trim=0cm 9.5cm 14cm 0cm, clip=true, width=0.20\textwidth]{horse6_our}\\ zoom on TOSCA mesh & zoom on \cite{Marinov2005} surface & zoom on LM solver & zoom on our surface\\ \includegraphics[trim=0cm 6cm 0cm 0cm, clip=true,height = 0.4\textwidth]{victoria2_tosca}& \includegraphics[trim=0cm 6cm 0cm 0cm, clip=true, height = 0.4\textwidth]{victoria2_1storder}& \includegraphics[trim=0cm 6cm 0cm 0cm, clip=true, height = 0.4\textwidth]{victoria2_LM} & \includegraphics[trim=0cm 6cm 0cm 0cm, clip=true, height = 0.4\textwidth]{victoria2_our}\\ TOSCA mesh, 45659 vertices& \cite{Marinov2005} subdiv. surface & LM solver & our subdiv. surface\\ \includegraphics[trim=2cm 18cm 1cm 0.5cm, clip=true, width=0.20\textwidth]{victoria2_tosca}& \includegraphics[trim=2cm 18cm 1cm 0.5cm, clip=true, width=0.20\textwidth]{victoria2_1storder}& \includegraphics[trim=2cm 18cm 1cm 0.5cm, clip=true, width=0.20\textwidth]{victoria2_LM} & \includegraphics[trim=2cm 18cm 1cm 0.5cm, clip=true, width=0.20\textwidth]{victoria2_our}\\ zoom on TOSCA mesh & zoom on \cite{Marinov2005} surface & zoom on LM solver & zoom on our surface\\ \end{tabular} \caption{Comparison to state-of-the-art: high-resolution input mesh (column 1) mesh and subdivision surfaces fitted with \cite{Marinov2005} (column 2) and our proposed model optimized with Levenberg-Marquard (column 3) and our proposed optimization algorithm (column 4). Comparing columns 2 and 4, shows the benefits of using a robust $q$-power distance function to fit thin surface structures like the lips or mesh eyelids without artifacts. Comparing columns 3 and 4 show the benefits of quadratic sequential program to avoid small-scale local minima close to the initialization that do not capture the shape details of the nose, lips, or mesh ears. All the subdivision surfaces have 1700 control vertices and are visualized by refining ${\mathbf{M}}_0$ 3 times.} \label{} \end{figure*} \begin{figure*} \centering \begin{tabular}{cccc} \includegraphics[width=0.20\textwidth]{michael13_tosca}& \includegraphics[width=0.20\textwidth]{michael13_1storder}& \includegraphics[width=0.20\textwidth]{michael13_LM}& \includegraphics[width=0.20\textwidth]{michael13_our}\\ TOSCA mesh, 52565 vertices& \cite{Marinov2005} subdiv. surface & LM solver & our subdiv. surface\\ \includegraphics[trim=2cm 14cm 1cm 0.5cm, clip=true, width=0.20\textwidth]{michael13_tosca}& \includegraphics[trim=2cm 14cm 1cm 0.5cm, clip=true, width=0.20\textwidth]{michael13_1storder}& \includegraphics[trim=2cm 14cm 1cm 0.5cm, clip=true, width=0.20\textwidth]{michael13_LM}& \includegraphics[trim=2cm 14cm 1cm 0.5cm, clip=true, width=0.20\textwidth]{michael13_our}\\ zoom on TOSCA mesh & zoom on \cite{Marinov2005} surface& zoom on LM solver & zoom on our surface\\ \end{tabular} \caption{Comparison to state-of-the-art: high-resolution input mesh (column 1) mesh and subdivision surfaces fitted with \cite{Marinov2005} (column 2) and our proposed model optimized with Levenberg-Marquard (column 3) and our proposed optimization algorithm (column 4). Comparing columns 2 and 4, shows the benefits of using the second-order, instead of a first-order, approximation of the squared distance to the surface to capture the curvature details of small structures like the mesh ears. Comparing columns 3 and 4 show the benefits of quadratic sequential program to avoid small-scale local minima close to the initialization that do not capture the shape details of the nose, lips, or mesh eyes. All the subdivision surfaces have 1700 control vertices and are visualized by refining ${\mathbf{M}}_0$ 3 times.} \label{} \end{figure*} \begin{table*} \vspace{-3cm} \thisfloatpagestyle{empty} \setlength\extrarowheight{-2pt} \centering \caption{Hausdorff distance (as a \textperthousand of the bounding box diagonal) between the input mesh and a subdivision surface with 1000 and 1700 control vertices. Columns 1--3 and 6--8 show the importance of each term in the energy model, while columns 4--5 and 9--10 compare our algorithm to state-of-the-art model of \cite{Marinov2005} and to the Levenberg-Marquardt (LM) optimization advocated by \cite{Taylor2016a,Cashman2013,Jaimez2017}. } \footnotesize \begin{tabular}{| l | c c c c c | c c c c c |} \hline & \multicolumn{5}{|c|}{1000 control vertices} & \multicolumn{5}{|c|}{1700 control vertices } \\ mesh & {$q=2$} & {$\alpha=0$} & {full model} & {\cite{Marinov2005}} & {LM solver} & {$q=2$} & {$\alpha=0$} & {full model} & {\cite{Marinov2005}} & {LM solver} \\ \hline cat0 & 0.180 & 0.176 & \textbf{0.158} & 0.209 & 0.237 & 0.113 & \textbf{0.103} & 0.103 & 0.123 & 0.141 \\ cat1 & 0.251 & 0.237 & \textbf{0.218} & 0.295 & 0.327 & 0.160 & 0.155 & \textbf{0.148} & 0.192 & 0.192 \\ cat10 & 0.213 & 0.201 & \textbf{0.019} & 0.246 & 0.251 & 0.132 & 0.123 & \textbf{0.121} & 0.155 & 0.184 \\ cat2 & 0.436 & 0.390 & \textbf{0.352} & 0.475 & 0.572 & 0.272 & 0.256 & \textbf{0.240} & 0.300 & 0.366 \\ cat3 & 0.210 & 0.200 & \textbf{0.183} & 0.259 & 0.263 & 0.128 & 0.127 & \textbf{0.126} & 0.155 & 0.165 \\ cat4 & 0.254 & 0.247 & \textbf{0.232} & 0.303 & 0.345 & 0.168 & 0.166 & \textbf{0.154} & 0.195 & 0.209 \\ cat5 & 0.219 & 0.208 & \textbf{0.201} & 0.254 & 0.286 & 0.139 & 0.131 & \textbf{0.129} & 0.166 & 0.165 \\ cat6 & 0.390 & 0.390 & \textbf{0.359} & 0.454 & 0.443 & 0.260 & 0.257 & \textbf{0.240} & 0.313 & 0.313 \\ cat7 & 0.437 & 0.391 & \textbf{0.373} & 0.468 & 0.483 & 0.268 & 0.253 & \textbf{0.246} & 0.304 & 0.444 \\ cat8 & 0.233 & 0.209 & \textbf{0.197} & 0.251 & 0.261 & 0.156 & 0.145 & \textbf{0.139} & 0.168 & 0.189 \\ cat9 & 0.283 & 0.263 & \textbf{0.248} & 0.336 & 0.356 & 0.179 & 0.171 & \textbf{0.166} & 0.204 & 0.238 \\ centaur0 & 0.448 & 0.432 & \textbf{0.400} & 0.490 & 0.589 & 0.305 & 0.303 & \textbf{0.285} & 0.347 & 0.357 \\ centaur1 & 0.508 & 0.449 & \textbf{0.424} & 0.507 & 0.526 & 0.323 & 0.321 & \textbf{0.298} & 0.361 & 0.374 \\ centaur2 & 0.632 & 0.578 & \textbf{0.553} & 0.653 & 0.733 & 0.402 & 0.403 & \textbf{0.381} & 0.459 & 0.571 \\ centaur3 & 0.443 & 0.446 & \textbf{0.417} & 0.489 & 0.597 & 0.306 & 0.310 & \textbf{0.289} & 0.352 & 0.361 \\ centaur4 & 0.545 & 0.488 & \textbf{0.478} & 0.550 & 0.651 & 0.336 & 0.346 & \textbf{0.320} & 0.393 & 0.394 \\ centaur5 & 0.474 & 0.466 & \textbf{0.441} & 0.521 & 0.563 & 0.319 & 0.325 & \textbf{0.304} & 0.371 & 0.422 \\ david0 & 0.274 & 0.250 & \textbf{0.232} & 0.273 & 0.402 & 0.174 & 0.171 & \textbf{0.163} & 0.187 & 0.203 \\ david1 & 0.403 & 0.377 & \textbf{0.370} & 0.445 & 0.645 & 0.275 & 0.262 & \textbf{0.255} & 0.296 & 0.502 \\ david10 & 0.396 & 0.329 & \textbf{0.324} & 0.378 & 0.453 & 0.218 & 0.216 & \textbf{0.199} & 0.251 & 0.283 \\ david11 & 0.477 & 0.434 & \textbf{0.421} & 0.479 & 1.004 & 0.313 & 0.293 & \textbf{0.284} & 0.351 & 0.410 \\ david12 & 0.409 & 0.363 & \textbf{0.360} & 0.391 & 0.465 & 0.257 & 0.233 & \textbf{0.219} & 0.272 & 0.381 \\ david13 & 0.349 & 0.341 & \textbf{0.321} & 0.381 & 0.647 & 0.238 & 0.227 & \textbf{0.216} & 0.253 & 0.304 \\ david6 & 0.297 & 0.295 & \textbf{0.272} & 0.333 & 0.343 & 0.185 & 0.186 & \textbf{0.177} & 0.211 & 0.288 \\ dog0 & 0.302 & 0.291 & \textbf{0.272} & 0.349 & 0.424 & 0.195 & 0.184 & \textbf{0.180} & 0.225 & 0.286 \\ dog1 & 0.433 & 0.407 & \textbf{0.398} & 0.470 & 0.624 & 0.269 & 0.264 & \textbf{0.250} & 0.323 & 0.381 \\ dog10 & 0.440 & 0.378 & \textbf{0.369} & 0.469 & 0.539 & 0.244 & 0.246 & \textbf{0.232} & 0.289 & 0.295 \\ dog2 & 0.360 & 0.333 & \textbf{0.311} & 0.425 & 0.421 & 0.214 & 0.207 & \textbf{0.202} & 0.249 & 0.303 \\ dog3 & 0.384 & 0.377 & \textbf{0.362} & 0.453 & 0.463 & 0.273 & 0.253 & \textbf{0.242} & 0.297 & 0.402 \\ dog5 & 0.336 & 0.327 & \textbf{0.309} & 0.391 & 0.449 & 0.209 & 0.205 & \textbf{0.193} & 0.244 & 0.266 \\ dog6 & 0.365 & 0.335 & \textbf{0.328} & 0.408 & 0.455 & 0.221 & \textbf{0.206} & 0.206 & 0.247 & 0.391 \\ dog7 & 0.444 & 0.429 & \textbf{0.413} & 0.496 & 0.628 & 0.288 & 0.280 & \textbf{0.268} & 0.331 & 0.389 \\ dog8 & 0.458 & 0.435 & \textbf{0.408} & 0.523 & 0.598 & 0.288 & 0.287 & \textbf{0.279} & 0.340 & 0.362 \\ gorilla1 & 0.497 & \textbf{0.379} & 0.398 & 0.457 & 0.753 & 0.340 & 0.274 & \textbf{0.274} & 0.312 & 0.449 \\ gorilla14 & 0.854 & \textbf{0.688} & 0.719 & 0.885 & 1.348 & 0.599 & \textbf{0.486} & 0.489 & 0.572 & 0.776 \\ gorilla5 & 0.711 & 0.568 & \textbf{0.546} & 0.689 & 0.880 & 0.446 & \textbf{0.380} & 0.383 & 0.444 & 0.706 \\ gorilla8 & 0.793 & 0.650 & \textbf{0.621} & 0.775 & 1.080 & 0.530 & \textbf{0.449} & 0.451 & 0.538 & 0.863 \\ horse0 & 0.372 & 0.366 & \textbf{0.355} & 0.431 & 1.252 & 0.250 & 0.246 & \textbf{0.233} & 0.290 & 0.317 \\ horse10 & 0.351 & 0.323 & \textbf{0.315} & 0.375 & 1.223 & 0.226 & 0.225 & \textbf{0.216} & 0.274 & 0.294 \\ horse14 & 0.370 & 0.348 & \textbf{0.343} & 0.422 & 0.457 & 0.246 & 0.248 & \textbf{0.240} & 0.298 & 0.346 \\ horse15 & 0.352 & \textbf{0.331} & 0.336 & 0.399 & 0.478 & 0.230 & 0.230 & \textbf{0.212} & 0.273 & 0.281 \\ horse17 & 0.379 & 0.356 & \textbf{0.352} & 0.414 & 0.448 & 0.257 & 0.250 & \textbf{0.239} & 0.297 & 0.342 \\ horse5 & 0.416 & 0.374 & \textbf{0.364} & 0.430 & 0.484 & 0.260 & 0.264 & \textbf{0.250} & 0.314 & 0.370 \\ horse6 & 0.353 & 0.350 & \textbf{0.340} & 0.411 & 0.424 & 0.233 & 0.237 & \textbf{0.222} & 0.283 & 0.309 \\ horse7 & 0.351 & 0.333 & \textbf{0.327} & 0.398 & 0.435 & 0.238 & 0.233 & \textbf{0.224} & 0.284 & 0.284 \\ michael0 & 0.309 & 0.276 & \textbf{0.266} & 0.320 & 0.392 & 0.185 & 0.189 & \textbf{0.176} & 0.210 & 0.257 \\ michael1 & 0.376 & 0.340 & \textbf{0.333} & 0.396 & 0.480 & 0.246 & 0.226 & \textbf{0.214} & 0.256 & 0.322 \\ michael10 & 0.417 & 0.373 & \textbf{0.346} & 0.418 & 0.618 & 0.267 & 0.256 & \textbf{0.240} & 0.290 & 0.384 \\ michael11 & 0.665 & 0.576 & \textbf{0.547} & 0.681 & 0.879 & 0.443 & \textbf{0.409} & 0.412 & 0.475 & 0.904 \\ michael12 & 0.436 & 0.365 & \textbf{0.344} & 0.431 & 0.507 & 0.268 & 0.262 & \textbf{0.255} & 0.297 & 0.394 \\ michael13 & 0.438 & 0.371 & \textbf{0.364} & 0.422 & 0.698 & 0.283 & 0.267 & \textbf{0.255} & 0.296 & 0.382 \\ michael15 & 0.440 & 0.387 & \textbf{0.369} & 0.449 & 0.639 & 0.283 & 0.263 & \textbf{0.258} & 0.294 & 0.427 \\ michael16 & 0.403 & 0.366 & \textbf{0.349} & 0.416 & 0.552 & 0.263 & 0.259 & \textbf{0.249} & 0.296 & 0.459 \\ michael18 & 0.431 & 0.371 & \textbf{0.346} & 0.422 & 0.594 & 0.266 & 0.255 & \textbf{0.245} & 0.285 & 0.449 \\ michael19 & 0.417 & 0.381 & \textbf{0.367} & 0.433 & 0.517 & 0.259 & 0.254 & \textbf{0.243} & 0.294 & 0.363 \\ michael2 & 0.578 & 0.476 & \textbf{0.470} & 0.537 & 0.647 & 0.358 & 0.325 & \textbf{0.318} & 0.365 & 0.527 \\ michael3 & 0.426 & 0.356 & \textbf{0.346} & 0.413 & 0.564 & 0.262 & 0.249 & \textbf{0.233} & 0.279 & 0.278 \\ michael4 & 0.414 & 0.364 & \textbf{0.350} & 0.415 & 0.600 & 0.264 & 0.249 & \textbf{0.239} & 0.282 & 0.386 \\ michael5 & 0.366 & 0.326 & \textbf{0.313} & 0.384 & 0.454 & 0.222 & 0.220 & \textbf{0.206} & 0.251 & 0.342 \\ michael6 & 0.384 & 0.363 & \textbf{0.349} & 0.411 & 0.453 & 0.248 & 0.234 & \textbf{0.227} & 0.265 & 0.325 \\ michael7 & 0.474 & 0.394 & \textbf{0.365} & 0.450 & 0.645 & 0.287 & 0.273 & \textbf{0.258} & 0.311 & 0.387 \\ victoria1 & 0.321 & \textbf{0.267} & 0.281 & 0.330 & 0.365 & 0.194 & 0.176 & \textbf{0.166} & 0.205 & 0.298 \\ victoria10 & 0.278 & 0.275 & \textbf{0.252} & 0.314 & 0.567 & 0.165 & 0.166 & \textbf{0.156} & 0.202 & 0.239 \\ victoria17 & 0.370 & \textbf{0.305} & 0.309 & 0.357 & 0.436 & 0.203 & 0.190 & \textbf{0.174} & 0.222 & 0.310 \\ victoria2 & 0.302 & 0.267 & \textbf{0.243} & 0.322 & 0.363 & 0.164 & 0.157 & \textbf{0.150} & 0.189 & 0.240 \\ victoria21 & 0.537 & \textbf{0.449} & 0.457 & 0.532 & 1.168 & 0.322 & 0.293 & \textbf{0.281} & 0.337 & 0.449 \\ victoria24 & 0.311 & 0.293 & \textbf{0.277} & 0.354 & 0.416 & 0.188 & 0.173 & \textbf{0.169} & 0.205 & 0.329 \\ victoria25 & 0.283 & 0.258 & \textbf{0.233} & 0.312 & 0.377 & 0.152 & 0.144 & \textbf{0.137} & 0.174 & 0.236 \\ victoria4 & 0.296 & 0.280 & \textbf{0.267} & 0.317 & 0.390 & 0.174 & 0.167 & \textbf{0.160} & 0.197 & 0.243 \\ wolf0 & 0.429 & 0.427 & \textbf{0.414} & 0.477 & 0.510 & 0.331 & 0.334 & \textbf{0.320} & 0.355 & 0.431 \\ wolf1 & 0.410 & 0.419 & \textbf{0.389} & 0.455 & 0.517 & 0.298 & 0.315 & \textbf{0.293} & 0.335 & 0.367 \\ wolf2 & 0.443 & 0.463 & \textbf{0.427} & 0.501 & 0.570 & 0.330 & 0.345 & \textbf{0.327} & 0.363 & 0.471 \\ \hline \end{tabular} \label{} \end{table*} \begin{table*} \vspace{-3cm} \setlength\extrarowheight{-2pt} \centering \caption{Execution time (in seconds) to fit a subdivision surface with 1000 and 1700 control vertices. Columns 1--3 and 6--8 show how each additional term in the energy model slows the optimization, while columns 4--5 and 9--10 compare our technique to state-of-the-art model of \cite{Marinov2005} and to the Levenberg-Marquardt (LM) optimization algorithm advocated by \cite{Taylor2016a,Cashman2013,Jaimez2017}. } \footnotesize \begin{tabular}{| l | c c c c c | c c c c c |} \hline & \multicolumn{5}{|c|}{1000 control vertices} & \multicolumn{5}{|c|}{1700 control vertices } \\ \hline mesh & {$q=2$} & {$\alpha=0$} & {full model} & {\cite{Marinov2005}} & {LM solver} & {$q=2$} & {$\alpha=0$} & {full model} & {\cite{Marinov2005}} & {LM solver} \\ \hline cat0 & 32.7 & 34.7 & 56.0 & \textbf{20.8} & 408.0 & 24.6 & 44.5 & 36.0 & \textbf{18.6} & 346.5 \\ cat1 & 37.1 & 41.2 & 44.0 & \textbf{12.9} & 345.2 & 34.5 & 43.3 & 52.6 & \textbf{15.1} & 367.5 \\ cat10 & 24.7 & 36.5 & 34.9 & \textbf{18.2} & 403.3 & \textbf{30.3} & 50.0 & 38.7 & 37.4 & 544.6 \\ cat2 & 23.5 & 33.9 & 43.6 & \textbf{21.3} & 450.0 & 31.3 & 44.3 & 42.3 & \textbf{21.0} & 399.9 \\ cat3 & 28.1 & 36.6 & 47.4 & \textbf{16.4} & 333.7 & 41.5 & 51.8 & 32.4 & \textbf{18.6} & 382.5 \\ cat4 & 30.6 & 43.4 & 46.7 & \textbf{16.5} & 344.1 & 28.6 & 34.8 & 54.7 & \textbf{20.6} & 384.0 \\ cat5 & 18.7 & 38.8 & 41.6 & \textbf{17.4} & 547.1 & 29.6 & 47.3 & 33.4 & \textbf{17.1} & 372.9 \\ cat6 & 24.9 & 43.9 & 40.7 & \textbf{18.0} & 365.1 & 28.5 & 48.3 & 41.1 & \textbf{16.1} & 332.3 \\ cat7 & 30.1 & 23.9 & 40.9 & \textbf{14.3} & 402.6 & 26.6 & 50.9 & 51.4 & \textbf{20.3} & 468.0 \\ cat8 & 34.7 & 42.1 & 35.9 & \textbf{17.4} & 267.5 & 27.4 & 48.7 & 44.1 & \textbf{20.8} & 378.9 \\ cat9 & 22.6 & 35.7 & 56.6 & \textbf{19.1} & 361.4 & 35.5 & 29.7 & 47.2 & \textbf{17.0} & 468.4 \\ centaur0 & 20.7 & 16.5 & 21.5 & \textbf{8.2} & 165.4 & 16.2 & 22.1 & 24.0 & \textbf{15.8} & 185.1 \\ centaur1 & 12.0 & 21.9 & 16.5 & \textbf{8.9} & 279.3 & 23.8 & 22.0 & 33.7 & \textbf{11.7} & 374.1 \\ centaur2 & 16.0 & 15.9 & 26.0 & \textbf{9.9} & 209.2 & 19.2 & 24.9 & 24.1 & \textbf{10.4} & 171.3 \\ centaur3 & 18.3 & 14.1 & 25.9 & \textbf{8.5} & 192.4 & 18.8 & 21.9 & 26.4 & \textbf{10.2} & 258.7 \\ centaur4 & 16.9 & 24.0 & 26.9 & \textbf{8.8} & 267.1 & 24.2 & 28.3 & 27.9 & \textbf{11.3} & 307.5 \\ centaur5 & 14.3 & 17.0 & 28.1 & \textbf{8.8} & 254.2 & 18.9 & 25.0 & 25.9 & \textbf{9.8} & 197.1 \\ david0 & 59.5 & 59.4 & 86.9 & \textbf{39.3} & 602.6 & 64.8 & 56.1 & 63.4 & \textbf{40.2} & 473.5 \\ david1 & 58.8 & 56.4 & 70.1 & \textbf{34.9} & 447.5 & 60.1 & 55.2 & 63.0 & \textbf{36.0} & 799.9 \\ david10 & 44.7 & 47.1 & 72.1 & \textbf{37.6} & 589.8 & 59.0 & 59.8 & 79.9 & \textbf{40.2} & 611.4 \\ david11 & 68.1 & 58.2 & 128.5 & \textbf{37.3} & 683.6 & 59.2 & 76.5 & 100.4 & \textbf{33.1} & 713.5 \\ david12 & 63.4 & 52.0 & 61.7 & \textbf{38.8} & 528.8 & 63.8 & 53.2 & 74.9 & \textbf{45.6} & 622.2 \\ david13 & 61.0 & 45.7 & 86.8 & \textbf{37.4} & 807.0 & 58.8 & 58.4 & 108.8 & \textbf{43.0} & 832.3 \\ david6 & 50.9 & 42.6 & 71.3 & \textbf{30.7} & 592.7 & 54.2 & 60.7 & 63.2 & \textbf{33.7} & 551.9 \\ dog0 & 20.7 & 25.4 & 48.8 & \textbf{14.0} & 325.0 & 23.2 & 31.1 & 39.6 & \textbf{14.6} & 321.8 \\ dog1 & 22.9 & 27.4 & 30.1 & \textbf{15.7} & 281.2 & 39.2 & 39.6 & 41.8 & \textbf{16.8} & 368.4 \\ dog10 & 21.4 & 30.5 & 39.6 & \textbf{12.8} & 283.4 & 31.0 & 31.4 & 32.8 & \textbf{15.0} & 348.5 \\ dog2 & 28.1 & 27.2 & 38.9 & \textbf{14.3} & 251.6 & 29.3 & 45.7 & 48.0 & \textbf{18.8} & 323.8 \\ dog3 & 21.1 & 34.0 & 32.3 & \textbf{16.1} & 303.0 & 28.3 & 27.1 & 34.1 & \textbf{15.4} & 311.2 \\ dog5 & 31.7 & 28.2 & 44.8 & \textbf{20.4} & 271.4 & 34.3 & 48.8 & 34.8 & \textbf{17.5} & 207.0 \\ dog6 & 35.4 & 31.3 & 53.0 & \textbf{18.8} & 293.2 & 24.8 & 45.5 & 49.4 & \textbf{18.5} & 378.7 \\ dog7 & 21.0 & 28.3 & 38.6 & \textbf{15.2} & 427.1 & 44.0 & 47.4 & 45.0 & \textbf{22.2} & 355.7 \\ dog8 & 20.4 & 27.4 & 41.8 & \textbf{14.8} & 304.5 & 28.5 & 45.4 & 37.6 & \textbf{15.7} & 366.5 \\ gorilla1 & 36.8 & 33.0 & 61.2 & \textbf{23.6} & 531.9 & 45.8 & 48.3 & 64.4 & \textbf{34.1} & 416.0 \\ gorilla14 & 32.8 & 41.3 & 52.9 & \textbf{21.0} & 519.9 & 41.6 & 72.1 & 65.6 & \textbf{25.3} & 620.7 \\ gorilla5 & \textbf{37.9} & 47.2 & 55.1 & 38.3 & 565.8 & 28.3 & 55.3 & 63.8 & \textbf{24.3} & 837.3 \\ gorilla8 & 46.5 & 44.5 & 59.6 & \textbf{19.9} & 436.8 & 37.8 & 64.6 & 69.5 & \textbf{26.9} & 695.7 \\ horse0 & 29.6 & 22.5 & 46.4 & \textbf{12.9} & 98.6 & 22.0 & 42.0 & 34.0 & \textbf{13.0} & 234.2 \\ horse10 & 23.1 & 26.7 & 47.0 & \textbf{12.6} & 120.0 & 34.9 & 36.7 & 32.1 & \textbf{11.8} & 302.6 \\ horse14 & 17.5 & 26.2 & 31.8 & \textbf{13.4} & 261.4 & 30.9 & 45.8 & 45.9 & \textbf{13.2} & 406.2 \\ horse15 & 17.5 & 21.6 & 34.1 & \textbf{12.0} & 489.1 & 20.1 & 24.3 & 40.4 & \textbf{16.5} & 252.8 \\ horse17 & 16.9 & 22.6 & 21.4 & \textbf{13.9} & 233.1 & 24.9 & 35.5 & 36.8 & \textbf{13.1} & 316.4 \\ horse5 & 17.2 & 28.9 & 30.5 & \textbf{13.2} & 322.4 & 29.3 & 35.0 & 37.5 & \textbf{16.5} & 346.2 \\ horse6 & 23.2 & 34.8 & 25.4 & \textbf{11.2} & 250.3 & 30.2 & 44.5 & 38.8 & \textbf{13.1} & 254.3 \\ horse7 & 28.0 & 24.0 & 38.6 & \textbf{11.1} & 194.6 & 19.7 & 29.8 & 28.0 & \textbf{12.7} & 290.6 \\ michael0 & 63.1 & 54.9 & 74.1 & \textbf{37.4} & 632.1 & 60.2 & 54.9 & 76.2 & \textbf{40.9} & 599.8 \\ michael1 & 64.5 & 55.0 & 86.9 & \textbf{47.2} & 588.5 & 56.3 & 56.5 & 63.7 & \textbf{44.5} & 909.7 \\ michael10 & 54.4 & 63.1 & 108.5 & \textbf{38.4} & 476.4 & 69.4 & 51.8 & 76.5 & \textbf{29.3} & 517.2 \\ michael11 & 44.1 & 80.9 & 115.8 & \textbf{29.3} & 583.0 & 54.3 & 56.9 & 84.3 & \textbf{31.3} & 762.0 \\ michael12 & 79.7 & 75.2 & 100.4 & \textbf{37.2} & 607.9 & 77.8 & 61.6 & 118.6 & \textbf{40.5} & 573.2 \\ michael13 & 53.7 & 70.3 & 103.5 & \textbf{39.0} & 553.5 & 52.2 & 51.5 & 71.6 & \textbf{39.4} & 546.4 \\ michael15 & 66.4 & 61.5 & 121.4 & \textbf{34.5} & 690.1 & 62.5 & 53.3 & 68.3 & \textbf{38.2} & 443.0 \\ michael16 & \textbf{46.8} & 50.9 & 65.4 & 47.8 & 495.6 & 80.9 & 69.7 & 72.0 & \textbf{56.1} & 571.5 \\ michael18 & 59.0 & 60.1 & 98.9 & \textbf{32.9} & 311.1 & 55.9 & 69.6 & 78.6 & \textbf{34.9} & 472.5 \\ michael19 & 53.6 & 52.5 & 58.6 & \textbf{34.6} & 473.3 & 71.0 & 65.9 & 60.5 & \textbf{38.0} & 535.3 \\ michael2 & 55.3 & 54.4 & 55.5 & \textbf{35.4} & 764.9 & 61.2 & 47.2 & 68.9 & \textbf{35.7} & 476.2 \\ michael3 & 51.8 & 57.4 & 95.9 & \textbf{36.2} & 452.0 & 52.8 & 49.6 & 60.7 & \textbf{38.3} & 658.7 \\ michael4 & 48.6 & 47.5 & 76.3 & \textbf{33.7} & 569.7 & 76.4 & 52.2 & 70.8 & \textbf{36.3} & 576.4 \\ michael5 & 77.6 & 59.3 & 106.4 & \textbf{27.4} & 468.1 & 57.1 & 53.4 & 57.7 & \textbf{38.9} & 1117.8 \\ michael6 & 63.6 & 39.5 & 68.5 & \textbf{26.5} & 389.0 & 68.5 & 67.0 & 65.8 & \textbf{37.6} & 439.6 \\ michael7 & 55.4 & 65.5 & 96.4 & \textbf{35.0} & 508.0 & 62.0 & 57.8 & 62.4 & \textbf{40.5} & 489.5 \\ victoria1 & 49.7 & 49.4 & 52.5 & \textbf{29.3} & 439.4 & 30.3 & 38.9 & 52.5 & \textbf{27.5} & 381.9 \\ victoria10 & 41.8 & 36.8 & 40.3 & \textbf{24.4} & 527.8 & 59.3 & 60.4 & 57.3 & \textbf{30.3} & 405.8 \\ victoria17 & 39.2 & 64.7 & 55.6 & \textbf{27.7} & 582.8 & 45.0 & 43.9 & 72.0 & \textbf{30.6} & 534.7 \\ victoria2 & 53.5 & 38.0 & 45.8 & \textbf{27.6} & 448.4 & 48.0 & 57.0 & 67.9 & \textbf{27.6} & 584.8 \\ victoria21 & 35.1 & 69.0 & 80.2 & \textbf{27.0} & 321.3 & 37.5 & 42.9 & 68.9 & \textbf{24.5} & 444.0 \\ victoria24 & 40.7 & 46.2 & 84.5 & \textbf{22.7} & 366.5 & 27.6 & 47.2 & 76.1 & \textbf{26.8} & 549.3 \\ victoria25 & 49.9 & 52.9 & 57.7 & \textbf{23.6} & 308.4 & 47.3 & 79.1 & 65.5 & \textbf{28.5} & 412.3 \\ victoria4 & 46.5 & 43.7 & 60.4 & \textbf{27.5} & 393.3 & 41.6 & 46.4 & 58.7 & \textbf{23.7} & 478.4 \\ wolf0 & 7.4 & 12.1 & 11.5 & \textbf{6.8} & 64.2 & 10.6 & 16.7 & 9.1 & \textbf{5.7} & 114.0 \\ wolf1 & \textbf{6.5} & 8.5 & 9.1 & 8.2 & 78.5 & 17.0 & 22.8 & \textbf{15.8} & 16.6 & 137.9 \\ wolf2 & 7.0 & 8.0 & 9.7 & \textbf{5.7} & 52.9 & \textbf{8.0} & 11.6 & 16.7 & 8.1 & 106.7 \\ \hline \end{tabular} \label{} \end{table*} \subsection*{Experiments on Shape Analsysis}\label{} Figure~\ref{} compares the wave kernel signature ~\cite{aubry-et-al-11} at a surface point $w_x$ with three different discretizations: a fine triangular mesh, a coarse triangular mesh, and our subdivision surface. This signature is a shape descriptor that assigns to each point $x\in S$ a function $w_x$ that depends on the value of the $\Delta$-eigenfunctions and their eigenvalues at $x$. The point signature of coarse mesh is different from the signature of fine mesh, which is well matched by the subdivision surface. The functional map framework~\cite{Ovsjanikov2012} formulates the problem of matching points of two shapes as matching smooth functions over them. The problem is invariant to isometric transformations by representing smooth function with $\Delta$-eigenfunctions and by finding a linear map between these eigenfunctions. Figure~\ref{} shows the computed matching with respect to two different shape representations: a fine mesh with 27894 vertices and a subdivision surface with 1700 control vertices. Our experiments show comparable matching results with a slight loss of accuracy at the tails of some cats, but this is compensated by a surface representation and matching problem of 10\% of the original size. To visualize the results, we show a template shape with a color map that interprets the three spatial coordinates of every point as RGB color information and transfer this color map to each target shape with the matching estimated with functional maps~\cite{Ovsjanikov2012}. In this experiment, using a coarse triangular mesh to do the shape matching simply fails. \begin{figure*} \centering \begin{tabular}{ccccc} \includegraphics[height = 0.17\textwidth]{figMatch-fine-VORONOI-0} & \includegraphics[height = 0.17\textwidth]{figMatch-fine-VORONOI-3} & \includegraphics[height = 0.17\textwidth]{figMatch-fine-VORONOI-5} & \includegraphics[height = 0.17\textwidth]{fig5-fine-VORONOI-out} \\ Fine mesh & Matched shape & Matched shape & Matched shape \\ \\ \includegraphics[height = 0.17\textwidth]{figMatch-sds-0} & \includegraphics[height = 0.17\textwidth]{figMatch-sds-3} & \includegraphics[height = 0.17\textwidth]{figMatch-sds-52} &\includegraphics[height = 0.17\textwidth]{fig5-sds-out110}\\ Subdivision surface & Matched shape & Matched shape & Matched shape \\ \end{tabular} \caption{Shape matching results of the functional map framework~\cite{Ovsjanikov2012}. After the matching is determined, the color information of the target shape is transferred to the matched shape with the mapping estimated by functional maps\cite{Ovsjanikov2012}.} \label{} \end{figure*} Figures \ref{} compare geodesics computed by the Heat method \cite{Crane2013} with different surface representations fitted to the Kinect pointclouds of Figures \ref{} and \ref{}. As a proxy for ground-truth, we use a high-resolution triangular mesh obtained by the Poisson reconstruction method~\cite{Kazhdan2013} and compare it to two compressed surface representations of the same size: a low-resolution triangular mesh obtained by quadratic edge collapse~\cite{Garland1997} of the high-resolution mesh and our subdivision representation. Our subdivision surface is comparable to the high-resolution mesh and can be used to compressed both the shape and its geodesics functions. In contrast, the low-resolution triangular mesh provides a discretization too coarse to represent the small scale details of the surface and its geodesic. To highlight the loss of small-scale structure on the geodesic, we represent the level lines of the geodesic at high and low resolution and demonstrate how the geodesic estimated from the coarse triangular mesh is unable to capture the high-frequency information of the geodesic (as estimated from the high-resolution mesh) while our surface representation captures both low and high frequency information accurately. \begin{figure*} \centering \begin{tabular}{cccc} \includegraphics[width=0.20\textwidth]{lotion_Poisson.png}& \includegraphics[width=0.20\textwidth]{lotion_Poisson_geodesic.png} & \includegraphics[width=0.20\textwidth]{lotion_Poisson_geodesiclevels.png}& \includegraphics[width=0.20\textwidth]{lotion_Poisson_geodesiclevels_sl.png} \\ triangular mesh from \cite{Kazhdan2013} & approximate geodesic & 30 geodesic level lines & 5 geodesic 5 level lines \\ 12928 vertices & $g$ & $\cos( 30 \pi g )$ & $\cos( 5 \pi g )$ \\ \includegraphics[width=0.20\textwidth]{lotion_collapsedPoisson.png}& \includegraphics[width=0.20\textwidth]{lotion_collapsedPoisson_geodesic.png}& \includegraphics[width=0.20\textwidth]{lotion_collapsedPoisson_geodesiclevels.png} & \includegraphics[width=0.20\textwidth]{lotion_collapsedPoisson_geodesiclevels_sl.png} \\ edge-collapsed mesh \cite{Kazhdan2013}& approximate geodesic & 30 geodesic level lines & 5 geodesic 5 level lines \\ 250 vertices & $g$ & $\cos( 30 \pi g )$ & $\cos( 5 \pi g )$ \\ \includegraphics[width=0.20\textwidth]{lotion_sds.png} & \includegraphics[width=0.20\textwidth]{lotion_sds_geodesic.png}& \includegraphics[width=0.20\textwidth]{lotion_sds_geodesiclevels.png} & \includegraphics[width=0.20\textwidth]{lotion_sds_geodesiclevels_sl.png}\\ % subdivision surface & approximate geodesic & 30 geodesic level lines & 5 geodesic 5 level lines \\ 250 control vertices & $g$ & $\cos( 30 \pi g )$ & $\cos( 5 \pi g )$ \\ \end{tabular} \caption{Computation of approximate geodesics by the Heat method \cite{Crane2013} with different surface representations fitted to a Kinect pointcloud. Column 1: representation of the surface with a high-resolution mesh obtained with Poisson reconstruction~\cite{Kazhdan2013} (row 1), a low-resolution triangular mesh obtained by edge collapse of the high-resolution mesh (row 2), and with our subdivision surface (row 3). Column 2: approximate geodesic $g$ estimated with the Heat method \cite{Crane2013} on each surface representation. Columns 3 and 4: level lines of the geodesics visualized as $\cos( \varpi g)$, with $\varpi = 30\pi$ in column 3 and $\varpi = 5\pi$ in column 4. Our subdivision surface is comparable to the surface obtained with the Poisson reconstruction method \cite{Kazhdan2013} and can faithfully represent geodesics at both low and high resolution even though it has 2\% of its vertices. In contrast, compressing the state-of-the-art mesh with the quadratic edge collapse method of \cite{Garland1997} looses all the small scale details of the surface and its geodesics. This is highlighted by representing the level lines of the geodesic on the mesh at high resolution (column 3) where we show how the discretization associated with a coarse triangular mesh representation cannot estimate the high-frequency information of the geodesic accurately.} \label{} \end{figure*} \begin{figure*} \centering \begin{tabular}{cccc} \includegraphics[width = 0.23\textwidth]{thinker_Poisson.png}& \includegraphics[width = 0.23\textwidth]{thinker_Poisson_geodesic.png} & \includegraphics[width = 0.23\textwidth]{thinker_Poisson_geodesiclevels.png}& \includegraphics[width = 0.23\textwidth]{thinker_Poisson_geodesiclevels_sl.png} \\ triangular mesh from \cite{Kazhdan2013} & approximate geodesic & 30 geodesic level lines & 5 geodesic 5 level lines \\ 20932 vertices & $g$ & $\cos( 30 \pi g )$ & $\cos( 5 \pi g )$ \\ \includegraphics[width = 0.23\textwidth]{thinker_collapsedPoisson.png}& \includegraphics[width = 0.23\textwidth]{thinker_collapsedPoisson_geodesic.png}& \includegraphics[width = 0.23\textwidth]{thinker_collapsedPoisson_geodesiclevels.png} & \includegraphics[width = 0.23\textwidth]{thinker_collapsedPoisson_geodesiclevels_sl.png} \\ edge-collapsed mesh \cite{Kazhdan2013} & approximate geodesic & 30 geodesic level lines & 5 geodesic 5 level lines \\ 250 vertices & $g$ & $\cos( 30 \pi g )$ & $\cos( 5 \pi g )$ \\ \includegraphics[width = 0.23\textwidth]{thinker_sds.png} & \includegraphics[width = 0.23\textwidth]{thinker_sds_geodesic.png}& \includegraphics[width = 0.23\textwidth]{thinker_sds_geodesiclevels.png} & \includegraphics[width = 0.23\textwidth]{thinker_sds_geodesiclevels_sl.png}\\ % subdivision surface & approximate geodesic & 30 geodesic level lines & 5 geodesic 5 level lines \\ 250 control vertices & $g$ & $\cos( 30 \pi g )$ & $\cos( 5 \pi g )$ \\ \end{tabular} \caption{Computation of approximate geodesics by the Heat method \cite{Crane2013} with different surface representations fitted to a Kinect pointcloud. Column 1: representation of the surface with a high-resolution mesh obtained with Poisson reconstruction~\cite{Kazhdan2013} (row 1), a low-resolution triangular mesh obtained by edge collapse of the high-resolution mesh (row 2), and with our subdivision surface (row 3). Column 2: approximate geodesic $g$ estimated with the Heat method \cite{Crane2013} on each surface representation. Columns 3 and 4: level lines of the geodesics visualized as $\cos( \varpi g)$, with $\varpi = 30\pi$ in column 3 and $\varpi = 5\pi$ in column 4. Our subdivision surface is comparable to the surface obtained with the Poisson reconstruction method \cite{Kazhdan2013} and can faithfully represent geodesics at both low and high resolution even though it has 2\% of its vertices. In contrast, compressing the state-of-the-art mesh with the quadratic edge collapse method of \cite{Garland1997} looses all the small scale details of the surface and its geodesics. This is highlighted by representing the level lines of the geodesic on the mesh at high resolution (column 3) where we show how the discretization associated with a coarse triangular mesh representation cannot estimate the high-frequency information of the geodesic accurately.} \label{} \end{figure*} \section{Conclusions}\label{} We have presented a method to compress surfaces with a representation well suited for shape analysis. To this purpose, we fit a subdivision surface to low-level surface representations and develop standard shape analysis operators with it. Our surface fitting model is robust to noise and outliers and can handle both clean input meshes and noisy point cloud obtained from real data. The resulting subdivision surfaces let us represent smooth shapes at high accuracy with a fraction of the variables used in triangular meshes and use the same subdivision schemes to parametrize functions over the surface and compute tangent spaces or curvatures analytically. Our experiments show how the smoothness of the representation reduces the size of the Laplace-Beltrami operator and its eigenfunctions without loss of accuracy and allows us to compute wave-kernel signatures, approximate geodesics, and shape matches in a compact representation.

\section{Introduction and main results} In this paper, we are interested with a class of periodic stationary sets in $\mathbb{R}^3$ for the following energy functional \begin{equation}\label{} {\mathcal P}_\g(\O):= |\partial \O|+{\g } \int_{\O}\int_{\O}{G}_{ \k}(|x-y|) dxdy, \end{equation} where $\k,\g>0$ and $G_\k(r)=\frac{1}{r } e^{-\k r}$ is the repulsive Yukawa potential (or the \textit{screened} repulsive Coulomb potential). The Yukawa potential $\overline{G}_\k(x)=G_\k(|x|)$ satisfies \begin{equation} -\D \overline{G}_\k+\k^2 \overline{G}_\k= 4\pi \d_0\qquad \textrm{ in $\mathbb{R}^3$.} \end{equation} It plays an important role in the theory of elementary particles, see \cite{Yukawa}. For this energy ${\mathcal P}_\g$, comprising short-range interaction (surface tension) and attractive Yukawa potential, a stationary set $\O$ may separate into well ordered disjoint components. This gives rise to formation of patterns in the limit of several nonlocal reaction-diffusion models from physics, biology and chemistry. Indeed, systems with competing short range and Coulomb-type attractive interaction provides sharp interfaces to the Ohta-Kawasaki model of block polymers and the activator-inhibitor reaction-diffusion system, see for for instance \cite{GMP,SA, OS} and the references therein. The paper of Muratov \cite{Muratov2002} contains a rigorous derivation of \eqref{} from the general mean-field free energy functional, with $\g \approx \k^2$ and $\g,\k\to0$. Equilibrium patterns such as spots, stripes and the annuli, together with morphlogical instabilities have been also studied in \cite{Muratov2002}. At certain energy levels, \eqref{} provides the sharp interface for the Ohta-Kawasaki model of diblock copolymer melt, see Muratov \cite{Muratov} and the work of Goldmann, Muratov and Serfaty in \cite{Serfaty2013,Serfaty2014} for related variant of $\Gamma$-convergence on the 2-dimensional torus. In \cite{Oshita}, Oshita showed that \eqref{} is the $\G$-limit for a class of activator-inhibitor reaction-diffusion system, see also Petrich and Goldstein in \cite{PD}, for the two dimensional case. The question of minimizing ${\mathcal P}_\g$ has been also raised by H. Kn\"{u}pfer, C. Muratov and M. Novaga in \cite{KMN}, for physical and mathematical perspectives. They speculated that bifurcations from trivial to nontirival solutions can be expected. Moreover they predict the "pearl-necklace morphology" (or string-of-pearl) exhibited by long polyelectrolyte molecules in poor solvents, \cite{DR, FAM}. In a geometric point of view, this is alike the cylinder-to-sphere transition, and known as Delauney surfaces (or unduloids). The unduloids are constant mean curvature surfaces of revolution interpolating between a straight cylinder to a string of tangent round spheres. Here, we will consider the limit configuration, and we prove existence of stationary sets for ${\mathcal P}_\g$ formed by a periodic string of nearly round balls.\\ It is known, see for instance \cite{Muratov2002, FFMMM}, that a stationary set (or critical point) $\O$ for the functional ${\mathcal P}_\g$ satisfies the following Euler-Lagrange equation \begin{equation} \label{} H_{\partial\O}(x)+\g \int_{\O}G_\k(|x-y|) dy=Const. \qquad \textrm{ for all } x\in \partial\O, \end{equation} where $H_{\partial\O}$ is the mean curvature (positive for a ball) of $\partial\O$. Here and in the following, for every set $\O$ with $C^2$ boundary (not necessarily bounded), we define the function $$ {\mathcal H}_\O:\partial\O\to \mathbb{R} $$ given by \begin{equation}\label{} {\mathcal H}_{\O}(x)= H_{\partial\O}(x)+\g \int_{\O}G_\k(|x-y|) dy,\qquad \textrm{ for every $x\in \partial\O$}. \end{equation} In this paper, \textit{equilibrium patterns} are sets $\O\subset\mathbb{R}^3$ for which ${\mathcal H}_\O\equiv Const$ on $\partial \O$.\\ Our aim is to construct nontrivial unbounded equilibrium patterns $\O$. These sets are multiply periodic and bifurcate from lattices of round \textit{spheres}, straight \textit{cylinders} and \textit{slabs}. In particular, we obtain stationary sets for ${\mathcal P}_\g$ made by lattices of near-spheres centered at any $M$-dimensional Bravais lattice. When $M=1$, these are adjacent nearly-spherical sets centered at a straight line with equal gaps between them. The cylinders, result from the 2-dimensional existence of spherical patterns with the corresponding Yukawa potential given by a modified Bessel function. These configurations appear for all $\g\in (0,\g_N)$, for some positive constant $\g_N$. Finally, we obtain lamellar structures (modulated slabs) bifurcating from parallel planes. The bifurcations of these modulated slabs occur as long as $ \frac{\k\sqrt{\k^2+1}}{\sqrt{\k^2+1}-\k}<2\pi \g <(\k^2+1)^{3/2}$. We first describe our doubly periodic slabs. We consider domains of the form, \begin{equation} \label{} \O_{v_p}=\left\{ (t, z)\in \mathbb{R}^{2}\times \mathbb{R} \,:\, -v_p(t )<z<v_p(t ) \right\} \subset\mathbb{R}^3, \end{equation} where $v_p: \mathbb{R}^2 \to (0,\infty)$ is $2\pi\mathbb{Z}^2$-periodic and satisfies some symmetry properties. Indeed, we let $C^{k,\gamma}(\mathbb{R}^2)$ denotes the space of $C^k(\mathbb{R}^2)$ bounded functions $u$, with bounded derivatives up to order $k$ and with $D^k u $ having finite H\"older seminorm of order $\a\in (0,1)$. The space $C^{k,\a}_p(\mathbb{R}^2) $ denotes the space of functions in $C^{k,\a}(\mathbb{R}^2)$ which are $2\pi\mathbb{Z}^2$-periodic. Finally, we define \begin{align} \label{} C^{k,\a}_{p,{\mathcal R}} &=\left\{v_p \in C^{k,\a}_p(\mathbb{R}^2)\,: \, \textrm{ $v_p( t_1,t_2)=v_p(t_2,t_1)=v_p(-t_1,t_2)$, for all $ (t_1,t_2)\in \mathbb{R}^2$} \right\}. \nonumber \end{align} The spaces $C^{k,\a}_{p,{\mathcal R}}$ and $C^{k,\a}_{p}(\mathbb{R}^2)$ are equipped with the standard H\"older norm of $C^{k,\a} (\mathbb{R}^2) $. Our first main result is the following. \begin{Theorem}\label{} Let $\k>0$. Then for every $\g$ satisfying \begin{equation}\label{} \frac{1}{2\pi} \frac{1}{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+1 }} }< \g < \frac{1}{2\pi} \frac{ 1}{ \frac{1}{(\k^2+1 )^{3/2}} }, \end{equation} there exists $\l_*=\l_*(\k,\g)>0$ and a smooth curve $$ (-s_0,s_0)\to \mathbb{R}_+\times C^{2,\a}_{p,{\mathcal R}}, \qquad s\mapsto (\l_s, v_p_{s}) $$ with the properties that $\l_0=\l_*$ and $\O_{v_p_s}$ is an equilibrium pattern satisfying, for all $s\in (-s_0,s_0)$, $$ {\mathcal H}_{\O_{v_p_{s}}} (x)=2\pi\g\,\textrm{arcsinh}(2\l_s/\k) \qquad \textrm{ for every $x\in \partial\O_{v_p_{s}}$.} $$ Moreover, for every $s \in (-s_0,s_0)$ and $(t_1,t_2)\in \mathbb{R}^2$, $$ v_p_{s}(t_1,t_2 )=\l_s+s\Bigl( \cos(t_1)+\cos(t_2)+v_{s}(t_1,t_2 )\Bigr), $$ with a smooth curve $ (-s_0,s_0) \to C^{2,\alpha}_{p,{\mathcal R}}(\mathbb{R}^2)$, $s\mapsto v_{s}$ satisfying $$ \int_{[-\pi,\pi]^2} v_{s}(t) \cos( t_j) \,dt = 0 \qquad \text{for $j=1,2$,} $$ and $v_{0}\equiv 0$. \end{Theorem} The domains $\O_{v_p_{s}}$ in Theorem \ref{} bifurcates from the same slab $\{-\l_*<z<\l_*$. For $\g$ satisfying $\frac{1}{2\pi} \frac{1}{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+1 }} }< \g $ (see the lower bound in \eqref{}), the constant $\l_*$ is the unique zero of the function \begin{equation} \label{} \l\mapsto 1 - 2\pi \g \left\{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+1 }} - \frac{\exp(-2\l\sqrt{\k^2+1}) }{ \sqrt{\k^2+1}} - \frac{\exp(-2\l\k) }{ \k}\right\}. \end{equation} The Crandall-Rabonowitz bifurcation theorem is the main tool we have used here to prove the existence of the branches $s\mapsto (\l_s,v_p_s)$. It provides a bifurcation result from a \textit{ unique simple zero eigenvalue}.\\ We now discuss the condition \eqref{}. If $ \g\leq \frac{1}{2\pi} \frac{1}{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+1 }} } $ then the only solutions bifurcating from $\O_{\l}$, for $\l>0$, are the trivial ones. This follows from the fact that all eigenvalues of the lienearization of $v_p\mapsto {\mathcal H}_{\O_v_p}$ at $\l$ are positive. However, equilibrium patterns with different morphology (undulated cylinders), boundary of Delauney-type surfaces, exist for $ \g< \frac{1}{2\pi} \frac{1}{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+1 }} } $, see \cite{MMF}. Now if $\g \geq \frac{1}{2\pi} \frac{ 1}{ \frac{1}{(\k^2+1 )^{3/2}} }$, we do not know the number of zero eigenvalues of the linearized operator of $v_p\mapsto {\mathcal H}_{\O_v_p}$ about the straight slab $\O_{\l_*}$, while the uniqueness of the trivial eigenvalue is crucial in our analysis. We describe next our (modulated) lamellae. For $\e\in \mathbb{R}_*$ close to zero, we consider the lamellae made by the slices of the slabs $\O_{\psi}$ above: \begin{equation} \label{} {\mathcal L}_{\psi}^\e=\O_{\psi}+\frac{1}{|\e|} \mathbb{Z} e_3=\bigcup_{p\in \mathbb{Z}}\left( \O_{\psi}+\left(0,0,\frac{p}{|\e|}\right)\right), \end{equation} where $\O_\psi$ is as defined in \eqref{}. It is easy to see that for small $|\e|$, the set ${\mathcal L}_{\psi}^\e $ is as smooth as $\psi$. Moreover the straight lamellae ${\mathcal L}^\e_1$ is an equilibrium pattern. More precisely, $ {\mathcal H}_{{\mathcal L}^\e_\l}\equiv Const$ for every $\l>0$ and $|\e|<\frac{1}{2\l}$. We establish the existence of nontrivial periodic patterns bifurcating from ${\mathcal L}^\e_{\l_*} $, with $\l_*$ given by Theorem \ref{}. \begin{Theorem}\label{} Let $\k,\g,\l_s>0$, $v_p_s$ and $s_0>0$ be as in Theorem \ref{}. Then there exists $\e_0>0$ and a smooth branch $$(-\e_0,\e_0) \times (-s_0,s_0)\to \mathbb{R}_+\times C^{2,\a}_{p,{\mathcal R}}(\mathbb{R}^2), \qquad \e\mapsto ( \l_{\e,s}, \psi_{\e,s})$$ such that the set ${\mathcal L}_{\psi_{\e,s}}^\e:= \O_{\psi_{\e,s}}+\frac{1}{|\e|} \mathbb{Z} e_3$ defined in \eqref{} is an equilibrium pattern, with $$ {\mathcal H}_{{\mathcal L}_{\psi_{\e,s}}^\e } (x)= {\mathcal H}_{{\mathcal L}_{\l_{\e,s}}^\e } \qquad \textrm{ for every $x\in \partial {\mathcal L}_{\psi_{\e,s}}^\e $}, $$ where, $\l_{0,s}=\l_s$, $\psi_{0,s}=v_p_s$. Moreover, for every $s\in (s_0,s_0)$, $\e\in (-\e_0,\e_0)$ and $(t_1,t_2)\in \mathbb{R}^2$, $$ \psi_{\e,s}(t_1,t_2)= \l_{\e,s}+s \Bigl( \cos(t_1)+\cos(t_2) +v_{\e, s}(t_1,t_2)\Bigr) $$ with $$ \int_{[-\pi,\pi]^2}v_{\e,s}(t) \cos(t_j) dt=0\qquad \textrm{ for $j=1,2$.} $$ \end{Theorem} As mentioned earlier, we are also interested on the cylindrical and spherical stationary set for the energy ${\mathcal P}_\g$. These surfaces are obtained by perturbing balls in $\mathbb{R}^N$, with $N=2,3$, centered at any $M$-dimensional Bravais lattice for $1\leq M\leq N$. The perturbation parameter is a scaling parameter which makes the distance between the lattice points sufficiently large. \\ Let $M\in\mathbb{N}$ with $1\leq M\leq 3$. Let $\left\{ \textbf{a}_1;\dots;\textbf{a}_M\right\}$ be a basis of $\mathbb{R}^M$. We consider the $M$-dimensional {Bravais lattice} in $\mathbb{R}^3$: $$ \scrL^M=\left\{\sum_{i=1}^M k_i \textbf{a}_i\,:\, k=(k_1,\dots,k_M)\in \mathbb{Z}^M \right\}\subset \mathbb{R}^3. $$ Here and in the following, we identify $\mathbb{R}^M$ with $\mathbb{R}^M\times \{0_{\mathbb{R}^{3-M}}\}\subset \mathbb{R}^3$. Obviously $\scrL^1=c\mathbb{Z}$, for some nonzero real number $c$. We define the perturbed cylinder $$ C_v_p:= \left\{( r,\th,t)\in \mathbb{R}_+\times S^1\times \mathbb{R} \,:\, r<v_p(\th) \right\}, $$ where $ (0,\infty)$ is an even function. For $\e>0$ small, we consider the lattice of perturbed cylinders centered at the lattice points $\scrL^M$ given by \begin{equation} \label{} {\mathcal C}_{v_p}^\e:=C_v_p+ \frac{1}{\e} \scrL^M. \end{equation} We obtain the following result. \begin{Theorem}\label{} Let $\scrL^M$ be an $M$-dimensional Bravais lattices, with $M=1,2$. Then there exists $ \g_2>0$ such that for every $\g\in (0, \g_2)$, there exists a smooth curve $(0,\e_0)\to C^{2,\a}(S^1)$, $\e\mapsto v_p_\e$ such that for every $\e\in (0,\e_0)$, the lattice of perturbed cylinders $ {\mathcal C}_{v_p_\e}^\e$ given by \eqref{} is an equilibrium pattern, with $$ {\mathcal H}_{ {\mathcal C}_{v_p_\e}^\e}(x)= {\mathcal H}_{ C_{1}}(\cdot)\qquad \textrm{ for every $x\in \partial {\mathcal C}_{v_p_\e}^\e$.} $$ Moreover $v_p_\e=1+\o_\e$ with $\o_\e$ even and $\|\o_\e\|_{C^{2,\a}(S^1) }\leq e^{-\frac{c}{\e}}$ as $\e\to 0$, for some positive constant $c$ depending only on $\k,\a,\g$ and $\scrL^M$. When $M=1$, the function $\o_\e$ is not constant. \end{Theorem} We emphasize that Theorem \ref{} results from a two-dimensional study. Indeed spherical equilibrium patterns in 2-dimension with the corresponding 2-dimensional screened Coulomb potential (given by the second of modified Bessel function $2 K_0(\k r)$) yields cylinderical patterns in 3-dimension with the Yukawa potential $G_\k(r)$. In \cite{CO}, Chen and Oshita showed that among all 2-dimensional Bravais lattices of near-balls, the hexagonal one provides least possible energy. \\ We expect that the perturbation $\o_\e$ is not constant on $S^1$. We are able to prove this fact only in the case of lower dimensional lattice, $M=1$, see Section \ref{s:Remarks} below. The fact that the perfect 2-dimensional lattice of cylinders $ {\mathcal C}_{v_p}^\e:=C_1+ \frac{1}{\e} \scrL^2$ is not an equilibrium pattern remains an open question. However, numerical simulations give evidence that they are not constant, at least in the hexagonal configuration, see Muratov \cite{Muratov2002}, for $0<\g=O(\k^2)\ll1$. \\ To introduce the sphere lattices, we consider the $M$-dimensional Bravais $\scrL^M$ as above, with $M\leq 3$. Here, $ S^{2}$ is the $2$-dimensional unit sphere. We consider the perturbed ball $$ B_\phi^3:= \{(r, \th )\in \mathbb{R}_+\times S^{2}\,:\, 0<r<\phi(\th )\}, $$ where $\phi: S^{2}\to (0,\infty)$ is an even function. For $\e$ small enough, we consider the following disjoint union of perturbed balls \begin{equation} \label{} {\mathcal B}_\phi^{\e} := B_{\phi}^3+\frac{1}{\e} \scrL^M. \end{equation} We have the following result. \begin{Theorem}\label{} There exists $ \g_3>0$ such that for every $\g\in (0, \g_3)$, there exists a smooth curve $(0,\e_0)\to C^{2,\a}(S^{2})$, $\e\mapsto \phi_\e$ such that for every $\e\in (0,\e_0)$, the set ${\mathcal B}_{\phi_\e}^\e$ defined by \eqref{} is an equilibrium pattern, with $$ {\mathcal H}_{ {\mathcal B}_{\phi_\e}^{\e}} (x)= {\mathcal H}_{ B^3_1} (\cdot)\qquad \textrm{ for every $x\in \partial {\mathcal B}_{\phi_\e}^{\e}$.} $$ Moreover ${\phi_\e}=1+w_\e$ with $\o_\e$ even and $\|w_\e\|_{C^{2,\a}(S^{2}) }\leq e^{-\frac{c}{\e}}$ as $\e\to 0$, for some positive constant $c$ depending only on $\k,\a,\g$ and $\scrL^M$. When $M=2$, the function $w_\e$ is not constant. \end{Theorem} The constants $\g_N$, for $N=2,3,$ in Theorem \ref{} and Theorem \ref{} depend only on the eigenvalues of the linearized operator of the map $w\mapsto {\mathcal H}_{B^{N}_w}$, see Lemma \ref{} below. In the case of near-sphere lattices also, we are able to prove that $w_\e$ is not constant on $S^{N-1}$ only in the lower dimensional lattices, $M\leq 2$, see Section \ref{s:Remarks} below. It is an open question to know whether for $M=3$, the function $w_\e$ is not constant i.e. ${\mathcal H}_{{\mathcal B}^\e_1}$ is not constant on $\partial B_1$. Related to results in the present paper are the works, for $\k=0$, of Ren and Wei in their series of papers \cite{RW-1,RW-2,RW-3,RW-4,RW-5,RW-6,RW-7,RW-8,RW-9,RW-10,RW-11}. Among many others, they build equilibrium patterns of type lamellar, ring/cylindrical, and spot/spherical solutions. See also \cite{KR-1, KR-2} and the references therein. \\ Next, we emphasize some similarities between equilibrium patterns derived here and the structures of Constant Nonlocal Mean Curvatures (CNMC) sets recently discovered. The notion of nonlocal mean curvature was recently introduced by Caffarelli and Souganidis in \cite{Caff-Soug2010}. See Caffarelli, Roquejoffre, and Savin~\cite{Caffarelli2010}, it arises from the Euler-Lagrange equation for the fractional perimeter functional which in turn is a $\Gamma$-limit of some nonlocal phase transition problems, see \cite{SO}.\\ The long range interaction in the expression of ${\mathcal H}_\O$ is, up to an additional constant, a weighted average of all "+1" in $\O$ and all "-1" from outside $\O$, since, \begin{equation}\label{ -tau} {\mathcal H}_\O(x)= H_{\partial\O}(x)+\frac{\g}{2} \int_{\mathbb{R}^3}(1_{\O}(y)-1_{\mathbb{R}^3\setminus\O}(y))G_\k(|x-y|) dy+ \overline{c}, \end{equation} for $x\in \partial\O$, where $\overline{c}=\frac{1}{2}\int_{\mathbb{R}^3}G_\k(|x|)dx=\frac{2\pi}{\k^2}$. Similarly, for $\a\in (0,1)$, the fractional or nonlocal mean curvature function ${\mathcal H}_\O^\a:\partial\O\to \mathbb{R}$ is given by $$ {\mathcal H}_\O^\a(x):=- \frac{1-\a}{|B^2|} \int_{\mathbb{R}^3}(1_{\O}(y)-1_{\mathbb{R}^3\setminus\O}(y))|x-y|^{-3-\a} dy $$ (positive if $\O$ is a ball). The factor $ \frac{1-\a}{|B^2|}$ implies that $H_{\partial \O}^\a(x)$ converges to the mean curvature $ H_{\partial \O}(x)$ as $\a\to1$, see \cite{Davila2014B,Abatangelo}. We can parallel the recently constructed CNMC sets and the equilibrium patterns obtained here. Indeed, by \cite{Cabre2015A, CFW}, CNMC bands exist, and from which 1-periodic CNMC slabs can be derived. Modulated CNMC slabs are found in \cite{MT}. Finally from \cite{Cabre2015B}, there exist CNMC near-sphere lattices from which CNMC lattices of near-cylinders can be derived. \\ Finally, we point out that resemblance between phase structures in diblock copolymer melts, and triply periodic constant mean curvature surfaces was also observed several years ago, see e.g. \cite{BF,TAHH}.\\ \section{Preliminary and notations} We let $K_\nu$ be the second kind of modified Bessel function of order $\nu \in \mathbb{R}$. See e.g. \cite{El}, for $\nu\in \mathbb{R}$, there holds $$ K_\nu'(r)=-\frac{\nu}{r}K_\nu(r) -K_{\nu-1}(r) $$ and $K_{-\nu}=K_{\nu}$ for $\nu<0$. Therefore \begin{equation} \label{} K_0'(r)=-K_1(r), \qquad (rK_1)'(r)= -r K_{0}(r). \end{equation} See also \cite{El}, for $\nu\geq 0$, \begin{equation}\label{} K_{\nu}(r)\sim \frac{\sqrt{\pi }}{\sqrt{2}} r^{-1/2}e^{-r} \qquad \textrm{ as $r\to +\infty$} \end{equation} and \begin{equation}\label{} K_{0}(r)\sim -\log (r), \qquad K_{\nu}(r)\sim \frac{1}{2}\G(\nu) (r/2)^{-\nu} \qquad\textrm{ as $r\to 0$.} \end{equation} We now collect some useful formula, where most of them can be found in \cite{GR}. In the following, we let $\a,\b,\d\geq 0$ and $\k>0$. By \cite[page 491, ET I 16(27)]{GR}, we have \begin{equation}\label{} \int_{\mathbb{R}} \cos(\b r) G_\k\left(\sqrt{r^2+\a ^2} \right)dr = 2 K_0(\a \sqrt{\k^2+\b^2}). \end{equation} See also \cite[page 719, WA 425(10)a, MO 48]{GR}, we have \begin{equation}\label{} \int_{\mathbb{R}}\cos(\b r) K_0(r\sqrt{\k^2+\a^2})dr= \frac{\pi}{\sqrt{\k^2+\b^2+\a^2 }}. \end{equation} By computing, we have \begin{equation}\label{} \int_{\mathbb{R}^2} G_\k\left(|x|\right)dx=\frac{2\pi}{\k}. \end{equation} From \cite[page 722, ET I 56(43)]{GR}, we get \begin{equation}\label{} \int_{\mathbb{R}} \cos(\b r) K_0\left( \sqrt{r^2+\d^2}\sqrt{\a^2+\k^2} \right) dr = \pi\frac{\exp(-\d \sqrt{\b^2+\a^2 +\k^2})}{\sqrt{\b^2+\a^2 +\k^2}}. \end{equation} Combining \eqref{} and \eqref{}, we obtain % \begin{equation}\label{} \int_{\mathbb{R}^2} G_\k\left(\sqrt{|x|^2+\d^2 }\right) dx = 2 \pi\frac{\exp(-\d \k) }{ \k}. \end{equation} For a finite set ${\mathcal N}$, we let $|{\mathcal N}|$ denote the length (cardinal) of ${\mathcal N}$. It will be understood that $ |\emptyset|=0$. Let $Z$ be a Banach space and $U$ a nonempty open subset of $Z$. If $T \in C^{k}(U,\mathbb{R})$ and $u \in U$, then $D^kT(u)$ is a continuous symmetric $k$-linear form on $Z$ whose norm is given by $$ \|D^{k}T ({u}) \|= \sup_{{u}_{1}, \dots, {u}_{k}\in Z } \frac{|D^{k} T ({u})[u_1,\dots,u_k]| }{ \prod_{j=1}^k \| {u}_{j} \|_{ Z }} . $$ If, $L: Z\to \mathbb{R}$ is a linear map, we have \begin{equation} \label{} D^{|{\mathcal N}|}(L T_2 )({u})[u_i]_{i\in{\mathcal N}}= L({u}) D^{|{\mathcal N}|} T_2({u})[u_i]_{i\in{\mathcal N}} + \sum_{j\in{\mathcal N}} L({u}_j) D^{|{\mathcal N}|-1} T_2({u}) [u_i]_{\stackrel{i \in {\mathcal N}}{ i\neq j}}. \end{equation} % We let $T$ be as above, $V \subset \mathbb{R}$ open with $T(U) \subset V$ and $g:V \to \mathbb{R}$ be a $k$-times differentiable map. The Fa\'{a} de Bruno formula states that \begin{equation} \label{} D^k( g\circ T)(u)[u_1,\dots,u_k]= \sum_{\Pi\in\scrP_k} g^{ (\left|\Pi\right|)}(T(u)) \prod_{P\in\Pi} D^{\left|P\right| }T(u)[u_j]_{j \in P} , \end{equation} for $u, u_1,\dots,u_k \in U$, where $\scrP_k$ denotes the set of all partitions of $\left\{ 1,\dots, k \right\}$, see e.g. \cite{FaadeBruno-JW}. \section{ Perodic slabs}\label{s:const-slab} \subsection{Doubly periodic slabs}\label{} In this section we construct equilibrium patterns with period cell given by a doubly periodic slab. That is domains of the form $$ \O_v_p:=\left\{ (t, z)\in \mathbb{R}^{2}\times \mathbb{R} \,:\, -v_p(t)<z<v_p(t) \right\} \subset\mathbb{R}^3, $$ where $v_p: \mathbb{R}^2 \to (0,\infty)$, with $v_p\in C^{2,\a}(\mathbb{R}^2)$. Recall that for $x\in \{-1,1\} $, the mean curvature is given by $$ H_{\partial\O_v_p}((t, v_p(t)x) )= -\text{div} \frac{\n v_p(t)}{\sqrt{1+|\n v_p(t)|^2}} . $$ Moreover, by a change of variable, the Yukawa interaction takes the form \begin{align*} &\int_{\O_v_p }G_\k(|(t, v_p(t)x)-X|) dX=\int_{\mathbb{R}^2} \int_{-v_p(s)}^{v_p(s)} G_\k(|(t, v_p(t)x)-(s,z)|)dz ds\\ &= \int_{\mathbb{R}^2} \int_{-1}^1v_p(s) G_\k(|(t, v_p(t)x)-(s,v_p(s)y)|) dy ds\\ &= \int_{\mathbb{R}^2} \int_{-1}^1v_p(s) G_\k(|(t, v_p(t))-(s,v_p(s)y)|) dy ds. \end{align*} Next, we define the open set $$ {\mathcal O}:=\{ v_p \in C^{2,\alpha}(\mathbb{R}^2)\,:\, v_p>0 \} $$ and the map ${\mathcal F}: {\mathcal O}\to C^{0,\a}(\mathbb{R}^2)$ by \begin{align} &{\mathcal F}(v_p)(t):= -\text{div} \frac{\n v_p(t)}{\sqrt{1+|\n v_p(t)|^2}} +\g \int_{\mathbb{R}^2}\int_{-1 }^1v_p(s)G_\k(|(t, v_p(t) )-(s,v_p(s)y)|) dyds,\label{} \end{align} so that, for every $x\in \{-1,1\}$ and $t\in \mathbb{R}^2$, $$ {\mathcal H}_{\O_v_p}((t,v_p(t)x))= {\mathcal F}(v_p)(t). $$ Our aim is to apply the Crandall-Rabinowitz bifurcation theorem to solve the equation $$ {\mathcal F}(\l+ u)=const $$ for $\l >0$ and $u\in C^{2,\a}(\mathbb{R}^2)$ is small. We will achieve this by solving the equation \begin{equation} \label{ solve-slab} {\mathcal F}(\l+ u)={\mathcal F}(\l). \end{equation} We note that by some computation and \eqref{}, we have \begin{equation}\label{} {\mathcal F}(\l)= \g \l \int_{-1}^1\int_{\mathbb{R}^2}G_\k(\sqrt{|r|^2+\l^2(1-y)^2})dr dy=2\pi\g\,\textrm{arcsinh}(2\l/\k). \end{equation} Let us study the linearized operator $D{\mathcal F}(\l)$, for $\l>0$. We consider the map $$ {\mathcal O} \to C^{0,\a}(\mathbb{R}^2),\qquad {\mathcal F}_1(v_p)(t):= \int_{\mathbb{R}^2}v_p(s) \int_{-1 }^1G_\k((|t-s|^2+(v_p(t) -v_p(s)y)^2)^{1/2}) dyds . $$ We have the following result. \begin{Lemma}\label{} The map ${\mathcal F}: {\mathcal O}\to C^{0,\a}(\mathbb{R}^2)$ is smooth, with $$ \|D^k {\mathcal F}(v_p)\|\leq c (1+\|v_p\|_{C^{2,\a}(\mathbb{R}^2)})^c , $$ for some positive constant $c=c(\k,\a,\d,k)$. Moreover for every $v_p\in {\mathcal O}$ and $w\in C^{2,\a}(\mathbb{R}^2)$, \begin{align} \label{} D{\mathcal F}_1(v_p)[w](t) &= \int_{\mathbb{R}^2} (w(t-r)-w(t)) G_\k((|r|^2+ ( v_p(t) - v_p(t-r))^2)^{1/2}) dr\nonumber\\ &\quad+\int_{\mathbb{R}^2} (w(t-r) +w(t))G_\k((|r|^2+ ( v_p(t) + v_p(t-r))^2)^{1/2}) dr . \end{align} In particular, for every $\l>0$, \begin{align} \label{} &D{\mathcal F}(\l)[w](t) =-\D w(t) - \g \int_{\mathbb{R}^2} (w(t)-w(t-r)) G_\k\left(|r| \right)dr\nonumber \\ % &\quad +\g ( w(t)+w(t-r)) \int_{\mathbb{R}^2} G_\k\left((|r|^2+4 \l^2 )^{1/2} \right) dr . \end{align} \end{Lemma} \begin{proof} We notice that \begin{align} &{\mathcal F}(v_p)(t) =-\text{div} \frac{\n v_p(t)}{\sqrt{1+|\n v_p(t)|^2}} \nonumber\\ &\quad +\g \int_{\mathbb{R}^2}v_p(s) \int_{-1 }^1G_\k((|t-s|^2+(v_p(t) -v_p(s)y)^2)^{1/2}) dyds. \end{align} The regularity of the map $$ C^{2,\a}(\mathbb{R}^2)\to C^{0,\a}(\mathbb{R}^2),\qquad v_p\mapsto \text{div} \frac{\n v_p(t)}{\sqrt{1+|\n v_p(t)|^2}} $$ is easy, we therefore consider the second term. Fro every $\d>0$, we define $$ {\mathcal O}_\d=\{ v_p\in C^{2,\a}(\mathbb{R}^2)\:\, v_p>\d\}. $$ Let us first compute the formal derivative of this map which we will show that it is smooth. Once this is done we can easily deduce the regularity of ${\mathcal F}_1$ and hence of ${\mathcal F}$. We have \begin{align}\label{} D{\mathcal F}_1(v_p)[w](t)=\int_{\mathbb{R}^2}w(s) \int_{-1}^1f(y)dy ds+ \int_{\mathbb{R}^2} \int_{-1}^1 (w(s)y-w(t)) f'(y)dy ds, \end{align} where $$ f(y)=G_\k(|(t, v_p(t) )-(s,v_p(s)y)|) $$ which satisfy $$ f'(y)=v_p(s)(v_p(s)y- v_p(t) ) G_\k'(|(t, v_p(t) )-(s,v_p(s)y)|) . $$ Integration by parts give \begin{align*} \int_{-1}^1& (w(s)y-w(t) ) f'(y)dy=-\int_{-1}^1 w(s) f(y)dy+ (w(s)\th-w(t) ) f(\th)\Big|^{\th=1}_{\th=-1}. \end{align*} Inserting this in \eqref{}, we deduce that \begin{align*} D{\mathcal F}_1(v_p)[w](t)&=\int_{\mathbb{R}^2} (w(s)\th-w(t)) f(\th)ds \Big|^{\th=1}_{\th=-1}\\ &=\int_{\mathbb{R}^2} (w(s)-w(t)) f(1)ds-\int_{\mathbb{R}^2} (-w(s)-w(t)) f(-1)ds\\ &= \int_{\mathbb{R}^2} (w(s)-w(t)) G_\k(|(t, v_p(t) )-(s,v_p(s))|) ds\\ &\quad+\int_{\mathbb{R}^2} (w(s)+w(t)) G_\k(|(t, v_p(t) )-(s,-v_p(s))|) ds. \end{align*} This gives \eqref{} after the change of variable $s=t-r$. We now have \begin{align} \label{} D{\mathcal F}_1(v_p)[w](t) &= \int_{\mathbb{R}^2} (w(t-r)-w(t)) G_\k(|(|r|^2+ ( v_p(t) - v_p(t-r))^2)^{1/2}|) ds\nonumber\\ &\quad+\int_{\mathbb{R}^2} (w(t-r) +w(t)) G_\k(|(|r|^2+ ( v_p(t) + v_p(t-r))^2)^{1/2}|) ds\nonumber\\ &= \int_{\mathbb{R}^2} \frac{w(t-r)-w(t)}{|r|} G_{{\k}{|r|}}\left( \left(1+ \left(\L_0(v_p,t,r)\right)^2\right)^{1/2} \right) dr\nonumber\\ &\quad+\int_{\mathbb{R}^2} (w(t-r) +w(t)) G_\k((|r|^2+ ( v_p(t) + v_p(t-r))^2)^{1/2}) dr, \end{align} where $\L_0: C^{2,\a}(\mathbb{R}^2)\times \mathbb{R}^4\to \mathbb{R}$ is given by $$ \L_0(v_p,t,r)= \int_0^1\nv_p(t-\varrho r )\cdot \frac{r}{|r|} d\varrho. $$ For every fixed $w\in C^{2,\a}(\mathbb{R}^2)$, we can define ${\mathcal J}_i: {\mathcal O}_\d \to L^\infty(\mathbb{R}^2)$ by $$ {\mathcal J}_1(v_p )(t):= \int_{\mathbb{R}^2} \frac{w(t-r)-w(t)}{|r|} G_{{\k}{|r|}}\left(\left| \left(1+ \left(\L_0(v_p,t,r)\right)^2\right)^{1/2}\right|\right) dr $$ and $$ {\mathcal J}_2(v_p)(t):= \int_{\mathbb{R}^2} (w(t-r) +w(t)) G_\k(|(|r|^2+ ( v_p(t) + v_p(t-r))^2)^{1/2}|) dr. $$ Therefore by \eqref{} \begin{equation}\label{} D{\mathcal F}_1(v_p)[w]= {\mathcal J}_1(v_p )+{\mathcal J}_2(v_p ). \end{equation} Thanks to \eqref{} and the fact that the integrand in ${\mathcal J}_2$ is bounded away from zero, it becomes straightforward to see that the map ${\mathcal J}_2: {\mathcal O}_\d\to C^{0,\a}(\mathbb{R}^2)$, is smooth and for every $k\in \mathbb{N}$, \begin{equation} \label{} \|D^k {\mathcal J}_2(v_p)\|\leq c (1+\|v_p\|_{C^{2,\a}(\mathbb{R}^2)})^c\,\|w\|_{C^{2,\a}(\mathbb{R}^2) }, \end{equation} where $c=c(\k,\a,\d,k)>0$. Next, we prove the regularity of ${\mathcal J}_1: {\mathcal O}_\d\to C^{0,\a}(\mathbb{R}^2)$. For $|r|>0$, we consider the function $$ h_r(x)= G_{{\k}{|r|}}\left( \left(1+x^2\right)^{1/2} \right). $$ We observe for every $k\in \mathbb{N}$, there exists a constant $c=c(\k,k)$ such that $$ |h^{(k)}_{r}(x)|\leq c\, G_{{\k}{|r|}}(1/2)\qquad \textrm{ for every $x\in \mathbb{R}$, $r\in \mathbb{R}^{2}\setminus\{0\}$}. $$ Letting ${\mathcal K}_1= C^{2,\a}(\mathbb{R}^2)\times \mathbb{R}^4\to \mathbb{R}$, given by $$ {\mathcal K}_1(v_p,t,r):= G_{{\k}{|r|}}\left(\left| \left(1+ \left(\L_0(v_p,t,r)\right)^2\right)^{1/2}\right|\right). $$ We can use \eqref{}, $g(x)=h_{r}(x)$ and $ T(v_p)= {\mathcal K}(v_p,t,r)$ to get the following estimates. For every $(t,r)\in \mathbb{R}^2\times \mathbb{R}^2$, $$ \| D^k_v_p {\mathcal K}_1(v_p,t,r)\|\leq c (1+\|v_p\|_{2,\a})^cG_{{\k}{|r|}}(1/2) , $$ with $c=c(\a,\k,k)>0$. Since also $\int_{\mathbb{R}^2}|r|^{-1} G_{{\k}|r|}(1/2)<\infty$, by using \eqref{}, we get, for every $k\in \mathbb{N}$, $$ \|D^k {\mathcal J}_1(v_p)\|\leq c (1+\|v_p\|_{C^{2,\a}(\mathbb{R}^2)})^c \, \|w\|_{C^{2,\a}(\mathbb{R}^2) }, $$ with $c=c(\k,\a,\d,k)>0$. This then shows that ${\mathcal J}_1: {\mathcal O}_\d\to C^{0,\a}(\mathbb{R}^2)$. From this together with \eqref{} and \eqref{}, we deduce that $D{\mathcal F}_1: {\mathcal O}_\d\to {\mathcal B}(C^{2,\a}(\mathbb{R}^2), C^{0,\a}(\mathbb{R}^2) ) $ is smooth. Hence ${\mathcal F}$ is smooth in ${\mathcal O}_\d$ and $$ \|D^k {\mathcal F}_1(v_p)\|\leq c (1+\|v_p\|_{C^{2,\a}(\mathbb{R}^2)})^c , \qquad \textrm{ for some positive constant $c=c(\k,\a,\d,k)$.} $$ In addition, we have that \begin{align*} D^k{\mathcal F}_1(v_p)[w](t) &= \int_{\mathbb{R}^2} \frac{w(t-r)-w(t)}{|r|} D^k_v_p {\mathcal K}_1(v_p,t,r) dr\nonumber\\ &\quad+\int_{\mathbb{R}^2} (w(t-r) +w(t)) D^k_v_p {\mathcal K}_2(v_p,t,r) dr, \end{align*} where ${\mathcal K}_2(v_p,t,r)= G_\k(|(|r|^2+ ( v_p(t) + v_p(t-r))^2)^{1/2}|)$. Since $\d$ was arbitrarily chosen, this proves the smoothness of ${\mathcal F}_1$ on ${\mathcal O}$. Finally, letting $v_p\equiv \l$ in \eqref{}, we obtain \eqref{}. \hfill {$\square$}\goodbreak \medskip \end{proof} We now study the spectral properties of $D{\mathcal F}(\l)$. \begin{Lemma}\label{} For every $\l>0$ we define the linear operator \begin{equation} \label{} L_\l:=D{\mathcal F}(\l) : C^{2,\a}_{p,{\mathcal R}}\to C^{0,\a}_{p,{\mathcal R}}. \end{equation} Then the functions $e_k\in C^{2,\a}_{p,{\mathcal R}}$, given by \begin{equation} \label{} e_k(t):=\frac{1}{\pi}\prod_{i=1}^2\cos(k_i t_i), \end{equation} for $k\in \mathbb{N}^2$, are the eigenfunctions of $L_\l$ and moreover \begin{equation} \label{} L_\l(e_k)=\s_{\l,\g}(|k|) e_k, \end{equation} where $|k|=\sqrt{k_1^2+k_2^2}$ and for every $\ell\in \mathbb{N}$, \begin{align}\label{} \s_{\l,\g}(\ell) &= \ell^2 - 2\pi \g \left\{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+\ell^2 }} - \frac{\exp(-2\l\sqrt{\k^2+\ell^2 })}{ \sqrt{\k^2 +\ell^2 }} - \frac{\exp(-2\l\k )}{ \k} \right\} . \end{align} % In particular, \begin{equation}\label{} \partial_\l \s_{\l,\g}(1)<0 \qquad\textrm{ for every $\g,\l>0$ } \end{equation} and \begin{equation}\label{} \lim_{\ell\to \infty} \frac{ \s_{\l,\g}(\ell)}{\ell^2}= 1. \end{equation} \end{Lemma} \begin{proof} By direct computations using Fourier decomposition, we have \begin{align}\label{} \s_{\l,\g}(|k|) &= |k|^2 - \g \int_{\mathbb{R}^2} \left(1- \prod_{i=1}^2\cos(k_ir_i) \right) G_\k\left(|r| \right)dr \nonumber\\ % &\quad + \g \int_{\mathbb{R}^2}\left(1+\prod_{i=1}^2 \cos(k_i r_i)\right) G_\k\left((|r|^2+4 \l^2 )^{1/2} \right) dr . \end{align} % By \eqref{}, we have $$ \int_{\mathbb{R}}\cos(k_2 r_2) \int_{\mathbb{R}} \cos(k_1 r_1) G_\k\left((r^2_1+r_2^2)^{1/2} \right)dr_1 dr_2= 2\int_{\mathbb{R}}\cos(k_2 r_2) K_0(r_2\sqrt{\k^2+k^2_1})dr_2 $$ and by \eqref{} $$ \int_{\mathbb{R}}\cos(k_2 r_2) K_0(r_2\sqrt{\k^2+k^2_1})dr_2= \frac{2\pi}{2\sqrt{\k^2+|k|^2 }}. $$ We then deduce that $$ \int_{\mathbb{R}}\cos(k_2 r_2) \int_{\mathbb{R}} \cos(k r_1) G_\k\left((r^2_1+r_2^2)^{1/2} \right)dr_1 dr_2= \frac{2\pi}{\sqrt{\k^2+|k|^2 }}. $$ Moreover, taking $k=0$ in the above expression, we have $$ \int_{\mathbb{R}^2} G_\k\left(|r|\right)dr=\frac{2\pi}{\k}. $$ From the two qualities above, we deduce that \begin{equation}\label{} \int_{\mathbb{R}^2}\left(1- \prod_{i=1}^2 \cos(k_ir_i)\right) G_\k\left(|r| \right)dr=\frac{2\pi}{\k}- \frac{2\pi}{\sqrt{\k^2+|k|^2 }}. \end{equation} By \eqref{} and \eqref{}, we get \begin{equation}\label{} \int_{\mathbb{R}^2}\prod_{i=1}^2 \cos(k_i r_i) G_\k\left((|r|^2+4 \l^2 )^{1/2} \right) dr= 2 \pi\frac{\exp(-2\l\sqrt{|k|^2+\k^2})}{\sqrt{|k|^2+\k^2}} \end{equation} and letting $k=0$, we have $$ \int_{\mathbb{R}^2} G_\k\left((|r|^2+4 \l^2 )^{1/2} \right) dr = 2 \pi\frac{\exp(-2\l \k) }{ \k}. $$ Using this with \eqref{} and \eqref {} in \eqref{}, we get the desired expression of $\s_{\l,\g}(|k|)$ in \eqref{}. \hfill {$\square$}\goodbreak \medskip \end{proof} Next, we give some qualitative properties of $\s_{\l,\g} $ . \begin{Lemma}\label{} Let $\k>0$ and let $\g$ be such that \begin{equation} \label{} \frac{1}{2\pi} \frac{1}{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+1 }} }< \g < \frac{1}{2\pi} \frac{ 1}{ \frac{1}{(\k^2+1 )^{3/2}} }. \end{equation} Then there exists a unique $\l_*=\l_*(\g,\k)>0$ such that \begin{equation} \label{} \s_{\l_*,\g}(1)=0 , \end{equation} \begin{equation} \label{} \s_{\l_*,\g}(0)=\frac{4\pi\g \exp(-2\l_* \k)}{\k}>0 \end{equation} and \begin{equation} \label{} \s_{\l_*,\g}(\ell+1)> \s_{\l_*,\g}(\ell) \qquad \textrm{ for every $\ell\geq 1 $}. \end{equation} % \end{Lemma} \begin{proof} We first note that, by elementary calculations, the strict inequality $$ \frac{1}{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+1 }} }< \frac{ 1}{ \frac{1}{(\k^2+1 )^{3/2}} } $$ is equivalent to $$ (\k^2+1)^2-(\k^4+2\k^2)>0, $$ which holds true for every $\k>0$. Therefore we can always pick $\g$ as in \eqref{ In view of \eqref{}, we have $$ \s_{\l,\g}(1) = 1 - 2\pi \g \left\{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+1 }} - \frac{\exp(-2\l\sqrt{\k^2+1}) }{ \sqrt{\k^2+1}} - \frac{\exp(-2\l\k) }{ \k}\right\} . $$ It is easy to see that $$ \lim_{\l\to 0}\s_{\l,\g}(1)= 1+ \frac{4\pi \g }{\sqrt{\k^2+1 }}>0. $$ Now if $\g$ satisfies the lower bound in \eqref{}, we have that $$ \lim_{\l\to +\infty}\s_{\l,\g}(1)= 1 - 2\pi \g \left\{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+1 }} \right\} <0. $$ Therefore \eqref{} follows, for a unique $\l_*=\l_*(\g,\k)$ thanks to \eqref{}. We now prove \eqref{}, where the assumption on the upper bound in \eqref{} will be used. Indeed, we consider the smooth function $g_{\l,\g}:[1,\infty)\to \mathbb{R}$, given by $$ g_{\l,\g}(x)= x^2 - 2\pi \g \left\{ \frac{1}{\k}- \frac{1}{\sqrt{\k^2+x^2 }} - \frac{\exp(-2\l\sqrt{\k^2+x^2}) }{ \sqrt{\k^2+x^2}} - \frac{\exp(-2\l\k) }{ \k}\right\} . $$ Then $$ g'_{\l,\g}(x)=2x\left (1- \frac{\pi \g }{(\k^2+x^2 )^{3/2}} - \frac{\pi \g \exp(-2\l\sqrt{\k^2+x^2})}{(\k^2+x^2 )^{3/2}}- \frac{2\pi \g \l \exp(-2\l\sqrt{\k^2+x^2})}{\k^2+x^2 } \right). $$ We want $g'_{\l,\g}(x)>0$ for every $\l>0$ and $x\geq 1$, which holds true provided $$ \pi \g <\frac{ 1}{\sup_{\stackrel{\l>0}{ x\geq 1}} \left(\frac{1 }{(\k^2+x^2 )^{3/2}} + \frac{ \exp(-2\l\sqrt{\k^2+x^2})}{(\k^2+x^2 )^{3/2}}+ \frac{ 2\l \exp(-2\l\sqrt{\k^2+x^2})}{\k^2+x^2 } \right)}. $$ But since the (maximized) function in the denominator is decreasing in $x$ and $\l$, this is equivalent to $$ \pi \g <\frac{ 1}{ \frac{2}{(\k^2+1 )^{3/2}} } $$ which is our assumed upper bound in \eqref{}. It is now clear hat as long as $\g$ satisfies \eqref{}, we have that $g'_{\l,\g}(x)>0$ for every $\l>0$ and $x\geq 1$. This completes the proof of \eqref{} since $\s_{\l,\g}(\ell)=g_{\l,\g}(\ell)$ for every $\ell\geq 1$. \hfill {$\square$}\goodbreak \medskip \end{proof} We will consider the subspaces of Sobolev spaces which are invariant under the rotations $$ {\mathcal R}(2):=\left\{\left( \begin{array}{cc} 0&1\\ 1&0 \end{array}\right), \left( \begin{array}{cc} -1&0\\ 0&1 \end{array}\right) \right\}. $$ In the following, for $j\in \mathbb{N}$, we define \begin{equation*} \label{} H^j_{p,{\mathcal R}} := \Bigl \{v \in H^{j}_{loc}(\mathbb{R}^2) \::\: \text{$v$ is $2\pi\mathbb{Z}^2$-periodic and $v(M t)=v(t)$, $\,M\in{\mathcal R}(2),\, t\in \mathbb{R}^2$}\Bigl \} \end{equation*} and we put $L^2_{p,{\mathcal R}}:= H^{0}_{p,{\mathcal R}}$. Note that $L^2_{p,{\mathcal R}}$ is a Hilbert space with scalar product $$ (u,v) \mapsto \langle u,v \rangle_{L^2([-\pi,\pi]^2)} := \int_{[-\pi,\pi]^2} u(t)v(t)\,dt \qquad \text{for $u,v \in L^2_{p,{\mathcal R}}$.} $$ Then the functions $e_{k}$ in \eqref{} form an orthonormal basis for $L^2_{p,{\mathcal R}}$. Moreover, $H^j_{p,{\mathcal R}} \subset L^2_{p,{\mathcal R}}$ is characterized as the subspace of all functions $v \in L^2_{p,{\mathcal R}}$ such that $$ \sum_{k \in \mathbb{N} ^2} (1+|k|^2)^{j} \langle v, e_k \rangle_{L^2([-\pi,\pi]^2)} ^2 < \infty. $$ Therefore, $H^j_{p,{\mathcal R}}$ is also a Hilbert space with scalar product \begin{equation*} \label{} (u,v) \mapsto \langle u,v \rangle_{H^j_{p,{\mathcal R}}} := \sum_{k \in \mathbb{N}^2} (1+|k|^2)^{j} \langle u, e_k \rangle_{L^2([-\pi,\pi]^2)} \langle v, e_k \rangle_{L^2([-\pi,\pi]^2)} \qquad \text{for $u,v \in H^j_{p,{\mathcal R}}$.} \end{equation*} We consider the eigenscapces corresponding to $\s_{\l,\g}(\ell)$, given by \begin{equation} \label{} V_\ell:= \textrm{span}\{ e_k \;:\; |k|=\ell \} \end{equation} and $P_\ell: L^2_{p,{\mathcal R}}\to V_\ell $ be the $L^2$-orthogonal projection on $V_\ell$. We will denote by $V_\ell^\perp$ to $L^2$-complement of $V_\ell$, which is given by $$ V_\ell^\perp:=\{v\in V_\ell\,:\, P_\ell v=0\}. $$ Obviously, the direct-sum $ \bigoplus_{\ell\in\mathbb{N}}V_\ell$ provides a complete decomposition of $L^2_{p,{\mathcal R}}$. More precisely for any $v\in L^2_{p,{\mathcal R}}$, we can write it as $$ v=\sum_{\ell\in \mathbb{N}} P_\ell v. $$ Moreover, by \eqref{}, we have \begin{equation}\label{eq:-Lv-Fpurier-slab} L_\l(v) =\s_{\l,\g}(\ell) v \qquad \textrm{ for all $v\in V_\ell$}. \end{equation} It is easy to see that \begin{equation}\label{} V_1\cap C^{0,\a}_{p,{\mathcal R}} =\textrm{span}\{\overline{v}\}=\mathbb{R} \overline v\qquad \textrm{ with }\,\,\,\overline{v}(t_1,t_2)= \cos(t_1) + \cos(t_2). \end{equation} We also consider the nonlinear operator ${\mathcal F}$ defined in \eqref{}, and we note that ${\mathcal F}$ maps ${\mathcal O} \cap C^{2,\a}_{p,{\mathcal R}}$ into $C^{0,\a}_{p,{\mathcal R}}$. Consider the open set $$ {\mathcal U}:= \{(\lambda,u) \in \mathbb{R} \times C^{2,\a}_{p,{\mathcal R}} \::\: \lambda>0, \: u> - \lambda \} \subset \mathbb{R} \times C^{2,\a}_{p,{\mathcal R}}. $$ The proof of our main theorem will be completed by applying the Crandall-Rabinowitz bifurcation theorem to the smooth nonlinear operator \begin{equation} \label{} {\mathcal G}: {\mathcal U} \subset \mathbb{R} \times C^{2,\a}_{p,{\mathcal R}} \to C^{0,\a}_{p,{\mathcal R}}, \qquad {\mathcal G}(\lambda,u) = {\mathcal F}(\lambda+u)-{\mathcal F}(\l). \end{equation} We have the following \begin{Proposition}\label{propPhi} Let $\g$ satisfy \eqref{}. Then there exists a unique $\lambda_*=\l_*(\g,\k)>0$ such that the linear operator $ L_{\l_*} = D_u {\mathcal G}(\l_*,0) $ has the following properties. \begin{itemize} \item[(i)] The kernel $N(L_{\l_*} )$ of $L_{\l_*} $ is given by \begin{equation} \label{} N(L_{\l_*} )=V_1=\mathbb{R}\overline{v} , \qquad \overline{v}(t_1,t_2)= \cos(t_1) + \cos(t_2). \end{equation} \item[(ii)] The range $R(L_{\l_*} )$ of $L_{\l_*} $ is given by $$ R(L_{\l_*} )=C^{0,\a}_{p,{\mathcal R}}\cap V_1^\perp . $$ \item[(iii)]Finally, \begin{equation} \label{} \partial_\lambda \Bigl|_{\lambda= \lambda_*} L_\lambda\overline{v}=\partial_\lambda \Bigl|_{\lambda= \lambda_*} \s_{\l,\g}(1)\overline{v} \;\not \in \; R(L_{\l_*}). \end{equation} \end{itemize} \end{Proposition} \noindent{{\bf Proof. }} For $\g$ satisfying \eqref{}, by Lemma~\ref{}, there exists a unique $\lambda_*=\l_*(\g,\k)>0$ such that $\sigma_{\lambda_*,\g}(1) =0$ and $\sigma_{\lambda_*,\g}(\ell) >0 $ for all $\ell\in \mathbb{N}\setminus\{1\}$. Hence by \eqref{eq:-Lv-Fpurier-slab}, we have that $L_{\l_*} v=0$ if and only if $v\in V_1$. This with \eqref{} prove (i).\\ We now prove (ii). We pick $f \in C^{0,\a}_{p,{\mathcal R}}\cap V_1^\perp\subset L^2_{p,{\mathcal R}}\cap V_1^\perp$. Then using Fourier decomposition, \eqref{eq:-Lv-Fpurier-slab}, \eqref{} and \eqref{}, we have a unique $w\in H^2_{p,{\mathcal R}}\cap V_1^\perp$, such that $$ L_{\l_*} w=f \qquad \textrm{ in $\mathbb{R}^2$}. $$ The proof of (ii) finishes once we show that $w\in C^{2,\a}_{loc}(\mathbb{R}^2)$. By Morrey's embedding, $w\in H^2_{loc}(\mathbb{R}^2)\subset C^{0,\a}_{loc}(\mathbb{R}^2)$, for every $\a\in (0, 1)$. We then have \begin{equation} \label{} -\D w=\g w(t)\int_{\mathbb{R}^2} G_\k\left(|r| \right)dr- \g w(t) \int_{\mathbb{R}^2} G_\k\left((|r|^2+4 \l^2 )^{1/2} \right) dr- F(t), \end{equation} where % $$ F(t):=\int_{\mathbb{R}^2} w(t-r) G_\k\left(|r| \right)dr+\int_{\mathbb{R}^2} w(t-r) G_\k\left((|r|^2+4 \l^2_* )^{1/2} \right) dr . $$ Then for every $t\in \mathbb{R}^2$ \begin{align*} |F(t)|&\leq \|w\|_{C^{0,\a}(\mathbb{R}^2)}\left( \int_{\mathbb{R}^2} G_\k\left(|r| \right)dr+\int_{\mathbb{R}^2} G_\k\left((|r|^2+4 \l^2_* )^{1/2} \right) dr \right) \\ &\leq C \|w\|_{C^{0,\a}(\mathbb{R}^2)}. \end{align*} It is also easy to see that for $t, \overline t\in \mathbb{R}^2$ \begin{align*} |F(t)-F(\overline t)|& \leq C\|w\|_{C^{0,\a}(\mathbb{R}^2)} |t-\overline t|^\a . \end{align*} We then obtain that $F\in C^{0,\a}(\mathbb{R}^2)$, for every $\a\in (0,1)$. Now by \eqref{} and standard elliptic regularity theory, $w\in C^{2,\a}_{loc}(\mathbb{R}^2)$ for every $\a\in (0,1)$. Hence (ii) follows.\\ It remains to prove~\eqref{}, which follows from \eqref{} and the identity \begin{equation}\label{Dlambda} \partial_\lambda \Bigl|_{\lambda= \lambda_*} L_\lambda \overline{v} = \partial_\lambda \Bigl|_{\lambda= \lambda_*}\sigma_{\l,\g}(1)\overline{v} . \end{equation} \hfill {$\square$}\goodbreak \medskip \subsection{Proof of Theorem \ref{}} Thanks to \eqref{propPhi}, we can apply the Crandall-Rabinowitz Theorem (see \cite[Theorem 1.7]{M.CR}), to have \begin{Proposition}\label{} There exists ${s_0}>0$, only depending on $ \k$ and $\g$, and a smooth curve \begin{equation} \label{} (-{s_0},{s_0}) \to (0,\infty)\times C^{2,\a}_{p,{\mathcal R}}, \qquad s \mapsto (\lambda_s,u_s) \end{equation} such that \begin{enumerate} \item[(i)] For every $s \in (-{s_0},{s_0})$ \begin{equation}\label{} {\mathcal G}(\lambda_s, u_s)= {\mathcal F}(\l_s+ u_s)-{\mathcal F}(\l_s)=0 . \end{equation} \item[(ii)] $\lambda_0= \lambda_*$. \item[(iii)] $u_s = s \bigl( \overline{v} + v_s\bigr)$ for all $s \in (-s_0,s_0)$, with a smooth curve $$ (-{s_0},{s_0}) \to C^{2,\a}_{p,{\mathcal R}}, \qquad s \mapsto v_s $$ satisfying $v_0=0$ and $\int_{[-\pi, \pi]^2}v_s(t)\overline{v}(t)dt=0$. \end{enumerate} \end{Proposition} \section{Periodic lamellae and proof of Theorem \ref{}} % For $\e\in \mathbb{R}\setminus\{0\}$, with $|\e|$ small, we define $$ {\mathcal L}_{\psi}^\e=\O_{\psi}+\frac{1}{|\e|} \mathbb{Z} e_3=\bigcup_{p\in \mathbb{Z}}\left( \O_{\psi}+\left(0,0,\frac{p}{|\e|}\right)\right), $$ where as above $\psi\in C^{2,\a}_{p,{\mathcal R}}\cap {\mathcal O}$. For $x\in \{-1,1\} $, the mean curvature takes the form \begin{equation} H_{\partial {\mathcal L}_{\psi}^\e}((t, \psi(t)x) )= -\text{div} \frac{\n \psi(t)}{\sqrt{1+|\n \psi(t)|^2}} , \end{equation} and by a change of variable, the nonlocal part takes the form \begin{align*} \int_{{\mathcal L}_{\psi}^\e }&G_\k(|(t, \psi(t)x)-\xi|) d\xi = \int_{\mathbb{R}^2} \int_{-1}^1\psi(s) G_\k(|(t, \psi(t)x)-(s,\psi(s)y) |) dy ds \\ &\quad + \sum_{p\in \mathbb{Z} _*} \int_{\mathbb{R}^2} \int_{-1}^1\psi(s) G_\k(|(t, \psi(t)x)-(s,\psi(s)y)-(0,0,{|\e|}^{-1} {p})|) dy ds. \end{align*} Now using the fact that $ \mathbb{Z} _*=-\mathbb{Z} _*$ and the change of variable $y$ to $-y$, we see that $$ \int_{{\mathcal L}_{\psi}^\e }G_\k(|(t,- \psi(t))-\xi|) d\xi=\int_{{\mathcal L}_{\psi}^\e }G_\k(|(t, \psi(t))-\xi|) d\xi. $$ This implies, in particular, that for every $x\in \{-1,1\}$, $t\in \mathbb{R}^2$ and $|\e|<\frac{1}{2\l}$, $$ H_{\partial {\mathcal L}_{\psi}^\e}((t, \l x) )= \g \l \sum_{p\in \mathbb{Z} } \int_{\mathbb{R}^2} \int_{-1}^1 G_\k\left(( |r|^2+ (\l(1-y)-{|\e|}^{-1} {p} )^2 )^{1/2} \right) dy dr. $$ Namely ${\mathcal L}_{\l}^\e$ is an equilibrium pattern for every $\l>0$, with $|\e|<\frac{1}{2\l}$. Recalling \eqref{}, we get \begin{align}\label{} & {\mathcal H}_{ {\mathcal L}_{\psi}^\e}((t, \psi(t)x) )=H_{\partial {\mathcal L}_{\psi}^\e}((t, \psi(t)) )+ \int_{{\mathcal L}_{\psi}^\e }G_\k(|(t, \psi(t))-\xi |) d\xi \nonumber \\ % & ={\mathcal F}(\psi)(t) + \sum_{p\in \mathbb{Z} _*} \int_{\mathbb{R}^2} \int_{-1}^1\psi(s) G_\k\left(\left|(t, \psi(t))-(s,\psi(s)y)-(0,0,{|\e|}^{-1} {p})\right|\right) dy ds, \end{align} where ${\mathcal F}: {\mathcal O}\to C^{0,\a}(\mathbb{R}^2)$ is as in the previous section, is given by \begin{align}\label{} &{\mathcal F}(\psi)(t):= -\text{div} \frac{\n \psi(t)}{\sqrt{1+|\n \psi(t)|^2}} +\g \int_{\mathbb{R}^2}\int_{-1 }^1\psi(s)G_\k(|(t, \psi(t) )-(s,\psi(s)y)|) dyds . \end{align} We now define $\widehat {\mathcal F}:\mathbb{R}\setminus\{0\}\times {\mathcal O}\to C^{0,\a}(\mathbb{R}^2)$ by \begin{align}\label{} \widehat {\mathcal F}(\e,\psi )(t):=\sum_{p\in \mathbb{Z} _*} \int_{\mathbb{R}^2} \int_{-1}^1\psi(s) G_\k\left(\left|(t, \psi(t))-(s,\psi(s)y)-(0,0,{|\e|}^{-1} {p} )\right|\right) dy ds . \end{align} By similar arguments as in the proof of Lemma \ref{}, we can deduce that \begin{align*} \label{} &D \widehat {\mathcal F}(\e,\psi ) [w](t) \\ &= \sum_{p\in \mathbb{Z}_*} \int_{\mathbb{R}^2} (w(t-r)-w(t)) G_\k((|r|^2+ ( \psi(t) - \psi(t-r)+|\e|^{-1}p)^2)^{1/2}) dr\nonumber\\ & + \sum_{p\in \mathbb{Z}_*} \int_{\mathbb{R}^2} (w(t-r) +w(t))G_\k((|r|^2+ ( \psi(t) + \psi(t-r)+|\e|^{-1}p)^2)^{1/2}) dr . \end{align*} Let ${\mathcal U}_\l$ be a neighborhood of $\l>0$ in ${\mathcal O}$. Then there exists $\overline{\e},c>0$, only depending on ${\mathcal U}_\l$ and $\k$, such that for $|\e|\in (0, \overline{\e})$ and $\psi\in {\mathcal U}_\l$, we have \begin{align*} G_\k((|r|^2+ ( v_p(t) - v_p(t-r)+|\e|^{-1}p)^2)^{1/2}) &\leq G_\k\left( \left( |r|^2+c^2 {|\e|}^{-2} {p^2} \right)^{1/2} \right). % \end{align*} Therefore by a change of variable and \eqref{}, if $|\e|\in (0, \overline{\e})$ and $\psi\in {\mathcal U}_\l$, then \begin{align*} \| D \widehat{\mathcal F}(\e, v_p)[w]\|_{C^{0,\a}(\mathbb{R}^2)} &\leq 4\|w\|_{C^{0,\a}(\mathbb{R}^2) }\sum_{p\in \mathbb{Z}_*} \int_{\mathbb{R}^2}G_\k\left( \left( |r|^2+c^2 {|\e|}^{-2} {p^2} \right)^{1/2} \right) dr\\ % &\leq 4 c\|w\|_{C^{0,\a}(\mathbb{R}^2) } \sum_{p\in \mathbb{Z}_*} |\e|^{-1}|p| \int_{\mathbb{R}^2}G_{\k c\frac{|p|}{ |\e|}}\left( \left( |t|^2+1 \right)^{1/2} \right) dt \\ &\leq 4 c\|w\|_{C^{0,\a}(\mathbb{R}^2) } |\e| \sum_{p\in \mathbb{Z}_*} |p|^{-1} \exp\left(-\k c {|p|}{ |\e|^{-1}} \right) \\ &\leq 4 c\|w\|_{C^{0,\a}(\mathbb{R}^2) }|\e| \exp\left(-\frac{\k}{2} c { |\e|^{-1}} \right) . \end{align*} Up to decreasing $\overline{\e}$ and choosing a smaller neighborhood of $\l$, we can easily see that $\widehat {\mathcal F}$ extends to a smooth function $(-\overline{\e},\overline{\e})\times {\mathcal U}_\l \to C^{0,\a}(\mathbb{R}^2)$ with $D^k\widehat {\mathcal F}(0,\psi )=0$ for all $\psi\in {\mathcal U}_\l$ and $k\in \mathbb{N}$. In view of \eqref{} and \eqref{}, we define $${\mathcal N}_\l: (-\overline{\e},\overline{\e}) \times {\mathcal U}_\l\cap C^{2,\a}_{p,{\mathcal R}}\to C^{0,\a}_{p,{\mathcal R}}$$ by \begin{equation} \label{} (\e,\psi)\mapsto {\mathcal N}_\l(\e,\psi)(t):= {\mathcal F}(\psi)(t)+\g\widehat{\mathcal F}(\e,\psi)(t)={\mathcal H}_{ {\mathcal L}_{\l }^\e}((t, \psi (t)) ), \end{equation} In the following, we let $\g$ satisfy \eqref{}, Instead of applying the Crandall-Rabinowitz Theorem \cite[Theorem 1.7]{M.CR}, we will use the same argument of its proof, mainly based on the Implicit Function Theorem. More precisely, we apply the latter theorem to solve the equation ${\mathcal M}=0$, with the function $$ {\mathcal M}: (-\overline{\e},\overline{\e})\times (- \l_* /8, \l_* /8) \times (1/2,3/2) \times C^{2,\a}_{p,{\mathcal R}}\to C^{0,\a}_{p,{\mathcal R}}, $$ given by \begin{align}\label{} {\mathcal M}(\e,s,\d, v)(t)&:= \frac{{\mathcal N}_{\l_*}(\e, \d \l_*+ s (\overline{v}+v))(t)-{\mathcal N}_{\l_*}(\e, \d \l_* ) }{s}\\ % &=\int_0^1D_\psi{\mathcal N}_{ \l_*}(\e, \d\l_*+\varrho s(\overline{v}+v)) )[\overline{v}+v ](t)d\varrho. \nonumber \end{align} It is clear that ${\mathcal M}$ is smooth in $(-\overline{\e},\overline{\e})\times (- \l_* /8, \l_* /8) \times (1/2,3/2) \times B_{C^{2,\a}_{p,{\mathcal R}}}(0, 1 /8)$. Now from Proposition \ref{propPhi}(i), we have that ${\mathcal M}(0,0,1,0)= L_{\l_*} (\overline{v})=0$. We also have that $$ D_v{\mathcal M}(0,0,1,0) = L_{\l_*} : C^{2,\a}_{p,{\mathcal R}}\cap V_1^\perp\to C^{0,\a}_{p,{\mathcal R}}\cap V_1^\perp$$ is an isomorphism by Proposition \ref{propPhi}(ii). Moreover, by Proposition \ref{propPhi}(iii), $$ D_\d {\mathcal M}(0,0, 1,0): \mathbb{R} \to V_1$$ is an isomorphism. From these, we can now apply the implicit function theorem to get $s_0,\e_0 >0$ and a smooth curve $$ (-\e_0,\e_0)\times (-s_0,s_0)\to (1/2,3/2) \times C^{2,\a}_{p,{\mathcal R}}\cap V_1^\perp,\qquad ( \e,s)\mapsto (\d_{\e,s}, v_{\e,s}) $$ with $ \d_{0,0}=1$ and $v_{0,0}=0$ such that $$ {\mathcal M}(\e, s, \d_{\e,s},v_{\e,s})=0 \qquad \textrm{ on $C^{0,\a}_{p,{\mathcal R}}$}. $$ From this together with \eqref{} and \eqref{}, we conclude that, for every $t\in \mathbb{R}^2$, \begin{align} \label{} {\mathcal F}(\l_{\e,s}+ s (\overline{v}+v_{\e,s}))(t)+\g\widehat{\mathcal F}(\e,\l_{\e,s}+ s (\overline{v}+v_{\e,s}))(t)&= {\mathcal F}(\l_{\e,s})+\g\widehat{\mathcal F}(\e, \l_{\e,s}) % \end{align} where we have set $\l_{\e,s}:= \d_{\e,s}\l_*$. This ends the first part of the proof of Theorem \ref{}. It follows from the uniqueness, inherent to the construction, that $\l_{0,s}=\l_s $ and $v_{0,s}= v_s $, for all $s\in (-s_0,s_0)$, where $\l_s$ and $v_s$ are given by Theorem \ref{}. \hfill {$\square$}\goodbreak \medskip \subsubsection{Proof of Theorem \ref{} (completed)} The existence of the smooth curve of equilibrium patterns ${\mathcal L}_{\psi_{\e,s}}^\e$ follows from the identity \eqref{}. Obviously $\psi_{0,s}=v_p_s$ as a consequence of the uniqueness form the Crandall-Rabinowitz theorem. \hfill {$\square$}\goodbreak \medskip \section{Periodic lattice of near-cylinders and near-spheres }\label{s:latt-Cyl-Sph} \subsection{Some preliminary and notations}\label{} Let $M\in\mathbb{N}$ with $1\leq M\leq N$. Let $\left\{ \textbf{a}_1;\dots;\textbf{a}_M\right\}$ be a basis of $\mathbb{R}^M$. We recall from the first section the $M$-dimensional {Bravais lattice} in $\mathbb{R}^N$: \begin{equation} \scrL^M=\left\{\sum_{i=1}^N k_i \textbf{a}_i\,:\, k=(k_1,\dots,k_M)\in \mathbb{Z}^N\right\}\subset \mathbb{R}^N. \end{equation} Throughout this paper, we will put $\scrL_*^M:= \scrL^M \setminus \{0\}$ and $\mathbb{Z}^M_*:= \mathbb{Z}^M \setminus \{0\}$. We let $S^{N-1}$ be the unit sphere on $\mathbb{R}^N$. We denote by ${\mathcal S}_k$ the finite dimensional subspace of spherical harmonics of degree $k$ on $S^{N-1}$, and by $(Y^i_{k})_{k\in\mathbb{N},\, i=1,\dots, n_k}$ an orthonormal basis of ${\mathcal S}_k$. It is well known that $n_0=1$ with $Y^1_0(\th)=\frac{1}{\sqrt{|S^{N-1}|}}$ and $n_1=N$ with $Y^i_1(\th)=\frac{\th_i}{\sqrt{|B^N|}}$, for each $i=1,\dots,N$. We recall that $-\D_{S^{N-1}} Y = k(k+N-2) Y $, for every $Y\in {\mathcal S}_k$. Moreover for a function $f: (0,2)\to \mathbb{R}_+$, with $\int_{0}^2f(t)(t(2-t))^{\frac{N-3}{2}}dt<\infty$, by the Funck-Hecke formula, see\cite[page 247]{El}, for $k\geq1$ and $\th\in S^{N-1}$, we have $$ \int_{S^{N-1}}(Y(\th)-Y(\s))f(|\th-\s|)\, d\s=\mu_k Y(\th), \qquad\textrm{ for every $Y\in {\mathcal S}_k$}, $$ for some real number $\mu_k\geq0$. From this, we deduce that $$ \mu_k=\frac{1}{2}\int_{S^{N-1}}\int_{S^{N-1}}(Y^i_{k}(\th)-Y^i_{k}(\s))^2f(|\th-\s|)\, d\s d\th \qquad \textrm{ for every $i=1,\dots, n_k$}. $$ For a function $w\in L^2(S^{N-1} )$, we can decompose it as $$ w=\sum_{k\in \mathbb{N}}\sum_{i=1}^{n_k} w_k^i Y_k^i,\qquad \textrm{ with } \qquad w_k^i=\int_{S^{N-1}}w(\s) Y_k^i(\s)\, d\s. $$ For $j\in \mathbb{N}$, we define the Sobolev spaces \begin{equation} H^j(S^{N-1} )=\left\{w\in L^2(S^{N-1})\,:\, \sum_{k\in \mathbb{N}}\sum_{i=1}^{n_k}(1+|k|^2)^{{j}} (w_k^i )^2 <\infty \right\} \end{equation} and \begin{equation} H^j_{\textbf{e}}(S^{N-1} )=\{w\in H^j(S^{N-1} )\,:\, \textrm{ $w$ is even} \}. \end{equation} We define also H\"older spaces $$ C^{j,\alpha}_{\textbf{e}}( S^{N-1}):= \biggl\{ \phi \in C^{j,\alpha}( S^{N-1}), \quad \textrm{$\phi(-\th)=\phi(\th)$ }\biggl\} $$ endowed with usual H\"older norm of $C^{j,\alpha}( S^{N-1})$. Following \cite{Cabre2015B}, for a fixed $e \in S^{N-1} $ and a Lipschitz continuous map of rotations $S^{N-1} \mapsto SO(N)$, $\th \mapsto R_\theta$ with the property that \begin{equation} \label{} S_e:= \{\th \in S^{N-1}\::\: \th \cdot e \ge 0\} \subset \{\th \in S^{N-1} \::\: R_{\th}e = \th \}. \end{equation} \begin{Example}[\cite{Cabre2015B}] \label{example-rotations} Consider the map $\th \mapsto R_\theta$ defined as follows. For $\theta \in S^{N-1} $ with $\th \cdot e \ge 0$, we let $R_\theta$ be the rotation of the angle $\arccos \th \cdot e$ which maps $e$ to $\th$ and keeps all vectors perpendicular to $\th$ and $e$ fixed. We then extend the map $\th \mapsto R_\theta$ to all of $S^{N-1} $ as an even map with respect to reflection at the hyperplane $\{\th \in \mathbb{R}^N \::\: \th \cdot e = 0\}$. \end{Example} It is easy to see that \begin{equation}\label{} |R_{\th}\s-\th |=|e-\s| \qquad \textrm{ for all $\th\in S_e$ and $\s\in S^{N-1} $}. \end{equation} Moreover, the Lipschitz property of the map $\th \mapsto R_\th$ implies that there is a constant $C>0$ with \begin{equation} \label{} \|R_{\th_1} - R_{\th_2}\| \le C |\th_1-\th_2| \qquad \text{for all $\th_1,\th_2 \in S^{N-1} $,} \end{equation} where, here and in the following, $\|\cdot\|$ denotes the usual operator norm. We put % $${\mathcal D}:= \{\phi \in C^{2,\a}(S^{N-1}):\: \phi>0\} .$$ For every $v_p\in {\mathcal D}$, we let \begin{equation}\label{} F_v_p : \mathbb{R}_+ \times S^{N-1} \to \mathbb{R}^N,\qquad F_v_p(r ,\s):= r \s v_p (\s ). \end{equation} The map $F_v_p$ defines a parameterization of the set \begin{equation} B_v_p^N:= F_{v_p}([0,1) \times S^{N-1}). \end{equation} We will denote $S^{N-1}_v_p:= \partial B_v_p^N= F(1,S^{N-1} )$. See e.g. \cite{Cabre2015B}, the unit outer normal of $B_v_p^N$ at a point $F_v_p(1,\sigma)$, $\sigma \in S^{N-1}$, is given by \begin{equation} \label{} \nu_v_p(F_v_p(1,\s)) = \frac{\phi(\s)\s -\n v_p(\s)}{\sqrt{\phi^2(\s) + |\nv_p(\s)|^2}} \end{equation} and for every continuous function $f$ on $\mathbb{R}^N$, we have \begin{equation} \label{} \int_{S^{N-1}_v_p} f(y) \,dV(y) = \int_{S^{N-1}} [f\circ F_v_p(1,\cdot )](\s) \G_v_p(\s) \, d\s \qquad \text{with}\quad \G_v_p= v_p^{N-2}\sqrt{ v_p^{2}+ |\n v_p|^2}. \end{equation} Moreover letting $$ {\mathcal A}(v_p)= \int_{S^{N-1}} \G_v_p(\s) \, d\s, $$ then differentiating ${\mathcal A}$, we get $$ D{\mathcal A}(v_p)[v]= \int_{S^{N-1}} {\mathcal H}(v_p)(\s)v(\s) \, d\s, $$ where $ {\mathcal H}(v_p)(\s)=H_{S^{N-1}_v_p }( v_p(\s)\s)$ is the mean curvature operator of $\partial B^N_v_p=S^{N-1}_\phi$ and is given by \begin{equation} \label{} {\mathcal H}(v_p) =- \text{div}\frac{ v_p^{N-2} \n v_p}{\sqrt{ v_p^{2}+ |\n v_p|^2}} +(N-2)\frac{v_p^{N-3} }{ \sqrt{ v_p^{2}+ |\n v_p |^2}}+\frac{v_p^{N-1}}{( v_p^{2} + |\n v_p |^2)^{-1/2}} . \end{equation} \subsubsection{From 2D near-balls to 3D near-cylinders}\label{} We define the perturbed cylinder \begin{equation} C_v_p:= \left\{( r,\th,t)\in \mathbb{R}_+\times S^{1}\times \mathbb{R} \,:\, r<v_p(\th) \right\}= B_v_p^2\times \mathbb{R}. \end{equation} We then consider the lattice of cylinders $$ {\mathcal C}_{v_p}^\e:=\bigcup_{p \in \scrL^M} \left(C_v_p+\frac{p}{|\e|}\right)= \bigcup_{p \in \scrL^M} \left(C_v_p+\left(\frac{p}{|\e|},0\right)\right) , $$ where $v_p: S \to (0,\infty)$ is an even and $\scrL^M$ is the $M$-dimensional Bravais lattice, with $M\leq 2$. \\ % We note that the integral of the Yukawa interaction $G_\k$ in the $t$ direction gives $2K_0$ which is the $2$-dimensional screened Coulomb potential. Our problem is therefore reduced to find spherical lattices of equilibrium pattern in $\mathbb{R}^2$. Indeed, for a point $Z_0=(v_p(\th)\th,t)\in \partial C_v_p$, the Yukawa interaction takes the form \begin{align*} \int_{ {\mathcal C}_{v_p}^\e}G_\k(|Z_0-Z|) dZ&= \sum_{p\in \scrL^M }\int_{ C_v_p} G_\k(|Z_0-Z-\frac{p}{|\e|} |)dZ\\ &=\int_{ {\mathcal C}_{v_p}^\e} G_\k(|Z_0-Z|)dZ+ \sum_{p\in \scrL ^M} \int_{ C_{v_p}} G_\k(|Z_0-Z-\frac{p}{|\e|} |)dZ\\ &= \int_{B_v_p^2} \int_{ \mathbb{R}} G_\k(|(v_p(\th)\th,t) -(y, s) -\frac{p}{|\e|}|) ds dy \\ &\quad+ \sum_{p\in \scrL^M } \int_{B_v_p^2} \int_{ \mathbb{R}} G_\k(|(v_p(\th)\th,t) -(y, s) |) ds dy . \end{align*} By \eqref{}, $$ \int_{\mathbb{R}} G_\k\left((r^2+\d^2)^{1/2} \right)dr= 2K_0( \k \d ), $$ where as before $K_0$ is the modified Bessel function of the second kind of order $0$. We then deduce that \begin{align*} \int_{ {\mathcal C}_{v_p}^\e}G_\k(|Z_0-Z|) dZ &=2 \int_{B_v_p^2} K_0(\k |(v_p(\th) \th)- y|) dy \\ &+2\sum_{p\in \scrL_* ^M} \int_{B_v_p^2} K_0\left(\k |(v_p(\th) \th)-y -\frac{p}{|\e|} | \right) dy. \end{align*} Since the mean curvature $H_{C_v_p}$ does not depend on $t$, from the computations above, ${\mathcal H}_{{\mathcal C}^\e_v_p}$ does not depend on $t$. We are therefore reduce to find near-ball lattices $B_v_p^2+\frac{1}{|\e|}\scrL^M$ in $\mathbb{R}^2$ with Yukawa potential (or screened Coulomb potential) $2K_0$. We will therefore treat simultaneously the 2D and 3D cases. For that, we define \begin{equation} \label{} G_{\k,N}(r):=\left\{\begin{array}{cc} 2K_0(\k r)&\qquad\textrm{ for $N=2$}\vspace{3mm}\\ G_\k(r)&\qquad\textrm{ for $N=3$}. \end{array} \right. \end{equation} \subsection{ Lattices of perturbed balls in 2D and 3D }\label{} For $|\e|$ small and $\phi\in {\mathcal D} $, we then define $$ {\mathcal B}_\phi^{\e} := \bigcup_{p \in \scrL^M} \Bigl( B_\phi^N + \frac{p}{|\e|}\Bigr)=B_{\phi}^N+\frac{1}{|\e|} \scrL^M. $$ We will consider in this section the Yukawa potential $G_{\k,N}$ defined in \eqref{}. For $Z_0=\th \phi(\th) \in \partial B_\phi^N$, we have \begin{equation}\label{} {\mathcal H}_{{\mathcal B}_\phi^{\e} } (Z_0)= H_{\partial B^N_\phi}(Z_0)+ \g \int_{{\mathcal B}_\phi^{\e}}G_{\k,N}(|Z_0-Z|) dZ. \end{equation} The Yukawa interaction takes the form \begin{align*} \int_{{\mathcal B}_\phi^{\e}}G_{\k,N}(|Z_0-Z|) dZ&= \sum_{p\in \scrL ^M}\int_{ B_\phi^N} G_{\k,N}(|Z_0-Z-\frac{p}{|\e|} |)dZ\\ &=\int_{ B_\phi^N} G_{\k,N}(|Z_0-Z|)dZ+ \sum_{p\in \scrL_* ^M} \int_{ B_\phi^N} G_{\k,N}(|Z_0-Z-\frac{p}{|\e|} |)dZ. \end{align*} We define the following maps \begin{equation} {\mathcal D}\to C^{0,\alpha}( S^{N-1}),\qquad {\mathcal F}_{0}(\phi)(\th):= \int_{ B_\phi^N} G_{\k,N}(|Z_0-Z|)dZ, \end{equation} \begin{equation} \mathbb{R}\setminus\{0\}\times {\mathcal D}\to C^{0,\alpha}( S^{N-1}), \qquad {\mathcal F}_{p}(\e, \phi)(\th):= \int_{ B_\phi^N} G_{\k,N}(|Z_0-Z-\frac{p}{|\e|} |)dZ \end{equation} and \begin{equation} {\mathcal F}: \mathbb{R}\setminus\{0\}\times {\mathcal D}\to C^{0,\alpha}( S), \qquad {\mathcal F} (\e, \phi)(\th):=\sum_{p\in \scrL_* ^M} \int_{ B_\phi^N} G_{\k,N}(|Z_0-Z-\frac{p}{|\e|} |)dZ. \end{equation} % % We observe that for any neighborhood ${\mathcal U}$ of $1$ in ${\mathcal D}$, there exist constants $C, \e_0>0$ (depending only on $\k$ and $\scrL^M$) such that for every $|\e|\in (0,\e_0)$ and any $ p\in \scrL_*^M$ $$ |{\mathcal F}_{p}(\e, \phi)(\th) | \leq C |p|^ {-1/2} |\e|^ {1/2} \exp\left(\frac{- \k |p|}{2|\e|} \right) ,\qquad\textrm{ for every $v_p\in {\mathcal U}$ and $\th\in S^{N-1}$} $$ and \begin{equation}\label{es:-cFp-eps} \|{\mathcal F}(\e, \phi) \|_{C^{0,\a}( S^{N-1})} \leq \exp(-c |\e|^{-1}) ,\qquad\textrm{ for every $v_p\in {\mathcal U}$ and $\th\in S^{N-1}$,} \end{equation} for some $c>0$ depending only on $\a,N,\k$ and $\scrL^M$.\\ It follows that ${\mathcal F}$ extends smoothly in a neighborhood of $(0,1)$ in $\mathbb{R}\times {\mathcal D}$. Now we define the mean curvature operator ${\mathcal H}: {\mathcal D}\to C^{0,\alpha}( S^{N-1})$ (computed in \eqref{}) by \begin{equation} {\mathcal H}(\phi)(\th)=H_{\partial B_\phi^N}(\phi(\th)). \end{equation} We define ${\mathcal Q}: \mathbb{R}\times {\mathcal D}\to C^{0,\a}( S^{N-1})$ by \begin{equation} \label{} {\mathcal Q}(\e,\phi):={\mathcal H}(\phi)+\g {\mathcal F}_{0}(\phi)+\g {\mathcal F}(\e, \phi) \end{equation} and we note that from \eqref{}, $$ {\mathcal H}_{{\mathcal B}_\phi^{\e} }(\th\phi(\th))={\mathcal Q}(\e,\phi)(\th). $$ Our aim is to apply the implicit function theorem to solve the equation \begin{equation} \label{ solve-sphere-lattice} {\mathcal Q}(\e,\phi) -{\mathcal Q}(0,1)=0. \end{equation} Let us then study the linearization of $D_\phi{\mathcal Q}(0,1)$. \begin{Lemma}\label{Regularity -cF-0-lattice-sph-cyl} The map ${\mathcal Q}: {\mathcal D}\to C^{0,\a}( S^{N-1})$ is smooth. Moreover for every $\l>0$ and $w\in C^{2,\a}( S^{N-1})$, we have \begin{align} D_\phi {\mathcal Q}(0,1)[w](\th )& =-\D_{S^{N-1}} w(\th)-(N-1)w(\th) - \g \int_{S^{N-1}}(w(\th)-w( \s)) G_{\k,N}\left( |\th -\s| \right) d\s \nonumber \\ &\quad + \g w(\th) \int_{S^{N-1}}(1-\th\cdot \s ) G_{\k,N}\left( |\th -\s| \right) d\s. \label{} % \end{align} \end{Lemma} The proof of this result is not immediate (at least for $N=2$). We therefore postpone its proof to Section \ref{} below. \begin{Lemma}\label{} We define the linear operator $$L:=D_u{\mathcal Q}(0,1):C^{2,\a}( S^{N-1}) \to C^{0,\a}( S^{N-1}). $$ Then the spherical harmonics $Y_k^i$ are the eigenfunctions of $L$ and moreover $$ L(Y_k^i(\th))=\s_{\g}(k) Y_k^i(\th), $$ with eigenvalues \begin{align} \s_{\g}(k) &= \l_k- \l_1 - \g ( \mu_k- {\mu}_1) , \end{align} where $\l_k=k(k+N-2)$ and, for $i=1,\dots, n_k$, $$ \mu_k =\frac{1}{2}\int_{S^{N-1}\times S^{N-1}}(Y_k^i(\th)-Y_k^i(\s))^2 G_{\k,N}( |\th-\s|) d\th d\s. $$ \end{Lemma} \begin{proof} Using Fourier decomposition in spherical harmonics, see Section \ref{}, we easily deduce that \begin{align*} \s_{\g}(k) = \l_k-(N-1) +\g \int_{S^{N-1}}(1-\th\cdot \s ) G_{\k,N}\left(|\th -\s| \right) d\s \\ % - \frac{\g}{2} \int_{S^{N-1}\times S^{N-1}}(Y_k(\th)-Y_k(\s))^2 G_{\k,N}\left( |\th-\s | \right) d\th d\s, \end{align*} for every $Y_k\in {\mathcal S}_k$. We have that \begin{align*} \int_{S^{N-1}}(1-\th\cdot \s ) G_{\k,N}\left(|\th -\s| \right) d\s&= \frac{1}{2}\sum_{i=1}^N \int_{S^{N-1}}(\th_i-\s_i )^2 G_{\k,N}\left(|\th -\s| \right) d\s\\ % &=\frac{ |B^N|}{2}\sum_{i=1}^N \int_{S^{N-1}}(Y^i_1(\th)-Y^i_1(\s) )^2 G_{\k,N}\left(|\th -\s| \right) d\s\\ % &=\frac{N |B^N|}{2} \int_{S^{N-1}}(Y^i_1(\th)-Y^i_1(\s) )^2 G_{\k,N}\left(|\th -\s| \right) d\s. \end{align*} Since the integral on the left hand side above does not depend on $\th$, integrating the above equality and using that $|S^{N-1}|=N|B^N|$, we get \begin{align*} \int_{S^{N-1}\times S^{N-1} }&(1-\th\cdot \s ) G_{\k,N}\left(|\th -\s| \right) d\s d\th\\ % &= \frac{ 1}{2} \int_{S^{N-1}\times S^{N-1} }(Y^i_1(\th)-Y^i_1(\s) )^2 G_{\k,N}\left(|\th -\s| \right) d\s d\th\\ % &= \mu_1. \end{align*} % % \hfill {$\square$}\goodbreak \medskip \end{proof} The following result will be useful in order to study the invertibility of the linearized operator on the space of even function. \begin{Lemma}\label{} We let \begin{equation} \label{} \g_N:=\min \left( \frac{\l_1}{ {\mu}_1}, \inf_{k\geq 2}\frac{\l_k-\l_1}{\mu_k-\mu_1} \right). \end{equation} Then for every $\g\in (0,\g_N)$ \begin{equation} \label{} \s_{\g}(k)> 0\qquad \textrm{ for all $k\geq 2$}, \qquad\textrm{ }\qquad \s_{\g}(0)<0 \end{equation} and \begin{equation}\label{} \lim_{k\to \infty} \frac{ \s_{ \g}(k)}{k^2}= 1. \end{equation} \end{Lemma} \begin{proof} We observe that, for $N=3$, $$ \mu_k\leq \mu_k^0:= \frac{1}{2}\int_{S^{N-1}\times S^{N-1}}(Y_k^i(\th)-Y_k^i(\s))^2 G_{0,N}( |\th-\s|) d\th d\s, $$ and $\mu_k^0$ is the eigenvalues of the well known \textit {hypersingular Riesz operator on the sphere} (\cite[Section 7]{FFMMM}). In this case, we have that $\lim_{k\to\infty}\frac{\l_k}{\mu_k^0}=\infty$, see e.g. \cite[Section 7]{FFMMM}.\\ For $N=2$, we have $Y_k^1(\cos(x),\sin(x))=\frac{\cos(k x)}{\pi^{1/2}}$ and $Y_k^2(\cos(x),\sin(x))=\frac{\cos(k x)}{\pi^{1/2}}$. Moreover $K_0(\k r)\leq C |\log(\k r)|$. Therefore, we can estimate \begin{align*} \mu_k^0&\leq C \int_{0}^{2\pi} (\cos(k x)-\cos(k y))^2|\log(|\k \sqrt{1-\cos(x-y)|})|dxdy\\ &\leq C \int_{0}^{\pi} \int_{0}^{\pi} |\log(\k \sqrt{2}|\sin(x-y)|)|dxdy, \end{align*} with $C$ independent on $k$. Since $\mu_i\neq\mu_j$ for $i\neq j,$ we can therefore define $$ \g_N:=\min \left( \frac{\l_1}{\mu_1}, \inf_{k\geq 2}\frac{\l_k-\l_1}{\mu_k-\mu_1} \right). $$ It follows that for every $\g\in (0, \g_N)$, we have $\s_\g(0)<0$ and $\s_\g(k)>0$ for every $k\geq 2$. \hfill {$\square$}\goodbreak \medskip \end{proof} We then have the following result. \begin{Lemma}\label{} Let $\g\in (0, \g_N)$. Then the linear map $${\mathcal L}_{\textbf{e}}:=L\big|_{C^{2,\a}_{\textbf{e}}(S^{N-1})}: C^{2,\a}_{\textbf{e}}(S^{N-1}) \to C^{0,\a}_{\textbf{e}}(S^{N-1}) $$ is an isomorphism. \end{Lemma} \begin{proof} By Lemma \ref{} the elements of the Kernel of $L$ correspond with the eigenvalue $\s_\g(1)$ which is the eigenvalues corresponding to the first non-constant spherical harmonics $Y_1^1$ and $Y_1^2$ which are odd. This implies that the kernel of ${\mathcal L}_{\textbf{e}}$ is $ \{0\}$. By Lemma \ref{}, for every $f\in Y\in L^2_{\textbf{e}}(S^{N-1})$, there exists a unique $w\in H^2_{\textbf{e}}(S^{N-1})$ such that $L(w)=f$. Since $2\leq N\leq 3$, by Morrey's embedding theorem, $w\in C^{0,\a}(S^{N-1})$ for every $\a\in (0,1)$.\\ We define $$ F(\th)= \int_{S^{N-1}}(w(\th)-w(\s)) G_{\k,N}( |\th -\s| ) d\s $$ so that \begin{align}\label{} L(w) =-\D_{S^{N-1}} w(\th) - \l_1 w(\th ) +2\g \mu_1 w(\th ) - 2 \g F(\th)= f . \end{align} Then for every $\th\in S^{N-1}$, \begin{equation} \label{} |F(\th)|\leq C \|w\|_{C^{0,\a}(S^{N-1})} \int_{S}|\th -\s|^\a G_{\k,N}( |\th -\s| ) d\s\leq C \|w\|_{C^{0,\a}(S^{N-1})} . \end{equation} We fix $e \in S^{N-1}$ and $R_\th$ be given by Example \ref{example-rotations}. Thanks to \eqref{}, a change of variable gives $$ F(\th)= \int_{S^{N-1}}(w(\th)-w(R_\th \s)) G_{\k,N}( |e -\s| ) d\s. $$ By \eqref{} and \eqref{}, for all $\th_1,\th_2\in S_e$ and $\s\in S^{N-1}$, we have $$ |(w(\th_1)-w(R_{\th_1} \s))-(w(\th_2)-w(R_{\th_2} \s)) |\leq C \|w\|_{C^{0,\a}(S)} \int_{S} \min( |e -\s|^\a, |\th_1-\th_2|^\a ). $$ We assume that $|\th_1-\th_2 |\leq 1/4$. Then by \eqref{}, we have \begin{align*} &|F(\th_1)-F(\th_2)|\leq C \|w\|_{C^{0,\a}(S^{N-1})}\int_{S^{N-1}} \min( |e -\s|^\a, |\th_1-\th_2|^\a ) |e -\s|^{-1} d\s\\ &\leq C \|w\|_{C^{0,\a}(S^{N-1})} \int_{ |e -\s| <|\th_1-\th_2|} |e -\s|^{-1+\a} d\s\\ &\quad+ C \|w\|_{C^{0,\a}(S^{N-1})} |\th_1-\th_2|^\a \int_{1> |e -\s|\geq |\th_1-\th_2|} |e -\s|^{-1} d\s\\ &\leq C \|w\|_{C^{0,\a}(S^{N-1})} |\th_1-\th_2|^\a(-\log(|\th_1-\th_2 |)). \end{align*} From this and \eqref{}, it then follows that $F\in C^{0,\b}(S^{N-1})$ for every $\b\in (0,1)$. Now from \eqref{}, we can apply standard elliptic regularity estimates to deduce that $w\in C^{2,\b}_{\textbf{e}}(S^{N-1})$. \hfill {$\square$}\goodbreak \medskip \end{proof} \subsubsection{Proof of Theorem \ref{} and \ref{} (completed)} We recall that the function $${\mathcal Q}: \mathbb{R} \times C^{2,\a}_{\textbf{e}}(S^{N-1})\cap {\mathcal D} \to C^{0,\a}_{\textbf{e}}(S^{N-1}) $$ is well-defined and smooth in a neighborhood of $(0, 1)\in \mathbb{R}\times C^{2,\a}_{\textbf{e}}(S^{N-1})\cap {\mathcal D}$. Moreover, by Lemma \ref{}, the linear map $D_v_p {\mathcal Q}(0, 1)={\mathcal L}_e : X \to Y $ is an isomorphism. We can therefore apply the implicit function theorem to obtain a constant $\e_0 >0$ and a smooth curve \begin{equation} (-\e_0,\e_0) \longrightarrow X ,\qquad \e\mapsto \o_{\e} \end{equation} such that $\|\o_\e\|_{C^{2,\a}(S^{N-1})}\to 0$ as $\e\to 0$ and \begin{equation} \label{} {\mathcal Q}(\e, \o_{\e})={\mathcal Q}(0, 1) \qquad \textrm{ on $ C^{0,\a}_{\textbf{e}}(S^{N-1})$.} \end{equation} Moreover it is easy to see from \eqref{es:-cFp-eps} and \eqref{} that \begin{equation} \label{} \|\o_\e\|_{C^{2,\a}(S^{N-1})}\leq \|{\mathcal L}_e^{-1} ({\mathcal F}(\e,1+\o_\e))\|_{ C^{2,\a}(S^{N-1})} \leq e^{-\frac{c}{|\e|} }, \end{equation} for some positive constants $c$. \hfill {$\square$}\goodbreak \medskip \subsection{ Remarks on the perturbations of near-cylinders and near-balls}\label{s:Remarks} The aim of this section is to show that the perturbation $v_p_\e$ and $\phi_\e$ from Theorem \ref{} and \ref{} are not constant in the respective lower dimensional lattice $M=1$ (for cylinder) and $1\leq M\leq 2$ (for the balls). This is equivalent to show that ${\mathcal C}_{1}^\e$ and ${\mathcal B}_{1}^\e$ are not equilibrium patterns, for small $\e$. % \begin{Proposition} Let $v_p_\e$ and $\phi_\e$ be given by Theorem \ref{} and \ref{}, respectively, and let $1\leq M\leq N$, with $N=2,3$. Then $$ v_p_\e(\th)=1- {\mathcal L}_e^{-1}(U_\e)(\th)+ \e^{5/2} O\left( \sum_{p\in\scrL_*^M} {|p|^{-3/2}} e^{ - \k\frac{|p|}{\e}} \right) $$ and $$ \phi_\e(\th)=1- {\mathcal L}_{\textbf{e}}^{-1}(V_\e)(\th)+ \e^2 O\left(\sum_{p\in\scrL_*^M} |p|^{-1}e^{-\k\frac{|p|}{\e} }\right). $$ Here $({\mathcal L}_{\textbf{e}})^{-1}: C^{0,\a}_{\textbf{e}}(S^{N-1})\to C^{2,\a}_{\textbf{e}}(S^{N-1})$ denotes the inverse of ${\mathcal L}_{\textbf{e}}$ (see \ref{}). Moreover $$ U_\e(\th)= \sum_{p\in\scrL_*^M} \cosh\left(\k \frac{\th\cdot p }{|p|}\right) \xi_\e(|p|),\qquad V_\e(\th)= \sum_{p\in\scrL_*^M} \cosh\left(\k \frac{\th\cdot p }{|p|}\right) \z_\e(|p|), $$ with $$ \xi_\e(\ell)= c_2\frac{\sqrt{\pi}}{\sqrt{2}} \frac{\e^{3/2}}{\ell^{5/2}} e^{ -\k \frac{\ell}{\e}},\qquad \xi_\e(\ell)= c_3 \frac{\e}{\ell} e^{ -\k \frac{\ell}{\e}}, $$ where $$ c_N:= \int_{B^N}e^{-\k y\cdot e_1}\,dy. $$ Moreover if $M=1$ then $v_p_\e$ is not constant and if $M\leq 2$ then $\phi_\e$ is not constant. \end{Proposition} \noindent{{\bf Proof. }} We consider the function $f_\e: S^1\to \mathbb{R}$, given by $$ f_\e (\th)=\int_{{\mathcal C}_1^\e\setminus C_1}G_\k(|\th-y-\frac{p}{\e}|) dy=2 \sum_{p\in \scrL^M_*} \int_{B_1^2}K_0(\k |\th-z-\frac{p}{\e}|) dz. $$ For $\e>0$, we write $$ |\th-z-\frac{p}{\e}|=\frac{|p|}{\e}(1+m_\e)^{1/2}=\frac{|p|}{\e}\left(1+\frac{1}{2}m_\e+ O(\e^2/|p|)\right), $$ where $$ m_\e=\frac{\e^2|y-\th|+2\e(y-\th)\cdot p}{|p |^2}. $$ Therefore $$ |\th-z-\frac{p}{\e}|=\frac{|p|}{\e}\left(1+ \frac{\e(y-\th)\cdot p }{|p|^2}+ O(\e^2/|p|)\right). $$ By \eqref{}, $$ K_0(r)\sim \frac{\sqrt{\pi }}{\sqrt{2}} r^{-1/2}e^{-r} \qquad \textrm{ as $r\to +\infty$} . $$ We then deduce that, as $\e\to 0$, \begin{align*} \frac{2\sqrt{2 }}{\sqrt{\pi }} K_0( |\th-z-\frac{p}{\e}|) &=\frac{|p|^{-1/2}}{\e^{-1/2}}\left(1- \frac{\e(y-\th)\cdot p }{2|p|^2}+ O(\e^2/|p|)\right)e^{ - \k \frac{|p|}{\e}} e^{-\k \frac{(y-\th)\cdot p }{|p|} } e^{O(\e)}\\ &=\left( \frac{\e^{3/2}}{|p|^{5/2}} e^{ - \k \frac{|p|}{\e}} e^{-\k \frac{(y-\th)\cdot p }{|p|} } + O( {\e^{5/2}}{|p|^{-3/2}} e^{ -\k \frac{|p|}{\e}})\right)e^{O(\e)} \\ &=\left( \frac{\e^{3/2}}{|p|^{5/2}} e^{ - \k\frac{|p|}{\e}} e^{-\k \frac{y\cdot p }{|p|} } e^{\k\frac{\th\cdot p }{|p|} } + O( {\e^{5/2}}{|p|^{-3/2}} e^{ - \k\frac{|p|}{\e}}) \right) . \end{align*} We define $$ c_2:= \int_{B^2}e^{-\k \frac{y\cdot p }{|p|} }\,dy= \int_{B^2}e^{-\k y\cdot e_1}\,dy $$ and \begin{equation} \label{} \xi_\e(\ell)= c_2\frac{\sqrt{\pi}}{\sqrt{2}} \frac{\e^{3/2}}{\ell^{5/2}} e^{ -\k \frac{\ell}{\e}} . \end{equation} To see that $f_\e$ is not constant, it suffices to check that the map $$ \th\mapsto U_\e(\th):= \sum_{p\in\scrL_*^M} e^{-\k \frac{\th\cdot p }{|p|} } \xi_\e(|p|) $$ is not constant. Since $\scrL_*^M=-\scrL_*^M$, we deduce that $$ U_\e(\th):= \sum_{p\in\scrL_*^M} \cosh \left(\k \frac{\th\cdot p }{|p|} \right) \xi_\e(|p|). $$ We then conclude that \begin{equation} \label{} f_\e(\th)= \int_{{\mathcal C}_1^\e\setminus C_1}G_\k(|\th-y-\frac{p}{\e}|) dy= U_\e(\th)+\e^{5/2} O\left( \sum_{p\in\scrL_*^M} {|p|^{-3/2}} e^{ - \k\frac{|p|}{\e}} \right). \end{equation} Provided $M=1$, we may assume that $\scrL^1=\mathbb{Z}\times \{0\}\subset\mathbb{R}^2$ the subspace of $\mathbb{R}^2= \textrm{span}\{e_1,e_2\}$ with $e_1=(1,0)$ and $e_2=(0,1)$. Therefore we have that $e_2\cdot p=0$ and thus $$ U_\e(e_2)=\sum_{p\in\mathbb{Z}_*} \xi_\e(|p|). $$ However $$ U_\e(e_1)= \cosh(\k)\sum_{p\in\mathbb{Z}_*} \xi_\e(|p|). $$ This implies that $U_\e(e_1)> U_\e(e_2)$. That is ${\mathcal H}_{ \scrC_1^\e}(e_1)\neq {\mathcal H}_{ \scrC_1^\e}(e_2)$, for $\e>0$ small.\\ We now turn to the sphere lattices and provide Taylor expansion of $\phi_\e$ in Theorem \ref{}, and prove that it is not constant for $M\leq 2$. We adopt the same strategy as a above. We consider the function $F_\e: S^2\to \mathbb{R}$, given by $$ F_\e (\th)= \int_{{\mathcal B}_1^\e\setminus B_1^2}G_\k(|\th-y |) dy= \sum_{p\in \scrL^M_*} \int_{B_1^2}G_\k( |\th-z-\frac{p}{\e}|) dz. $$ Using the Taylor expansions as above, we get \begin{align*} G_\k( |\th-z-\frac{p}{\e}|)&=\frac{\e}{|p|}\left(1- \frac{\e(y-\th)\cdot p }{|p|^2}+ O(\e^2/|p|)\right) e^{-\k\frac{|p|}{\e}\left(1+ \frac{\e(y-\th)\cdot p }{|p|^2}+ O(\e^2/|p|)\right) }\\ % &= \frac{\e}{|p|}e^{-\k\frac{|p|}{\e} } e^{O(\e)} e^{-\k (y-\th)\cdot \frac{p}{|p|}} (1+O(\e^2/|p|))\\ % &= \frac{\e}{|p|} e^{-\k\frac{|p|}{\e} } e^{-\k (y-\th)\cdot \frac{p}{|p|}}+ O\left(\frac{ \e^2}{|p|} e^{-\k\frac{|p|}{\e} }\right). \end{align*} We define $$ c_2:= \int_{B^3} e^{-\k y\cdot \frac{p}{|p|}}dy = \int_{B^3} e^{-\k y\cdot e_1}dy $$ and \begin{equation} \label{} \z_\e(\ell):= c_1 \frac{\e}{\ell} e^{ -\k \frac{\ell}{\e}} . \end{equation} Letting $$ V_\e(\th):= \sum_{p\in\scrL_*} e^{\frac{\th\cdot p }{|p|} \k} \z_\e(|p|), $$ it follows that \begin{equation} \label{} F_\e(\th)=\sum_{p\in\scrL_*^M} G_\k( |\th-z-\frac{p}{\e}|)=V_\e(\th)+\e^2 O\left(\sum_{p\in\scrL_*^M} |p|^{-1}e^{-\k\frac{|p|}{\e} }\right). \end{equation} Since $\scrL_*^M=-\scrL_*^M$, we deduce that $$ V_\e(\th):= \sum_{p\in\scrL_*^M} \cosh \left(\k\frac{\th\cdot p }{|p|} \right) \z_\e(|p|)=\sum_{\ell\in \mathbb{N}_*} \sum_{p\in E_\ell}\cosh(\k\th\cdot \frac{p}{\ell})\z_\e(\ell), $$ where $E_\ell:=\{p\in \scrL_*^M\,:|p|=\ell\}$. We are therefore reduced to prove that $V_\e$ is not constant. Now if $M\leq 2$, then $e_3\cdot p=0$ and thus $$ V_\e(e_3)=\sum_{\ell \in \mathbb{N}_*}| E_\ell|\z_\e(\ell). $$ We have $$ V_\e(e_1):= \sum_{\ell\in \mathbb{N}_*} \sum_{p\in E_\ell}\cosh(\k e_1\cdot \frac{p}{\ell})\z_\e(\ell). $$ It is clear that for some $\ell\in\mathbb{N}_*$, we must have $\sum_{p\in E_\ell}\cosh(\k e_1\cdot \frac{p}{\ell})>| E_\ell|$, since $e_1 $ cannot be perpendicular to both $\textbf{a}_1$ and $\textbf{a}_2$. It follows from the construction, via the Implicit Function Theorem, of near-cylinder and the near- sphere lattices that, $v_p_\e$ in Theorem \ref{} has the following Taylor expansion $$ v_p_\e=1- {\mathcal L}_{\textbf{e}}^{-1}(U_\e)+ \e^{5/2} O\left( \sum_{p\in\scrL_*} {|p|^{-3/2}} e^{ - \k\frac{|p|}{\e}} \right). $$ and $\phi_\e$ in Theorem \ref{} Taylor expands as $$ \phi_\e =1- {\mathcal L}_{\textbf{e}}^{-1}(V_\e)+ \e^2 O\left(\sum_{p\in\scrL_*} |p|^{-1}e^{-\k\frac{|p|}{\e} }\right). $$ Here ${\mathcal L}_{\textbf{e}}$ is given by Lemma \ref{} (with the corresponding dimensions $N=2,3$) and $$ U_\e(\th)= \sum_{p\in\scrL_*^M} \cosh\left(\k \frac{\th\cdot p }{|p|}\right) \xi_\e(|p|),\qquad V_\e(\th)= \sum_{p\in\scrL_*^M} \cosh\left(\k \frac{\th\cdot p }{|p|}\right) \z_\e(|p|), $$ with $\xi_\e$ and $\z$ defined in \eqref{} and \eqref{} respectively. The above expressions of $v_p_\e$ and $\phi_\e$ follow from \eqref{} and \eqref{} without any restriction on $M$, the dimension of the lattices. \hfill {$\square$}\goodbreak \medskip \section{Appendix: Proof of Lemma \ref{Regularity -cF-0-lattice-sph-cyl} }\label{} We prove the regularity of the map ${\mathcal Q}: {\mathcal D}\to C^{0,\a}(S^{N-1})$. We note that ${\mathcal H}$ and ${\mathcal F}$ are smooth in ${\mathcal D}$ and $(-\e_0,\e_0)\times {\mathcal D}$, for small $\e_0$, respectively. We will only consider ${\mathcal F}_{0,N}$ in the following. Recalling the notations in Section \ref{}, we have the following result. \begin{Lemma}\label{} The map ${\mathcal F}_{0,N}: {\mathcal D}\to C^{0,\a}( S^{N-1})$ is smooth. Moreover for every $\phi\in {\mathcal D}$ and $w\in C^{2,\a}( S^{N-1})$, \begin{align}\label{} D {\mathcal F}_{0,N}( \phi)[w](\th) =\int_{S^{N-1}}{(w(\s)\s-w(\th)\th)\cdot \nu_\phi(\s)}G_{\k,N}(|F_\phi(\th)-F_\phi(\s) | )\G_\phi(\s)\, d\s , % \end{align} where $F_\phi:=F_\phi(1, \cdot)$. \end{Lemma} We first make a formal proof of \eqref{}. Then we prove regularity estimates for the map $\phi\mapsto D{\mathcal F}_{0,N}(\phi)[w]$, for every fixed $w$. This will end the proof of the lemma.\\ For every $\d>0$, we define \begin{equation} {\mathcal D}_\d:=\{\phi\in C^{2,\a}(S^{N-1})\,:\, \phi>\d\}, \end{equation} % Let $\phi\in {\mathcal D}_\d$ and $w\in C^{2,\a}(S^{N-1})$ such that $\|w\|_{C^{2,\a}(S^{N-1})}< \d/2$. We recall our definition of $F_v_p $ in \eqref{} and $S_\phi^{N-1}:=\partial B^N_\phi$ for every $\phi \in {\mathcal D}_\d$. % Using polar coordinates, \eqref{} and \eqref{}, we have \begin{align*} D {\mathcal F}_{0,N}( \phi)[w](\th)& = \frac{d}{dt}\Big|_{t=0} \int_{S^{N-1}}\int_{0}^{\phi(\s)+tw(\s)}G_{\k,N}(|F_{\phi+t w}(\th)-r\s |)r^{N-1} dr d\s\\ % &= \int_{S^{N-1}} G_{\k,N}(|F_\phi(\th)-\s \phi(\s) |)\phi^{N-1} w(\s) d\s\\ % & + \int_{S^{N-1}}\int_{0}^{\phi(\s) }G_{\k,N}'(|F_\phi(\th)-r\s |) w(\th) \th\cdot (\th\phi(\th)-\s r))r^{N-1} dr\\ % % & =\int_{S^{N-1}}{\s w(\s)\cdot \nu_\phi(\s)}G_{\k,N}(|F_\phi(\th)-F_\phi(\s) | )\G_\phi(\s)\, d\s \\ % & - \int_{B_\phi^{N}}\n_y G_{\k,N}(|F_\phi(\th)-y | )\cdot \th w(\th ) \, dy , % \end{align*} where $\G_\phi, \nu_\phi$ are defined in Section \ref{}. Using the divergence theorem in the last integral, we obtain \begin{align*} & D {\mathcal F}_{0,N}( \phi)[w](\th) =\int_{S^{N-1}}{(w(\s)\s-w(\th)\th)\cdot \nu_\phi(\s)}G_{\k,N}(|F_\phi(\th)-F_\phi(\s) | )\G_\phi(\s)\, d\s , % \end{align*} which is the expression in \eqref{}. We will now prove that ${\mathcal F}_{0,N} $ is smooth by deriving regularity estimates for $ D {\mathcal F}_{0,N}( \phi)[w] $ computed above.\\ By \eqref{} and \eqref{}, we have $$ (\th-\s) \cdot \nu_{\phi}(F_\phi(\s)) \, \G_\phi(\s)= - (\th-\s) \cdot\n\phi(\s) \phi^{N-2 }(\s) +(\th-\s) \cdot\s \phi^{N-1 }(\s), $$ so that \begin{align} D {\mathcal F}_{0,N}( \phi)[w](\th) & =\int_{S^{N-1}}(w(\s)-w(\th) )G_{\k,N}(|F_\phi(\th)-F_\phi(\s) | )\phi^{N-1}(\s) \, d\s \nonumber \\ % &+{w(\th)}\int_{S^{N-1}}(\th-\s )\cdot \n\phi (\s)G_{\k,N}(|F_\phi(\th)-F_\phi(\s) | ) \phi^{N-2 }(\s) \, d\s \\ &+\frac{w(\th)}{2}\int_{S^{N-1}}|\th-\s |^2 G_{\k,N}(|F_\phi(\th)-F_\phi(\s) | ) \phi^{N-1 }(\s) \, d\s, \nonumber % \end{align} where we have used $2(\th-\s) \cdot\s=-|\th-\s|^2$. To prove the regularity of $\phi\mapsto D {\mathcal F}_{0,N}( \phi)[w](\th) $, for a fixed $w\in C^{2,\a}(S^1)$, it is enough to consider the first term. In fact, we will prove a stronger result than needed. \\ For $\b\in (0,1)$ and $\d>0$, we consider the open sets $$ {\mathcal A}:=\{\phi\in C^{1,\b}(S^{N-1})\,:\, \phi>0\} $$ and $$ {\mathcal A}_\d:=\{\phi\in C^{1,\b}(S^{N-1})\,:\, \phi>\d\}. $$ We define the function (the first term in the expression of $ D {\mathcal F}_{0,N}( \phi)[w](\th) $) given by \begin{equation}\label{} {\mathcal A}_\d \to L^\infty(S^{N-1}),\qquad \phi\mapsto h(\phi):= \int_{S^{N-1}}(w(\s)-w(\th) )G_{\k,N}(|F_\phi(\th)-F_\phi(\s) | )\phi^{N-1}(\s) \, d\s. \end{equation} Recalling the notations in Section \ref{}, by a change of variable, for every $\th\in S_e$, we get \begin{align}\label{} h(\phi)(\th) =\int_{S^{N-1}}(w(R_\th\s)-w(\th) )G_{\k,N}(|F_\phi(\th)-F_\phi(R_\th\s) | )\phi^{N-1}(\s) \, d\s . % \end{align} Once the following proposition is proved, Lemma \ref{} follows immediately. \begin{Proposition}\label{Propo-reg-h} For every $\e\in (0,\b)$, the map $h:{\mathcal A}\subset C^{1,\b}(S^{N-1})\to C^{0,\b-\e}(S^{N-1})$ is of class $C^\infty$. Moreover % \begin{align*} \|D^k h(\phi)\| \leq c\left(1+ \|{\phi }\|_{C^{1,\b}(S^{N-1})} \right)^{c} \|w\|_{C^{\b}(S^{N-1})} , \end{align*} for some positive constant $c$ depending only on $\k,\b$ and $\e$. In addition if $N=3$, we can take $\e=0$. In particular ${\mathcal F}_{0,N}$ is of class $C^\infty$ in ${\mathcal D}$. \end{Proposition} We now distinguish the two case $N=2$ and $N=3$. We start with the \\ \noindent \textbf{Case $N=2$}.\\ \noindent Here, $h$ takes the form (recall \eqref{}) \begin{align*} h(\phi)(\th) & =\int_{\mathbb{R}}\int_{S^{1}}(w(R_\th\s)-w(\th) )G_{\k}(|(F_\phi(\th),0)-(F_\phi(R_\th\s) ,0)+t^2 e_3 | )\phi (\s) \, d\s dt. % \end{align*} We have $$ |F_\phi(\th)-F_\phi( \s)|^2 +t^2=\L_0(\phi,\th, \s)^2+\phi( \s)\phi(\th)|\th-\s|^2+ t^2. $$ Then with the change of variable $\rho=\frac{t}{|\th-\s|}$, we get $$ h( \phi)(\th) =\int_{\mathbb{R}}\int_{S^{1}}(w(R_\th\s)-w(\th) ){\mathcal K}(\phi,\th,R_\th\s, \rho)\phi (R_\th\s ) \, d\s d\rho, % % $$ where ${\mathcal K}:{\mathcal A}_\d\times S^{ 1}\times S^{ 1}\times\mathbb{R}\to \mathbb{R}$ is given by $$ {\mathcal K}(\phi,\th,\s,\rho ):= G_{|\th-\s|\k}\left(\L_0(\phi,\th,\s )^2+\phi(\s)\phi(\th) +\rho^2 \right), $$ with $$ \L_0(\phi, \th,\s)=\frac{ \phi(\th)-\phi( \s) }{|\th-\s|}. $$ See e.g. \cite{Cabre2015B}, there exists a constant $C>0$ such that for all $\s \in S$, all $\th,\th_1,\th_2 \in S_e$ and all $\phi_1,\phi_2 \in C^{1,\b}(S^1)$ we have \begin{equation} \label{} | \L_0(\phi, R_\th \s, \th)|\leq C \|\phi\|_{C^{1,\b}(S)}\|\phi\|_{C^{1,\b}(S^1)} . \end{equation} and \begin{align} | \L_0(\phi, R_{\theta_1} \s,\th_1)&- \L_2(\phi, R_{\theta_2} \s, \th_2)| \label{ &\leq C \|\phi\|_{C^{1,\b}(S)} |\theta_1-\theta_2|^\b. \nonumber \end{align} We follow the arguments as in \cite{Cabre2015B}, where a more singular kernel was considered. \begin{Lemma}\label{} For every $e\in S^1, k\in \mathbb{N},\d>0$, there exists a constant $c=c(\k,k,\d,\a)>0$ such that for $\phi\in{\mathcal A}_\d$, $\rho\in \mathbb{R}$, $\th, \th_1,\th_2\in S_e$ and $ \s\in S^1$, we have $$ \|D^k_\phi{\mathcal K}(\phi,\th,R_\th\s,\rho) \|\leq c (1+\|\phi\|_{C^{1,\b}(S)})^c \exp(-\k|\s-e|(\d^2+\rho^2)) $$ and \begin{align*} \|D^k_\phi{\mathcal K}(\phi,\th_1,R_{\th_1}\s,\rho )&- D^k_\phi {\mathcal K}(\phi,\th_2,R_{\th_2}\s,\rho)\|\nonumber\\ &\leq c (1+\|\phi\|_{C^{1,\b}(S^1)})^c\, |\th_1-\th_2|^\b\, \exp(-\k|\s-e|(\d^2+\rho^2)). \end{align*} \end{Lemma} \noindent{{\bf Proof. }} We have $$ {\mathcal K}(\phi,\th,R_\th\s,\rho ):= G_{|\s-e|\k}\left(\L_0(\phi,\phi,\th,R_\th\s )^2+\phi(R_\th \s)\phi(\th) +\rho^2 \right). $$ In view of the arguments in \cite{Cabre2015B}, it suffices to have some estimates related $$x\mapsto G_{a}(( x+\rho^2)^{1/2})=\frac{\exp(-a( x+\rho^2)^{1/2})}{( x+\rho^2)^{1/2}}$$ for every $x>\d^2 $, $\rho\in\mathbb{R}$ and $a\in (0,2\k)$. Now elementary computations show that \begin{equation} \label{} \partial^m_x G_{a}(( x+\rho^2)^{1/2})\leq c_{\k,\d,m} \exp(-a( x+\rho^2)^{1/2}) \qquad \textrm{ for all $x>\d^2 $, $\rho\in \mathbb{R}$, $a\in (0,2\k)$.} \end{equation} As in \cite{Cabre2015B}, combining this estimate together with \eqref{}, \eqref{}, and eventually with the Fa\'{a} de Bruno formula (see e.g. \cite{FaadeBruno-JW}), we get the estimates in the lemma. \hfill {$\square$}\goodbreak \medskip In the following, for a function $f: S^1 \to \mathbb{R}$, we use the notation $$ [f; \theta_1,\theta_2]:= f(\theta_1)-f(\theta_2)\qquad \text{ for $\theta_1,\theta_2 \in S^1$,} $$ and we note the obvious equality \begin{equation} \label{} [fg; \theta_1,\theta_2] = [f ;\theta_1,\theta_2]g(\theta_1) +f(\theta_2)[g;\theta_1,\theta_2] \quad \text{for $f,g: S^1 \to \mathbb{R}$, $\theta_1,\theta_2 \in S^1$.} \end{equation} The next step is to have estimates of all possible candidates for the derivatives of $h$. This is the purpose of the following \begin{Lemma}\label{} For $\b\in (0,1)$ and $\d>0$, we let ${\phi } \in {\mathcal A}_\d$, with $\d>0$. Let $k\in \mathbb{N} $. Let also $W,w\in C^{\b}(S^1)$ and ${\psi}_1,\dots, {\psi}_k \in C^{1,\b}(S^1)$. Define the functions $\widetilde{h} : S^1 \to \mathbb{R}$ by \begin{align*} \widetilde{h}(\th )&=\int_{{S^1}} (w(R_\th\s)-w(\th)) W(R_\th \s)\, D_\phi^k {{\mathcal K}}(\phi,R_\th \s,\th,\rho)[\psi_1,\dots,\psi_k] \, d\s d\rho, \end{align*} Then for every $\e\in (0,\b)$ there exists a constant $c=c(N,\b,k,\d,\e)$ such that \begin{align}\label{} \|\widetilde{h}(\th ) \|_{C^{\b-\e}(S^1)} & \leq c\left(1+ \|{\phi }\|_{C^{1,\b}(S^1)} \right)^{c} \prod_{i=1}^k \|{\psi}_i\|_{C^{1, \b}} \|w\|_{C^{\b}(S^1)} \|W\|_{C^{\b}(S^1)} . \end{align} \end{Lemma} \noindent{{\bf Proof. }} Let $\th_1,\th_2 \in S^1_e$ and $\s\in S^1$. We first note that \begin{equation} \label{easy-est-G} |W(R_{\th_1} \s)- W(R_{\th_2}\s)| \le C \|W\|_{C^{\b}(S^1)} |R_{\th_1} \s - R_{\th_2} \s|^\b \le C \|W\|_{C^{\b}(S)} |\th_1-\th_2|^\b, \end{equation} \begin{align} \label{} |(w(R_{\th_1} \s))-w(\th_1))-( w(R_{\th_2}\s)-w(\th_1))| \le C \|w\|_{C^{\b}(S^1)}\min( | \s-e|^\b, |\th_1-\th_2|^\b) \end{align} and \begin{align} \label{} |(w(R_{\th} \s))-w(\th)) | \le C \|w\|_{C^{\b}(S^1)} | \s-e|^\b . \end{align} Using inductively \eqref{} together with Lemma \ref{}, (\ref{easy-est-G}), \eqref{} and \eqref{}, we get the estimate \begin{align}\label{intermed-est-F_1} &|[\widetilde{h};\th_1,\th_2 ]| \leq c\left(1+ \|\phi \|_{C^{1,\b}(S^1)}^{c} \right)\|w \|_{C^{\b}(S^1)}\|W\|_{C^{\b}(S^1)} \prod_{i=1}^k \|\psi_i\|_{C^{1,\b}(S^1)} \nonumber\\ % & \quad \times \left( |\th_1-\th_2|^\b\!\! \int_{\mathbb{R}} \int_{S^{1}} | \s-e|^\b \exp(-\k |\s-e|\rho) d\s d\rho\right. \nonumber\\ &\left.\quad+ \int_{\mathbb{R}} \int_{S^{1}}\min( | \s-e|^\b, |\th_1-\th_2|^\b) \exp(-\k |\s-e|\rho) d\s d\rho\right) \nonumber\\ &= c\left(1+ \|\phi \|_{C^{1,\b}(S^1)}^{c} \right)\|w \|_{C^{\b}(S^1)}\|W\|_{C^{\b}(S^1)} \prod_{i=1}^k \|\psi_i\|_{C^{1,\b}(S^1)} \times \nonumber\\ % &\left( |\th_1-\th_2|^\b\!\! \int_{\mathbb{R}} \int_{S^{1}} \frac{\exp(-\k r) }{| \s-e|^{1-\b}} d\s dr + \int_{\mathbb{R}} \int_{S^{1}}\frac{\min( | \s-e|, |\th_1-\th_2|) \exp(-\k r)}{ |\s-e|} d\s dr \right). \end{align} For all $\th_1,\th_2\in S$, we then have \begin{align*} \int_{S^{1}}\frac{\min( | \s-e|^\b, |\th_1-\th_2|^\b)}{ |\s-e|} d\s &\leq \int_{ |e-\s| \leq |\th_1-\th_2|} | \s-e|^{\b-1} d\s\\ &\quad + |\th_1-\th_2|^\b \int_{ |\th_1-\th_2|\leq |e-\s| \leq 2}|\s-\th|^{-1} d\s\\ &\leq c |\th_1-\th_2|^\b+ c |\th_1-\th_2|^\b |\log(|\th_1-\th_2| )| . \end{align*} We thus deduce from (\ref{intermed-est-F_1}) that for all $\e\in (0,1)$ \begin{align*} |[\widetilde{h};\th_1,\th_2 ]| & \leq c |\th_1-\th_2|^{\b-\e} \left(1+ \|\phi \|_{C^{1,\b}(S^1)}^{c} \right)\|w \|_{C^{\b}(S^1)} \|W\|_{C^{\b}(S^1)} \prod_{i=1}^k \|\psi_i\|_{C^{1,\b}(S^1)}. \end{align*} % % \hfill {$\square$}\goodbreak \medskip \subsubsection{Completion of the proof of Proposition \ref{Propo-reg-h} for $N=2$.} We consider \begin{align*} \overline{h}(\th)&=\sum_{i=1}^k\int_\mathbb{R} \int_{S^1}(w(R_\th\s)-w(\th)) \phi_i(R_\th\s) D^k_\phi {\mathcal K}(\phi,\th,R_\th\s,\rho)[\psi_1,\dots,\psi_k]d\s d\rho\\ &+ \sum_{i=1}^k\int_\mathbb{R} \int_{S^1}(w(R_\th\s)-w(\th)) \psi_i(R_\th\s) D^{k-1}_\phi{\mathcal K}(\phi,\th,R_\th\s,\rho)[\psi_j]_{ \stackrel{j=1,\dots,k}{j\neq i}} d\s d\rho. \end{align*} Then by Lemma \ref{}, we have \begin{align}\label{} \|\overline{h} \|_{C^{\b-\e}(S^1)} & \leq c\left(1+ \|{\phi }\|_{C^{1,\b}(S^1)} \right)^{c} \prod_{i=1}^k \|{\psi}_i\|_{C^{1, \b}} \|w\|_{C^{\b}(S^1)} . \end{align} We can now follow step by step the arguments in \cite{Cabre2015B} to deduce that $h:{\mathcal A}_\d\subset C^{1,\b}(S^1)\to C^{0,\b-\e}(S^1)$ is of class $C^\infty$. Moreover, by \eqref{}, we get % $$ D^k h(\phi)[\psi_1,\dots,\psi_k](\th)=\overline{h}(\th). $$ Now, we pick a $w\in C^{1,\b}(\mathbb{R})$, with $\|w \|_{C^{1,\b}(\mathbb{R})}<\d/2$. Then for $\phi\in {\mathcal A}_{\d/2}$, we have \begin{align*} [{\mathcal F}_{0,N}(\phi+ w)-{\mathcal F}_{0,N}(w)- h(\phi )](\th)&= \int_0^1 [h(\phi+ w) - h(\phi )](\th)\, d\varrho\\ &= \int_0^1 \varrho\int_0^1 D h(\phi+ \rho \varrho w)[w] (\th)\,d\rho d\varrho, \end{align*} recall the expression of $h$ in \eqref{}. Since $\phi+\rho \varrho w\in {\mathcal A}_{\d}$ for $\rho,\varrho\in (0,1)$, it follows from Lemma \ref{Propo-reg-h} that \begin{align*} \|{\mathcal F}_{0,N}(\phi+ w)-{\mathcal F}_{0,N}(\phi)- &h(\phi )\|_{C^{0,\a}(S^{N-1})}\leq \int_0^1 \int_0^1 \|h(\phi+ \rho \varrho w)[w]\|_{C^{0,\b}(S^{N-1})}\, d \rho d\varrho\\ &\leq c \, \|w\|_{C^{0,\b}(\mathbb{R})}^2(1+\|\phi\|_{C^{1,\b}(S^{N-1})}+\|w\|_{C^{0,\b}(S^{N-1})} )^c. \end{align*} From this, we conclude that ${\mathcal F}_{0,N}$ is differentiable on ${\mathcal A}_{\d/2}$ with $$ {\mathcal F}_{0,N}(v_p)[w]=h(v_p). $$ The $C^\infty$-character of ${\mathcal F}_{0,N}$ on ${\mathcal A}$ now follows from the one of $h$ and the fact that $\d$ is an arbitrary positive number. \hfill {$\square$}\goodbreak \medskip \subsubsection{Completion of the proof of Proposition \ref{Propo-reg-h} for $N=3$.} As we will see, this case is much easier because \begin{equation} \label{} \int_{S^2}|\s-e|^{-1}d\s<\infty. \end{equation} Here, $h$ takes the form (recall \eqref{}) \begin{align} h(\phi)(\th) & = \int_{S^{2}}\frac{w(R_\th\s)-w(\th) }{ |\s-e|}{\mathcal K}_3(\phi,\th,R_\th\s )\phi^2 (R_\th\s ) \, d\s , \nonumber \\ % % \end{align} where ${\mathcal K}_3:{\mathcal A}_\d\times S^{ 2}\times S^{2} \to \mathbb{R}$ is given by $$ {\mathcal K}_3(\phi,\th,\s):= G_{|\th-\s|\k}\left(\L_0(\phi, \th,\s )^2+\phi(\s)\phi(\th) \right), $$ with, as above, $$ \L_0(\phi, \th,\s)=\frac{ \phi(\th)-\phi( \s) }{|\th-\s|}. $$ Moreover by \eqref{}, \eqref{} and \eqref{}, we have the estimates $$ \|D^k_\phi{\mathcal K}_3(\phi,\th,R_\th\s ) \|\leq c (1+\|\phi\|_{C^{1,\b}(S^2)})^c $$ and \begin{align*} \|D^k_\phi{\mathcal K}_3(\phi,\th_1,R_{\th_1}\s )&- D^k_\phi {\mathcal K}(\phi,\th_2,R_{\th_2}\s )\leq c (1+\|\phi\|_{C^{1,\b}(S^2)})^c |\th_1-\th_2|^\b, \end{align*} which holds for every $e\in S^2, k\in \mathbb{N},\d>0$, $\phi\in {\mathcal A}_\d$, $\rho\in \mathbb{R}$, $\th, \th_1,\th_2\in S_e$ and $ \s\in S^2$. We define \begin{align*} \hat{h}(\th)&=\sum_{i=1}^k \int_{S^2}\frac{w(R_\th\s)-w(\th)}{|\s-e|} \phi(R_\th\s) D^k_\phi {\mathcal K}_3(\phi,\th,R_\th\s)[\psi_1,\dots,\psi_k]d\s\\ &+ \sum_{i=1}^k \int_{S^2}\frac{w(R_\th\s)-w(\th)}{|\s-e|} \psi_i(R_\th\s) D^{k-1}_\phi{\mathcal K}_3(\phi,\th,R_\th\s)[\psi_j]_{ \stackrel{j=1,\dots,k}{j\neq i}} d\s. \end{align*} Then, from the estimates of the kernel ${\mathcal K}_3$ above and \eqref{}, we can deduce that \begin{align}\label{} \|\hat{h} \|_{C^{\b}(S^2)} & \leq c\left(1+ \|{\phi }\|_{C^{1,\b}(S^2)} \right)^{c} \prod_{i=1}^k \|{\psi}_i\|_{C^{1, \b}} \|w\|_{C^{\b}(S^2)} . \end{align} Now similar arguments as in the case $N=2$, show that $h:{\mathcal A}_\d\to C^{0,\b}(S^2)$ is of class $C^\infty$ and $D^k h[\psi_1,\dots,\psi_k](\phi)(\th)= \hat h(\th)$. The proof of Proposition \ref{Propo-reg-h} is then completed. \hfill {$\square$}\goodbreak \medskip

\section{Introduction} It is widely thought that eruptive solar flares are powered by the energy released during magnetic reconnection \citep[e.g.,][]{shibata1995}. The standard thinking is that turbulent photospheric motions lead to the twisting and braiding of magnetic fields in the corona and that this topological complexity is ultimately dissipated through magnetic reconnection. The process of how energy stored in the magnetic field is transferred to the plasma, however, is poorly understood. For example, it is not clear if energy transport occurs through the formation of shocks from field lines retracting from the reconnection region, the acceleration of particles in the current sheet, the dissipation of waves, or through some combination of these, and perhaps other mechanisms. The lack of progress on these issues stems from the difficulty of observing magnetic reconnection directly. Theory suggests that it occurs on very small spatial scales \citep[e.g.][]{zweibel2009}. Furthermore, the plasma processed through the current sheet is from the relatively tenuous corona. This combination of small volumes and low densities should lead to low emission measure and make it challenging to observe. In this paper we report on the structure and evolution of a current-sheet-like feature that formed in the wake of a solar eruption and flare that occurred on the west limb of the Sun on September 10, 2017. There is no commonly accepted observational signature of a current sheet, but, as we will see, this event has all of the expected features. The eruption begins with the heating of a filament low in the corona. As the filament erupts it forms a cavity that expands as it moves outward. Bright, high-temperature emission develops at the base of the eruption and eventually evolves into a classic post-flare loop arcade with a cusp-like shape at its top. Finally, a long, narrow, linear structure forms from the top of the cusp and follows the path of the erupting filament. It is this feature that we assume is the current sheet where magnetic reconnection is occurring. Here we discuss observations of the event from the EUV Imaging Spectrometer (EIS; \citealt{culhane2007}) on \textit{Hinode} \citep{kosugi2007} and the Atmospheric Imaging Assembly (AIA; \citealt{lemen2012}) on the \textit{Solar Dynamics Observatory} (\textit{SDO}). At the time of the eruption EIS was executing a flare watch study consisting of short exposures over a relatively wide field of view that extended up to about $1.15R_{\odot}$. We use the ratio of the EIS \ion{Fe}{24} 255.10\,\AA\ to \ion{Fe}{23} 263.76\,\AA\ lines to infer temperatures of 15--20\,MK in the current sheet. The highest temperatures occur near the base of the current sheet in the cusp-shaped feature at the top of the post-flare loop arcade and decline with height. Many previous studies had suggested elevated temperatures in this region relative to the brightest emission in the flare arcade \citep[e.g.,][]{tsuneta1997,warren1999}. Here we report the first temperature measurements that have extended into the current sheet just above the arcade. Comparisons with the AIA observations in this region indicate that the distribution of temperatures is relatively narrow, as it appears bright in the 193 and 131 channels that are dominated by high temperature emission lines and dim in the channels dominated by lower temperature lines. This is confirmed by a formal inversion of the observed intensities to compute the differential emission measure distribution. Many previous studies of non-thermal line broadening in high temperature emission lines have utilized spatially unresolved observations \citep[e.g.,][]{doschek1980,antonucci1984,mariska1993}. These studies have found that the largest non-thermal velocities generally occur early in a flare, sometime even before any hard X-ray emission is detected \citep[e.g.,][]{alexander1998,harra2001}. We use spatially resolved observations of the EIS \ion{Fe}{24} 192.04\,\AA\ line to measure high temperature line broadening. We find that the strongest non-thermal velocities occur early in the flare, consistent with earlier results. Further, we find that non-thermal broadening increases with height in the current sheet. The current sheet can be observed in the AIA 211 and 171 channels, which do not have strong contributions from emission lines at flare temperatures \citep{odwyer2010}. Thus the emission from the current sheet in these channels is likely to be from thermal bremsstrahlung. Comparisons with the emission from other EIS wavelengths and AIA channels dominated by Fe line emission can be used to infer elemental abundances. We find that the observations are consistent with a coronal composition in the current sheet. Several previous eruptive flares have shown evidence for a current sheet, but such observations are relatively rare. Almost all previous spectroscopic observations appear to have been taken with the Ultraviolet Coronagraph Spectrometer (UVCS; \citealt{kohl1995}) on the \textit{Solar and Heliospheric Observatory} (\textit{SoHO}). \citet{ciaravella2002}, \citet{ko2003}, \citet{ciaravella2008}, and \citet{schettino2010} report on UVCS observations of four different events, all observed at heights above about $1.5R_{\odot}$. Some spectroscopic observations from EIS near the base of the current sheet have been reported by \cite{hara2008}. The properties of the current sheet have also been studied using imaging observations \citep[e.g.,][]{savage2010,seaton2017,zhu2016}. Our results are generally consistent with these previous studies and we will discuss this in detail in the final section of the paper. \section{Observations} \begin{figure}[t!] \centerline{\includegraphics[width=0.5\textwidth]{f01.eps}} \caption{\textit{GOES} soft X-ray observations for the X8.3 flare that occurred on 10 September 2017 near 16\,UT. \textit{GOES} indicates the presence of relatively high temperature emission in the flare, but the spatial distribution of this hot plasma cannot be determined from these integrated observations. Top: \textit{GOES} light curves for the 1--8 and 0.5--4\,\AA\ channels. Dashed lines indicate the assumed background levels. The red ticks at the top of the plot indicate the start and stop time for each of the EIS rasters taken during this period. Middle and bottom: The temperature and emission measure derived from the \textit{GOES} 0.5--4 to 1--8\,\AA\ ratio.} \label{} \end{figure} In this paper we focus on the spatially resolved EIS and AIA observations of the September 10, 2017 X flare. We begin, however, with a discussion of some context soft X-ray irradiance data from the X-Ray Sensor (XRS; see, for example, \citealt{garcia1994}) on the \textit{Geostationary Operational Environmental Satellite} (\textit{GOES}). Light curves from the 1--8 and 0.5--4\,\AA\ channels are shown in Figure~\ref{}. The peak in the 1--8 channel occurred at 16:06:28 at a level of X8.3 ($8.3\times10^4$\,W~m$^{-2}$). The ratio of the two XRS channels can be used to infer a response-weighted average temperature \cite[e.g.,][]{white2005}. This analysis is also shown in Figure~\ref{} and indicates a peak temperature of about 29\,MK at 15:56:49. The spatial distribution of this hot plasma, however, cannot be determined from these integrated observations. \begin{figure*}[t!] \centerline{\includegraphics[angle=90,width=\textwidth]{f02.eps}} \caption{A series of AIA 193\,\AA\ images for the flare. This sequence shows the eruption of the filament and the formation of a linear, current-sheet-like feature behind the cavity. AIA 193\,\AA\ includes contributions from \ion{Fe}{12} and \ion{Fe}{24} and shows both the million degree corona and high temperature emission from the flare. The field of view shown here is $268\arcsec\times166\arcsec$ centered at (1068\arcsec, -143\arcsec). Note that these are composite images constructed from combining long and short exposure time images. The boundary between the two can be seen in many of the images. The wedge pattern seen in many of the images is the diffraction of emission from the brightest part of the flare off of the mesh supporting the front entrance filters.} \label{} \end{figure*} AIA is a set of multi-layer telescopes capable of imaging the Sun at high spatial resolution (0.6\arcsec\ pixels) and high cadence (typically 12\,s). EUV images are available at 94, 131, 171, 193, 211, 304, and 335\,\AA. AIA images are also available at UV and visible wavelengths, but they are not used in this analysis. During a flare, exposure times for the EUV channels are generally reduced in an alternating long-short pattern. For this event, the exposure time for the 193 channel, for example, was consistently 2\,s until 15:44:31, when the exposure time of alternating images was reduced to better accommodate the intense emission from the flare. Exposure times in 193 went as low as 4.9\,ms. The exposure times in the other EUV channels were handled similarly, except for 171, which used a fixed exposure time throughout. For display purposes we have created composite long-short images, which show the brightest parts of the flare from the short exposure and the fainter emission using the long exposure. Note that all of the AIA data presented here were downloaded from the Stanford JSOC and processed with the SolarSoftware (SSW; \citealt{freeland1998}) routine \verb+aia_prep+ using the default settings. Composite 193 images showing the eruption, post-flare loop arcade, and current sheet are presented in Figure~\ref{}. We note that the bright emission in the current sheet does not form simultaneously. Instead, the increase in intensity begins at the lowest heights and propagates outward between approximately 16:00 and 16:10\,UT. To quantify this evolution we have created a height-time plot of the intensity in a narrow region parallel to the current sheet. This is shown in Figure~\ref{}, and yields a speed of about 288\,km~s$^{-1}$ for the propagation of the intensity radially along the current sheet. \begin{figure}[t!] \centerline{ \includegraphics[width=0.48\textwidth]{f03.eps}} \caption{Bottom: The height-time plot of intensity along the current sheet showing how it brightens with time. The dashed white line corresponds to a velocity of 0.4\arcsec\ s$^{-1}$, or 288\,km~s$^{-1}$, and represents the speed at which the intensity brightening propagates radially along the current sheet. Top: an example AIA 193 image showing the position of the ``slit.'' The time for this image is indicated by the black vertical line in the bottom panel. A movie version of this figure is available with the electronic version of the manuscript. } \label{} \end{figure} One instrumental artifact that is worthy of mention here is the scattering of incident photons off of the mesh used to support the Al filters (see, for example, \citealt{poduval2013} and \citealt{lin2001}). These filters are needed to block visible light from reaching the detector but also must be very thin to allow for EUV emission to pass. To increase their durability they are supported by a thin mesh. This diffraction can be seen as the wedge-pattern emanating from the brightest emission in Figure~\ref{}. As we will see, EIS has a similar design and also shows this effect. The impact of this scattering on the observations will be discussed in the Appendix. \begin{figure*}[t!] \centerline{\includegraphics[angle=90, width=\textwidth]{f04.eps}} \caption{A time series of EIS \ion{Fe}{24} 255.10\,\AA\ rasters. The field of view shown here is $240\arcsec\times174\arcsec$ centered at (1069\arcsec, -138\arcsec). Note that some of the brightest pixels early in the flare are saturated. As with the AIA images, the wedge pattern seen in many of the rasters is the diffraction of the brightest part of the flare off of the mesh supporting the front entrance filters.} \label{} \end{figure*} \begin{figure}[t!] \centerline{\includegraphics[width=0.5\textwidth]{f05.eps}} \caption{The theoretical ratio of \ion{Fe}{24} 255.10\,\AA\ to \ion{Fe}{23} 263.76\,\AA\ as a function of temperature computed using the CHIANTI ionization fractions. The temperatures corresponding to ratios of 0.1, 1, and 10. are highlighted. } \label{} \end{figure} EIS is a high spatial and spectral resolution imaging spectrograph. EIS observes two wavelength ranges, 171--212\,\AA\ and 245--291\,\AA, with a spectral resolution of about 22\,m\AA\ and a spatial pixel size of about 1\arcsec. Solar images can be made by stepping the slit over a region of the Sun and taking an exposure at each position. Because of telemetry constraints, EIS is often restricted to saving only narrow spectral windows from each exposure. This restriction becomes particularly severe when high cadence is desired. Fortunately, for this event EIS received additional telemetry and was able to continuously run an observing sequence that saved 15 spectral windows over a field of $240\arcsec\times304\arcsec$ using the 2\arcsec\ slit and 3\arcsec\ steps between exposures. The exposure time at each position was fixed at 5\,s and the total time for each raster was 535\,s. The sequence used to observe this flare began at 05:44 on September 10, 2017 and ended at 16:53, just after the peak of the event. Observations resumed at 18:33 in a ``flare hunting'' mode where EIS executed a very low telemetry study until a flare was detected by the \textit{Hinode} X-Ray Telescope (XRT; \citealt{golub2007}) and then branched to the high telemetry raster. These observations from late in the flare will be discussed in a separate paper. All of the EIS level0 data from this time were processed using \verb+eis_prep+ with the default settings. We fit each spectral feature of interest in each raster with a Gaussian. There are approximately 30 emission lines available in these observations. Of primary interest here are \ion{Fe}{24} 192.04\,\AA, 255.10\,\AA\ and \ion{Fe}{23} 263.76\,\AA. As we will show, the temperatures in the current sheet are relatively high and the current sheet is not observed in the lower temperature emission lines, such as \ion{Ca}{17} 192.858\,\AA. Since the exposure time is fixed, the brightest features near the peak of the flare are saturated. This is common for the \ion{Fe}{24} 192.04\,\AA\ line, which occurs near the peak effective area for EIS. The effective area is smaller in the long wavelength channel and the \ion{Fe}{24} 255.10\,\AA\ and \ion{Fe}{23} 263.76\,\AA\ lines are less affected by saturation. All of the EIS rasters in the \ion{Fe}{24} 255.10\,\AA\ line from near the peak of the flare are displayed in Figure~\ref{}. \begin{figure*}[t!] \centerline{% \includegraphics[width=0.33\textwidth]{f06.eps} \includegraphics[width=0.33\textwidth]{f07.eps} \includegraphics[width=0.33\textwidth]{f08.eps} } \caption{The temperature in the flare arcade and current sheet as a function of space and time as determined from the EIS \ion{Fe}{24}/\ion{Fe}{23} ratio. The temperature peaks at the base of the current sheet and declines with height. The highest temperatures are observed early in the event. Note that the color scale is different for each ratio map to emphasize the relative differences in temperature observed at that time. The intensity, \ion{Fe}{24}/\ion{Fe}{23} ratio, and derived temperatures along the dotted line tracing the current sheet are shown in the lower panels} \label{} \end{figure*} \subsection{Current Sheet Temperature} We now turn to a discussion of the temperature structure of the current sheet. The ratio of the \ion{Fe}{24} 255.10\,\AA\ to \ion{Fe}{23} 263.76\,\AA\ emission lines observed by EIS provides good sensitivity to temperature in the 6 to 22\,MK range. This is illustrated in Figure~\ref{}, where we show the theoretical ratio computed using CHIANTI v8.0.2 \citep{delzanna2015b,dere1997}. Figure~\ref{} shows the value of this ratio over the EIS field of view for three different times during the flare. In each case the peak ratio, and therefore the peak temperature, occurs at intermediate heights, essentially in the cusp region above the brightest emission in the flare. The temperature declines with height above the limb in the current sheet. This is clearly seen in the lower panels of Figure~\ref{}, which show the intensity, ratio, and temperature as a function of height in the current sheet. In each case the temperature of the brightest emission in the flare arcade is about 12\,MK, rises to 20--25\,MK just above this region, and then slowly declines back to about 12\,MK at the largest heights for which the ratio can be measured. Saturation affects the brightest emission early in the flare. We can, however, infer the temperature in this region examining the diffracted signal, which show ratios in the brightest region of the flare of approximately 2--3. This corresponds to temperatures of 12--14\,MK. There are significant limitations to using a single pair of emission lines to infer a temperature. It does not, for example, yield any information on the distribution of temperatures. Fortunately, the AIA EUV channels are also sensitive to high temperature flare emission and we can combine the EIS and AIA observations to compute a well constrained temperature distribution. A detailed discussion of the AIA temperature response is given by \citet{odwyer2010}. Briefly, at temperatures above 10\,MK the 193 channel is dominated by \ion{Fe}{24} 192.04\,\AA\ (which peaks at $\log T = 7.25$), 131 by \ion{Fe}{21} 128.75\,\AA\ (7.05), and 94 by \ion{Fe}{18} 93.93\,\AA\ (6.85). At flare temperatures the 335 channel is affected by crosstalk from the 131 channel and has a significant contribution from \ion{Fe}{21} 128.75\,\AA\ (see \citealt{boerner2012} Section 2.2). The 211 and 171 channels do not have a significant contribution from emission lines formed above 10\,MK and instead are dominated by continuum emission at these temperatures. As will be discussed in the next subsection, we can exploit this to infer the composition in the current sheet. A set of AIA EUV images taken near 16:41:50 is shown in Figure~\ref{}. The current sheet is bright in 193 and 131 but faint in 94, suggesting that the distribution of temperatures is relatively narrow, as there does not appear to be much emission from \ion{Fe}{18}. To quantify this further we investigate the emission measure structure implied by the EIS and AIA observations. While the current sheet is clearly evident in the AIA images, there is still a contribution from foreground and background emission from the million degree corona in these channels. To subtract this contribution we have taken slices across the current sheet and fit the resulting intensities with a combination of a Gaussian and second order polynomial. The peak of the Gaussian is taken as the intensity in the current sheet at that height. This is illustrated in Figure~\ref{} where we show the intensities for a single slice across the current sheet. We have also done this with the emission in the EIS channels. The EIS \ion{Fe}{24} and \ion{Fe}{23} emission lines are not affected by foreground or background emission. They are, however, affected by the diffracted signal from the mesh. This is about a 10\% effect for both lines at the lowest heights, and the diffracted signal becomes smaller with height. The intensities can be used to compute emission measure loci curves, \begin{equation} \textrm{EM}(T) = \frac{I_{obs}}{\epsilon_\lambda(T)}, \end{equation} where $I_{obs}$ is an observed intensity and $\epsilon_\lambda(T)$ is the corresponding emissivity or temperature response curve. These curves form an envelope on the emission measure distribution and are useful for investigating the temperature structure of the plasma without performing a formal inversion. For the AIA channels we use the SSW routine \verb+aia_get_response+ with the ``evenorm'' and ``timedepend'' options to compute the temperature responses and to correct for the decline in sensitivity over the mission. To bring the EIS \ion{Fe}{24} 255.10\,\AA\ into agreement with AIA 193 we multiply the intensities of the EIS long wavelength lines by a factor of 1.8 (see \citealt{delzanna2013} and \citealt{warren2014b} for a discussion of the EIS calibration). An example set of EM loci curves is also shown in Figure~\ref{}. The general confluence of curves near a temperature of 16\,MK suggests a relatively narrow temperature distribution, consistent with the simple interpretation of the raw images. The largest discrepancy is for the 335 channel, the origin of which is unclear. Interpreting the emission from this channel has proven problematic in a number of different analyses (see, for example, \citealt{boerner2014}). \begin{figure*}[t!] \centerline{\includegraphics[angle=90,width=\textwidth]{f09.eps}} \caption{AIA images near 16:41:50. The current sheet is bright in 193 (\ion{Fe}{24}) and 131 (\ion{Fe}{21}) but relatively weak in 94 (\ion{Fe}{18}). This suggests that the temperature in the current sheet is relatively high (10--20\,MK) and that the distribution of temperatures is relatively narrow. Despite the absence of high-temperature line emission in the 211 and 171 channels, there is some evidence of the current sheet at these wavelengths. As discussed in the text, this appears to be from continuum emission.} \label{} \end{figure*} \begin{figure*}[t!] \centerline{ \includegraphics[angle=90,width=\textwidth]{f10.eps} } \centerline{ \includegraphics[width=0.345\textwidth]{f11.eps} \includegraphics[width=0.655\textwidth]{f12.eps} } \caption{Emission measure analysis of the current sheet observed by EIS and AIA near 16:41:50 at a position of (1028\arcsec,-143\arcsec). The observed intensities are approximately consistent with an isothermal temperature of about 16\,MK, as indicated by the red dot. An MCMC emission measure inversion (red line) is also consistent with a narrow temperature distribution. Also shown are 250 Monte Carlo MCMC calculations (thin black lines). The paired panels each show a $60\arcsec\times60\arcsec$ region near the base of the current sheet and the intensities along a slice perpendicular to the current sheet. Intensities used in the emission measure analysis are derived from the peak of a Gaussian fit to data along the slice.} \label{} \end{figure*} Finally, it is possible to use the intensities to perform a formal inversion and estimate the differential emission measure distribution (DEM). For this we use the ``MCMC'' method as described by \citet{kashyap1998}. This is a Bayesian method that assumes no functional form for the DEM. It also provides estimates of the error in the DEM by recalculating the emission measure using perturbed values for the intensities. The resulting calculation is also shown in Figure~\ref{} and confirms the narrow distribution inferred from the images. We have computed background subtracted intensities, emission measure loci curves, and DEMs for a number of times and positions in the flare. The resulting temperatures are consistent with those derived from the simple EIS \ion{Fe}{24} to \ion{Fe}{23} ratio and the more detailed analysis from Figure~\ref{}. The measured temperatures are always approximately 15--20\,MK and the temperature distributions are always narrow. \subsection{Current Sheet Abundances} \begin{figure*}[t!] \centerline{\includegraphics[angle=90,width=\textwidth]{f13.eps}} \caption{Line widths and intensities derived from the EIS \ion{Fe}{24} 192.04\,\AA\ line. The field of view shown here is $240\arcsec\times174\arcsec$ centered at (1069\arcsec, -138\arcsec). The broadest profiles tend to occur early in the event and broadening appears to increase with height above the arcade. Line widths also decrease with time during the event. Non-thermal velocity calculations at selected positions (marked by the crosses) are shown in Figure~\ref{}.} \label{} \end{figure*} Measurements of elemental abundances hold potential clues to how plasma in the solar atmosphere is heated. It is now recognized that elemental abundances are not fixed but vary from feature to feature and that these variations are organized by first ionization potential (FIP; for a review see \citealt{laming2015}). There is some evidence that high temperature flare emission has essentially a photospheric abundance \citep[e.g.,][]{warren2014}, suggesting that this plasma originates deep in the chromosphere and is evaporated into the corona before it has time to fractionate. Other studies indicate that the composition in flares can be complex, sometimes showing a photospheric composition, but other times showing FIP enhancements or even an inverse FIP effect (see \citealt{doschek2015} and \citealt{doschek2017} and references therein). As mentioned previously, the AIA 211 and 171 channels do not include strong emission lines formed at temperatures above 10\,MK. As can be seen in Figures~\ref{} and \ref{}, however, the current sheet can be seen at these wavelengths. This is likely to be from continuum emission at these wavelengths (see \citealt{odwyer2010}), which makes the magnitude of the intensities measured in these channels independent of the Fe abundance. We have recalculated the AIA temperature response curves assuming a photospheric composition and confirmed this. The other EIS wavelengths and AIA channels dominated by line emission scale linearly with the Fe abundance. The response curves shown in Figure~\ref{} are all computed assuming a coronal composition \citep{feldman1992}. Thus the agreement of the AIA 211 and 171 emission measure loci curves with those from the other channels indicate that the composition in the current sheet is close to coronal. This suggests that the plasma in the current sheet originated in the corona. \subsection{Current Sheet Turbulence} \begin{figure*}[t!] \centerline{\includegraphics[angle=90,width=\textwidth]{f14.eps}} \centerline{\includegraphics[angle=90,width=\textwidth]{f15.eps}} \centerline{\includegraphics[angle=90,width=\textwidth]{f16.eps}} \caption{EIS \ion{Fe}{24} 192.04 line profiles and non-thermal velocities as a function of height for various times during the early part of the event. The observed Gaussian width, observed temperature, and the inferred non-thermal velocity for each profile is indicated in each panel. The EIS instrumental width (FWHM) is approximately 70.4\,m\AA\ at this position. } \label{} \end{figure*} We now turn to the measurement of non-thermal velocities in the current sheet. For this we consider \ion{Fe}{24} 192.04\,\AA. The EIS effective area is relatively high at this wavelength and the line can generally be measured at the largest heights observed in the raster. Maps of the Gaussian line width at several times during the flare are shown in Figure~\ref{}. These maps clearly show that the widest profiles are observed early in the flare, that the line widths appear to increase with height above the limb in the current sheet, and that the line broadening diminishes with time during the event. The observed line width results from a combination of thermal broadening, instrumental broadening, and non-thermal turbulent motions of the plasma. This can be expressed as \begin{equation} W^2_{obs} = W_{inst}^2 + 4\ln(2)\left(\frac{\lambda}{c}\right)^2\left(v_T^2 + v_{NT}^2\right), \end{equation} where $\lambda$ is the wavelength of the emission, $c$ is the speed of light, and $v_T^2=2k_BT/M$ is the thermal velocity, where $k_B$ is the Boltzmann constant, $T$ is the temperature, and $M$ is the ion mass, and $v_{NT}$ is the non-thermal velocity. To account for the temperature dependence in the calculation we use the temperatures derived from the \ion{Fe}{24} to \ion{Fe}{23} ratio shown in Figure~\ref{}. Since the temperatures are somewhat noisy, we use values interpolated from a simple exponential fit to the observed temperatures profiles. Non-thermal velocity calculations are presented in Figure~\ref{} for eight positions along the current sheet observed in the 16:09, 16:18, and 16:27 rasters. The combination of increasing observed line width and declining temperature lead to a significant rise in the non-thermal velocity with height in the current sheet. In the 16:09 raster, for example, we see a rise from about 87\,km~s$^{-1}$ to about 152\,km~s$^{-1}$ over the observed length of the current sheet. At the other times the non-thermal velocities are reduced, but not as much as is suggested in the maps of the line widths. Since the temperature is also declining with time, the non-thermal velocity does not vanish over this period. In the 16:27 raster, for example, the values are 73\,km~s$^{-1}$ and 101\,km~s$^{-1}$ at the lowest and largest heights, respectively. \section{Summary} We have presented an analysis of the current sheet observed during the September 10, 2017 X8.3 flare with EIS and AIA. This analysis shows that while the temperature in the current sheet is much higher than that of the surrounding corona, the highest temperatures are observed in the cusp of the flare arcade and the temperature in the current sheet declines with height. We find that the distribution of temperatures in the current sheet is relatively narrow and is locally well described as an isothermal plasma with a temperature of 15--20\,MK. Further, the observations show that the intensity enhancement in the current sheet begins at the lowest heights and propagates outward at a speed of about 288\,km~s$^{-1}$. The spatial dependence of the temperature in the flare arcade suggests that some heating occurs in the current sheet itself, but that additional energy must be released after the field has reconnected. One possibility is that that energy is released as field lines relax and become more dipolar after reconnection. This process has been well observed \citep[e.g.,][]{mckenzie1999,savage2011} and there is some indication that it leads to significant energy release \citep[e.g.,][]{scott2016,hanneman2014,guidoni2011}. Of course, energetic particle acceleration is also likely to play a role in energy transport during this time. The observed intensities in the AIA 211 and 171 channels, which do not have significant contributions from high temperature line emission, are consistent with a coronal composition. This suggests a coronal source for the plasma in the current sheet. This is consistent with the observations of very high temperature flare plasma observed at soft X-ray wavelengths \citep{caspi2010}. Finally, we observe strong non-thermal broadening of approximately 70--150\,km~s$^{-1}$ in the current sheet. This broadening increases with height above the limb and declines with time. This is generally consistent with the previous observations of non-thermal velocities in flares \citep[e.g.,][]{doschek1980,antonucci1984,mariska1993}, although these spatially unresolved measurements are likely to have been dominated by the brightest emission. Our results are also consistent with the trend of increasing broadening with height observed in previous EIS observations of high temperature emission just above the flare arcade \citep{doschek2014,hara2008}. Observations of current sheets have been presented previously in the literature. \citet{ciaravella2002}, \citet{ko2003}, \citet{ciaravella2008}, and \citet{schettino2010} have all reported on the properties of linear structures that form in the wake of a coronal mass ejection as observed by UVCS. Despite observing the current sheet at much larger heights and generally much later in the event than we have, their results are largely consistent with what we observe here. Each of the studies finds evidence for high temperature emission, as \ion{Fe}{18}, \ion{Si}{12}, and \ion{Ca}{14} are detected. Their temperatures, however, are generally lower than what we report here for the region just above the flare arcade. \citet{ciaravella2002}, \citet{ko2003}, and \citet{ciaravella2008} all report FIP enhancements of varying magnitudes in the current sheet. \citet{schettino2010} and \citet{ciaravella2008} report strong broadening in the \ion{Fe}{18} line. \citet{seaton2017} used AIA observations of a flare to infer temperatures in a current sheet and also find evidence for high temperature emission. They do find much broader temperature distributions that peak at somewhat lower temperatures than we find, but their current sheet is relatively faint and they lacked spectroscopic data, so their inversions are not as well constrained. \citet{zhu2016} also find temperature distributions in a current sheet that are somewhat broader and cooler than what we observe. The current sheet emission in their event appears to be relatively intense, but lacking spectroscopic observations the DEMs are still not likely to be well constrained. Magnetic reconnection is widely believed to play a major role in all flaring activity and thus extended current sheets should be present in all coronal mass ejections. It is curious then that the literature on the physical properties of current sheets is so sparse. As mentioned in the introduction, the obvious argument is that the current sheet should be thin and relatively tenuous and so have a low emission measure. Our observations, however, show a current sheet that has a high emission measure and is relatively thick. In Figure~\ref{} we see an isothermal emission measure of approximately $10^{30}$\,cm$^{-5}$, orders of magnitude higher than the background corona \citep[e.g.,][]{warren2009}. Figure~\ref{} also shows an observed width of over 3,000\,km for both EIS and AIA, much larger than the spatial resolution of either instrument (see \citealt{brooks2012} for measurements of loops observed with EIS and AIA). The AIA studies mentioned in the previous paragraph also show relatively wide current sheets when they are observed. This suggests that significant heating in the current sheet may only occur under very specific conditions.

\section{Introduction} Most of the stars in the local universe formed out of tremendous cool gas reservoirs ($M_{\rm gas} \sim 10^{10-11} {\rm M {\odot} }$, $T \sim10-100$ K) in the interstellar medium (ISM) of high redshift (${\rm 1 < z < 3.5}$) galaxies with intense star-formation (SF) \citep{carilli13,madau14}. Massive, dusty star forming galaxies at high-$z$ (DSFGs; $M_{\rm dust} \sim 10^{8-9} {\rm M_{ \odot} }$) are typically gas-rich galaxies selected via their bright observed (sub)-mm fluxes (also known as SMGs). The rest-frame far-IR (FIR)-mm luminosity associated with the thermal dust emission \citep{efstathiou00, johnson13} (re-radiated far-UV radiation) traces the total star-formation (SF) activity, while the extreme star-formation rates (SFRs) in these IR luminous galaxies are likely due to a sustained supply of cool gas from the intergalactic medium (IGM). The dense molecular ingredients of the ISM thereby limits the timescale for extended starburst (SB) activity, with short-lived SB episodes of 10's-100's of Myr. These are believed to often include gas-rich mergers that induce star formation via tidal torques, which drive gas infall and subsequent collapse \citep{hern92}. The most active SB galaxies at $z \sim 2$ contribute key insights into galaxy evolution and structure formation, as their massive gas reservoirs play a key role in the bulk stellar mass growth in their ISM environments, and as a result are believed to be the progenitors to massive elliptical/spheroidal galaxies and clusters at low-$z$ \citep{casey14}. The SMG population can be accounted for by major or minor-merger dominated starbursts \citep{baugh05,swinbank08} in some semi-analytic models. Others suggest that the observed population is a heterogeneous mix of early and late stage major mergers and blending of passive star-forming disc galaxies. The brightest SMGs are likely early-stage mergers, exchanging a significant amount of molecular material for continued star formation \citep{hayward12, narayanan15}. SMGs typically have high gas mass fractions, ${\rm M_{gas} / M_{\star} }$, up to 80\% \citep{carilli13}. CO line measurements are vital for directly probing the fuel for these star-forming galaxies, i.e. the total molecular gas mass, at the peak of the co-moving SFR density (${\rm z \sim 2-3}$). The gas accretion history of growing dark matter (DM) halos in cosmological simulations \citep{keres05} agrees well with the observed evolution of the CO luminosity function, as \citet{decarli16} find a peak redshift for CO luminous galaxies at $z \sim 2$, comparable to the peak of the co-moving star formation rate density. The CO (J$=1\rightarrow 0$) transition accounts for both the dense and most diffuse molecular gas, and has traditionally been calibrated to trace the bulk H$_{\rm 2}$ gas mass (via collisional excitation with the H$_{\rm 2}$ gas). The observed CO line luminosity, L'$_{\rm CO}$, to ${\rm H_{2}}$ mass conversion factor, $\alpha_{\rm CO}$, \citep{carilli13, bolatto13}, is calibrated to this transition, making observations of CO($1-0$) important for determining the total ${\rm H_{2}}$ mass content at high-$z$. The number of high-$z$ sources with galaxy integrated CO($1-0$) detections is sparse \citep[see][]{scoville17,carilli13}, although it is accumulating \citep[e.g.][]{carilli02, hainline06, riechers11a,harris12, thomson12, riechers13,fu13, aravena13, sharon16, decarli16, decarli16b, huynh17}, with approximately 60 to date. Resolved imaging of this lowest rotational transition of CO \citep[e.g.][]{riechers11a,lestrade11} in high-$z$ SMGs indicates that the total molecular gas can extend up to 30 kpc for merging systems. Only the most active star forming sources with apparent $L_{IR} \ge10^{12-14} L_\odot$ at ${\rm z > 1}$ can be observed at this fundamental CO rotational transition. These apparent luminosities are often due to strong lensing. The strong lensing effect (usually with magnification factor, $\mu = 10-30X$) \citep[e.g.][]{bussmann13, bussmann15, geach15, spilker16a}, yields shorter integration times to provide secure detections of the molecular gas in both strongly lensed, intrinsically bright and faint, but highly magnified, normal star-forming systems. The far-IR/sub-mm \textit{Herschel}\ Astrophysical Terahertz Large Area Survey (H-ATLAS) \citep{eales10} and the Herschel Multi-tiered Extragalactic Survey (HerMES) \citep{oliver12}, together covering about 650${\rm \deg}^{2}$, and the 2500 ${\rm \deg}^{2}$ mm South Pole Telescope \citep[\textit{SPT},][]{carlstrom11} have paved the way forward in discovering a rare population of gravitationally lensed DSFGs \citep[e.g.][]{negrello10, planck7, wardlow13, vieira13, weiss13, negrello17}, as well as an intrinsically bright, unlensed population. \citet{canameras15} and \citet{harrington16} have exploited the all-sky sensitivity of \textit{Planck}\ to find the most luminous high redshift galaxies currently known in the \textit{Planck}\ era -- all of which are gravitationally lensed \citep[also][]{herranz13,fu12, combes12}. Here we present galaxy integrated, CO($1-0$) measurements of seven $z >$1 galaxies using the Green Bank Telescope (\textit{GBT}). This is a pilot study for a larger program to identify a large sample of extremely luminous high-$z$ SMGs identified by the all-sky \textit{Planck}\ survey. In our original pilot study leading to the sample in this work, our goal was to identify sources that have the probability to be gravitationally lensed given their high flux densities in the 3 SPIRE bands of 250, 350, 500\micron \, \citep[e.g. the ${\rm S_{500} \, {\rm or} \, S_{350} \ge 100 \, mJy}$][]{negrello10, ivison11}. We have previously obtained one J$_{\rm up} = 2-4$ transition for all seven of the sources presented in this study using the Redshift Search Receiver (RSR) on the \textit{LMT}. The majority of these sources have apparent ${\rm \mu L_{IR} > 10^{14.0-14.5} L_\odot}$ making them some of the most luminous sources currently known \citep{harrington16, canameras15}. Our goals in this study are to confirm the \textit{LMT}\ CO redshift, measure the CO($1-0$) line emission to constrain our estimate of the ${\rm H_{2}}$ masses and begin analysing the CO spectral line energy distributions (CO SLEDs). In \S~2 we review our sample selection and previous observations described in detail in \citet{harrington16}, and in \S~3 we outline our CO($1-0$) observations using the VErsatile GBT Astronomical Spectrometer (VEGAS) instrument on the \textit{GBT}. Measured and derived gas properties using the CO($1-0$) line emission and supplementary \textit{LMT}\ CO data is found in \S~4, followed by a discussion in \S~5. Finally, we conclude our study in \S~6. We adopt a $\Lambda$ CDM cosmology with ${\rm H_{0} = 70 \, km s^{-1} Mpc^{-1} }$ with ${\rm \Omega _{m} = 0.3}$, and ${\rm \Omega _{\Lambda} = 0.7}$ throughout this paper. \section{Sample} In a search for the most extreme, and thus rare, star-forming galaxies at $z > 1$, we exploit the full-sky sub-mm coverage offered by the the \textit{Planck}\ Catalog of Compact Sources (PCCS). The highest frequency observed by \textit{Planck}\ (857 GHz / 350 \micron) contains a dataset of $\sim$ 24,000 point source objects \citep{planck13}. From this dataset we limit our searches to point sources at Galactic latitude ${\rm |b| > 30\ensuremath{^\circ}}$ to minimize the Galactic source contamination. This filtered sample is then cross-correlated with the combined catalogs of three \textit{Herschel}\ large area surveys: \textit{Herschel}\ Multi-tiered Extragalactic Survey \citep[HerMES,][]{oliver12}, \textit{Herschel}\ Stripe 82 Survey \citep[HerS-82,][]{viero14}, and the dedicated \textit{Planck}\ follow-up \textit{Herschel}\ DDT ``Must-Do" Programme: ``The \textit{Herschel}\ and \textit{Planck}\ All-Sky Source Snapshot Legacy Survey''. The details of our selection method can be found in \citet{harrington16} for the \textit{Planck}\ - \textit{Herschel}\ counterparts with $ S_{\rm 350} \ge 100 {\rm mJy}$ in our initial follow-up during the Early Science Campaign 2 for the \textit{LMT}. In brief, we cross-matched \textit{Planck}-\textit{Herschel} \, counterparts within 150". In total there were 350 \textit{Herschel}\ counterparts to 56 \textit{Planck}\ sources within 150". The higher spatial resolution of \textit{Herschel}\ allowed us to pinpoint the position of the \textit{Planck}\ point sources, enabling follow up studies. For 8/11 galaxies observed with the \textit{LMT}\ we detected a single, compact source using the AzTEC 1.1mm camera. Subsequently we detected a strong CO line with the RSR spectrometre. We make use of the 3 SPIRE bands of \textit{Herschel}\ (250, 350, 500 \micron) and the additional \textit{LMT}\ observations to derive apparent ${\rm \mu L_{IR} > 10^{13.0-14.5} L_\odot}$ at ${\rm z_{CO} = 1.33 - 3.26}$ (see Table~\ref{}). The current sample in this \textit{GBT}\ study consists of observations of only seven of the original eight targets. \section{\textit{GBT}\ Observations} \begin{table*} \caption{Sources and \textit{GBT}\ Observations Summary} \label{} \begin{tabular}{lcccccccc} \hline Source ID & RA & DEC &$\mu L_{IR}^\dagger$ & \textit{GBT}\ RX & Dates & Int. Time (On-Source) & ${\rm < T_{sys}} >$ \\ \hline & J2000 & J2000 & ($10^{14} L_\odot)$ & & 2016 & mins & K \\ \hline PJ142823.9 & 14h28m23.9s & +35d26m20s & $0.19\pm0.04$& Q & 2/12;2/19 & 336 & 100 \\ PJ160722.6 &16h07m22.6s & +73d47m03s & $0.14\pm0.03$& Q &2/12;2/19 & 216 & 75 \\ \hline PJ105353.0 & 10h53m53.0s & +05d56m21s & $2.9\pm0.4$ & ${\rm K_{a} }$ &3/30 & 100 & 40 \\ PJ112714.5 & 11h27m14.5s & +42d28m25s & $1.1\pm0.2$ & ${\rm K_{a} }$ & 3/30 & 84 & 40 \\ PJ120207.6 &12h02m07.6s & +53d34m39s & $1.4\pm0.3$ & ${\rm K_{a} }$ &3/30 & 80 & 40 \\ PJ132302.9 & 13h23m02.9s & +55d36m01s & $1.2\pm0.2$ & ${\rm K_{a} }$ &3/30 & 96 & 40 \\ PJ160917.8 & 16h09m17.8s & +60d45m20s & $2.0\pm0.4$ & ${\rm K_{a} }$ &3/26 & 92 & 35 \\ \hline \end{tabular} [1] Q band receiver frequency coverage: $38.2-49.8$GHz. [2] ${\rm K_{a} }$ band receiver frequency coverage: $26.0-40$GHz. [3]$^\dagger$ $L_{IR}$ is the far-infrared luminosity integrated between 8-1000 \micron\ . \end{table*} Based on our RSR spectroscopy, two of our sources have redshifted CO($1-0$) (i.e. rest-frame ${115.27}$ GHz) line emission in the range of the Q band receiver ($38.2-49.8$ GHz) on the \textit{GBT}. The other five sources fall within the ${\rm K_{a} }$ band receiver ($26.0-40$ GHz). We used the low-resolution 1500 MHz bandwidth mode of the backend spectrometre, VEGAS. The raw channel resolution corresponds to 1.465 MHz, or ${\rm \sim 16}$ km s$^{-1}$ in ${\rm K_{a}}$ band, using 1024 channels. Observations between February and March, 2016, took place in typical weather conditions. For both $Q$ and ${\rm K_{a} }$ band observations, we used a SubBeamNod procedure, nodding the 8m \textit{GBT} \, sub-reflector every 6 seconds between each receiver feed for an integration time of four minutes. In most cases, this 4 minute procedure was repeated back-to-back for up to an hour to achieve the ON source integration times presented in Table ~\ref{}. The atmosphere becomes highly variable at the frequencies within $Q$ and ${\rm K_{a} }$ band, therefore we observed pointing sources roughly every hour. The routine pointing and focus procedures allowed us to assess the best azimuth and elevation corrections, as well as the best focus values for the peak line strength measurements. After total-power switching for the standard ${\rm (ON-OFF)/OFF}$ GBTIDL calculations, we include the observatory's atmospheric model, which tracks zenith opacity as a function of frequency and time. Each scan is corrected for the atmospheric time and frequency variations on the sky, given zenith opacity ${\rm \tau_{sky} }$ and elevation, EL: ${\rm T_{antenna} = T_{sys} \times e^{\frac{\tau_{sky}}{sin(EL)}} \times \frac{(ON-OFF)}{OFF} }$. The elevation ranges for Q and ${\rm K_{a}}$ band spanned $33-84 \deg$ and $28-73 \deg$ , respectively. The typical ${\rm T_{sys}}$ values ranged from 67-134 K in Q band, and 30 - 45 K in ${\rm K_{a}}$ band. To convert the measured antenna temperature in $T_A^{*}$ to flux density we use the calibration factor derived for \textit{GBT}: Q band scales as ${\rm 1 K / 1.7}$ Jy, ${\rm K_{a}}$ band as ${\rm 1 K / 1.6 Jy}$. \begin{figure*} \begin{center} \includegraphics[width=15.0cm]{LMTGBT-f.png} \end{center} \vspace*{-5mm} \caption{The RSR CO spectra (yellow) for all 7 galaxies \citep{harrington16} are scaled by ${\rm J_{\rm up}^{2}}$ and overlaid (red) onto the GBT CO spectra (yellow) in this study. The comparable line widths and spectral features are coincident. PJ160918 has both its CO($4-3$) and ($3-2$) lines compared to the ($1-0$) line emission.} \label{} \end{figure*} We used a high-pass filter to remove very low-frequency ripples in the overall baseline without removing the line emission. The width of the high-pass filter was at least twice the expected full-width at half maximum (FWHM) of the CO line based on our \textit{LMT}\ RSR spectra. We utilised Gaussian smoothing to decrease the resolution by a factor of 4, resulting in a 5.86 MHz (${\rm \sim 50 \, km s^{-1} }$ for ${\rm K_{a}}$) channel resolution. In this smoothing step, each channel was treated as independent to avoid correlations in the noise of neighboring channels. As the high-pass filter removes the low-frequency ripples, and not the mid-frequency baseline ripples, we then fit and remove a baseline (${\rm n_{poly} = 2-3}$) to the emission free regions of the spectra. The resulting spectra can be seen in Fig. \ref{}. We adopt a 30\% total uncertainty given a 15-20\% flux calibration error, typical 5-10\% pointing/focus drifts and atmospheric losses and a conservative 10-15\% for the baseline removal due to the variations across the bandpass at the observed frequencies. \section{Results: CO ($1-0$) Line Properties} We detect CO($1-0$) at ${\rm S_{\rm peak}/N_{\rm channel} > 7}$ from each of our seven targets at the expected redshifts. We first derive the observed central frequency by fitting a single Gaussian to the CO($1-0$) line emission, confirming the exact redshifts of these \textit{Planck}-\textit{Herschel}\ identified galaxies, which had previously been derived using only one CO line from the \textit{LMT} (Tab.~\ref{}). The spectroscopic redshifts span from $1.33 < z < 3.26$. Our new \textit{GBT}\ measurements further support our previous redshift determinations from the combination of panchromatic photometry (WISE-11 and 22 \micron, \textit{Herschel}\ SPIRE 250, 350, and 500 \micron, AzTEC 1100 \micron\ and NVSS/FIRST radio) and single CO line observations (see Appendix A of \citep{harrington16}). We find that the CO($1-0$) lines show nearly identical profiles and widths as the J$_{\rm up} = 2-4$ CO lines, with FWHM = [375--740 km s$^{-1}$] (see Fig.~\ref{}). It is unlikely that there is a significant amount of gas that excites the CO($1-0$) but not, e.g. the CO($3-2$). Therefore, the similar line emission FWHM and line profiles suggests that both transitions are tracing co-spatial volumes. We calculated the line luminosity, ${\rm L'_{CO(1-0)}}$, using Eq.~(3) by \citet{solomon97},$$L'_{\rm CO} = 3.25\times10^7 S_{\rm CO}\Delta V \nu_{\rm obs}^{-2} D_{\rm L}^2(1+z)^{-3}$$ with $S_{\rm CO}\Delta V$ in Jy km s$^{-1}$, $\nu_{\rm obs}$ in GHz, and $D_{\rm L}$ in Mpc (see Table~\ref{}). Since the CO lines are not exactly Gaussian, we also integrated the spectra within ${\rm \pm 1500}$ km s$^{-1}$ to compute $S_{\rm CO}\Delta V$. The corresponding Gaussian derived values of $S_{\rm CO}\Delta V$ are the same within 1-sigma. Some of the measured apparent $L'_{\rm CO}$ are amongst the brightest, if not the brightest, for all $z > 1$ QSO, SMG, LBG, including the SPT DSFGs \citep{aravena16}, as well as $ z < 1$ ULIRGs, nearby spirals, and low-$z$ starburst galaxies \citep{carilli13}. These bright apparent luminosities suggest that our galaxies have been strongly magnified. \section{Discussion} \subsection{CO Spectral Line Energy Distributions (SLEDs)} \begin{figure*} \includegraphics[width=15.0cm]{sled.png} \caption{Here we plot the velocity-integrated line intensity ratios of J$_{\rm up}/$J$ = 1$, normalised to the CO($1-0$) derived integrated flux for the current sample. Our seven galaxies (red diamonds) are within the spread for average SMGs \citep{bothwell13} (yellow), and can be compared to the low-$z$ (U)LIRG population \citep{pap12} (blue), and the Milky Way center \citep{fixsen99} (gray). All regions contain the dispersion between the 25th and 75th percentile of the distribution of the CO($1-0$) normalised integrated flux. Yellow stars show the average QSO values out to J $= 6$ \citep{carilli13}.} \label{} \end{figure*} In Fig~\ref{} we plot the ratio of the line integrated intensity of the higher-J CO $S_{\rm CO}\Delta V$ to our CO($1-0$) $S_{\rm CO}\Delta V$. All of our galaxies show sub-thermalised excitation conditions. Up to $J \le 3$, we find these values to be consistent with both the lower end of the SMG excitation distribution \citep{bothwell13,carilli13} and the upper end of the MW \citep{fixsen99}. The uncertainty of the MW measurements and the physical intrinsic SMG dispersion overlap for $J \le 4$. Without higher-J CO lines, where SMGs and the MW differ strongly, it is a challenge to disentangle which of these two ISM conditions dominate our galaxies. We parametrized these CO SLEDs in terms of brightness temperature, or CO line luminosity, ratios, $$r_{up,1} = \frac{T_{B} (J_{upper} > 1)}{ T_{B} (1-0)} = \frac{L'_{CO} (J_{ upper} > 1)}{L'_{CO} (1 -0)}.$$ For two sources with only $r_{21}$ we found $< r_{21} > = 0.92$, while the remaining five sources have $< r_{31} > = 0.34$. Finally, for the one source with a CO($4-3$) line observation, we found $r_{41} = 0.18$, similar to what has been reported in \citet{hainline06} for an SMG of similar redshift (i.e. $z \sim 3.3-3.5$). As noted in \citep{frayer11}, there is a wide-range in the observed $r_{31} = 0.1 - 1$ in the local starburst population. The fact that our galaxies fall in the lower end of the SMG excitation distribution, while being some of the most luminous sources currently known, suggests that they may not be exceptional SMGs, but more strongly magnified sources. In \citet{harrington16} we showed that all of these sources fall within the parameter space for SF powered luminosity (rather than AGN) in a mid-IR to far-IR color-color diagnostic plot \citep{kirkpatrick13}. Using their CO SLED we can further rule out the presence of a powerful QSO in our galaxies, as typical QSO host galaxies with powerful AGN activity often exhibit thermalised line ratios out to ${\rm CO(4-3)}$ \citep[e.g][]{riechers06,weiss07}. However, we caution that most QSO hosts with a good coverage in the CO SLED are strongly lensed objects selected in the optical/NIR. This may result in a bias towards the excitation conditions within the central region. \citet{sharon16} show there is a statistical similarity between the $r_{31}$ values reported for SMGs and AGN in their sample. However, the line ratios in their sample have a global average (AGN and SMG) 3 times higher for $r_{31}$ (in fact close to thermalised: $<r_{31}> = 0.9$) as compared to our sources. This suggests that their sample might consist of hybrid SMG/AGN galaxies. Our CO SLEDs are currently limited out to J $ = 3$ or $4$, therefore we cannot rule out the presence of an AGN. \begin{table*} \caption{Summary of CO($1-0$) VEGAS Measurements} \label{} \begin{tabular}{@{}lccccccc} \hline ID & $\nu_{obs}$ & $z_{CO(1-0)}$ & $\Delta V$ & $S\Delta V^\dagger$ & $S_{\nu}$ & $\mu L'_{CO}$ \\ & (GHz) & & (km/s) & (Jy km/s) & Peak (mJy) & ($10^{10}$ K km s$^{-1}$ pc$^{2}$) \\ \hline \\ PJ105353.0 & 28.7712 & $3.0053 \pm 0.00016$ & $ 738\pm 38 $ & $4.3\pm1.3$ &$6.2\pm1.9$ & $170\pm60$ \\ PJ112714.5 & 35.6248 & $2.2352\pm0.00006$ & $736\pm20$ & $7.4\pm2.2$ & $9.3\pm2.8$ & $178\pm63$ \\ PJ120207.6 & 33.4970 & $2.442\pm0.00007$ & $602\pm21$ & $5.5\pm1.7$ & $8.9\pm2.7$ & $154\pm55$\\ PJ132302.9 & 33.7350 & $2.4165\pm0.00006$ & $540\pm17$ & $4.7\pm1.4$ & $8.4\pm2.5$ &$ 129\pm46$ \\ PJ142823.9 & 49.5766 & $1.3254\pm0.00005$ & $436 \pm25$ & $1.2\pm0.4$ & $2.9\pm0.9$ & $11\pm4$ \\ PJ160722.6 & 46.4115 & $1.4838\pm0.00006$ & $374 \pm27$ & $1.0\pm0.3$ & $2.4\pm 0.7$ & $12\pm4$ \\ PJ160917.8 & 27.0911 & $3.2567\pm0.00014$ & $705 \pm 31$ & $7.6\pm2.3$ & $9.8\pm2.9$ & $343\pm121$\\ \hline \end{tabular} [1] $T_A^{*}$ to flux density using the \textit{GBT}\ : Q band is 1K/1.7 Jy, Ka band is 1K/1.6 Jy. [2] The reported redshifts correspond to the values obtained after velocity offset corrections. The line widths reported indicate the FWHM values after correcting for the instrumental resolution. This correction was on average less than 1\% due to large observed line widths. [3] $^\dagger$ The integrated value obtained within the interval of ${\rm \pm 1500}$ km s$^{-1}$ from the center. \end{table*} \begin{table*} \caption{Gas Properties} \label{} \begin{tabular}{@{}lccccccc} \hline ID & $L(IR) / L'_{\rm CO(1-0)}$ & $r_{\rm up,1}$ & $\mu M_{\rm ISM}$ & $\mu M_{\rm H2}$ & $\tau_{\rm depl}$-CO & $\tau_{\rm depl}$-ISM \\ & $ L_{\odot} / {\rm (K km s^{-1} pc^{2})}$ & & ($10^{10} {\rm M_\odot}$) & $(\alpha_{\rm CO} 4.3,0.8)$($10^{10}$ ${\rm M_\odot}$)& $ (\alpha_{\rm CO} 4.3,0.8)$ (Myr) & (Myr) \\ \hline PJ105353.0 & $ 170\pm65 $ & $0.36\pm0.13 [r_{31}] $ & $624\pm156$ & $[732,136]\pm[259, 54]$ & $[239,44]$ &203 \\ PJ112714.5 & $ 62\pm25$ & $0.29\pm0.10 [r_{31}] $ &$160\pm40$ & $[764,142]\pm[270, 27]$ & $[636,118]$ & 133 \\ PJ120207.6 & $91\pm38$ & $0.4\pm0.14 [r_{31}] $ & $452\pm113$ & $ [663,123]\pm[234, 30]$& $[442,82]$ & 302 \\ PJ132302.9 & $93\pm36$ & $0.31\pm0.11 [r_{31}] $ & $215\pm54$ & $[557,103]\pm[196, 23]$& $[425,79]$ & 164 \\ PJ142823.9 & $168\pm69$ & $0.88\pm0.36 [r_{21}] $ & $97\pm24$ & $[49,9]\pm[17, 4]$ & $[228,42]$ & 452 \\ PJ160722.6 & $121\pm50$ & $0.95\pm0.38 [r_{21}] $ & $52\pm13$ & $[50,9]\pm[18, 5]$& $[331,62]$ &344\\ PJ160917.8 & $58\pm24$ & $0.35\pm 0.13 [r_{31}] $ & $465\pm116$ & $[1473,274]\pm[521, 85]$& $[694,129]$ & 219 \\ PJ160917.8 & $-$ & $0.18\pm0.06 [r_{41}] $ & $-$& $-$ & $-$ & $-$\\ \hline \end{tabular} [1] Unknown lensing amplification $\mu$ is reflected in the derived CO luminosity and $H_2$\ mass. ISM masses were calculated following \citet{scoville16}, scaled to our AzTEC 1.1mm photometry with a fixed dust temperature of 25 K. \end{table*} \subsection{Ratio of IR Luminosity to CO Line Luminosity} The observable ${\rm \mu L_{IR} / \mu L'_{CO(1-0)}}$ ratio serves as a proxy for SF efficiency (SFE), and stands independent of the unknown magnification factor \footnote{We assume, without high angular resolution imaging of the two luminosity sources, that the CO($1-0$) emitting region and the pervasive dust content are on average co-spatial.}. The integrated IR emission (8-1000$\micron$) reflects the bulk star-forming activity, while the CO line luminosity indicates the amount of gas supplying the ongoing star formation. Using the value of this ratio we place our sample in the context of SB versus typical star-forming galaxies at different $z$, IR and CO line luminosity \citep[Fig~\ref{}; e.g.][]{genzel10}. We measure the ${\rm \mu L_{IR} / \mu L'_{CO(1-0)}}$ ratio as ${\rm (58-170) \, L_\odot / (K \, km s^{-1} pc^{2})}$, with ${\rm < 110 \pm 22 > \, {\rm L_\odot} / K}$ km s$^{-1}$ pc$^{-2}$ (see Fig~\ref{}). The average value of our galaxies is closer to 140 ${\rm L_\odot / (K km s^{-1} pc^{2})}$ observed in SB galaxies, rather than 30 ${\rm L_\odot / (K km s^{-1} pc^{2})}$ observed in typical star-forming galaxies \citep{sv05, genzel10, frayer11}. From this we conclude that the $L_{IR} / L'_{CO(1-0)}$ values obtained for this subset of \textit{Planck}-\textit{Herschel}\ sources have enhanced ratios with respect to typical star-forming galaxies, as expected from their large apparent $L_{\rm IR}$ \citep{canameras15, harrington16}. We note that our sample exhibits slightly lower ratios on average compared to both the highly excited system, HFLS-3 \citep[$z \sim 6$][]{riechers13}, as well as the lensed SPT DSFGs \citep{aravena16} (Fig~\ref{}). Roughly half of the strongly lensed, dusty \textit{Herschel}\ galaxies \citep{harris12} are consistent with our sample and lie within the yellow shaded region for SB systems. In contrast, more than half of the SPT sources have excess $L_{IR}$-to-$L'_{CO(1-0)}$ \citet{aravena16}, although the dispersion is similar for both H-ATLAS and SPT samples. Our seven \textit{Planck}-\textit{Herschel}\ sample are unusual as they lack a similar dispersion. This may reflect the ability of the all-sky sub-mm sensitivy and coverage of the \textit{Planck}\ survey in detecting the rare galaxies that are, on average, more strongly lensed than the similarly selected H-ATLAS sample. While the SPT lensed galaxies are a similar population at high-$z$ with comparable $L_{IR}$, due to different selection methods (350 \micron\ versus mm), the average redshift of their sample is significantly shifted towards a higher value compared to our sample: $< z > = 3.9$ and $< z > = 2.3$, respectively. At such a high redshift, $z \sim 4$, a MW type galaxy would be subject to non-negligible dust heating due to the CMB \citep{dacunha13}, and may contribute to the higher $L_{IR}$-to-$L'_{CO(1-0)}$ values observed in the SPT sample. At $z \sim 4$, the CMB temperature is also a sufficient background to radiatively excite the cool reservoirs of CO, particularly the J$= 1 \rightarrow 0$ ground state rotational transition, resulting in a dimming of the observed CO line emission. Because (sub)-mm flux measurements are made against the CMB, the contrast in the CO ($1-0$) line integrated intensity via collisional excitation (typically with H$_{\rm 2}$ molecules), compared with the radiatively excitated CO gas from the CMB background becomes more severe beyond $z = 4$. About 80\% of the CO ($1-0$) emission can be recovered against the CMB at $z = 2-3$, but only 50-60\% just beyond $z = 4$ if there would be a gas kinetic temperature of 40 K \citep{dacunha13}. We caution that the effects of the CMB alone cannot account for the differences observed in these luminosity-luminosity ratios, as the H-ATLAS and SPT sample have a similar spread in their $L_{IR}$-to-$L'_{CO(1-0)}$ values. The similar redshift range of the 12 H-ATLAS sample compared to the sample of 7 \textit{Planck}-\textit{Herschel}\ galaxies in this study reveals that the CMB effects cannot explain this offset. The excitation conditions of a multi-phase, multiple gas component ISM are also expected to change for each galaxy. One would expect that the density and kinetic temperature of the CO ($1-0$) emitting gas (and the gradients across the galaxy) to factor into the total attenuation of the CO ($1-0$) line emission \citep{tunnard16,tunnard17} and any self-shielding. As the intense star-forming conditions during the redshifts indicated in these three samples (SPT, H-ATLAS, \textit{Planck}-\textit{Herschel}) will give rise to a dynamic set of ISM conditions, these varying gas excitation conditions will therefore have non-negligible effects in the observed $L_{IR}$-to-$L'_{CO(1-0)}$ values. \begin{figure*} \includegraphics[width=15.0cm]{lirlco-sept-b.png} \caption{Here we present the ${\rm L_{IR} / L'_{CO(1-0)}}$ ratios of our sample compared with known, lensed \textit{Herschel}\ and SPT DSFGs \citep{harris12, aravena16}, the highly excited HFLS-3 \citep{riechers13} and the median for all SMGs (${\rm 125 \pm 50 L_\odot / K km s^{-1} pc^{-2} }$) compiled in the literature by \citet{frayer11} (shaded yellow). We plot 2$\sigma$ boundaries taken from \citet{genzel10} for starburst (140 ${\rm L_{\odot} / K km s^{-1} pc^{-2}}$) and typical star-forming galaxies (30 ${\rm L_{\odot} / K \, km s^{-1} pc^{-2}}$). The average for our seven targets in this study is ${\rm 110 \pm 22 \, L_{\odot} / (K km s^{-1} pc^{2})}$. } \label{} \end{figure*} \subsection{Total gas mass from $L'_{CO(1-0)}$} CO is the second most abundant molecule in the ISM after the highly abundant molecular hydrogen, ${\rm H_{2}}$, and the CO($1-0$) line emission is the most direct proxy for ${\rm H_{2}}$ as it traces even the most diffuse gas. Our galaxy integrated CO($1-0$) line luminosity is converted to a total molecular gas mass assuming an $\alpha_{\rm CO}$ conversion factor \citep[see review by][]{bolatto13}. It is common to use a standard ULIRG conversion, i.e. $\alpha_{\rm CO} = 0.8$, for star bursting SMG/DSFGs at high-$z$, although we reference a standard Galactic value in Table~\ref{}. The similarity of the ${\rm L_{IR} / L'_{CO(1-0)}}$ ratios observed in our sample and those of local ULIRGs seems to further support the use of a starburst-like $\alpha_{\rm CO}$ conversion factor, even if the centrally compact, concentrated nuclei in local ULIRGs may not be representative of the entire ISM environments in our galaxies. We found $\mu M_{\rm H2} = (0.9-27.4) \times 10^{11} {\rm M_\odot (\alpha_{\rm CO}/0.8)} $, which are amongst the largest apparent gas contents measured at high-$z$, even if a lensed magnification of an order of magnitude is taken into account \citep[see][]{carilli13}. We also compare our CO-based gas mass to the ISM gas mass estimates using the empirical calibration from measured rest-frame dust continuum \citep[e.g.][]{scoville16, scoville17}. Using our AzTEC 1.1mm photometry ($\nu_{\rm obs} = 272$ GHz $\rightarrow$ rest-frame 250-470$\micron$), we compute the ISM mass using Eq. 14 of \citet{scoville17}. The ISM masses we report scatter predictably around the values obtained from a ULIRG or Galactic conversion factor, suggesting that the value for $\alpha_{\rm CO}$ varies intrinsically from galaxy to galaxy. Later in \S~5.5 we will revist this empirical calibration to compare the CO line luminosity to the specific luminosity at rest-frame 850\micron. \begin{figure*} \includegraphics[width=15.0cm]{codust-sept.png} \caption{We compare our measured $L'_{CO(1-0)}$ to rest-frame $L_{\rm \nu}(353 GHz/850\micron\ )$ in our sample to the low-$z$ star-forming galaxies \citep{dale12, young95}, local ULIRGs \citep{mentuch12, chu17, sanders89, sanders91, solomon97}, $z \sim 2$ SMGs \citep{greve03, harris10, ivison11, carilli11, harris12, riechers11a, lestrade11,thomson12, ivison13, fu13, aravena13,thomson15}, and lensed SPT galaxies \citep{aravena16} with global measurements of CO($1-0$) -- or CO($2-1$) for some SPT sources, where we used $r_{21}= 0.9$ when applicable. We overplot the best fit linear relation from \citet{scoville17}: $L'_{\rm CO(1-0)} = 3.02\times10^{-21} L_{\rm \nu 850}$. } \label{} \end{figure*} \subsection{Gas Depletion Time Scales} The amount of time for a galaxy to consume its total molecular gas, given its current galaxy integrated star formation rate, is its so-called depletion time, or gas consumption time scale, $\tau_{\rm depl} = \mu M_{H2} / \mu SFR$. This inverse SFE reflects the nature of the SF activity of a galaxy, and is a measure that stands independent of the magnification factor in the same way for the ${\rm L_{IR}}$-to-CO(${\rm 1-0}$) line luminosity ratios above. To derive our SFR estimates we used the integrated 8-1000$\micron$ SED and the empirical calibration \citep{kennicutt98} to convert L$_{\rm IR}$ to SFR -- adopting a Kroupa IMF. The values we obtain are, uncorrected for magnification amplification, ${\rm \sim 1500 - 30700 \, M_\odot yr^{-1}}$ \citep{harrington16}. Combined with the CO-based gas masses reported in \S~5.3, this suggests a depletion time scale of $\tau_{\rm depl} \sim 80 \, {\rm Myr}$ These actively evolving galaxies represent a special mode of rapid starburst activity. This is consistent with short gas depletion times observed on the order of $\tau_{\rm depl} = 10-100 \, {\rm Myr}$ \citep[e.g.][]{genzel15,bethermin16,aravena16,scoville16}, and also with typical galaxy-galaxy crossing time ($\sim 100$Myr; \citep{scoville16}). The rapid $\tau_{\rm depl}$ in these galaxies at high-$z$ rival the $\tau_{\rm depl} = 2.2$ Gyr timescales for normal star-forming galaxies at $z=0$ \citep{leroy13}. \subsection{Global Gas to Dust Comparison} The ratio of measured $L'_{CO(1-0)}$ to rest-frame specific luminosity at 850\micron\ serves as a foundation for converting the optically thin Rayleigh-Jeans dust continuum, observed in the (sub)-mm, into total ISM mass \citep{scoville14, scoville16, scoville17}. To infer the rest-frame 850\micron\ of our galaxies, and thus to compare them to the empirical relation, we use the far-IR SED model fit procedure described by \citet{harrington16}, fitting the \textit{Herschel}\ SPIRE 250-500 \micron\ and AzTEC 1.1mm photometry with a modified blackbody \citep[Eq. 14][]{yun02} (Fig.~\ref{}). Several of $z \sim$ 2-3 galaxies lie above the empirical calibration obtained by \citet{scoville17}. In \citet{scoville17} the SED analyses was redone using the published sub-mm photometry and CO($1-0$) line emission for the 30 normal low-$z$ star-forming galaxies \citep{dale12, young95}, 12 low-$z$ ULIRGs \citep{mentuch12, chu17, sanders89, sanders91, solomon97}, and 30 SMGs \citep{greve03, harris10, ivison11, carilli11, harris12, riechers11a, lestrade11,thomson12, ivison13, fu13, aravena13,thomson15} at comparable redshifts to our sample. This empirical relation, based primarily on galaxies with solar metallicities, was recently validated using $\sim$ 70 main-sequence, low-$z$ star-forming galaxies \citep{hughes17}. Without optical or FIR fine-structure emission lines we cannot directly constrain the metallicities of our sample. However, we do not expect these galaxies to have sub-solar metallicities given their large apparent dust masses ($ \mu M_{d} = [0.1-2] \times 10^{10} {\rm M_{\odot} }$) and given the empirical mass-metallicity relationship out to high-$z$ \citep{geach11,saintonge16}. The SMG/DSFG population predominantly falls below the 1:1 relation, making our small sample the first to populate the upper envelope--which corresponds to a higher amount of observed CO gas per unit 850\micron\ dust emission. The highest value of $L'_{\rm CO} / L_{\rm 850}$ observed in the SMG population compiled by \citet{scoville17} is the 350\micron\ selected source in \citet{ivison11}. Two of our galaxies are above the observed scatter, while three of our galaxies exhibit extreme CO luminosities compared to their rest-frame dust luminosity. A larger sample is undoubtedly required to further unveil if, as suggested by our sample, there is a larger intrinsic scatter at high-$z$, particularly at $\log(L'_{CO(1-0)}) > 10.5 {\rm \, K \, km \,s^{-1} \, pc^{2}}$ and $\log(L_{\rm \nu 850}) > 31.5 \, {\rm erg s^{-1} Hz^{-1}}$. To compare to the SPT-DSFGs with J$\leq 2$ CO line detections \citep{aravena16}, we take their 18 galaxies with consistent sampling of 0.25-1.4mm photometry, similar to our 0.25-1.1mm data, and fit their FIR-mm SEDs as described above. Those SPT galaxies with only CO($2-1$) were converted to $L'_{\rm CO(1-0)}$ using an $r_{21}= 0.9$. The relatively high $L'_{\rm CO} / L_{\rm 850}$ ratios observed in our galaxies indicate larger gas-to-dust mass (GDMR) ratios than observed in previous samples (Fig.~\ref{}). Converting the AzTEC 1.1mm continuum measurement into dust mass, assuming $T_{\rm d} = 25$K, we found GDMRs in the range [40-200] using the CO-based gas mass ($\alpha_{\rm CO} = 0.8$), compared with the average GDMR of $\sim 230$ from the 1.1mm derived ISM mass \citep{scoville17}. This range is both consistent, though slightly larger, than observed in local galaxies with solar metallicities \citep{leroy11a,draine07}. Assuming instead a Galactic $\alpha_{\rm CO} = 4.3$, we would infer extremely high GDMRs (up to 1100), only observed in local, greatly metal-poor dwarf galaxies, e.g., the blue compact dwarf, I Zwicky 18, with $1/50$ solar metallacity \citep{annibali15}. The assumption of $\alpha_{\rm CO}$, as well as the choice of dust temperature in the ISM mass calculations ultimately determines the derived GDMRs. \section{Conclusions} Using VEGAS on the \textit{GBT}, we have successfully measured the CO($1-0$) line emission for seven of the most gas-rich SMGs/DSFGs studied to date. The key results of this study can be summarised as: \begin{itemize} \item We have confirmed the previously determined spectroscopic redshifts reported by \citet{harrington16} \item The linewidths/profiles for the low-J RSR and CO($1-0$) VEGAS measurements are nearly identical; therefore the emitting regions are likely co-spatial, with ${\rm < FWHM > = 590}$ km s$^{-1}$, \item The CO SLEDs of the galaxies in our sample are indicative of a gas component with sub-thermal excitation conditions: CO line luminosity ratios of $< r_{21} > = 0.92$ (2 sources), $< r_{31} > =0.34$ (5 sources), and $r_{41} = 0.18$ (1 source) \item We find enhanced $ L_{\rm IR} / L'_{\rm CO(1-0)}$ ratios with respect to normal star-forming systems, as we report an average value of ${\rm 110 \pm 22 \, L_\odot / (K km s^{-1} pc^{2})}$, comparable to the median of other well studied SMGs. \end{itemize} With the CO($1-0$) line emission we place tight constraints on the total molecular gas mass, and allow future CO SLED analyses to benefit from having the fundamental rotational transition observed. The large gas masses obtained are $ \mu M_{H2} = (0.9 - 27) \times 10^{11} (\alpha_{\rm CO}/0.8) {\rm M_\odot }$. The average gas depletion time we find is $\tau_{\rm depl} \sim 80 {\rm Myr}$. These extremely luminous IR galaxies (with $L_{IR} \ge 10^{13-14} L\odot$) exhibit rapid depletion timescales, and we are likely capturing this light from a relatively short-lived starburst episode. \section{Acknowledgements} The quality of this manuscript, and the scientific considerations, have been improved significantly after discussions with N. Scoville, L. Liang and M. Sargent. All authors would like to thank the referee for their kind considerations of how this paper can be developed so as to make its most useful contribution to the field. KCH and DB would like to also thank Karen O'Neil and Toney Minter for organizing, and supporting our travel to, the first \textit{GBT}\ remote observing school at the Green Bank Telescope. KCH would like to express great thanks to D. Frayer and all of the control room operators (specifically Amber and Donna) for making the observations run as smooth as possible. The Green Bank Observatory is a facility of the National Science Foundation operated under cooperative agreement with Associated Universities, Inc. KCH and DB heartily acknowledge our financial support granted to us by UMass Amherst's Astronomy Department, while KCH much appreciated the ongoing support from the UMass Amherst Commonwealth Honors College Research Grant which led to the travel to the \textit{GBT}\ remote observing school in January, 2016. KCH, AK and EFJA acknowledge the support by the Collaborative Research Center (CRC) 956, subproject A1, funded by the Deutsche Forschungsgemeinschaft (DFG). KCH acknowledges the financial support from the CRC-956 student exchange program. This opportunity has enabled KCH to engage in useful discussions and supervision with D. Riechers, who provided strong guidance in the final stages of this manuscript. Support for BM was provided by the DFG priority program 1573 'The physics of the interstellar medium'. KCH is a member of, and receives financial support for his research by, the International Max-Planck research school for Astronomy and Astrophysics at the Universities of Bonn and Cologne and through the Max-Planck-Institut f\"{u}r Radioastronomie.

\section{Conclusions and outlook} \label{sec::conclusion} We performed a freed-isobar analysis of COMPASS data for the \process channel, based in the analysis published in ref.~\cite{big::compass}, using an extended set of 12 freed waves. In this analysis model, we identified continuous ambiguities---zero modes---and showed a way to resolve them using additional constraints on the dynamic isobar amplitudes. With this method, we were able to extract the dynamic isobar amplitude for the spin-exotic $1^{-+}1^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Pw$ wave from the data, shown in \cref{fig::result}. Even though this dynamic isobar amplitude is dominated by the \PrhoM resonance, it shows significant deviations from a simple Breit-Wigner shape. As next steps, we will extend the analysis to all bins in the three-pion mass and the squared four-momentum transfer. We will also free the dynamic isobar amplitudes of other interesting waves, for example in the $\JPC=3^{++}$ or $4^{++}$ sectors. \section{Ambiguities in the freed-isobar approach} \label{sec::freedIsobar} With increasing number of freed-isobar waves, the flexibility of the analysis model drastically increases, since one complex-valued free parameter appears for every \mTwoPi bin of every freed-isobar wave. For some combination of freed waves, this flexibility allows to approximate continuous ambiguities in the analysis model, that have to be identified an resolved. Such ambiguities will be called zero modes in this article. A more detailed discussion of zero modes can be found in ref.~\cite{freed::isobar}. \subsection{Appearance of zero modes} Since in the freed-isobar approach, the dynamic isobar amplitudes of freed waves can take any shape---up to the finite \mTwoPi bin width---the zero modes correspond to specific shapes $\zero_i\lr{\mTwoPi}$ of these dynamic amplitudes, that cancel exactly to zero at evert point \ps in phase space: \begin{equation} \sum_{i\,\in\,\text{waves}}\angular_i\lrps\zero_i\lr{\mTwoPi}+\boseText = 0 ,\end{equation} where the sum runs over all waves that are affected by the particular zero mode. Since the cancellation is exact, a shift of the isobar dynamic amplitudes $\shape_i\lr{\mTwoPi}$ by these shapes $\zero_i\lr{\mTwoPi}$ does not change the overall amplitude and thus the intensity. The $\zero_i\lr{\mTwoPi}$ constitute a continuous ambiguity in a freed-isobar analysis: \begin{equation*} \intensAbs{\sum_{i\,\in\,\text{waves}}\angular_i\lrps\shape_i\lr{\mTwoPi}+\boseText}=\intensAbs{\sum_{i\,\in\,\text{waves}}\angular_i\lrps\LR{\shape_i\lr{\mTwoPi}+\zeroCoeff\zero_i\lr{\mTwoPi}}+\boseText} .\end{equation*} This ambiguity is represented by the arbitrary complex-valued coefficient \zeroCoeff. \subsection{Resolution of zero modes} \label{sec::zeroModeResolution} The presence of zero modes may affect the result of a freed-isobar PWA and has to be accounted for, since the extracted dynamic isobar amplitudes $\binVal^{\text{fit},b}$ can be shifted away from the physical ones $\binVal^{\text{phys},b}$: \begin{equation} \binVal^{\text{phys},b}_{i}=\binVal^{\text{fit},b}_i+\zeroCoeff\binVal^{0,b}_i ,\end{equation} where $\binVal^{0,b}_i$ is the shape $\zero_i\lr{\mTwoPi}$ of the zero mode for wave $i$ evaluated at the \mTwoPi bin $b$. To remove the effect of the zero mode, we have to determine \zeroCoeff such, that the resulting combination $\binVal^{\text{fit},b}_i+\zeroCoeff\binVal^{0,b}_i$ resembles the physical solution. This is done, by performing a second fit step, where the following $\chi^2$ function is minimized: \begin{equation}\label{eqn::chi2} \chi^2\lr{\zeroCoeff}=\sum_{i,b}\sum_{j,c}\delta_i^b\lr{\zeroCoeff}\lr{\mathbf{C}^{-1}}_{i,j}^{b,c}\delta_j^c\lr{\zeroCoeff} ,\end{equation} where $\mathbf{C}_{i,j}^{b,c}$ is the covariance matrix of the $\binVal^{\text{fit},b}_i$ obtained by the minimizing algorithm and the differences $\delta_i^b\lr{\zeroCoeff}$ are given by: \begin{equation} \delta_i^b\lr{\zeroCoeff} = \binVal^{\text{fit},b}_i+\zeroCoeff\binVal_i^{0,b}-\binVal_i^{\text{model},b} .\end{equation} The values $\binVal_i^{\text{model},b}$ are the values of a model for the dynamic isobar amplitude of wave $i$ evaluated at the corresponding \mTwoPi bin. Such a model can, for example, be a Breit-Wigner amplitude, if the corresponding wave is known to be dominated by a single resonance, as is true for the \PrhoM waves with $\LR{\Ppi\Ppi}_{1^{--}}$ isobar, or for the \PfTwoM isobar for waves with $\LR{\Ppi\Ppi}_{2^{++}}$ isobar. This second fit adjusts only one complex-valued parameter, while the freed-isobar fit extracts a complex-valued parameter for every \mTwoPi bin. Therefore, most of the information obtained with the freed-isobar method is kept, even in this second fit step is performed. To demonstrate, that resolving zero-mode ambiguities cannot change the resulting dynamic isobar amplitudes arbitrarily, we generated a Monte Carlo data set, analyzed it with the freed-isobar method, and resolved the zero-mode ambiguities using a Breit-Wigner parameterization with three different values for mass and width as constraints. The results of this study are shown in \cref{fig::wrongFixing}. It can be seen, that the resulting position of the \PrhoM peak (blue points) is nearly independent of the chosen constraint (gray line) and that using the correct constraint yields the Monte Carlo input shape. \begin{figure} \includegraphics[width=.32\textwidth]{./paramsOnIntens.pdf} \includegraphics[width=.32\textwidth]{./paramsOffIntens.pdf} \includegraphics[width=.32\textwidth]{./paramsFarIntens.pdf} \caption{Results of the freed-isobar fit before (red) and after resolving of the zero mode ambiguity (blue) with different constraints (gray line). The shape used to generate the Monte Carlo data is indicated by light blue points.\label{fig::wrongFixing}} \end{figure} \subsection{Example: spin-exotic wave} \label{sec::spinexotic} One example for the appearance of a zero mode is the freed spin-exotic $1^{-+} 1^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Pw$ wave. In this case, the zero mode is contained within this single wave. Therefore, the shape of the zero-mode has to fulfill: \begin{equation}\label{eqn::exoticZeroCondition} \angular_{1^{-+}}\lrps\zero_{1^{-+}}\lr{\mTwoPi}+\boseText=0 .\end{equation} The angular amplitude for the spin-exotic wave is given by: \begin{equation} \angular_{1^{-+}}\lrps\propto\vec p_1\times\vec p_3 ,\end{equation} where $\vec p_{1,3}$ are the three-momenta of the two \Ppim in the final state in the rest frame of the three-pion system. Inserting this into \cref{eqn::exoticZeroCondition} and using the antisymmetry of the cross product, we obtain: \begin{equation}\label{eqn::exoticZeroCondition2} 0=\zero_{1^{-+}}\lr{\mTwoPi}\vec p_1\times \vec p_3 + \zero_{1^{-+}}\lr{\mTwoPiBose}\vec p_3\times \vec p_1 = \vec p_1\times\vec p_3\LR{\zero_{1^{-+}}\lr{\mTwoPi}-\zero_{1^{-+}}\lr{\mTwoPiBose}} ,\end{equation} where \mTwoPiBose is the Bose symmetrized invariant two-pion mass with final-state particles 1 and 3 interchanged. \Cref{eqn::exoticZeroCondition2} is fulfilled, if $\zero\lr{\mTwoPi}$ is a constant, which is the shape of the zero mode in this wave. To resolve the ambiguity caused by this zero mode we use the method introduced in \cref{sec::zeroModeResolution}, using a Breit-Wigner parameterization with fixed mass and width for the \PrhoM resonance as constraint. To check consistency we try a second approach, where we leave the mass and width of this parameterization as additional free parameters in the minimization of \cref{eqn::chi2}. As final value for \zeroCoeff, we use the weighted average of both approaches. \section{Freed-isobar partial wave analysis} \label{sec::introduction} Due to the large data set collected by COMPASS, more elaborate analysis methods become applicable. One of these methods is the so called freed-isobar PWA. This approach is an extension of the conventional three-body PWA, that relies on the isobar model. In this model, the process is modeled as a two step decay of an intermediary state $\PX$ into one of the final state \Ppim and another intermediary state \Pxi, the isobar. The isobar subsequently decays into \Ppip \Ppim. Within this model the intensity distribution $\intens\lrps$ for the $3\Ppi$ final state is modeled as coherent superposition of waves with given angular momentum quantum numbers and known isobar resonances. These quantum numbers are the spin \Jj and eigenvalues under parity and generalized charge conjugation---\PP and \Cc---, the spin projection \Mm, and reflectivity \eps of \PX, as well as the analogous quantum numbers $\JPC_\Pxi$ of the isobar. With these quantum numbers defining a partial wave, we can write the intensity distribution: \begin{equation} \intens\lrps=\intensAbs{\sum_{i\,\in\,\text{waves}}\prodAmp_i\left[\angular_i\lrps\shape_i\lr{\mTwoPi} + \boseText\right]} ,\end{equation} as a function of the phase space variables \ps. The complex-valued production amplitudes $\prodAmp_i$ describe the intensities and relative phases of the partial waves, the angular amplitudes $\angular_i\lrps$ describe their angular distributions, and the dynamic amplitudes $\shape\lr{\mTwoPi}$ describe the dependence of the amplitudes of the $\Ppip\Ppim$ system on \mTwoPi. The production amplitudes are free parameters in a likelihood fit to the measured \ps distribution and the angular amplitudes are fully determined by the appearing angular-momentum quantum numbers. However, the dynamic amplitudes are not determined by first principle, but are a necessary model input; for the most simple case of isolated $\Ppip\Ppim$ resonances, Breit-Wigner amplitudes are commonly used. Due to computational reasons, the dynamic amplitudes cannot have any free fit parameters. To avoid the necessity for this model input in the freed-isobar approach, the dynamic amplitudes are replaced by sets of step-like functions, that approximate the dynamic amplitude by a set of constants $\binVal^b$ within narrow bins $b$ of \mTwoPi \cite{first::fi}: \begin{equation} \shape_i\to\sum_{b\,\in\,\text{bins}}\binVal_i^b\step^b\lr{\mTwoPi},\quad\text{where}\quad\step^b\lr{\mTwoPi}=\begin{cases}1 & \text{if \mTwoPi in $b$},\\ 0 & \text{otherwise}. \end{cases} \end{equation} In this approach, the step-like functions appear like ``normal'' partial waves in the fit. This allows us to use the same likelihood method as in the conventional PWA and determine the dynamic isobar amplitudes encoded by the free parameters $\binVal^b$ directly from the data. Such a set of steplike functions will be denoted will be denoted by $\LR{\Ppi\Ppi}_{\JPC_\Pxi}$, where $\JPC_\Pxi$ are the \JPC quantum numbers of the freed isobar. Partial waves with steplike dynamic isobar amplitudes will be called freed-isobar waves, waves with fixed parameterizations of the dynamic isobar amplitudes will be called fixed-isobar waves. Within the freed-isobar approach, models with any combination of freed- and fixed-isobar waves are possible. \section{Extended freed-isobar PWA} \label{sec::newStudy} Based on the first freed-isobar analysis published in ref.~\cite{big::compass}, we performed an extended freed-isobar analysis with twelve freed waves: \begin{equation}\begin{aligned} 0^{-+}0^+\LR{\Ppi\Ppi}_{0^{++}}\Ppi\Sw\quad &\ 1^{++}0^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Sw\ &\ 2^{-+}0^+\LR{\Ppi\Ppi}_{0^{++}}\Ppi\Dw\quad &\ 2^{-+}0^+\LR{\Ppi\Ppi}_{2^{++}}\Ppi\Sw\\ 0^{-+}0^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Pw\quad &\ 1^{++}1^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Sw\ &\ 2^{-+}0^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Pw\quad &\ 2^{-+}1^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Pw\\ 1^{++}0^+\LR{\Ppi\Ppi}_{0^{++}}\Ppi\Pw\quad &\ 1^{-+}1^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Pw\ &\ 2^{-+}0^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Fw\quad &\ 2^{++}1^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Dw .\end{aligned}\end{equation} To minimize potential leakage effects 11 of the freed waves were chosen to be the eleven waves with the larges intensity in the conventional analysis in ref.~\cite{big::compass}. In addition, the spin-exotic $1^{-+}1^+\PrhoM\Ppi\Pw$ wave was freed, since it is a wave of major interest in the analysis. With this choice of freed waves, the analysis model has 12 fixed- and 72 freed-isobar waves. Due to the high flexibility of this model, five zero modes appear: One in the $\JPC\Mrefl=0^{-+}$ sector, one in the $1^{++}0^+$ sector, two in the $2^{-+}0^+$ sector, and one on the spin-exotic wave, as discussed in \cref{sec::spinexotic}. In this article, we only show results for the spin-exotic wave, therefore only the zero mode in this wave has to be resolved, which we do as described in \cref{sec::spinexotic}. The result for the dynamic isobar amplitude for the spin-exotic wave is shown in \cref{fig::result}. The resulting dynamic isobar amplitude is dominated by the \PrhoM resonance, which justifies the use of this resonance as constraint for resolving the zero-mode ambiguity. This is also supported by the fact, that resolving the ambiguity with floating \PrhoM mass and width yields compatible results. These are shown in light blue in \cref{fig::result}. Even though the dynamic isobar amplitude is dominated by the \PrhoM resonance, we find, that the exact shape differs significantly from a pure Breit-Wigner shape, since the low mass tail rises slower and the high mass tail falls slower, than a pure Breit-Wigner amplitude. Reasons for this behavior can be rescattering effects with the third pion in the final state or non-resonant contributions to the process. \begin{figure} \includegraphics[width=.5\textwidth]{./1mp1p1mm_t3_rhoExample_scale.pdf} \includegraphics[width=.5\textwidth]{./1mp1p1mm_t3_rhoArgand_scale.pdf} \caption{Intensity distribution and Argand diagram for the dynamic-isobar amplitude for the $1^{-+}1^+\LR{\Ppi\Ppi}_{1^{--}}\Ppi\Pw$ wave in a bin of the three-pion mass and the squared four momentum transfer $t^\prime$.\label{fig::result}} \end{figure} \section{Freed-isobar PWA by COMPASS} \label{sec::oldStudy} A first freed-isobar analysis based on an extensive PWA of a data set of $46\times10^6$ events for the process $\process$ collected by COMPASS in 2008 was published in ref.~\cite{big::compass} using a set of 88 waves. In this first analysis, seven fixed-isobar waves are replaced by three freed-isobar ones, due to matching angular momentum quantum numbers: \begin{equation}\begin{aligned} &0^{-+}0^+\PfZeroS\Ppi\Sw,\ 0^{-+}0^+\PfZeroN\Ppi\Sw,\ 0^{-+}0^+\PfZeroF\Ppi\Sw\ &\to&\ 0^{-+}0^+\LR{\Ppi\Ppi}_{0^{++}}\Ppi\Sw\\ &1^{++}0^+\PfZeroS\Ppi\Pw,\ 1^{++}0^+\PfZeroN\Ppi\Pw\ &\to&\ 1^{++}0^+\LR{\Ppi\Ppi}_{0^{++}}\Ppi\Pw\\ &2^{-+}0^+\PfZeroS\Ppi\Dw,\ 2^{-+}0^+\PfZeroN\Ppi\Dw\ &\to&\ 2^{-+}0^+\LR{\Ppi\Ppi}_{0^{++}}\Ppi\Dw .\end{aligned}\end{equation} This first study yielded very promising results, that helped understand the $\LR{\Ppi\Ppi}_{0^{++}}$ wave and to eliminate the possibility that the \PaOneF resonance-like signal, first observed by COMPASS in ref.~\cite{a1::1420}, is an artifact of the parameterization of the \PfZeroN dynamic isobar amplitude. Inspired by these results, we made attempts to extend the freed-isobar approach to a larger set of partial waves.

\section{Introduction} Several applications are characterized by multi-physics phenomena, such as the interaction between an incompressible fluid and a compressible structure. The capability to perform real-time multi-physics simulations could greatly increase the applicability of computational methods in applied sciences and engineering. To reach this goal, reduced-order modelling techniques are applied in this paper. We refer the interested reader to \cite{Amsallem2010,LassilaQuarteroniRozza2012,Colciago2014,BallarinRozza2016} for some representative previous approaches to the reduction of fluid-structure interaction problems, arising in aeroelasticity \cite{Amsallem2010} or haemodynamics \cite{LassilaQuarteroniRozza2012,Colciago2014,BallarinRozza2016}. The reduction proposed in the current work is based on a POD--Galerkin approach. A difference with our previous work \cite{BallarinRozza2016}, where a reduced-order monolithic approach has been proposed, is related to the use of a partitioned reduced-order model, based on the semi-implicit operator-spitting approach originally employed in \cite{Fernandez2007} for the high-fidelity method. Extension to other methods (e.g. \cite{BadiaQuaini2008,Guidoboni2009}) is possible and object of forthcoming work. The formulation of the FSI problem is summarized in Section \ref{}, and its high-fidelity discretization is reported in Section \ref{sec:HF}. Two reduced-order models are proposed in Section \ref{}, and compared by means of a numerical test case in Section \ref{}. Conclusions and perspectives are summarized in the final section of the paper. \section{Problem formulation} \label{} In this section the formulation of the fluid-structure interaction (FSI) model problem is summarized. Let us consider the bidimensional fluid domain $\Omega = [0, L] \times [0, h_f]$. Its boundary is composed of a compliant wall $\Sigma = [0, L] \times \{h_f\}$ (top), fluid inlet section $\Gamma_{in} = \{0\} \times [0, h_f]$ (left) and fluid outlet section $\Gamma_{out} = \{L\} \times [0, h_f]$ (right), and a wall $\Gamma_{sym} = [0, L] \times \{0\}$ (bottom). For the sake of simplicity we assume low Reynolds numbers for the fluid problem and infinitesimal displacements for the compliant wall. Thus, in the following, we will consider \emph{unsteady Stokes} equations on a \emph{fixed} domain $\Omega$ for the fluid and a linear structural model for the compliant wall. In particular, we will further assume that the structure undergoes negligible horizontal displacements, so that the structural equations of the compliant wall can be described by a \emph{generalized string model} \cite{QuarteroniFormaggia2004,QuarteroniTuveriVeneziani2000}. The coupled fluid-structure interaction problem is therefore: for all $t \in (0, T]$, find fluid velocity $\u(t): \Omega \to \RR^2$, fluid pressure $p(t): \Omega \to \RR$ and structure displacement $\eta(t): \Sigma \to \RR$ such that \begin{equation} \begin{cases} \rho_f \partial_t \u - \div(\boldsymbol{\sigma}(\u, p)) = \vector{0} & \text{ in } \Omega \times (0, T],\\ \div \u = 0 & \text{ in } \Omega \times (0, T],\\ \u = \partial_t \eta\, \ensuremath{\vector{n}} & \text{ on } \Sigma \times (0, T],\\ \rho_s h_s \partial_{tt} \eta - c_1 \partial_{xx} \eta + c_0 \eta = - \boldsymbol{\sigma}(\u, p) \ensuremath{\vector{n}} \cdot \ensuremath{\vector{n}} & \text{ on } \Sigma \times (0, T]. \end{cases} \label{} \end{equation} Equations \eqref{ formulate the Stokes problem on the fixed fluid domain $\Omega$, having defined the fluid Cauchy stress tensor $\boldsymbol{\sigma}(\u, p) := - p \boldsymbol{I} + 2 \mu_f \boldsymbol{\varepsilon}(\u)$, $\boldsymbol{\varepsilon}(\u) := \frac{1}{2} (\grad\u + \grad^T \u)$, while \eqref{ is the equation of the structural motion of the compliant wall. Continuity of the velocity on the interface is guaranteed by \eqref{. The FSI system is completed by suitable initial and boundary conditions. In this paper we assume resting conditions at $t = 0$ and the following fluid boundary conditions on $\partial\Omega \setminus \Sigma$: \begin{equation} \begin{cases} \boldsymbol{\sigma}(\u, p) \ensuremath{\vector{n}} = - p_{in}(t) \ensuremath{\vector{n}} & \text{ on } \Gamma_{in} \times (0, T],\\ \boldsymbol{\sigma}(\u, p) \ensuremath{\vector{n}} = - p_{out}(t) \ensuremath{\vector{n}} & \text{ on } \Gamma_{out} \times (0, T],\\ \u \cdot \ensuremath{\vector{n}} = 0, \quad \boldsymbol{\sigma}(\u, p) \ensuremath{\vector{n}} \cdot \vector{\tau} = 0 & \text{ on } \Gamma_{sym} \times (0, T],\\ \end{cases} \label{} \end{equation} and the following structure boundary condition on $\partial\Sigma$: \begin{equation} \eta = 0 \quad \text{ on } \partial\Sigma \times (0, T]. \label{} \end{equation} Equations \eqref{ and \eqref{ prescribe pressure on inlet and outlet section, respectively; equation \eqref{ is a symmetry condition, arising from the consideration of this simplified 2D problem as a section of a 3D cylindrical configuration. Here $\ensuremath{\vector{n}}$ and $\vector{\tau}$ denote the outer unit normal to $\Omega$ and tangential vector to $\Sigma$, respectively. Finally, \eqref{} prescribes a clamped wall near both the inlet and outlet section of the fluid. The value of constitutive parameters $\rho_f$ (fluid density), $\rho_s$ (structure density), $\mu_f$ (fluid viscosity), $c_1$ and $c_0$ (structure constitutive parameters), $L$ (domain width), $h_f$ (fluid height), $h_s$ (structure thickness) will be further specified in Section \ref{}, as well as choices of $p_{in}(t)$, $p_{out}(t)$ and $T$ for which the FSI system \eqref{} is characterized by propagation of a pressure wave. \section{High-fidelity formulation: semi-implicit scheme} \label{sec:HF} In this section we summarize the high-fidelity discretization of the FSI system \eqref{}. An operator splitting approach, based on a Chorin-Temam projection scheme, is pursued. In particular, Robin-Neumann iterations are carried out in order to enhance the stability of the resulting algorithm. \subsection{A projection-based semi-implicit coupling scheme} We employ in this work a projection-based semi-implicit scheme, as proposed in \cite{Fernandez2007,Astorino2010}. The fluid equations are discretized in time using the Chorin-Temam projection scheme \cite{QuarteroniValli2008}. Thus, denoting by $\Delta t$ the time step length, $D_t f^{k+1} := \frac{f^{k+1} - f^{k}}{\Delta t}$ the first backward difference approximation of the time derivative of $f(t^{n+1})$ and $D_{tt} f^{k+1} = D_t(D_t f^{k+1})$, we consider the following semi-implicit time discretization of \eqref{}: for any $k = 1, \hdots, K = T/\Delta t$ \begin{enumerate} \item Explicit step (fluid viscous part): find\footnote{For the sake of an easier comparison with Remark 1 we employ here the $\widetilde{\cdot}$ notation for the velocity. However, since we will always employ the pressure Poisson formulation, the $\widetilde{\cdot}$ will be dropped in the next sections.} $\widetilde{\u}^{k+1}: \Omega \to \RR^2$ such that: \begin{equation} \begin{cases} \rho_f \frac{\widetilde{\u}^{k+1} - \widetilde{\u}^{k}}{\Delta t} - 2 \mu_f \div\, \boldsymbol{\varepsilon}(\widetilde{\u}^{k+1}) = - \grad p^k & \text{ in } \Omega,\\ \widetilde{\u}^{k+1} = D_t \eta^k\, \ensuremath{\vector{n}} & \text{ on } \Sigma. \end{cases} \label{} \end{equation} \item Implicit step: \begin{enumerate} \item[2.1.] Fluid projection substep: find $p^{k+1}: \Omega \to \RR$ such that: \begin{equation} \begin{cases} - \div(\grad p^{k+1}) = - \frac{\rho_f}{\Delta t} \div \widetilde{\u}^{k+1} & \text{ in } \Omega,\\ \frac{\partial}{\partial \ensuremath{\vector{n}}} \ensuremath{p}^{k+1} = - \rho_f D_{tt} \eta^{k+1} & \text{ on } \Sigma. \end{cases} \label{} \end{equation} \item[2.2.] Structure substep: find $\eta^{k+1}: \Sigma \to \RR$ such that: \begin{equation} \rho_s h_s D_{tt} \eta^{k+1} - c_1 \partial_{xx} \eta^{k+1} + c_0 \eta^{k+1} = - \boldsymbol{\sigma}(\widetilde{\u}^{k+1}, p^{k+1}) \ensuremath{\vector{n}} \cdot \ensuremath{\vector{n}} \quad \text{ on } \Sigma. \label{} \end{equation} \end{enumerate} \end{enumerate} The implicit step couples pressure stresses to the structure, and it is iterated until convergence. \begin{remark} \label{} In place of step 1 and 2.1 (pressure Poisson formulation) one could also consider the following pressure Darcy formulation: \begin{enumerate} \item[I.] Explicit step (fluid viscous part): find $\widetilde{\u}^{k+1}: \Omega \to \RR^2$ such that: \begin{equation*} \begin{cases} \rho_f \frac{\widetilde{\u}^{k+1} - \u^{k}}{\Delta t} - 2 \mu_f \div\, \boldsymbol{\varepsilon}(\widetilde{\u}^{k+1}) = \vector{0} & \text{ in } \Omega,\\ \widetilde{\u}^{k+1} = D_t \eta^k\, \ensuremath{\vector{n}} & \text{ on } \Sigma. \end{cases} \end{equation*} \item[II.] Implicit step: \begin{enumerate} \item[II.1.] Fluid projection substep: find $\u^{k+1}: \Omega \to \RR^2$ and $p^{k+1}: \Omega \to \RR$ such that: \begin{equation*} \begin{cases} \rho_f \frac{\u^{k+1} - \widetilde{\u}^{k+1}}{\Delta t} + \grad p^{k+1} = \vector{0} & \text{ in } \Omega,\\ \u^{k+1} \cdot \ensuremath{\vector{n}} = D_t \eta^{k+1} & \text{ on } \Sigma. \end{cases} \end{equation*} \item[II.2.] Structure substep: find $\eta^{k+1}: \Sigma \to \RR$ such that: \begin{equation*} \rho_s h_s D_{tt} \eta^{k+1} - c_1 \partial_{xx} \eta^{k+1} + c_0 \eta^{k+1} = - \boldsymbol{\sigma}(\widetilde{\u}^{k+1}, p^{k+1}) \ensuremath{\vector{n}} \cdot \ensuremath{\vector{n}} \quad \text{ on } \Sigma. \end{equation*} \end{enumerate} \end{enumerate} For the sake of a more efficient reduced-order model (see Section \ref{}) it is convenient to consider the pressure Poisson formulation (steps 1 and 2.1) rather than the pressure Darcy formulation (step I and II.1), \reviewerA{because the latter would require a larger system (comprised of both velocity and pressure) at step II.1}. \end{remark} Finally, in order to enhance the stability of the projection method we employ a Robin-Neumann coupling, as proposed in \cite{Astorino2010}. See also \cite{BadiaVergara2008,Fernandez2016} for related topics. Thus, we replace \eqref{ with \begin{equation} \frac{\partial}{\partial \ensuremath{\vector{n}}} \ensuremath{p}^{k+1} + \alpha_{Rob} \ensuremath{p}^{k+1} = - \rho_f D_{tt} \eta^{k+1} + \alpha_{Rob} \ensuremath{p}^{k, *} \quad \text{ on } \Sigma. \end{equation} being $\alpha_{Rob} > 0$ and $\ensuremath{p}^{k, *}$ an extrapolation of the pressure \reviewerA{(which will be defined in section 3.2)}. In particular, following \cite{Fernandez2013a,Fernandez2013b}, we choose $\alpha_{Rob} := \frac{\rho_f}{\rho_s h_s}$. \reviewerA{We remark that, due to the simplifying assumptions of this problem (linear structural model, fixed domain), the implicit step could have been solved in one shot, since it defines a linear system in $(p^{k+1}, \eta^{k+1})$. We still keep the Robin-Neumann coupling in this case in order to assess the capabilities of such procedure in a reduced-order setting, since it will be required in more general nonlinear problems.} \subsection{Space discretization of the high-fidelity formulation} Denote by $V = [H^1(\Omega)]^2$ the fluid velocity space (endowed with the $H^1$ seminorm), by $Q=L^2(\Omega)$ the fluid pressure space (endowed with the $L^2$ norm), and by $E=H^1(\Sigma)$ the structure displacement space (endowed with the $H^1$ seminorm). After having obtained a weak formulation of the semi-implicit formulation, we consider a finite element (FE) discretization for steps 1, 2.1 and 2.2. Second order Lagrange FE are employed for fluid velocity (step 1) and structural displacement (step 2.2), resulting in FE spaces $V_h \subset V$ and $E_h \subset E$, respectively, while fluid pressure is discretized by first order Lagrange FE, $Q_h \subset Q$. Thus, the corresponding Galerkin-FE formulation reads: for any $k = 1, \hdots, K$ \begin{enumerate} \item[1$_h$.] Explicit step (fluid viscous part): find ${\u}^{k+1}_h \in V_h$ such that: \begin{equation} \int_\Omega \frac{\rho_f}{\Delta t} \u^{k+1}_h \cdot \v_h \,d\ensuremath{\vector{x}} \,+ \int_\Omega 2 \mu_f \boldsymbol{\varepsilon}({\u}^{k+1}_h) : \grad\v_h \,d\ensuremath{\vector{x}} = \int_\Omega \frac{\rho_f}{\Delta t} \u^{k}_h \cdot \v_h \,d\ensuremath{\vector{x}} \, - \int_\Omega \grad p^k_h \cdot \v_h \,d\ensuremath{\vector{x}} \label{} \end{equation} for all $\v_h \in V_h$, subject to the coupling condition \begin{equation} {\u}^{k+1}_h = D_t \eta^k_h\, \ensuremath{\vector{n}} \quad \text{ on } \Sigma \times (0, T], \label{} \end{equation} and to the boundary condition ${\u}^{k+1}_h \cdot \ensuremath{\vector{n}} = 0$ on $\Gamma_{sym} \times (0, T]$. \item[2$_h$.] Implicit step: for any $j = 0, \hdots$, until convergence: \begin{enumerate} \item[2.1$_h$.] Fluid projection substep, with Robin boundary conditions: find $p^{k+1,j+1}_h \in Q_h$ such that: \begin{align*} \int_\Omega \grad p^{k+1,j+1}_h \cdot \grad q_h \,d\ensuremath{\vector{x}} &+ \int_\Sigma \alpha_{Rob}\, p^{k+1,j+1}_h \, q_h \,ds = \\ & - \int_\Omega \frac{\rho_f}{\Delta t} \div {\u}^{k+1}_h \, q_h \,d\ensuremath{\vector{x}} - \int_\Sigma \rho_f D_{tt} \eta^{k+1,j}_h \, q_h \,ds \\ & + \int_\Sigma \alpha_{Rob}\, p^{k+1,j}_h \, q_h \,ds \end{align*} for all $q_h \in Q_h$, subject to the boundary conditions $p^{k+1,j+1}_h = p_{in}(t)$ on $\Gamma_{in} \times (0, T]$ and $p^{k+1,j+1}_h = p_{out}(t)$ on $\Gamma_{out} \times (0, T]$. \reviewerA{Here the value $p^{k+1,j}_h$ has been chosen as pressure extrapolation for the Robin-Neumann coupling.} \item[2.2$_h$.] Structure substep: find $\eta^{k+1,j+1}_h \in E_h$ such that: \begin{align} \int_\Sigma \frac{\rho_s h_s}{\Delta t^2} \, \eta^{k+1,j+1}_h \, \zeta_h \,ds & + \int_\Sigma c_1 \partial_{x} \eta^{k+1,j+1}_h \, \partial_x \zeta_h \,ds \notag\\ &+ \int_\Sigma c_0 \eta^{k+1,j+1}_h \, \zeta_h \,ds = \int_\Sigma \frac{\rho_s h_s}{\Delta t^2} \eta^{k}_h \, \zeta_h \,ds \label{ & + \int_\Sigma \frac{\rho_s h_s}{\Delta t} D_t \eta^{k}_h \, \zeta_h \,ds - \int_\Sigma \boldsymbol{\sigma}({\u}^{k+1}, p^{k+1,j+1}) \ensuremath{\vector{n}} \cdot \zeta_h \ensuremath{\vector{n}} \,ds\notag \end{align} for all $\zeta_h \in E_h$, subject to the boundary conditions $\eta^{k+1,j+1}_h = 0$ on $\partial\Sigma$. \end{enumerate} \end{enumerate} The coupling condition \eqref{} is imposed strongly. We will further comment in Section \ref{} on the imposition of this condition at the reduced-order level. A relative error on the increments is chosen as stopping criterion for step $2_h$, that is the implicit step is repeated until \begin{equation*} \min\left\{\frac{\Norm{p^{k+1,j+1}_h - p^{k+1,j}_h}_{Q}}{\Norm{p^{k+1,j+1}_h}_{Q}}, \frac{\Norm{\eta^{k+1,j+1}_h - \eta^{k+1,j}_h}_{E}}{\Norm{\eta^{k+1,j+1}_h}_{E}}\right\} < \text{tol}, \end{equation*} for some prescribed tolerance $\text{tol}$. The solution $(p^{k+1,j^*}_h, \eta^{k+1,j^*}_h)$ at the iteration $j^*$ such that convergence is achieved is then denoted by $(p^{k+1}_h, \eta^{k+1}_h)$. \section{Reduced-order formulation: POD--Galerkin semi-implicit scheme} \label{} In this section we propose two Proper Orthogonal Decomposition (POD)--Galerkin semi-implicit reduced-order models (ROMs) for FSI system \eqref{}. The first ROM (FSI ROM 1) is built starting directly from steps $1_h$, $2.1_h$ and $2.2_h$, and performing a Galerkin projection. Special treatment will be devoted to the imposition of the coupling condition \eqref{}; unfortunately, this requires enlarging the reduced-order systems. The second ROM (FSI ROM 2) will exploit a simple change of variable for the fluid velocity to bypass this issue. In both cases, an offline-online computational decoupling is sought \cite{RozzaHuynhPatera2007}. \subsection{FSI ROM 1 approach} \subsubsection{Offline stage} \label{} During the offline stage, the solution of the high-fidelity problem $1_h$, $2.1_h$ and $2.2_h$ is computed. We then consider the following snapshot matrices \begin{align*} S_{\u} = [\underline{\mathbf{u}}^{1}_h | \hdots | \underline{\mathbf{u}}^{K}_h] \in \RR^{N_h^{\u} \times K},\\ S_{\ensuremath{p}} = [\underline{\mathbf{p}}^{1}_h | \hdots | \underline{\mathbf{p}}^{K}_h] \in \RR^{N_h^{\ensuremath{p}} \times K},\\ S_{\eta} = [\underline{\boldsymbol{\eta}}^{1}_h | \hdots | \underline{\boldsymbol{\eta}}^{K}_h] \in \RR^{N_h^{\eta} \times K}, \end{align*} where we denote with the underlined notation the vector of FE degrees of freedom corresponding to each solution. Here $N_h^{\u} = \text{dim}(V_h)$, $N_h^{\ensuremath{p}} = \text{dim}(Q_h)$ and $N_h^{\eta} = \text{dim}(E_h)$. Then, we carry out a \reviewerA{proper orthogonal decomposition} of each snapshot matrix; \reviewerA{the method of snapshots is used, and the snapshots are weighted with the inner product associated to their functional space.} Then, the first $N^{\u}$, $N^{\ensuremath{p}}$ and $N^{\eta}$ (respectively) left singular vectors, denoted by $\{\vector{\varphi}_i\}_{i = 1}^{N^{\u}}$, $\{\psi_j\}_{j = 1}^{N^{\ensuremath{p}}}$ and $\{\phi_l\}_{l = 1}^{N^{\eta}}$ (resp.), are then chosen as basis functions for the reduced spaces $V_N^{(1)}$, $Q_N^{(1)}$ and $E_N^{(1)}$ (resp.), i.e.\ \begin{equation*} V_N^{(1)} = \text{span}(\{\vector{\varphi}_i\}_{i = 1}^{N^{\u}}), \quad Q_N^{(1)} = \text{span}(\{\psi_j\}_{j = 1}^{N^{\ensuremath{p}}}), \quad E_N^{(1)} = \text{span}(\{\phi_l\}_{l = 1}^{N^{\eta}}). \end{equation*} \subsubsection{On the imposition of coupling condition \eqref{}} \label{} The major drawback of this approach is related to the fact that the reduced spaces $V_N^{(1)}$ and $E_N^{(1)}$ do not guarantee, in general, that the coupling conditions \eqref{} holds. To this end, during the online stage, we will resort to a weak imposition of \eqref{} by Lagrange multipliers. More precisely, we will enforce weakly the normal component of \eqref{} (i.e. ${\u}^{k+1}_h \cdot \ensuremath{\vector{n}} = D_t \eta^k_h$), while the tangential component of \eqref{} (i.e. ${\u}^{k+1}_h \cdot \vector{\tau} = 0$) is already imposed strongly, since it is homogeneous and all basis functions in $V_N$ satisfy it. Thus, during the offline stage we need to build an additional snapshot matrix of the Lagrange multipliers, in order to carry out a POD to obtain a reduced Lagrange multipliers space. The traction on the interface, which can be evaluated as the residual of \eqref{} for test functions $\v_h := v_h \ensuremath{\vector{n}}$ which do not vanish on the interface, is indeed the Lagrange multiplier to \eqref{}. Therefore, we build an additional snapshot matrix \begin{align*} S_{\lambda} = [\underline{\boldsymbol{\lambda}}^{1}_h | \hdots | \underline{\boldsymbol{\lambda}}^{K}_h] \in \RR^{N_h^{\u} \times K}, \end{align*} containing the FE degrees of freedom corresponding to the residual of \eqref{} for test functions $\v_h := v_h \ensuremath{\vector{n}}$ that do not vanish on the interface, compute a POD, and, similarly to the previous section, obtain a reduced space $L_N^{(1)}$ as the space spanned by the first $N_{\vector{\lambda}}$ left singular vectors. \reviewerA{During the POD computation by the method of snapshots we employ the $L^2$ inner product on the interface as weight.} We remark again here that the Lagrange multiplier approach is \emph{not} actually used during the high-fidelity solution of the FSI system (in favor of a strong imposition), but rather the snapshot matrix $S_{\lambda}$ is obtained as a post-processing of the obtained solution. In contrast, in the online stage the Lagrange multiplier approach will actually be used while solving the linear system associated to the reduced fluid viscous step, in order to impose weakly the coupling condition \eqref{}. \subsubsection{Online stage} \label{} A reduced-order approximation of the FSI problem is then obtained by means of a Galerkin projection over the reduced spaces $V_N^{(1)}, Q_N^{(1)}$ and $E_N^{(1)}$, respectively, treating the coupling condition \eqref{} with Lagrange multipliers in the reduced space $L_N^{(1)}$. Thus, the corresponding online stage of the POD--Galerkin method reads: for any $k = 1, \hdots, K$ \begin{enumerate} \item[1$_N^{(1)}$.] Explicit step (fluid viscous part), with weak imposition of coupling conditions through Lagrange multipliers: find $({\u}^{k+1}_N, \lambda^{k+1}_N) \in V_N^{(1)} \times L_N^{(1)}$ such that: \begin{align*} \begin{cases} \int_\Omega \frac{\rho_f}{\Delta t} \u^{k+1}_N \cdot \v_N \,d\ensuremath{\vector{x}} & + \int_\Omega 2 \mu_f \boldsymbol{\varepsilon}({\u}^{k+1}_N) : \grad\v_N \,d\ensuremath{\vector{x}} \\ & + \int_\Sigma \lambda^{k+1}_N \ensuremath{\vector{n}} \cdot \v_N \,ds = \int_\Omega \frac{\rho_f}{\Delta t} \u^{k}_N \cdot \v_N \,d\ensuremath{\vector{x}} \, \\ & - \int_\Omega \grad p^k_N \cdot \v_N \,d\ensuremath{\vector{x}},\\ \int_\Sigma {\u}^{k+1}_N \cdot \Upsilon_N \ensuremath{\vector{n}} \,ds & = \int_\Sigma D_t \eta^k_h\,\, \Upsilon_N \,ds, \end{cases} \end{align*} for all $(\v_N, \Upsilon_N) \in V_N^{(1)} \times L_N^{(1)}$. We note that the boundary condition ${\u}^{k+1}_N \cdot \ensuremath{\vector{n}} = 0$ on $\Gamma_{sym}$ is implicitly verified, since it is satisfied by any element in $V_N$. \item[2$_N^{(1)}$.] Implicit step: for any $j = 0, \hdots$, until convergence: \begin{enumerate} \item[2.1$_N^{(1)}$.] Fluid projection substep, with Robin boundary conditions: find $p^{k+1,j+1}_N \in Q_N^{(1)}$ such that: \begin{align*} \int_\Omega \grad p^{k+1,j+1}_N \cdot \grad q_N \,d\ensuremath{\vector{x}} &+ \int_\Sigma \alpha_{Rob}\, p^{k+1,j+1}_N \, q_N \,ds = \\ & - \int_\Omega \frac{\rho_f}{\Delta t} \div {\u}^{k+1}_N \, q_N \,d\ensuremath{\vector{x}} - \int_\Sigma \rho_f D_{tt} \eta^{k+1,j}_N \, q_N \,ds \\ & + \int_\Sigma \alpha_{Rob}\, p^{k+1,j}_N \, q_N \,ds \end{align*} for all $q_N \in Q_N^{(1)}$. The imposition of the boundary conditions $p^{k+1,j+1}_N = p_{in}(t)$ on $\Gamma_{in}$ and $p^{k+1,j+1}_N = p_{out}(t)$ on $\Gamma_{out}$, although not automatically prescribed by the reduced space $Q_N^{(1)}$, can be easily treated by a lifting in $2.1_h$ without the need to introduce an additional Lagrange multiplier for the pressure, since the values to be imposed value do not depend on any reduced space, rather are known functions. The details are omitted for the sake of brevity; the interested reader is referred to \cite{BallarinManzoniQuarteroniRozza2015} for more details. \item[2.2$_N^{(1)}$.] Structure substep: find $\eta^{k+1,j+1}_N \in E_N^{(1)}$ such that: \begin{align*} \int_\Sigma \frac{\rho_s h_s}{\Delta t^2} \, \eta^{k+1,j+1}_N \, \zeta_N \,ds & + \int_\Sigma c_1 \partial_{x} \eta^{k+1,j+1}_N \, \partial_x \zeta_N \,ds \\ &+ \int_\Sigma c_0 \eta^{k+1,j+1}_N \, \zeta_N \,ds = \int_\Sigma \frac{\rho_s h_s}{\Delta t^2} \eta^{k}_N \, \zeta_N \,ds \\ & + \int_\Sigma \frac{\rho_s h_s}{\Delta t} D_t \eta^{k}_N \, \zeta_N \,ds - \int_\Sigma \boldsymbol{\sigma}({\u}^{k+1}, p^{k+1,j+1}) \ensuremath{\vector{n}} \cdot \zeta_N \ensuremath{\vector{n}} \,ds \end{align*} for all $\zeta_N \in E_N^{(1)}$. The boundary condition $\eta^{k+1,j+1}_N = 0$ on $\partial\Sigma$ is implicitly verified. \end{enumerate} \end{enumerate} As for the high-fidelity model, a stopping criterion on the relative increment of the solution is employed to terminate step $2_N^{(1)}$. \begin{remark}[On efficient offline-online decoupling] \label{} The reduced-order problem $1_N^{(1)}$, $2.1_N^{(1)}$ and $2.2_N^{(1)}$ can easily account for an efficient offline-online decoupling, thanks to the linearity assumption in the problem formulation. For instance, the fluid mass term $\int_\Omega \frac{\rho_f}{\Delta t} \u^{k+1}_N \cdot \v_N \,d\ensuremath{\vector{x}}$ in $1_N^{(1)}$, is efficiently assembled at the end of the offline stage as \begin{equation*} M_N^{(1)} := (Z_N^{\u})^T \, M_h \, Z_N^{\u}, \end{equation*} and loaded during the online stage. Here $Z_N^{\u}$ is the matrix which contains the velocity basis functions $\{\vector{\varphi}_i\}_{i = 1}^{N^{\u}}$ as columns, and $M_h$ is the FE matrix corresponding to fluid mass term $\int_\Omega \frac{\rho_f}{\Delta t} \u^{k+1}_h \cdot \v_h \,d\ensuremath{\vector{x}}$ in $1_h$. One can carry out a similar computational procedure for all terms in the reduced formulation $1_N^{(1)}$, $2.1_N^{(1)}$ and $2.2_N^{(1)}$. In a more general (nonlinear, geometrical parametrized) setting one can resort to the empirical interpolation method \cite{BarraultMadayNguyenPatera2004} to recover an efficient offline-online splitting, as recently shown for FSI problems in \cite{BallarinRozza2016}. \end{remark} \begin{remark}[On supremizer enrichment: \reviewerA{the role of pressure}] In contrast to what is usually done in the reduced basis approximation of parametrized fluid dynamics problem (see \cite{RozzaVeroy2007,Rozza2013,Rozza2009}, and also \cite{BallarinManzoniQuarteroniRozza2015} for an extension to POD--Galerkin methods) and previous works on FSI \cite{BallarinRozza2016,Colciago2014,LassilaManzoniQuarteroniRozza2013a}, we do not employ in this case a supremizer enrichment of the velocity space \reviewerA{to enforce inf-sup stability of the mixed velocity-pressure formulation}. This is heuristically motivated by the fact that, even at the high-fidelity level, the Chorin-Temam scheme, in its pressure Poisson version, can be successfully applied to FE spaces that do not fulfill a $(V, Q)$-inf-sup condition \cite{guermond1998stability}, even though it may result in non-optimal error estimates. \end{remark} \begin{remark}[On supremizer enrichment: \reviewerA{the role of Lagrange multiplier}] \begin{comment} This is the fundamental reason for which we have chosen to use the pressure Poisson version in this work, rather than the pressure Darcy summarized in Remark \ref{}. Assuming, for simplicity, $N_{\u} = N_{\ensuremath{p}} = N_{\lambda} = N$, this would allow in principle to obtain a reduced order velocity problem of size $N_{\u} = N$, rather than $N_{\u} + N_{\ensuremath{p}} = 2N$. Unfortunately, this does not happen in practice for FSI ROM 1 because the advantage of the smaller velocity space dimension is squandered by the weak imposition of coupling condition \eqref{} through Lagrange multipliers, which increases again the size of the velocity problem $1_N^{(1)}$ to $N_{\u} + N_{\lambda} = 2N$. Moreover, since $1_N^{(1)}$ is again a saddle point problem, one should (in principle) enrich the reduced spaces to fulfill a $(V_N, L_N)$-inf-sup condition for the (velocity, Lagrange multiplier) problem $1_N^{(1)}$, with a further increase in dimensionality.\footnote{In practice, we do not carry out this enrichment since the numerical results proposed in Section \ref{} are numerically stable even without any enrichment.} The ROM proposed in the next section has been devised to overcome this limitation. \end{comment} \reviewerA{% Problem $1_N^{(1)}$ still features a saddle point structure. We remark that this structure is not due to the original problem, but rather due to our choice of coupling conditions by Lagrange multipliers. A drawback of ROM 1 is now apparent for what concerns the size of the reduced system $1_N^{(1)}$, which needs to be increased to $N_{\u} + N_{\vector{\lambda}}$ due to weak imposition of coupling condition. The ROM proposed in the next section has been devised to overcome this limitation, and results in a reduced explicit step of size $N_{\u}$. Moreover, a further increase in dimension would be required if we were willing to enrich the velocity space $V_N^{(1)}$ with supremizers corresponding to the inf-sup condition associated to problem $1_N^{(1)}$, i.e. solutions to \begin{equation*} \int_\Omega \nabla\vector{s}^k \cdot \nabla\vector{v} = \int_{\Sigma} \vector{\lambda}^k_h \cdot \vector{v} \quad \forall \vector{v} \in V,\\ \end{equation*} for all $k = 1, \hdots, K$. In this work we do not carry out such enrichment since it would further increase the size of the reduced explicit step; nevertheless, a detailed investigation of the stability of $1_N^{(1)}$ with and without enrichment by $\vector{s}^k$ is an ongoing task and will be presented in a forthcoming work.% } \end{remark} \subsection{FSI ROM 2 approach} \label{} As we have seen in the previous section, it is challenging to enforce \eqref{} at the reduced-order level. The second reduced-order model proposed in this paper overcomes these difficulties performing a change of variable for the fluid velocity, namely defining an auxiliary unknown $\vector{z}^{k+1}: \Omega \to \RR^2$ as \begin{equation} \vector{z}^{k+1} = \u^{k+1} - D_t \widehat{\eta}^k\, \ensuremath{\vector{n}}, \label{} \end{equation} where $\widehat{\eta}^k$ is the solution of the following \emph{harmonic extension} problem \begin{equation*} - \Delta \widehat{\eta}^k = 0 \quad \text{ in } \Omega, \end{equation*} subject to the following inhomogeneous boundary condition on the interface \begin{equation*} \widehat{\eta}^k = \eta^k \quad \text{ on } \Sigma, \end{equation*} and homogeneous boundary condition on the remaining boundaries. In this way, the coupling condition \eqref{} is equivalent to \begin{equation*} \vector{z}^{k+1} = \vector{0} \quad \text{ on } \Sigma, \end{equation*} for which no weak imposition by Lagrange multipliers is required. \subsubsection{Offline stage} \label{} During the offline stage, the solution of the high-fidelity problem $1_h$, $2.1_h$ and $2.2_h$ is first sought. Auxiliary unknowns $\vector{z}^{k+1}$ are then computed thanks to \eqref{}, for all $k = 0, ..., K - 1$. We then consider the following snapshot matrix \begin{align*} S_{\ensuremath{\vector{z}}} = [\underline{\mathbf{z}}^{1}_h | \hdots | \underline{\mathbf{z}}^{K}_h] \in \RR^{N_h^{\u} \times K}, \end{align*} and, similarly to Section \ref{}, retain the first $N_{\ensuremath{\vector{z}}}$ POD modes in the reduced space $V_N^{(2)}$. Reduced spaces for fluid pressure and structure displacement are defined as in Section \ref{}, $Q_N^{(2)} := Q_N^{(1)}$ and $E_N^{(2)} := E_N^{(1)} = \text{span}(\{\phi_l\}_{l = 1}^{N^{\eta}})$. Moreover, harmonically extend each $\{\phi_l\}_{l = 1}^{N^{\eta}}$ to $\{\widehat{\phi}_l\}_{l = 1}^{N^{\eta}}$. \subsubsection{Online stage} \label{} Similarly to Section \ref{}, a reduced-order approximation of the FSI problem is now obtained by means of a Galerkin projection over the reduced spaces $V_N^{(2)}, Q_N^{(2)}$ and $E_N^{(2)}$, respectively, that is: for any $k = 1, \hdots, K$ \begin{enumerate} \item[1$_N^{(2)}$.] Explicit step (fluid viscous part), with change of variable for the fluid velocity: find ${\ensuremath{\vector{z}}}^{k+1}_N \in V_N^{(2)}$ such that: \begin{align*} \int_\Omega \frac{\rho_f}{\Delta t} \ensuremath{\vector{z}}^{k+1}_N \cdot \v_N \,d\ensuremath{\vector{x}} & + \int_\Omega 2 \mu_f \boldsymbol{\varepsilon}({\ensuremath{\vector{z}}}^{k+1}_N) : \grad\v_N \,d\ensuremath{\vector{x}} = \\ & \int_\Omega \frac{\rho_f}{\Delta t} \u^{k}_N \cdot \v_N \,d\ensuremath{\vector{x}} \, - \int_\Omega \grad p^k_N \cdot \v_N \,d\ensuremath{\vector{x}}, \\ & - \int_\Omega \frac{\rho_f}{\Delta t} D_t \widehat{\eta}^k_N\, \ensuremath{\vector{n}} \cdot \v_N \,d\ensuremath{\vector{x}} - \int_\Omega 2 \mu_f \boldsymbol{\varepsilon}(D_t \widehat{\eta}^k_N\, \ensuremath{\vector{n}}) : \grad\v_N \,d\ensuremath{\vector{x}} \end{align*} for all $\v_N \in V_N^{(2)}$. We note that boundary and \emph{interface} conditions on ${\ensuremath{\vector{z}}}^{k+1}_N$ are implicitly verified, since they are homogeneous and satisfied by any element in $V_N^{(2)}$. Finally, for the sake of the implicit step, we define $\u^{k+1}_N$ as $\vector{z}^{k+1}_N + D_t \widehat{\eta}^k_N\, \ensuremath{\vector{n}}$. \item[2$_N^{(2)}$.] Implicit step: for any $j = 0, \hdots$, until convergence: \begin{enumerate} \item[2.1$_N^{(2)}$.] Fluid projection substep, with Robin boundary conditions: as in 2.1$_N^{(1)}$. \item[2.2$_N^{(2)}$.] Structure substep: as in 2.1$_N^{(2)}$. Finally, at convergence, harmonically extend $\eta^{k+1}_N$ to $\widehat{\eta}^{k+1}_N$. Note that this can be easily carried out as the linear combination of the harmonically extended displacement basis $\{\widehat{\phi}_l\}_{l = 1}^{N^{\eta}}$. \end{enumerate} \end{enumerate} \begin{remark}[On efficient offline-online decoupling] \label{} Similarly to Remark \ref{}, an efficient offline-online decoupling can be obtained also in this case. Thanks to the definition of the harmonically extended displacements basis $\{\widehat{\phi}_l\}_{l = 1}^{N^{\eta}}$ an efficient assembly of (e.g.) the right-hand side mass term $\int_\Omega \frac{\rho_f}{\Delta t} D_t \widehat{\eta}^k_N\, \ensuremath{\vector{n}} \cdot \v_N \,d\ensuremath{\vector{x}}$ can be obtained. We remark that this accounts for a negligible additional offline cost (solution of \reviewerA{$N^\eta$} harmonic extension problems) and no additional online cost, since the extension of $\eta^{k+1}_N$ to $\widehat{\eta}^{k+1}_N$ does not actually require the solution of reduced problem, but rather a linear combination of $\{\widehat{\phi}_l\}_{l = 1}^{N^{\eta}}$ \reviewerA{once the coefficients of the structural unknown have been computed}. \end{remark} \section{Numerical comparison} \label{} In this section we summarize the numerical results obtained by the proposed reduced-order models. The values of constitutive and geometrical parameters are summarized in Table \ref{} \cite{Formaggia2001}. The domain has been discretized with a $120 \times 10$ structured mesh, while the time-step is $\Delta t = 10^{-4} \text{ s}$. The final time is $T = 0.13 \text{ s}$, so that $K = 1300$. $T$ has been chosen to simulate the pressure wave propagation, just before wave reflection occurs. Numerical simulations are carried out using \emph{RBniCS} \cite{HesthavenRozzaStamm2015,RBniCS}, an open-source reduced order modelling library developed at SISSA mathLab, built on top of \emph{FEniCS} \cite{LoggMardalEtAl2012a}. \begin{table} \centering \begin{tabular}{|l|l||l|l||l|l|}\hline $\rho_f$ & $1 \text{ g/cm}^3$ & $\mu_f$ & $0.035 \text{ Poise}$ & $\rho_s$ & $1.1 \text{ g/cm}^3$ \\\hline $E_s$ & $0.75 \times 10^6 \text{ dyn/cm}^2$ & $\nu_s$ & $0.5$ & $c_1$ & $\frac{h_s E_s}{h_f^2 \ (1 - \nu_s^2)}$ \\\hline $c_0$ & $\frac{h_s E_s}{2\ (1+\nu_s)}$ & $L$ & $6 \text{ cm}$ & $h_f$ & $0.5 \text{ cm}$ \\\hline $h_s$ & $0.1 \text{ cm}$ & $p_{in}$ & $10^4 (1-\cos(\frac{2 \pi t}{0.005})) \boldsymbol{1}_{t < 0.005} \text{ dyn/cm}^2$ & $p_{out}$ & $0 \text{ dyn/cm}^2$\\\hline \end{tabular} \caption{Constitutive parameters for test case (from \cite{Formaggia2001}).} \label{} \end{table} \vspace{\baselineskip} \noindent \textbf{FSI ROM 1:} Fig. \ref{} shows the POD singular values and retained energy as a function of the number $N$ of POD modes for FSI ROM 1. It can be noticed that the decay of the singular values of fluid velocity and structure displacement is slower than the one of fluid pressure; moreover, the first POD mode of fluid pressure retains a larger energy ($\approx 36\%$) than the first modes of structure displacement ($\approx 15\%$) and fluid velocity ($\approx 12\%$). Accordingly, the (relative) error analysis (Fig. \ref{}) shows that the reduced solution converges to the high-fidelity one, and that (except for small $N$), the relative error on the pressure is smaller than the displacement one, which is smaller than the velocity relative error. Employing only $N=30$ POD modes out of the $K=1300$ snapshots, velocity, displacement and pressure relative errors are of the order of $10^{-4}$, $10^{-5}$ and $10^{-7}$, respectively. Figure \ref{} shows that the overall speedup for large values of $N$ is of at least two order of magnitudes. In particular, speedup for the \reviewerA{explicit} step is approximately constant for increasing $N$, \reviewerA{while the speedup for the implicit steps is increasing for larger $N$ since the number of required iterations for the implicit step decreases with $N$ (see Fig. \ref{ and c for the maximum and average number of iterations, respectively).} \begin{figure} \centering \subfloat[POD singular values.]{\includegraphics[width=0.48\textwidth]{ROM/1/1300/sigma}}\; \subfloat[POD retained energy.]{\includegraphics[width=0.48\textwidth]{ROM/1/1300/energy}} \caption{Results of the offline stage of FSI ROM 1: POD singular values and retained energy as a function of the number $N$ of POD modes for fluid velocity, fluid pressure, and solid displacement.} \label{} \end{figure} \begin{figure} \centering \subfloat[Error analysis.]{\includegraphics[width=0.48\textwidth]{ROM/1/error/error}\label{; \subfloat[Speedup analysis.]{\includegraphics[width=0.48\textwidth]{ROM/1/speedup/speedup}\label{}} \caption{Error analysis and speedup analysis of FSI ROM 1, as a function of the number $N$ of POD modes for fluid velocity, fluid pressure, and solid displacement.} \end{figure} \vspace{\baselineskip} \noindent \textbf{FSI ROM 2:} Fig. \ref{} shows a comparison of the offline stage for ROMs 1 and 2. From Fig. \ref{ it can be noticed that, after the change of variable, the first auxiliary velocity $\ensuremath{\vector{z}}$ POD mode of FSI ROM 2 retains approximately $1.5\%$ more energy than the first velocity $\u$ POD mode. The remaining variables are omitted since the pressure and displacement bases are the same as in FSI ROM 1. Moreover, Fig. \ref{} shows that the left-hand side matrix of the fluid viscous step $1_N^{(2)}$ (FSI ROM 2) is characterized by a condition number which is, for all $N$, at least 10 order of magnitude smaller than the one of $1_N^{(1)}$ (FSI ROM 1). The combination of these two remarks justifies the improvement in the error analysis for the velocity variables, shown in Fig. \ref{. On average, the relative error on the velocity unknown obtained by FSI ROM 1 is seven times the one obtained by FSI ROM 2. Relative errors for the remaining unknowns are omitted because they are comparable among FSI ROMs 1 and 2. \reviewerA{Moreover, we show in Fig. \ref{} the error analysis for the interface stress. The plot clearly shows that FSI ROM 2 provides a better approximation of the interface stress for $N > 20$.} Online performance (Fig. \ref{}) are comparable to the ones obtained by FSI ROM 1. \reviewerA{This is due to the fact that (i) the number of iterations to reach convergence in step $2_N^{(1)}$ and $2_N^{(2)}$ are the same as maximum values (Fig. \ref{}) and comparable on average (Fig. \ref{})}, and (ii) the time to solve the explicit step does not depend strongly on $N$ (Fig. \ref{), even though step $1_N^{(2)}$ (FSI ROM 2) requires the solution of a linear system of size $N \times N$ (rather than $2 N \times 2 N$ for step $1_N^{(1)}$ (FSI ROM 1)). \begin{figure} \centering \subfloat[POD singular values.]{\includegraphics[width=0.48\textwidth]{ROM/2/1300/sigma}}\; \subfloat[POD retained energy.]{\includegraphics[width=0.48\textwidth]{ROM/2/1300/energy}} \caption{Comparison of the offline stage of FSI ROMs 1 and 2: POD singular values and retained energy as a function of the number $N$ of POD modes for fluid velocity $\u$ (FSI ROM 1) and auxiliary fluid velocity $\vector{z}$ (FSI ROM 2).} \label{} \end{figure} \begin{figure} \centering \subfloat[Condition number of explicit step.]{\includegraphics[width=0.48\textwidth]{ROM/2/condition_number/condition_number}\label{ \subfloat[Maximum number of iterations of implicit step.]{\includegraphics[width=0.48\textwidth]{ROM/2/iterations/iterations}\label{; \subfloat[\reviewerA{Average number of iterations of implicit step.}]{\includegraphics[width=0.48\textwidth]{ROM/2/iterations/iterations_average}\label{}} \caption{Comparison of the condition number of the left-hand side matrix of $1_N^{(1)}$ and $1_N^{(2)}$ (top) and \reviewerA{of the maximum (bottom, left) and average (bottom, right) number of iterations required by $2_N^{(1)}$ and $2_N^{(2)}$}, as a function of the number $N$ of POD modes for fluid velocity, fluid pressure, and solid displacement.} \end{figure} \begin{figure} \centering \subfloat[Error analysis of FSI ROM 2.]{\includegraphics[width=0.48\textwidth]{ROM/2/error/error}}\; \subfloat[Comparison of the error for fluid velocity unknown of FSI ROMs 1 and 2.]{\includegraphics[width=0.48\textwidth]{ROM/2/error/error_comparison}} \caption{Error analysis of FSI ROM 2, as a function of the number $N$ of POD modes for fluid velocity, fluid pressure, and solid displacement.} \label{} \end{figure} \begin{figure} \centering \includegraphics[width=0.48\textwidth]{ROM/2/error/error_stress} \caption{\reviewerA{Error analysis of the interface stress for FSI ROMs 1 and 2, as a function of the number $N$ of POD modes for fluid velocity, fluid pressure, and solid displacement.}} \label{} \end{figure} \begin{figure} \centering \subfloat[Speedup analysis of FSI ROM 2.]{\includegraphics[width=0.48\textwidth]{ROM/2/speedup/speedup}}\; \subfloat[Comparison of the speedup for fluid velocity unknown of FSI ROMs 1 and 2.]{\includegraphics[width=0.48\textwidth]{ROM/2/speedup/speedup_comparison}} \caption{Speedup analysis of FSI ROM 2, as a function of the number $N$ of POD modes for fluid velocity, fluid pressure, and solid displacement.} \label{} \end{figure} \section{Conclusions and perspectives} \label{} Two semi-implicit reduced-order models for FSI problems have been proposed in this work, based on a POD--Galerkin approximation of an operator splitting semi-implicit high-fidelity scheme. FSI ROM 1 is a standard Galerkin projection over the reduced spaces generated by POD. No supremizer enrichment is required, thanks to the operator splitting approach. Even though FSI ROM 1 shows good performance in terms of error analysis, its major drawbacks (when compared to FSI ROM 2) are related to the weak imposition of \eqref{} by Lagrange multipliers. Numerical results of the previous section have shown that this is detrimental for several aspects of the ROM: increased system dimension of the fluid explicit step, increased condition number of the fluid explicit step, increased error for the velocity. FSI ROM 2, instead, stems from the idea that \eqref{} can be easily imposed in a reduced-order framework by performing the change of variable \eqref{}. In this way, all the detrimental effects of FSI ROM 1 are remedied. Moreover, better properties in terms of POD retained energy are also obtained. The combination of these two factors results in a better approximation of the fluid velocity. In particular, in FSI ROM 2, we try to separate in the fluid velocity the fluid-structure interaction component from the pure fluid part. \reviewerA{Even though FSI ROM 1 suffers several drawbacks when compared to FSI ROM 2 (especially for what concerns the increased condition number), the approach proposed by FSI ROM 1 can be more easily integrated with existing reduced order modelling capabilities for fluid problems, since it does not require to change the existing computations of fluid velocity basis functions. Nevertheless, especially for FSI ROM 1, a more detailed analysis of enrichment procedures shall be carried out to further investigate the stability of the resulting reduced problem, due to two saddle point structures to be taken into account (see Remarks 3 and 4).} Future work will concern ROMs that are better able to face the hyperbolic nature of the problem. A more efficient separation at the reduced-order level of the parabolic and hyperbolic components of the system may decrease the number of basis functions required to obtain an accurate reduced description of the FSI problem. \section*{Acknowledgements} We acknowledge the support by European Union Funding for Research and Innovation -- Horizon 2020 Program -- in the framework of European Research Council Executive Agency: H2020 ERC CoG 2015 AROMA-CFD project 681447 ``Advanced Reduced Order Methods with Applications in Computational Fluid Dynamics'' and the PRIN project ``Mathematical and numerical modelling of the cardiovascular system, and their clinical applications''. We also acknowledge the INDAM-GNCS project ``Tecniche di riduzione della complessit\`{a} computazionale per le scienze applicate'' and INDAM-GNCS young researchers project ``Numerical methods for model order reduction of PDEs''. { \small

\section{Introduction\label{}} When performing computer simulations of complicated systems, such as low-temperature plasmas, it is typically not immediately obvious whether the results are correct, or made invalid by simplistic physics models, weak algorithms, poor numerical convergence, software implemented or input files configured incorrectly, or a combination of subtly interacting factors. To distinguish spurious simulation results from physical ones, skeptical analysis and systematic scrutiny are thus prudent. The importance of verification, benchmarking and validation of plasma-simulation codes is therefore increasingly being recognized~\cite{Turner2013}. An early example of such an effort for low-temperature plasmas is the pioneering work of Surendra for a capacitively-coupled radiofrequency (RF) discharge~\cite{Surendra1995}, where simulation results from twelve different codes (including four particle-in-cell codes) were compared with each other (benchmarking) and with experiment (validation). There is an inherent conflict between benchmarking and validation. For a successful benchmarking exercise, where differences between codes are small and fully understood, it is preferable to have a simple simulation model. For a successful validation exercise, a realistic, and therefore more complex, simulation model might be necessary. In this paper we perform both validation and benchmarking. We first do validation with a fairly complete simulation model, and subsequently do additional benchmarking using much simpler simulation models to allow us to identify the source of differences between codes. Succesful verification (testing that equations are correctly solved) of main aspects of both codes had been done prior to the work reported here and will not be further discussed. We benchmark and validate the two particle-in-cell (PIC) codes EDIPIC~\cite{Sydorenko2006} and LSP~6.95~\cite{Clark2005} for a short (or obstructed) parallel-plate glow discharge in helium at 3.5~Torr~\cite{DenHartog1988,Lawler1991}, where the anode terminates the negative-glow region. The discharge had water-cooled aluminum electrodes with 3.2~cm diameter, located 0.62~cm apart, and was operated in the moderately abnormal regime. In preliminary experiments, a segmented cathode was used consisting of an inner cylindrical segment with 1.6~cm diameter and a close-fitting outer annular segment with outer diameter of 3.2~cm. When both segments were held at the same potential, the measured current densities on each were virtually identical. It was thus concluded that radial effects were negligible, and a one-dimensional approximation was well justified, where only the distance along the axis needs to be resolved numerically in simulations. The plasma in the device is sustained by secondary electrons emitted from the cathode surface, primarily by incident ions. The effective secondary-electron emission yield, with a value of around 0.3, is thus one of the most important physical parameters. However, as will be discussed in Section~\ref{sec:UQ} below, its value is not precisely known. The secondary electrons are rapidly accelerated in the cathode fall and create multiple electron/ion pairs through impact ionization. The electron density in the cathode-fall region is small. The negative-glow region has a quasi-neutral, cool plasma confined in a weak potential well and the electric field in this region is very weak. As stated in~\cite{DenHartog1988} and references therein, electron temperature in the negative glow is determined by a balance between thermalization via elastic collisions with neutrals and heating via Coulomb collisions with the fast electrons accelerated by the cathode fall. The main source of electrons for this relaxation process is the ionization within the negative glow by ``ballistic'' electrons originating in the cathode fall. Kinetic effects are thus important in both regions of the plasma, with the electrons from the cathode fall forming a significant tail on the electron energy distribution function in the negative glow. The location of the boundary between the cathode-fall and negative-glow regions depends on the applied voltage in a complex way. With increasing voltage, the cathode fall first widens, then narrows again. For the lowest and highest voltages, 173~V and 600~V, respectively, the cathode fall is approximately 0.38~cm wide. For the intermediate voltages, 211~V, 261~V and 356~V, respectively, the width of the cathode fall is approximately 0.28~cm. The electric field in the glow discharge was measured using various laser diagnostics, primarily optogalvanic detection of Rydberg atoms~\cite{Lawler1980,Doughty1984a}. Optogalvanic spectroscopy can be used when the neutral-neutral collision frequency is sufficiently high for Rydberg atoms to predominantly ionize, rather than relax radiatively, the latter process being utilized by laser-induced fluoresence. Optogalvanic spectroscopy measures spikes in discharge current generated by this increased ionization due to a resonant laser pulse. The resonance frequency depends on the Stark shift of degenerate excited levels and is therefore relatively straightforwardly related to the local DC electric field. By measuring the discharge current as the laser frequency is scanned, the strength of the electric field can thus be determined, in this case with a claimed accuracy of 1\%~\cite{DenHartog1989}. The electron density and temperature in the negative-glow region were inferred from Monte-Carlo simulations using optogalvanic-spectroscopy data as input. The one-dimensional nature of the configuration, together with the precisely measured electric field, makes this particular experiment attractive as a validation target for plasma-simulation codes, or "as a set of benchmark experiments for further modeling of the cathode region", as proposed by the authors~\cite{DenHartog1989}. Such a validation was done for the 211~V case by Parker, Hitchon and Lawler~\cite{Parker1993} using a code based on an algorithm known as the "convected scheme" for self-consistently solving the Poisson-Boltzmann equations in two-dimensional phase space (axial location and velocity)~\cite{Hitchon1989,Sommerer1989,Sommerer1991}. Excellent agreement was found for the electric field, but the electron density found by the code was lower than the value inferred from the optogalvanic-spectroscopy data. The Parker validation will be discussed in more detailed in Sections~\ref{sec:UQ} and \ref{}. It should be noted that there is also a strong theoretical motivation for choosing a cold-cathode glow discharge at several hundred volts for validation and benchmarking of numerical simulations. Although quite a simple system conceptually, it requires numerical treatment that must account for non-local kinetics with strongly anisotropic and/or non-Maxwellian velocity distributions. The system is strongly inhomogeneous with the dynamic range of electron energies and spatial densities covering about three orders of magnitude (between the cathode-fall and the negative-glow regions). Also the anisotropy has a strong variation along the axis of the device, as illustrated by Fig.~\ref{}, which concisely presents the structure of the electron component of the plasma in the phase space by mapping the anisotropy parameter $a=\langle\frac{2}{3}\sin^2\theta\rangle$, where $\theta$ is the angle between the velocity vector and the dicharge axis, versus the full energy (kinetic plus potential) and distance from the cathode. The angle brackets $\langle...\rangle$ denote averaging over the local distribution. For a fully isotropic distribution, $a=1$, while, for example, for a distribution uniform over a hemi-sphere $a=\frac{1}{3}$ and for a beam aligned with the electric field $a=0$. The discharge in this example is at 800~V in 3.5~Torr helium, with an electrode spacing of 1.2~cm. The map indicates that electrons accelerated to high energies in the cathode fall are strongly anisotropic, with a non-local distribution approximately defined by the value of the energy integral (which corresponds to the location the electron originates at in the cathode avalanche). There is also an intermediate group (below 150~eV), relaxed in momentum but not in energy, whose energy distribution "remembers" the non-local structure. The cold (0.1~eV) electrons in the negative glow, whose density is about 1000 times higher than in the cathode sheath, are not mapped. The theory can be found, e.g., in~\cite{Kolobov1992}. \begin{figure \begin{center} \includegraphics[width = 0.7\columnwidth, angle = -90]{anisotr_map.eps} \end{center} \caption{Anisotropy map for a 1.2~cm glow discharge at 800~V in 3.5~Torr helium (from previous, unpublished work, in which a similar obstructed glow discharge in helium was studied)} \label{} \end{figure} The remainder of this paper is organized as follows. The simulation models used for this benchmarking/validation exercise by the two PIC codes EDPIC and LSP~6.95 are described in Section~\ref{}. Uncertainty quantification for two of the key simulation-model parameters is the topic of Section~\ref{sec:UQ}. The results from the glow-discharge simulations are presented in Section~\ref{}. To identify the cause of some code discrepancies, a set of simplified cases of electron-beam injection into helium gas are simulated, as discussed in Section~\ref{}. A comparison between fixed-voltage and fixed-current glow-discharge simulation results is done in Section~\ref{}. The main findings and conclusions are presented in Section~\ref{}. \section{Elementary processes and simulation model\label{}} The simulation model is one-dimensional, with only the axial direction (the distance between the circular electrodes) resolved. The two codes EDIPIC and LSP~6.95 that are used for the benchmarking and validation exercise presented here are both particle-in-cell (PIC) codes with Monte-Carlo collision (MCC) models. For both codes the electric field is found by solving the Poisson equation with Dirichlet-Dirichlet boundary conditions on an equidistant mesh, discretized using finite differences with second-order discretization error. A spatial resolution of 10~$\mu{}m$ is found sufficient to resolve both the Debye length at peak density in the negative glow and the anode sheath. For both codes, the direct-implicit particle advance~\cite{Langdon1983} is used for the electrons for all the inter-code benchmark simulations. For LSP, the results are virtually identical with direct-implicit and explicit Boris particle advance, respectively. The typical number of macro particles launched per cell of each species are in the range of 8--40, with an initial density of 1--5$\times 10^{10} / cm^3$. Both codes are parallelized using MPI. LSP uses the same domain decomposition for fields and macroparticles. EDIPIC has a direct, serial field solve and the macroparticles in each MPI process are distributed over the whole computational domain. For a typical simulation, around a dozen processor cores can be efficiently used. The time step is limited by the cell transit time for an electron accelerated over the whole cathode fall, just over a picosecond for the high-voltage case. \if 0 The primary parameters for the glow-discharge simulations are summarized in Tab.~\ref{}. \begin{table \begin{center} \begin{tabular}{|l|c|c|c|} \hline & 173~V & 211~V & 600~V \\ \hline \hline Time step & \multicolumn{2}{|c|}{$2 ps$} & $1 ps$ \\ \hline Simulation time & \multicolumn{3}{|c|}{$\approx 120 \mu{}s$} \\ \hline Plasma length & \multicolumn{3}{|c|}{$0.62 cm$} \\ \hline Cell size & \multicolumn{3}{|c|}{$10 \mu{}m$} \\ \hline Helium gas pressure & \multicolumn{3}{|c|}{$3.5 Torr$} \\ \hline Helium gas density & $11.2 \times 10^{16}/cm^3$ & $10.8 \times 10^{16}/cm^3$ & $8.01 \times 10^{16}/cm^3$ \\ \hline Helium gas temperarure & $26.29 meV$ & $27.26 meV$ & $36.75 meV$ \\ \hline Initial electron density & \multicolumn{3}{|c|}{$10^{10}/cm^3$} \\ \hline Initial helium ion density & \multicolumn{3}{|c|}{$10^{10}/cm^3$} \\ \hline Initial electron temperature & \multicolumn{3}{|c|}{$5 eV$} \\ \hline Initial helium ion temperature & \multicolumn{3}{|c|}{$25.2 meV$} \\ \hline \hline \end{tabular} \end{center} \caption{Much too big\ldots} \label{} \end{table} \fi As mentioned in Section~\ref{}, ion-impact secondary-electron emission (iSEE) is a critical process for maintaining the discharge. The standard version of LSP has an iSEE model that is deterministic. When an ion macroparticle impinges on a wall, it will emit one electron macroparticle with numerical weight reduced by a factor equal to the iSEE emission yield. Using this model severly slows down the simulation by introducing ever more macroparticles with ever smaller numerical weight. This numerical inefficiency could be mitigated by merging lightweight particles. However, more seriously, we fail to achieve numerical convergence when the standard LSP iSEE model is configured to emit only every few time steps. The simulation results varies significantly and unexpectedly depending on whether the electrons are emitted every one, two or three time steps. Rather than trying to make this existing, deterministic iSEE emission model perform correctly, we implemented a new probabilistic model, similar to the one in EDIPIC. In these latter models, secondary electron macroparticles are emitted with a probability equal to the emission yield when an ion macroparticle hits a wall. For a typical simulation, with a quasineutral initial plasma and equal numbers of ion and electron macroparticles, the emitted electrons will then conveniently have the same numerical weight as the initial elecrons. The true value for the iSEE emission yield is difficult to estimate, as it tends to drift over time as the condition of the electrode surfaces changes due to plasma exposure. The parameter uncertainty of the emission yield is an important topic and is discussed in more detail in Section~\ref{sec:UQ} below. The predominant collisions are charge exchange of ions on neutrals; elastic, excitation and ionization for electrons on neutrals, and Coulomb collisions for electrons on electrons. For each charged-neutral collision, a random atom is sampled from the background gas and used as the collision target, taking into account the neutral temperature. The charge exchange (CX) cross section varies weakly with energy over the relevant energy range and is well approximated by Eq.~(3.9.4) in the text book by Smirnov~\cite{Smirnov2015}. The CX model in LSP was found to give unexpected result and was modified. Our new version of the CX model was then verified by reproducing analytic results for simulations of ion-beam injection into neutral gas with the field solve disabled (not shown). For LSP, an arbitrary number of excitation levels can be used, and for helium a table with cross sections for excitation to the lowest seven excited states for electron energies up to one keV is included in the code suite and is plotted in Fig.~\ref{}. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{XsectionsHe.eps} \end{center} \caption{Tabulated electron-neutral cross sections used in the simulations. Elastic in blue and ionization in red. The black graphs are for excitation collisions with dashed line for the seven lowest excitation levels and the solid line for a single, consolidated (energy-weighted) excitation cross section.} \label{} \end{figure} The provenance of this data is not known to us, but it is in reasonable agreement with data from well-known sources, except at energies above 200~eV, where the LSP cross sections seem to fall off too slowly. For the current simulations, this discrepancy is believed to occur at energies too high to significantly affect the simulation results. For EDIPIC a single, consolidated excitation cross section is needed. Such a consolidated cross section was calculated by summing up the energy-weighted lowest seven individual excitation cross sections, shown as the solid, black graph in Fig.~\ref{}. Two LSP simulations were done to compare the results using the original and consolidated excitation cross sections, respectively. The simplified excitation model is not found to noticably increase the error bar set by other uncertainties, most notably the experimental measurement error of the cross sections. The elastic, excitation and ionization collisions are all anisotropic. For elastic and excitation collisions, both codes use an anisotropy model recently developed specifically for helium~\cite{Khrabrov2012}, based on energy-dependent screened-Coulomb scattering~\cite{Okhrimovskyy2002, Berlenguer1999, Wentzel1927}: \begin{equation} \label{} \frac{1}{\sigma(E)} \frac{d\sigma}{d\Omega}(E,\theta) = \frac{1}{4\pi} \frac {1+\varepsilon} {\left(1+\varepsilon\sin^{2}\frac{\theta}{2}\right)^2}, \end{equation} where the normalised energy $\varepsilon$ is defined as $8E/E_{aniso}(E)$ and $E_{aniso}$ is the anisotropy parameter, on the order of several atomic energy units. For the electron-neutral ionization, the same screened-Coulomb model is applied, but with a constant (not energy-dependent) screening length treated as an adjustable parameter that determines the level of anisotropy for scattering on the given species of the target atoms. In adopting the above approach to scattering (of both primary and secondary electrons) in ionizing collisions, we follow previous work such as Refs.~\cite{Parker1993} and \cite{Boeuf1982}, where ad-hoc approximations were employed due to lack of reliable reference data. It should be noted that eventually we found the elastic-scattering model of Ref.~\cite{Khrabrov2012}, applied at corresponding post-collision energies, to work equally well as the adjustable model with the chosen value of $E_{aniso}$. Another model-related structural uncertainty, or inadequacy, of the current simulations is the neglect of atoms in metastable excited states. Metastables can, in principle, introduce associative ionization as well as create secondary electrons as they diffuse to the cathode and impinge. However, at least for the low and intermediate voltages cases, the associative ionization is only about 5\% of the total ionization~\cite{DenHartog1989}, which is within the error bar of the cross section. Additionally, it would require the implementation of a rather complex model, which could at best modestly benefit the validation results, at the expense of making the code benchmarking more complicated. Metastables are therefore excluded from the simulation models used by EDIPIC and LSP here. Unlike EDIPIC, LSP does not have null collisions and determines if a collision occurred for each macroparticle of a collisional species. To facilitate the benchmarking, EDIPIC's null-collision algorithm is disabled for the current simulations. A subtle structural uncertainty can be introduced by pseudo-random-number generators (PRNGs) with statistical deficiences. In EDIPIC, the standard PRNG is replaced with the Mersenne Twister~\cite{Matsumoto1998}, which has emerged as the most commonly used high-quality PRNG. In LSP, for expediency, the standard PRNG is instead replaced by drand48, which is part of the Portable Operating System Interface (POSIX) standard and an intrisic function available from the standard C library of the compilers used. \section{Parametric uncertainty quantification\label{sec:UQ}} Experimentally, the effective SEE emission yield $\gamma$ has a large parameter uncertainty, $\gamma \approx 0.32 \pm 0.10$. The value of the screening energy $E_{aniso}$ in the scattering model is also poorly known, but believed to be around 100~eV, based on the values of this parameter for elastic collisions at scattering energies where the ionization cross-section is near its maximum~\cite{Khrabrov2012}. To assess how sensitively the simulation results depend on the values of these two simulation-model parameters we perform a number of simulations with LSP of the highest-voltage (600~V) case. To reduce the execution time of the simulations, a relatively large time step of 5~ps was used, too large for full temporal convergence. The values of the computed cathode current are therefore slightly high, but the scaling with $\gamma$ and $E_{aniso}$ is nevertheless valid. The results are shown in Fig.~\ref{}. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{CurrentVsGamma.eps} \end{center} \caption{Cathode current computed with LSP vs. $\gamma$ for three different values of the $E_{aniso}$ parameter in the Okhrimovskyy anisotropy model for ionization~\cite{Okhrimovskyy2002}} \label{} \end{figure} As can be seen, the current grows rapidly with increasing values of $\gamma$. For example, increasing $\gamma$ by 50\% (from 0.22 to 0.33) increases the current by 154\% (from 0.50~mA/cm${}^2$ to 1.27~mA/cm${}^2$). In the simulations, $\gamma$ is therefore for practical purposes a free parameter that can be adjusted to reproduce the experimental cathode current. Changes to $E_{aniso}$ that are well within the error bar also have large impact on the cathode current, albeit less dramatic than for $\gamma$. Also $E_{aniso}$ can thus be used as a free parameter to adjust the current. For example, ($\gamma = 0.28, E_{aniso} = 100~eV$) result in almost the same current as ($\gamma = 0.30, E_{aniso} = 50~eV$). \section{Validation for glow discharge\label{}} Validation of EDIPIC and LSP is done for the glow discharge~\cite{DenHartog1989} using the simulation model described in Section~\ref{}. Three cases are simulated: the lowest voltage (173~V), one of the intermediate voltages (211~V) and the highest voltage (600~V). The 211~V case was simulated previously by Parker \emph{et al.}~\cite{Parker1993}, not with a PIC-MCC code, but with a code that solves the Poisson-Boltzmann equations using the so-called "convected scheme"~\cite{Hitchon1989}. For both the important parameters $\gamma$ and $E_{aniso}$, the experimental error bars are too large to offer any real guidance on which values to use in the simulations. We therefore, somewhat arbitrarily, choose $E_{aniso} = 100~eV$ for all the simulations and use $\gamma$ as a free parameter that is adjusted in increments of 0.01 to reproduce the cathode current reported from the experiment. For the 173~V case, EDIPIC reproduces the experimental cathode current of 0.19~mA/cm${}^2$ with $\gamma = 0.28$, whereas LSP gets closest to the experimental current for $\gamma = 0.29$. For LSP, temporal convergence is achieved with a time step of 2~ps. With 1~ps time step, the early simulation results are virtually identical as with double the time step, and are therfore not ran to steady state. The limitation on the time step seems to be imposed by the cell transit time for an electron accelerated over the full cathode fall. Steady state is reached after 120~$\mu{}s$, or 60~million time steps, as shown in Fig.~\ref{}. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{SteadyState.eps} \end{center} \caption{Integral of ionization rate (blue graph) and ion flux (green graph) at end of EDIPIC simulation of 173~V case. Approximate overlap is indicative of near steady-state solution, except in negative-glow region ($>$~0.35~cm).} \label{} \end{figure} It should be noted that this steady state does not apply to the accumulation of cold electrons in the negative glow, which occurs on a millisecond time scale and will be discussed in more detail below. Electric-field profiles from these simulations, as well as from the experiment (data digitized from Fig.~4 of Ref.~\cite{DenHartog1989}), are shown in Fig.~\ref{}. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{ElectricField173V.eps} \end{center} \caption{Electric field for the 173~V case. Black circles are experimental values, blue, green and red graphs are LSP results with decreasing time steps (10, 5 and 2~ps, respectively). The magenta graph is the EDIPIC result with 2~ps time step.} \label{} \end{figure} As can be seen, when the time step is small enough for temporal convergence, the two codes are in excellent agreement (red and magenta graphs). Compared to experiment, the codes predict an electric field that is about 10\% larger at the cathode and decreases faster than linearly away from it. For the 211~V case, both codes reproduce the experimental cathode current of 0.52~mA/cm${}^2$ with $\gamma = 0.28$. Like for the lower voltage, the 2~ps time step is sufficient for temporal convergence and 120~$\mu{}s$ for steady state. The results are shown in Fig.~\ref{}. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{ElectricField211V.eps} \end{center} \caption{Electric field for the 211~V case. Black circles are experimental values, blue, green and red graphs are LSP results with decreasing time steps (10, 5 and 2~ps, respectively). The magenta graph is the EDIPIC result with 2~ps time step.} \label{} \end{figure} As for the lower-voltage case, the two codes are in excellent agreement (red and magenta graphs), but overestimate the electric field at the cathode by about 15\%. For the 600~V case, EDIPIC requires $\gamma = 0.33$ to match the experimental cathode current of 1.50~mA/cm${}^2$, and LSP needs $\gamma = 0.36$. To reach temporal convergence, the time step has to be halved to 1~ps. The electric-field profile is shown in Fig.~\ref{}. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{ElectricField600V.eps} \end{center} \caption{Electric field for the 600~V case. Black circles are experimental values, blue, green and red graphs are LSP results with decreasing time steps (5, 2 and 1~ps, respectively). The magenta graph is the EDIPIC result with 1~ps time step.} \label{} \end{figure} As for the two lower-voltage cases, the codes agree (red and magenta graphs), but show better agreement with experiment, with the electric field at the cathode about 7\% too large. The main results from the glow-discharge validation are summarized in Tab.~\ref{}. \begin{table \begin{center} \begin{tabular}{|l|c|c|c|} \hline & 173~V & 211~V & 600~V \\ \hline \hline Exp. $\gamma$ & $\approx 0.32$ & $\approx 0.32$ & $\approx 0.32$ \\ \hline Exp. $E^0 (kV/cm)$ & 0.90 & 1.43 & 3.02 \\ \hline Exp. $d_c \, (cm)$ & 0.38 & 0.30 & 0.40 \\ \hline \hline EDIPIC $\gamma$ & 0.28 & 0.28 & 0.33 \\ \hline EDIPIC $E^0 (kV/cm)$ & 0.97 & 1.58 & 3.18 \\ \hline EDIPIC $d_c \, (cm)$ & 0.35 & 0.25 & 0.34 \\ \hline \hline LSP $\gamma$ & 0.29 & 0.28 & 0.36 \\ \hline LSP $E^0 (kV/cm)$ & 0.99 & 1.57 & 3.18 \\ \hline LSP $d_c \, (cm)$ & 0.35 & 0.25 & 0.34 \\ \hline \hline \end{tabular} \end{center} \caption{For 173~V the current density is 0.19~mA/cm${}^2$, for 211~V it is 0.52~mA/cm${}^2$ and for 600~V it is 1.50~mA/cm${}^2$. In the simulations, the emission yield $\gamma$ was adjusted in each case to reproduce the experimental current density. $E^0$ is the axial electric field at the cathode surface. $d_c$ is the width of the cathode-fall region. Experimental values from Tab.~1 of Ref.~\cite{DenHartog1988}.} \label{} \end{table} As a benchmarking exercise, the 173~V and 211~V cases are unqualified successes, but for the 600~V case a 10\% larger $\gamma$ was needed for LSP to compute the same current. As a validation exercise, the simulations are partially succesful. Both codes correctly reproduced the dependence of the cathode-fall width on voltage, with a minimum width at intermediate voltage. In the experiment the cathode-fall width was extrapolated from the electric-field measurements to 0.38~cm at 173~V, 0.30~cm at 211~V and 0.40~cm at 600~V. In the simulations we get 0.35~cm at 173~V, 0.24~cm at 211~V and 0.34~cm at 600~V. To compensate for the shorter cathode fall the electric-field values in the simulations are larger than in the experiment, by 7\%--15\% at the cathode surface. As can be seen in Figs.~\ref{}, \ref{} and \ref{}, the electric field in the negative-glow region is weak. To suppress statistical noise in the simulation data, temporal and spatial averaging is applied to the numerically converged LSP cases for each voltage. The resulting electric fields are shown in Fig.~\ref{}. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{TimeAveragedElectricField.eps} \end{center} \caption{LSP simulated averaged electric fields in the negative-glow region for 173~V (blue graph), 211~V (green graph) and 600~V (red graph)} \label{} \end{figure} As can be seen, even after averaging, the computed electric field is noisy in the negative glow. However, on the reasonable assumption that the true electric field decreases monotonically from cathode to anode, one can infer that the field-reversal point is well inside the negative-glow region, by 0.10--0.15~cm, or somewhere between 0.40--0.53~cm from the cathode. The corresponding shallow potential well in the negative glow, which confines the cold electrons, does not provide a more exact field-reversal location. The depth of the potential well is of the order of a volt and it has a very wide and flat minimum (not shown). The earlier simulations by Parker of the 211~V case using the "convected scheme"~\cite{Parker1993} showed excellent agreement with the experimentally measured electric field and found a cathode-fall width of 0.27~cm for $\gamma = 0.25$ and with a simplified model of ionization anisotropy. These earlier simulations did include the associative ionization of metastables, but did not include electron emission due to the same. All simulations (EDIPIC, LSP and convected scheme) correctly compute an electron temperature of between 0.1~eV and 0.2~eV inside the negative glow, but underestimate the density. The reported peak density from the experiment is about $5 \times 10^{11} / cm^3$. The EDIPIC result is about $2 \times 10^{11} / cm^3$, LSP less than $10^{11} / cm^3$ and convected scheme in the range of $1-5 \times 10^{11} / cm^3$ (the exact value was not reported). The cold-electron accumulation in the negative glow is primarily determined by trapping of electrons produced in ionization events within the plasma potential well, and of some of those accelerated in the cathode fall (which produce the ionization). Correctly simulating this process requires an accurate model of binary electron-electron Coulomb collisions as well as milliseconds of physical simulation time. The relevant (or at least, the longest one involved) time scale for establishing the density profile of cold plasma in the negative glow can be roughly estimated as the ambipolar diffusion time. Based on $T_e=0.1~eV$, helium ion mobility $\mu = 10~cm^2/(V s)$ and the length scale of 0.1~cm, this equilibration time is on the order of 10~ms. The EDIPIC binary-collsion model has recently been thoroughly validated as part of yet-unpublished efforts to simulate runaway electrons and breakdown and we are fairly confident in its accuracy. The procedure employed in EDIPIC is the Langevin approximation based on the Fokker-Planck kinetic equation, as developed by~\cite{Manheimer1997}: the effect of Coulomb collisions on simulation particles is represented as a drift-diffusion stochastic differential equation. However, even in very long (billion-time-step) simulations, it seems unlikely that EDIPIC would reproduce the very high density values reported from the experiment. In the LSP simulations the density in the negative glow is approximately 2.5 times less than in the EDIPIC simulations. A likely culprit is the LSP binary-collision model, which we have neither scrutinized nor tested. However, an inadequate binary-collision model is unlikely to explain the need for a larger effective SEE emission yield in LSP for the 600~V case. To try to identify the cause of this discrepancy between the codes, we perform a number of synthetic code benchmarks, some of which are presented in the next section. \section{Synthetic code benchmark for electron-beam injection\label{}} To facilitate the benchmarking specifically of the collision models in the two codes, simplified simulations are ran of injection of a mono-energetic electron beam into helium gas. These synthetic-benchmark simulations are using the same set up as for the 211~V discharge, but without initial plasma and with the electric-field solve and secondary emission disabled. Fig.~\ref{} shows the steady-state density profiles with a 100~mA/cm${}^2$ beam of 100~eV electrons injected from the right boundary and interacting with the helium gas only via artificially anisotropic (forward-scattering only) ionization collisions. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{beam-100eV-forward-ionization-only.eps} \end{center} \caption{100~eV electrons injected from the right boundary producing ionization with forward-scattering only. Blue graph is steady-state electron density for LSP with the standard random-number generator, red is with drand48 random numbers and green graph is with EDIPIC.} \label{} \end{figure} All other collision models are temporarily disabled. Both codes are run with the same number of macroparticles, but the LSP results have more statistical noise because EDIPIC does time averaging over multiple time slices. Taking this into account, there is no statistically significant discrepancy between the results. Similar simulations are run with only isotropic elastic and excitation collisions with results shown in Fig.~\ref{}. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{beam-100eV-excitPLUSelast-only-iso.eps} \end{center} \caption{100~eV electrons injected from the right boundary undergoing isotropic elastic and excitation collisions only. Blue graph is steady-state electron density for LSP with the standard random-number generator, red is with drand48 random numbers and green graph is with EDIPIC.} \label{} \end{figure} As with forward-scattering ionization only, simulations with isotropic elastic and excitation collisions only produce results that differ only in statistical noise for the two codes. However, when the elastic and excitation collisions are made anisotropic the results markedly differ, as can be seen in Fig.~\ref{}. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{beam-100eV-excitPLUSelast-only-okh100.eps} \end{center} \caption{100~eV electrons injected from the right boundary undergoing anisotropic (Okhrimovskyy model with $E_{aniso}$ = 100~eV) elastic and excitation collisions only. Blue graph is steady-state electron density for LSP with the standard random-number generator, red is with drand48 random numbers and green graph is with EDIPIC.} \label{} \end{figure} In this case, the Okhrimovskyy model with $E_{aniso}$ = 100~eV is used for the elastic and excitation collisions. The steady-state density is significantly lower for the EDIPIC simulation (green graph) than for the LSP simulations (blue and red graphs). There also is some difference between the LSP simulations with different random-number generators, with drand48 producing results in slightly better agreement with EDIPIC. Similar discrepancies for different random-number generators for both codes also in the more realistic glow-discharge validation simulations led us to abandon the existing random numbers in each code in favor of more modern, higher-quality ones (drand48 in LSP and the Mersenne Twister in EDIPIC). The increased density in LSP compared to EDIPIC in the presence of anisotropic collisions leads us to hypothesize that LSP spuriously partially isotropizes the electron velocity distribution. This would be consistent with the need to increase the SEE yield for the 600~V case to compensate. This hypothesis needs further testing to be confirmed. It is also possible that spurious, numerical velocity-space diffusion plays a role~\cite{Turner2006}. \section{Code benchmark with fixed current\label{}} The characteristics of the power supply used in the experiment are not known to us. For the glow-discharge simulations presented so far, the power supply is assumed to be an ideal voltage source. To assess the importance of non-zero internal resistance of the power supply, simulations are performed with both LSP and EDIPIC using their external-circuit models with a voltage source connected to the anode in series with an external resistor and the cathode connected to ground. The value of the internal resistance is made large enough for the power supply to approximate a current source. The circuit model in EDIPIC~\cite{Sydorenko2014} is based on the standard algorithm~\cite{Verboncoeur1993}, but it additionally incorporates the implicitness of surface charge deposited by electrons advanced in time using the direct implicit scheme~\cite{Friedman1981}. LSP has a less sophisticated circuit model that is found to be numerically unstable. The fixed-voltage LSP simulation that produced the Fig.~\ref{} 2-ps-time-step case is modified to keep the current density fixed at 190~$\mu{}A/cm^2$, but with the original LSP circuit model all electron macroparticles were lost within the first 100~ns. When the source code for the circuit model was scrutinized, it was found that all surface charge deposited by the plasma on the wall was immediately forced through the circuit model, implicitly neglecting the possibility of capacitive effects. On the hypothesis that the instability was due to neglect of the sheath capacitance, an external capacitor is added in parallel to the circuit model to compensate. The neglected sheath capacitance is estimated from the steady-state profiles of potential and charge density for the fixed-voltage case to about 540~fF/cm${}^2$. Fig.~\ref{} shows the time evolution of the number of electron macro particles, which is proportional to the volume integral of the electron density, in fixed-current LSP simulations for different values of external parallel capacitance. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{MacrosVsTime.eps} \end{center} \caption{Number of electron macro particles vs.~time for various LSP simulation of the 173~V / 190~$\mu{}A/cm^2$ case. Black graph is for fixed voltage. Blue, cyan and green graphs are for fixed current with external parallel capacitances of 50, 100 and 200~fF/cm${}^2$, respectively. The yellow graph is for 500~fF/cm${}^2$, close to the neglected sheath capacitance. The orange, red and magenta graphs are for 1, 2 and 5~pF/cm${}^2$, respectively.} \label{} \end{figure} For increasing values of external parallel capacitance, the macro-particle number at steady state initially increases, but then levels off for a capacitance equal to or greater than the actual sheath capacitance with the macro-particle number lower than for fixed voltage (see Fig.~\ref{}). The reason for the reduced electron density for fixed current compared to fixed voltage was found to be due to additional heating in the former case. For fixed current, the voltage oscillates at a frequency determined by the time constant for the RC circuit formed by plasma plus external circuit. This type of oscillation has been observed experimentally and explained theoretically in the parameter regime corresponding to the transition from a Townsend to a normal glow discharge~\cite{Kaganovich1994b,Melekhin1984}. With the true sheath capacitance, the RC oscillation frequency is about 300~kHz for the case we are simulating. The oscillation is collisionally damped through electron-electron collisions that heat the electrons. With the increase in thermal velocity, the electrons can transport current across the negative glow at reduced density. However, in a fixed-current simulation with EDIPIC, the electron density at steady state is essentially the same as for the fixed-voltage case, as shown in Fig.~\ref{}. \begin{figure \begin{center} \includegraphics[width = 0.9\columnwidth]{DensFixedCurrentVsVoltage.eps} \end{center} \caption{Comparison of electron densities in fixed-voltage (black graph) and fixed-current (red graph) EDIPIC simulations} \label{} \end{figure} In Fig.~\ref{}, the boundary between cathode fall and negative glow is closer to the cathode for fixed current than for fixed voltage. By comparing the electron-density profiles at other time slices, one finds that the boundary location is immobile for fixed voltage, but oscillates for fixed current around the same average location. The oscillation of the boundary location is consistent with a potential oscillation in the negative glow. Like for the fixed-current LSP simulations, the oscillation frequency is given by the RC time constant and the amplitude is also similar, about $\pm$10\%. Since no density reduction is associated with the voltage oscillation in EDIPIC, we conclude that the heating found in the LSP simulations is spurious and possibly caused by electron-electron collisionality being too large. We note the critical importance of code benchmarking in this case. If scrutinized in isolation, the LSP simulations results might have been difficult to identify as spurious. \section{Conclusions\label{}} The two PIC-MCC codes EDIPIC and LSP~6.95 are benchmarked and validated. When charged-neutral collisions are isotropic and electron-electron Coulomb collisions are negligible, the two codes are in excellent agreement. A short glow discharge~\cite{DenHartog1989} is chosen as the validation target because i) data is available from accurate electric-field measurements, ii) it is a simple configuration that can be accurately modeled as one-dimensional, and iii) it provides a very discerning test case for collision models. Validation of the two codes for this configuration is mostly successful. The dependence of the cathode-fall width on the applied voltage, with a minimum at the intermediate voltage, is reproduced in the simulations. The axial electric field in the simulations is 7--15\% too high at the cathode surface and drops off faster than in the experiment away from the cathode. A possible explanation is the model inadequacy introduced by neglecting associative ionization by metastables in the simulations. The simulated electron temperatures in the negative glow are correct within experimental error bars, but the densities are low. Also an earlier simulation of the intermediate-voltage case using the convected scheme reported a density lower~\cite{Parker1993} than the one claimed for the experiment. The larger values of the secondary-electron emission yield needed for the highest-voltage simulations might indicate a contribution from metastables in the experiment. Also ultraviolet photons and fast neutrals could play a role that warrants further investigation. The primary lesson from the work presented here is that it can be critically important to use state-of-the-art and well-tested collision models. Also, many common generators of pseudo random numbers are insufficient for simulations with Monte-Carlo collisions and can lead to subtly incorrect results. The Mersenne Twister~\cite{Matsumoto1998} provides a superior alternative and many implementations are freely available to be incorporated into PIC-MCC codes. Simulation of low-temperature plasmas is a non-trivial endeavor. Multiple physics models for collisions, wall interation and/or chemical processes are typically invoked and each model can have large parameter uncertainty. There is also often model uncertainty, i.e. it might be unknown beforehand what physical processes are essential and should be included in the simulation model. Inadequate physics models and algorithms can cause simulation results that might not be obviously unphysical, while still being subtly wrong. Code benchmarking can help indentify such cases. The work presented here should be seen as partial progress and part of an ongoing process to develop a set of representative benchmarking/validation problems and a more rigorous approach to simulation of low-temperature plasmas, with the goal of achieving quantitative results and ultimately a reliable predictive capability. Other groups have performed benchmarking and validation for RF discharges, including the pioneering work by Surendra~\cite{Surendra1995} and more recent work by Turner~\cite{Turner2013}. Similar efforts should be spent on other representative types of low-temperature plasmas and benchmarking/validation of all codes should become standard practice to make simulation a more reliable and useful tool in our field. \ack The information, data, or work presented herein was funded in part by the Advanced Research Projects Agency-Energy (ARPA-E), U.S. Department of Energy, under Award Number DE-AR0000298. The information, data, or work presented herein was funded in part by the Advanced Research Projects Agency-Energy (ARPA-E), U.S. Department of Energy, under Award Number DE-AR00000670. This work was made possible by funding from the Department of Energy for the Summer Undergraduate Laboratory Internship (SULI) program. This work is supported by the US DOE Contract No.~DE-AC02-09CH11466.\\ The digital data for this paper can be found at \url{ \bibliographystyle{iopart-num}

\section*{\refname}} \begin{document} \title{Production of neutral Higgs bosons associated with Z-boson and photon in 2HDM at future lepton colliders} \author{Nasuf Sonmez} \email{} \affiliation{Ege University, Faculty of Science, Physics Department, Izmir 35040, Turkey} \date{\today} \begin{abstract} Inspired by precision tests of the Standard Model in future lepton colliders, the numerical analysis of the following scattering processes, $e^+e^- \rightarrow Z h^0\gamma$ and $e^+e^- \rightarrow Z H^0 \gamma$, are carried at the tree level including all possible diagrams in the two-Higgs-doublet model (2HDM). This model has many free parameters, but the parameters which take part in the scattering amplitudes of these two processes are primarily the mixing angle parameter, $s_{\beta-\alpha}$, and the masses of the neutral Higgs bosons, ($h^0, H^0$). Therefore, measuring the production rates of $Z h^0\gamma$ and $Z H^0 \gamma$ final states open another test for the scalar sectors of the 2HDM. The numerical analysis is performed under the current experimental constraints. The production rates and the asymmetry in the forward-backward direction are presented as a function of the center-of-mass energy covering the future lepton colliders. The unpolarized cross-section gets up to 6.19 (4.86) fb at $\sqrt{s} = 350$ (500) GeV and 0.164 (0.157) fb at $\sqrt{s} = 350$ (500) GeV for $e^+e^- \rightarrow Z h^0\gamma$ and $e^+e^- \rightarrow Z H^0 \gamma$, respectively. The polarization of the incoming $e^+e^-$ beams are studied for various configurations, and it enhances the cross-section by a factor of 1.78 in both processes for $P_{e^+,e^-}=(+0.6,-0.8)$. \end{abstract} \pacs{} \maketitle \section{Introduction} \label{sec1} The last missing part namely Higgs boson, which is related with the electroweak symmetry breaking mechanism in the Standard Model (SM), was discovered at the LHC \cite{, , }. The production and the decay channels of the Higgs particle have been studied extensively since then \cite{, 2015PhRvL.114s1803A}. The discovery commenced questioning which model it belongs to. Many studies for its properties have been reported, and experiments are still continuing. All the results presented by ATLAS and CMS collaborations showed that the findings resemble a Higgs boson in the SM. It is possible that the SM is a low energy approximation of a larger theory which explains up to the Fermi scale. There are many proposals which extend the SM, and they have been studied extensively in the last decades. A simple extension of the SM is to enlarge only the scalar sector of the model, which is called the two-Higgs-doublet model; mainly, a new Higgs doublet is added to the theory. So that these two Higgs doublets can couple to matter and gauge bosons and give mass to leptons, quarks, and electroweak bosons. In the 2HDM, there are two charged Higgs bosons $(H^\pm)$ and three neutral Higgs bosons $(h^0,A^0,H^0)$ \cite{, }. The extra doublet generates new couplings and interactions. As a result, rich phenomenology arises, especially for Higgs physics. To be able to unravel the properties of this particle, the collision of fundamental particles other than protons is needed. The best candidate is the lepton, and a lepton collider is an excellent machine for studying the properties of the Higgs boson. Various proposals for future lepton colliders have been submitted around the world running at the center of mass energies between 240 - 1000 GeV; they are the Circular Electron-Positron Collider (CEPC) in China \cite{, }, the Future Circular Collider (FCC-ee) \cite{} at CERN \cite{}, and the International Linear Collider (ILC) in Japan \cite{}. In all these proposals, the planned facilities will produce millions of Higgses by electron-positron collisions. Thus, the properties of these Higgses can be revealed with high statistical precision. Two of the well-motivated channels at the $e^+e^-$-colliders are the production of Higgses with gauge bosons and fermions, namely the production of $ZH$ and $H\nu_e\bar{\nu}_e$. These channels will help to determine the couplings, $c_{HZZ}$ and $c_{HWW}$, and the Higgs total decay width \cite{, , , , , , 2018arXiv180604992Z}. Studying double Higgs-strahlung along with WW double-Higgs fusion makes it possible to determine the triple Higgs self-coupling ($c_{HHH}$) with an astonishing precision \cite{, , , , }. Measuring Higgs self-couplings is essential for reconstructing the Higgs potential in the SM. The complete reconstruction of the Higgs potential in the SM requires the determination of the quartic Higgs self-coupling $c_{HHHH}$ as well. This coupling can be reached directly in the triple Higgs production \cite{, }. There are some couplings which are not defined at tree-level but arise at loop-level in the SM, and they are called the anomalous trilinear Higgs couplings, $HZ\gamma$, $H\gamma\gamma$ and $HZZ\gamma$. These couplings are sensitive to new physics contributions through new massive particles propagating in the loops \cite{Cao2013, Han2013}. In all these processes, the couplings of Higgses with gauge bosons are significant to determine the other Higgs couplings. Since there are additional Higgs states in the 2HDM, the determination of the Higgs couplings becomes entangled. Many authors have studied the scattering process $e^+e^-\rightarrow ZH$, and measuring the cross-section is also one of the objectives which are shared among all three future colliders. Besides this channel, the $ZH$ could be produced in association with a high energetic photon. The process $e^+e^-\rightarrow ZH\gamma$ suits great for studying the couplings of the Higgs to neutral gauge bosons in the SM, and it is also part of the inclusive process $e^+e^-\rightarrow Z H+X$. There are three neutral Higgs bosons in the 2HDM, so there are three similar processes: $e^+e^- \rightarrow Z h^0 \gamma$ , $e^+e^- \rightarrow Z H^0 \gamma$, and $e^+e^- \rightarrow Z A^0 \gamma$. The relevant Higgs couplings for these processes are $c_{\{h^0, H^0\}ZZ}$ and $c_{\{h^0, H^0\} A^0Z}$, therefore, the situation is complicated compared to the SM. Additionally, the couplings $c_{h^0ZZ}$ and $c_{H^0ZZ}$ in the 2HDM influence the anomalous trilinear couplings. Compared to the $Z H$ channel, the production cross-section of $e^+e^- \rightarrow \hzg$ is considerably less, but they still have a moderate cross-section, and it is possible to study them in the colliders. It could be considered as another test for the Higgs couplings with an additional vector boson (photon) at the final state. The production of Higgs boson with additional vector bosons turns into a complicated problem in extended Higgs models. Therefore, the following questions need to be answered: "Which couplings are significant for these scattering processes?", "What are the relevant free parameters that take place in each scattering process?", moreover, "What could be extracted from these scattering processes in the context of 2HDM?". The scattering process $e^+e^-\rightarrow ZH\gamma$ in the SM was calculated before in ref. \cite{} at the loop level. The same process was studied in ref. \cite{} by taking into account the CP-conserving couplings between Higgs to gauge bosons within the effective Lagrangian framework and dimension-six operators, and the effects of these anomalous Higgs-gauge boson couplings were compared. This paper presents the calculation of $e^+e^- \rightarrow \hzg$ in the 2HDM, including all possible diagrams. The cross-section is calculated as a function of the center-of-mass (CM) energy, and comparison is carried out regarding the proposed colliders. Dependence of the cross-sections on the free parameters is presented. Several polarization configurations of the $e^+e^-$ beams are assumed. The forward-backward asymmetry between the Higgs boson and the photon is calculated, and a discussion is carried for the colliders assumed. The layout of this paper is organized as follows. In section \ref{sec2}, the theory of the 2HDM, all constraints, and assumptions in the computation are discussed. In section \ref{sec3}, expressions regarding the kinematics of the scattering and the total cross-section with polarized electron-positron beams are emphasized. In section \ref{sec4}, numerical results of the total cross-section with polarization and the asymmetry distributions are presented. The conclusion is drawn in section \ref{sec5}. \section{Theoretical framework of the 2HDM} \label{sec2} The 2HDM has been studied before, and a detailed introduction of the 2HDM framework was given by various authors \cite{, , , , }. Therefore, we only give a summary of the 2HDM, particularly the scalar potential and the free parameters in the model. The 2HDM is constructed by adding a second $SU(2)_L$ Higgs doublet with the same hypercharge ($Y=1$) to the scalar sector. Imposing a discrete symmetry, $\mathcal{Z}_2$, on the scalar potential helps to avoid flavor-changing-neutral-currents (FCNC) \cite{} at the tree-level. In this study, CP-conservation is considered. Accordingly, the scalar potential is defined as follows: \begin{eqnarray} V(\Phi_1,\Phi_2)&=&m_{11}^2 | \Phi_1|^2+m_{22}^2|\Phi_2|^2 - \left[ m_{12}^2 \Phi_1^{\dagger} \Phi_2 +h.c. \right] \nonumber \\ &+& \frac{\lambda_1}{2}| ( \Phi_1^{\dagger} \Phi_1 )^2 + \frac{\lambda_2}{2} ( \Phi_2^{\dagger} \Phi_2 )^2 +\lambda_3 | \Phi_1 |^2 | \Phi_2 |^2 \label{} \nonumber \\ &+&\lambda_4 | \Phi_1^{\dagger} \Phi_2 |^2 + \left [ \frac{\lambda_5}{2} (\Phi_1^{\dagger} \Phi_2)^2 + h.c. \right] \;, \end{eqnarray} where the coupling constants, $\lambda_i, (i=1,..,4)$, are real, the parameters $m_{12}^2$ and $\lambda_{5}$ can be complex, but they are taken real for simplicity, and $\Phi_i,\;(i=1,2)$ are the Higgs doublets. Following the prescription defined in ref. \cite{} the masses of all the extra Higgs bosons can be calculated as usual. First, the stationary conditions of the potential are applied, and several constraints are obtained. Second, these bound conditions are substituted into the scalar potential to eliminate $m_{11}$ and $m_{22}$. Then the scalar potential decomposes into a quadratic term plus cubic and quartic ones \cite{, , }. Finally, diagonalizing the quadratic mass terms, the physical Higgs states and their masses are obtained. The other terms, cubic and quartic ones, define the couplings and the interactions in the model. As a result, the free parameters of the model are the masses of the neutral Higgs bosons $(m_{h/H^0/A^0})$ and charged Higgs bosons ($m_{H^\pm}$), the ratio of the vacuum expectation values ($t_\beta=\tan\beta=v_1/v_2$), the mixing angle between the CP-even neutral Higgs states ($\alpha$) and the soft breaking scale of the discrete symmetry $m^2=m_{12}^2/(\sin\beta\cos\beta)$ \cite{}. The masses of charged and CP-odd Higgs states are defined as follows: \begin{equation} m^2_{A^0} =\frac{m_{12}^2}{\sin\beta\cos\beta} -\lambda_5 v^2, \;\; m^2_{H^\pm} =m^2_{A^0} + \frac{1}{2}(\lambda_5-\lambda_4)v^2\;. \end{equation} The masses of the CP-even states are defined below. \begin{equation} \begin{pmatrix} h^0 \\ H^0 \end{pmatrix} = \mathcal{R} \begin{pmatrix} m_{12}^2t_\beta +\lambda_1 v_1 ^2 & -m_{12}^2+\lambda_{345} v_1 v_2 \\ -m_{12}^2+\lambda_{345} v_1 v_2 & m_{12}^2/t_\beta +\lambda_2 v_2^2 \end{pmatrix} \mathcal{R}^T \end{equation} where $\mathcal{R}$ is a unitary rotation matrix which diagonalizes the CP-neutral Higgs mass matrix \cite{} as a function of $(\beta-\alpha)$, and $\lambda_{345}=\lambda_3+\lambda_4+\lambda_5$. After the discovery of the Higgs boson at the LHC, $h^0$ state was extensively defined as the SM-like Higgs boson with the same mass and couplings. This choice was called the exact alignment limit where $(\beta-\alpha)=\pm\pi/2$. The angle $s_{\beta-\alpha}=\sin(\beta-\alpha)$ defines the mixing between the CP-even Higgs states. In this study, the exact alignment limit is chosen, $s_{\beta-\alpha}=1$, consequently, $h^0$ becomes indistinguishable from the SM Higgs boson $H$. To suppress the FCNC at the tree-level in the Higgs-fermion interactions, the same discrete symmetry, which was imposed before on the scalar potential, can be extended to Yukawa sector. These interactions are written in four different and independent ways. The couplings between fermions and Higgses are named Type-I to -IV\cite{}. Setting a different Yukawa coupling scheme does not have a noticeable effect on the results. Therefore, the calculation is performed only considering the Type-I Yukawa coupling scheme, though the results are the same in the numerical precision for Type-II. \section{Machinery for the numerical analysis and the parameter space} \label{sec3} \subsection{Scattering processes and the cross-section} The machinery of the calculation and conventions of the scattering processes are presented here. Throughout this paper, the scattering processes are denoted as \begin{eqnarray} e^+ (k_1,\mu) +e^- (k_2,\nu)\;\rightarrow \; Z (k_3) + H_i (k_4)+ \gamma (k_5), \label{} \end{eqnarray} where $k_a$ $(a=1,...,5)$ are the four-momenta of the incoming positron and electron beam, the outgoing Z-boson, neutral Higgs-bosons, and photon. Additionally, spin polarizations of the positron and the electron are indicated by $\mu$ and $\nu$, respectively. Feynman diagrams which contribute to the process $e^+e^- \rightarrow \hzg$ at the tree-level are shown in figure \ref{fig1}. The 2HDM Lagrangian and a detailed phenomenological discussion were reported in refs. \cite{, , , }. The vertices were defined in \textsc{FeynArts} \cite{, }; the amplitudes are constructed using \textsc{FeynArts}. Since the computation is carried out at the tree-level, the \emph{Unitary gauge} suits better for the calculation. The simplification of the fermion chains, squaring amplitudes, and the numerical analysis are accomplished using the driver program given in \textsc{FormCalc} \cite{}. \begin{widetext} \onecolumngrid \begin{figure}[htbp] \centering \includegraphics[width=.75\textwidth]{epem_gh0Z0_2hdm.pdf} \caption{\label{fig1} The Feynman diagrams which contribute each of the scattering processes at the tree level are plotted. $H_i$ represents the neutral Higgs bosons $h^0$, $H^0$, and $A^0$. In the production of $ZA^0\gamma$, the intermediated propagators $A^0$ and $Z$ are exchanged with $h^0$ and $H^0$ boson, respectively.} \end{figure} \end{widetext} The relevant couplings for all the scattering processes are given in Table \ref{tab1}. The parameters $Y_{(1,2,3)}=(c_\alpha/s_\beta,s_\alpha/s_\beta,-\cot\beta)$ for Type-I and -IV, and they are $(-s_\alpha/c_\beta, c_\alpha/c_\beta,\tan\beta)$ for Type-II and -III. It can be seen that the most important parameter for all the scattering processes is $s_{\beta-\alpha}$. It is involved in $c_{\{h^0,H^0\}A^0Z}$ and $c_{\{h^0,H^0\}ZZ} $ couplings. Therefore, these couplings are the effective ones in each scattering process. The couplings between the incoming electron-positron and the neutral Higgs bosons, $c_{e^+e^- \{h^0, H^0, A^0\}}$, are also a function of the free parameters of the 2HDM through $Y_i$. However, the factor $m_e/m_w$ suppresses them heavily; thus, they do not contribute significantly. As a result, it can be said that choosing one Yukawa coupling scheme (Type-I to -IV) over the others makes no important difference in all the scattering processes, and thus the total scattering amplitudes are affected only by the mixing angle $s_{\beta-\alpha}$. \begin{table}[htbp] \centering \caption{\label{tab1} The couplings involved in all scattering processes. The weak angle is abbreviated as $s_w=\sin\theta_w$, $c_w=\cos\theta_w$, and $c_{\beta-\alpha}=\cos(\beta-\alpha)$.} \footnotesize \begin{tabular*}{80mm}{l@{\extracolsep{\fill}}l | lc} \hline\hline $c_{\{\gamma,Z\}e^+e^-}$ & & $i e\left \{ \begin{pmatrix} 1 \\ 1\\ \end{pmatrix}, \begin{pmatrix} c_{2w}/s_{2w} \\ -s_w/c_w\\ \end{pmatrix}\right \}$ \\ $c_{\{h^0,H^0,A^0\}e^+e^- }$ & & $-\frac{i e m_e }{2m_w s_w} \left\{ \begin{pmatrix} Y_1 \\ Y_1\\ \end{pmatrix}, \begin{pmatrix} Y_2 \\ Y_2\\ \end{pmatrix}, \frac{1}{i}\begin{pmatrix} Y_3 \\ -Y_3\\ \end{pmatrix} \right\}$ \\ $c_{\{h^0,H^0\}A^0Z}$ & & $\frac{e}{2c_w s_w}\{c_{\beta-\alpha},s_{\beta-\alpha},\}$ \\ $c_{\{h^0,H^0\}ZZ } $ & & $\frac{i e m_w}{c_w^2 s_w}\{s_{\beta-\alpha},c_{\beta-\alpha}\}$ \\ \hline \hline \end{tabular*} \end{table} The couplings given in Table \ref{tab1} are plotted in Figure \ref{fig2} as a function of $s_{\beta-\alpha}$. In Figure \ref{fig2}, Type-I Yukawa coupling scheme and $t_\beta=10$ are assumed. Though the numerical values of the couplings are the same for Type-I, only the couplings $c_{e^+e^- \{h^0,H^0,A^0\}}$ get lower. The highest contribution comes from $c_{ZZh^0}$ or $c_{ZZH^0}$ couplings, at the order of 65. The coupling $c_{ZZH^0}$ has the maximum value at $s_{\beta-\alpha}=0$, and it lowers when $s_{\beta-\alpha}\rightarrow1$, while the coupling $c_{ZZh^0}$ rises with $s_{\beta-\alpha}$. It should be noted that there is no $c_{ZZA^0}$ coupling in the 2HDM, and thus the cross-section of $e^+e^- \rightarrow Z A^0 \gamma$ is suppressed heavily. The couplings between fermion pairs and the neutral Higgses are at the order of $10^{-5}$, so they do not have much impact on the total amplitude. The couplings $c_{e^+e^-\gamma}$ and $c_{e^+e^-Z}$ are universal, and they are not a function of the free parameters of the 2HDM. Last, the couplings $c_{A^0 Z \{h^0, H^0\}}$ are also a function of $s_{\beta-\alpha}$, but their values are around $0.37$; one of them drops while the other one rises. In conclusion, the couplings $c_{ZZh^0}$ and $c_{ZZH^0}$ are the most effective ones, and the cross-section of all these three processes are influenced heavily by $s_{\beta-\alpha}$. \begin{figure}[htbp] \centering \includegraphics[width=0.48\textwidth]{plot_couplings_type1.pdf} \caption{\label{fig2} The couplings which take part in the scattering of all the channels are plotted as a function of $s_{\beta-\alpha}$. The $t_\beta=10$ and Type-I Yukawa coupling scheme are assumed.} \end{figure} The differential cross-section for three final state is defined as follows \cite{}: \begin{equation} \frac{d\sigma (s;\mu,\nu)}{dk^0_5 dk^0_3 d\cos\theta \, d\eta}=\frac{2}{ (4\pi)^4}\frac{1}{\Phi} \left(\frac{1}{4} \sum_{pol}{|\mathcal{M}(s;\mu,\nu)_{tot}|^2}\right), \label{} \end{equation} where the amplitude $\mathcal{M}_\text{tot}$ is squared and summed over the polarization vectors, $\Phi=2\sqrt{\lambda(s,m_e^2,m_e^2)}$ is the flux factor of the incoming $e^+e^-$ beams. The Monte-Carlo integration methods are required in the computation over the phase space of the final states, therefore, \textsc{CUBA} \cite{, } routines are used. The cross-section for an arbitrary degree of longitudinal beam polarization is defined as follows: \begin{equation} \label{} \sigma(P_{e^+},P_{e^-})=\frac{1}{4} \sum_{\alpha,\beta=\pm1} (1+\alpha P_{e^+})(1+\beta P_{e^-}){\sigma}_{\alpha\beta} \;, \end{equation} The $\sigma_{LR}$ in equation \ref{} represents the cross-section with a completely left-handed polarized ($P_{e^+} = -1=L$) positron beam and a right-handed polarized electron beam ($P_{e^-} = +1=R$). The cross-sections $\sigma_{RL}$, $\sigma_{LL}$ and $\sigma_{RR}$ are defined analogously. Note that due to the nature of the scattering process $\sigma_{LL}$ and $\sigma_{RR}$ don't make any important contribution, and they are neglected safely in equation \ref{}. Finally, due to the massless nature of the photon and the phase space of the outgoing particles, soft and collinear IR singularities appear naturally in the calculation. Therefore, motivated from photon energy resolution of the experimental apparatus (detector), we applied cut on the transverse momentum ($p_{T}^\gamma$) of the photon. As a result, the cross-section is calculated in a realistic environment. The cross-sections of each process are explored for varying the transverse momentum of the photon which is required to be $p_{T}^\gamma>p_{T,\text{cut}}^\gamma=(5,10,15,20,25)\text{ GeV}$. \subsection{Constraints and benchmark scenario} The first set of constraints comes from a theoretical point of view: stability, unitarity, and perturbativity. Stability ensures that the scalar potential is positive even at large values of the field \cite{el2007consistency, sher1989electroweak, deshpande1978pattern, , }. Unitarity ensures that the scattering amplitudes are flat at asymptotically high energies \cite{ginzburg2005tree}. Perturbativity is the condition where all the quartic couplings in the 2HDM need to be smaller than a particular value ($16\pi$). In this study, \textsc{2HDMC}(v1.7.0) \cite{} is used to test whether the benchmark points obey these theoretical constraints. The 2HDM was studied by many in the previous and still ongoing experiments, and numerous results have been piled up. All these results restrict the parameter space of the 2HDM from various angles. It was concluded in the previous section that the amplitudes are dependent mostly on the $s_{\beta-\alpha}$. The phase spaces in all the scattering processes are a function of the masses of the neutral Higgs bosons. Therefore, all the scattering processes are unaffected by all the experimental constraints, except the $s_{\beta-\alpha}$ and the masses of the neutral Higgs bosons. Charged currents are important for the flavor observables because they can make novel contributions. However, we do not regard the charged Higgs mass constraints, because it does not make any direct contribution to the process (see figure \ref{fig1} for couplings and intermediated particles). However, the value of $m_{H^\pm}$ most definitely affects the couplings in the model. Choosing up a wrong mass for the charged Higgs boson may not hold the theoretical and experimental constraints introduced in this section. Inspired by ref. \cite{} and the results presented by LHC \cite{}, we set masses of the extra Higgs bosons as $m_{A^0} = m_{H^\pm}$. This also satisfies the constraints on the oblique parameters \cite{, , , gunion2003cp, , }. It should be underlined that if the parameter $s_{\beta-\alpha}$ is fixed to unity, then the couplings of light Higgs ($h^0$) resemble the discovered SM Higgs boson. The parameter $t_\beta$ affects the Yukawa couplings, but they make no important contribution to the scattering processes; still, a value needs to be assigned. According to ref. \cite{}, the low $t_\beta$ values were strongly constrained by $\bar{\mathcal{B}}(B_s^0 \rightarrow \mu^+\mu^-)$ and neutral meson mixings $\Delta M_s$ in Type-I, so $t_\beta>1$ can be considered. The masses of the scalars are not constrained on large $t_\beta$ range compared to the Type-II. Thus, $t_\beta=10$ is picked. The calculation is performed following those experimental constraints. \section{Results and discussion} \label{sec4} In this section, the numerical results for the production of $Z H_i \gamma$ in an $e^+e^-$-collider are presented and discussed. The following SM parameters are taken from ref. \cite{}: $m_e=0.5109989\text{ MeV}$, $m_Z=91.1876\text{ GeV}$, $m_{W^\pm}=80.385\text{ GeV}$ and $\alpha=1/127. 944$. We used the current known value for the Higgs mass $m_H=125.09\text{ GeV}$ \cite{2015PhRvL.114s1803A}. \subsection{Cross-section distributions} The numerical analysis is carried out for unpolarized and polarized incoming beams. If it is not stated otherwise, we set $m_{H_i}=150\text{ GeV}$ and $t_\beta=10$ in the computation. In figure \ref{fig3}, the cross-section distributions are given for various polarizations as a function of the CM energy. These polarization configurations were discussed before in various refs. such as \cite{}, we adopted $(P_{e^+},P_{e^-})=(+0.6,-0.8)$ and $(+0.3,-0.8)$ in this study. The vertical dashed lines indicate the proposed energies of each lepton collider. In figure \ref{fig3} (left), the distributions $e^+e^-\rightarrow Zh^0\gamma$ are plotted, while on (right) the same distributions for $e^+e^-\rightarrow ZH^0\gamma$ are shown. Since the $h^0$-state behaves like the SM Higgs boson in the exact alignment limit ($s_{\beta-\alpha}=1$), the distributions in figure \ref{fig3} (left) are in good agreement with the SM tree-level results presented in refs. \cite{, }. In this limit, results are similar with the SM results as expected. \begin{widetext} \onecolumngrid \begin{figure}[htbp] \centering \includegraphics[width=0.48\textwidth]{plot_1Z0h0g_cba0.pdf}\hfill \includegraphics[width=0.48\textwidth]{plot_2Z0H0g_cba2.pdf} \caption{ \label{fig3} The distributions of the cross-section are plotted for various polarization configurations of the incoming $e^+e^-$ beams. The calculation is carried for $p_{T,\text{cut}}^\gamma=10\text{ GeV}$ on both figures. (left): The cross-section distributions are for $e^+e^- \rightarrow Z h^0 \gamma$. The exact alignment limit $s_{\beta-\alpha}=1$ is assumed. (right): The cross-section distributions are for $e^+e^- \rightarrow Z H^0 \gamma$. The mixing angle of the neutral Higgs bosons is taken as $c_{\beta-\alpha}=0.2$ ($s_{\beta-\alpha}\approx 0.98$).} \end{figure} \end{widetext} The unpolarized cross-section reaches up to $6.19 \text{ fb}$ around the planned collision energy of the FCC-ee collider ($\sqrt{s}=350\text{ GeV}$), and it slowly declines at higher $\sqrt{s}$ values. At the ILC ($\sqrt{s}=500\text{ GeV}$), the unpolarized cross-section gets $\sigma_{UU}\sim 4.86\text{ fb}$. The proposed CEPC running at $\sqrt{s}=240\text{ GeV}$ could still probe the process, but not as significant as the others, the unpolarized cross-section is around $1.11 \text{ fb}$. If the future colliders are capable of polarizing the incoming electron and positron beams, the cross-section will be enhanced dramatically. The polarized cross-sections are also added in both graphics given in figure \ref{fig3}. When the positron beam is completely left-handed and the electron beam is right-handed, $\sigma_{LR}$ is around $9.72 \text{ fb}$. In the opposite case where polarization is set to $(+1,-1)$, then $\sigma_{RL}\sim 15.04\text{ fb}$. Two other polarization configurations are also plotted in figure \ref{fig3} where the peaks are around the FCC-ee operation range, $\sigma_{(+0.3,-0.8)}\sim 9.14\text{ fb}$ and $\sigma_{(+0.6,-0.8)}\sim 11.02\text{ fb}$. We can see that polarizing the beams enhances the cross-section up to a factor of 1.8 in (+0.6, -0.8) configurations. The mixing angle $s_{\beta-\alpha}$ plays an important role in the calculation, the cross-section of $e^+e^- \rightarrow Z H^0 \gamma$ vanishes in the exact alignment limit ($s_{\beta-\alpha}=1$). Therefore, we let the $s_{\beta-\alpha}$ deviate from unity by only $\sim2\%$, thus the parameter $c_{\beta-\alpha}=0.2$ is assumed in figure \ref{fig3} (right) for the process $e^+e^- \rightarrow Z H^0 \gamma$. The cross-section distributions for possible polarizations of the $e^+e^-$ are presented in figure \ref{fig3} (right). Compared to the previous scattering process, the cross-section of $e^+e^- \rightarrow Z H^0 \gamma$ is smaller. The unpolarized cross-section is around $\sigma_{UU}\sim 0.164\text{ fb}$ for the planned FCC-ee ($\sqrt{s}=350\text{ GeV}$) project. In addition to this, the cross-section of the completely polarized beams gets $\sigma_{RL}\sim 0.398\text{ fb}$ and $\sigma_{LR}\sim 0.258\text{ fb}$ at $\sqrt{s}=350\text{ GeV}$. The distributions of the cross-section as a function of energy have the same trend as the previous process. The cross-section peaks around $\sqrt{s}\sim 400\text{ GeV}$, then it decreases rapidly moving to higher energies due to the suppression of 1/s. Due to the kinematical threshold of the process (the total mass of $H^0$-boson, Z-boson, and $\gamma$ with the transverse energy of 10 GeV), the CEPC is only capable of measuring the process $e^+e^- \rightarrow Z H^0 \gamma$ with $m_{h^0}\lesssim m_{H^0}\lesssim 140\text{ GeV}$ mass constraint. The CEPC has limits measuring this particular parameter space. For the ILC ($\sqrt{s}=500\text{ GeV}$), the cross-section reaches $\sigma_{UU}\sim0.157\text{ fb}$. The polarized cross-section values reach up to $\sigma(+0.3,-0.8)=0.231\text{ fb}$ and $\sigma(+0.6,-0.8)=0.279\text{ fb}$. The polarization of the $e^+e^-$-beams raises the cross-section by a factor of 1.8 compared to the unpolarized values. \begin{figure}[htbp] \centering \includegraphics[width=0.45\textwidth]{plot_cba_analysis_sqrt350GeV.pdf} \caption{ \label{fig4} The cross-section distributions of $e^+e^- \rightarrow Z h^0 \gamma$ and $e^+e^- \rightarrow Z H^0 \gamma$ are plotted as a function of $s_{\beta-\alpha}$ at $\sqrt{s}=350\text{ GeV}$. The figure shows the cross-sections in departing from the exact alignment limit. The calculation is carried for $p_{T,\text{cut}}^\gamma=10\text{ GeV}$ and $t_\beta=10$ on both processes. } \end{figure} The readers might wonder how the production cross-section of these two processes differ when one departs from the exact alignment limit ($s_{\beta-\alpha}=1$). In figure \ref{fig4}, the distributions of $e^+e^- \rightarrow Z h^0 \gamma$ and $e^+e^- \rightarrow Z H^0 \gamma$ at $\sqrt{s}=350\text{ GeV}$ are plotted for the unpolarized incoming beams. The vertical dashed line points the $s_{\beta-\alpha}=0.98$ value. The scattering process $e^+e^- \rightarrow Z h^0 \gamma$ gets lower while $s_{\beta-\alpha}\rightarrow0$, and $\sigma(e^+e^- \rightarrow Z h^0 \gamma)$ is $\sim 5.92\text{ fb}$ at $s_{\beta-\alpha}=0.98$. It is lowered by $4.36\%$ compared to $s_{\beta-\alpha}=1$. Contrary, the process $e^+e^- \rightarrow Z H^0 \gamma$ gets higher for lower $s_{\beta-\alpha}$ values as expected. Deviating from the exact alignment limit by $\sim2\%$ enhances the cross-section; it is $\sigma(e^+e^- \rightarrow Z H^0 \gamma)\sim 0.18\text{ fb}$ at $s_{\beta-\alpha}=0.98$. In table \ref{tab2}, the cross-section values of the $e^+e^- \rightarrow Z h^0 \gamma$ and the $e^+e^- \rightarrow Z H^0 \gamma$ at different CM energies (and colliders) are given for different cuts on the photon's transverse momentum ($p_{T,\text{cut}}^\gamma$). It can be seen that raising the variable $p_{T,\text{cut}}^\gamma$ chops the cross-sections. The effect is most dramatic for the CEPC due to the collider's low CM energy. If the photon's transverse momentum is set as low as $p_{T,\text{cut}}^\gamma=5\text{ GeV}$ the cross-section becomes as high as 3.45 fb for the $e^+e^- \rightarrow Z h^0 \gamma$, unfortunately the $e^+e^- \rightarrow Z H^0 \gamma$ is not available kinematically. Since the $p_{T,\text{cut}}^\gamma$ is a specific parameter of the measuring apparatus, an electromagnetic calorimeter with low energy resolution will increase the acceptance, and this will result in a higher number of events. Considering the other lepton collider proposals, FCC-ee and ILC, the cut on the $p_{T}^\gamma$ is not vital, but raising the cut lowers the cross-section as expected. \begin{tiny} \begin{table}[htbp] \centering \caption{\label{tab2} The unpolarized cross-section for $e^+e^- \rightarrow Z h^0 \gamma$ and $e^+e^- \rightarrow Z H^0 \gamma$ processes with varying cuts on photon's transverse energy ($p_{T,\text{cut}}^\gamma$). All energies are given in GeV.} \footnotesize \begin{tabular*}{90mm}{c@{\extracolsep{\fill}}| ccc | ccc} \toprule Collider & CEPC & FCC-ee & ILC & CEPC & FCC-ee & ILC \\ $\sqrt{s}$ & (240) & (350) & (500) & (240) & (350) & (500) \\ \hline\hline $p_{T,\text{cut}}^\gamma$ & \multicolumn{3}{|c|}{$\sigma_{e^+e^- \rightarrow Z h^0 \gamma}$ (fb)} & \multicolumn{3}{c}{$\sigma_{e^+e^- \rightarrow Z H^0 \gamma}$ (fb)} \\ \cline{2-7} 5 & 3.450 & 9.189 & 6.536 & - & 0.266 & 0.223 \\ 10 & 1.077 & 5.948 & 4.674 & - & 0.164 & 0.157 \\ 15 & 0.288 & 4.348 & 3.714 & - & 0.115 & 0.123 \\ 20 & 0.027 & 3.345 & 3.091 & - & 0.085 & 0.101 \\ 25 & - & 2.644 & 2.641 & - & 0.064 & 0.085 \\ \hline \end{tabular*} \end{table} \end{tiny} The future lepton colliders could be used to study these two scattering processes. If the total luminosity of the future lepton colliders is expected to be at the order of 2000 $\text{fb}^{-1}$ \cite{}, then according to Table \ref{tab2} the total number of events at $\sqrt{s}=350\text{ GeV}$ (FCC-ee) with $p_{T}^\gamma>10\text{ GeV}$ are expected to be $\approx 11.9\text{k}$ and $\approx 330$ for $e^+e^- \rightarrow Z h^0 \gamma$ and $e^+e^- \rightarrow Z H^0 \gamma$ processes, respectively. The process $e^+e^- \rightarrow Z h^0 \gamma$ has a remarkable cross-section, but the process $e^+e^- \rightarrow Z H^0 \gamma$ has a considerably smaller one. The branching ratios of the final state particles and the acceptance of the detector are necessary to estimate the precise number of events that can be expected in the detector. Therefore, detector simulations for signal and background channels are required. The process $e^+e^- \rightarrow Z H^0 \gamma$ can be used to study machine learning algorithms due to the expected small number of events. Lastly, the cross-section of $e^+e^- \rightarrow Z A^0 \gamma$ is at order of $10^{-13}\text{ fb}$ due to missing $c_{ZZA^0}$ coupling. The process cannot be studied in the future lepton colliders. Therefore, the results are not presented. \subsection{Polarization of the incoming beams} The cross-section contours are given in figure \ref{fig5} for possible polarization configurations of the incoming beams. The ratio of $\sigma(P_{e^+},P_{e^-})/\sigma_{UU}$ is plotted as a function of ($P_{e^+},P_{e^-}$) for the process $e^+e^- \rightarrow Z h^0 \gamma$ at $\sqrt{s}=500\text{ GeV}$. The cross-section is obtained with $p_{T,\text{cut}}^\gamma=10\text{ GeV}$, $t_\beta=10$, $s_{\beta-\alpha}=0.98$ and $m_{H^0}=150\text{ GeV}$. The same distribution is obtained for the $e^+e^- \rightarrow Z H^0 \gamma$ because the similar Feynman diagrams contribute to the total amplitude. \begin{figure}[htbp] \centering \includegraphics[width=0.48\textwidth]{plot10_cba0.pdf} \caption{\label{fig5} The ratio of the cross-sections ($\sigma(P_{e^+},P_{e^-})/\sigma_{UU}$) as a function of the polarization of $e^+e^-$ beams. The computation assumes $p_{T,\text{cut}}^\gamma=10\text{ GeV}$, $t_\beta=10$, $s_{\beta-\alpha}=0.98$ and $\sqrt{s}=500\text{ GeV}$. } \end{figure} If the future colliders are able to polarize the positron beam right-handed and electron beam left-handed, then the cross-section values will be higher. It can be seen that in figure \ref{fig5}, the enhancement is raised to 2.25 at the left top corner for both processes. Additionally, the computations showed that the ratio $\sigma(P_{e^+},P_{e^-})/\sigma_{UU}$ doesn't change with the parameter $s_{\beta-\alpha}$. The contours remain universal because the ratio $\sigma(P_{e^+},P_{e^-})/\sigma_{UU}$ is not a function of energy. Thus, the ratio distribution as a function of the polarization of incoming beams stay unchanged at higher CM energies. \subsection{Asymmetry distributions} We explored the forward-backward asymmetry distribution between the neutral Higgs boson and the photon. Since the weak interaction violates parity symmetry, it produces asymmetric contributions to angular observables. Therefore, $A_{\text{FB}}$ could be a good observable for testing type of electroweak model for the Higgs bosons. The asymmetry for each processes is defined as follows: \begin{equation} A_{FB}=\frac{\int_0^1\frac{d\sigma}{d\cos\theta_{H_i\gamma}}d\cos\theta_{H_i\gamma}-\int_{-1}^0\frac{d\sigma}{d\cos\theta_{H_i\gamma}}d\cos\theta_{H_i\gamma}}{\int_0^1\frac{d\sigma}{d\cos\theta_{H_i\gamma}}d\cos\theta_{H_i\gamma}+\int_{-1}^0\frac{d\sigma}{d\cos\theta_{H_i\gamma}}d\cos\theta_{H_i\gamma}}. \end{equation} As usual the $p_T^\gamma>10\text{ GeV}$ cut is applied in the calculation. In figure \ref{fig6} (left), the asymmetry is plotted as a function of CM energies for unpolarized incoming beams, free parameters are given in the caption. In the computation, we obtained that in both processes most of the events are accumulated more in the backward direction than the forward direction. As a result, the forward-backward asymmetry distributions become negative for both processes. It gets a value of -0.38 around $\sqrt{s}=240\text{ GeV}$ (at the CEPC) for the $e^+e^- \rightarrow Z h^0 \gamma$, and at higher CM energies the asymmetry reaches a peak around $\sqrt{s}=270\text{ GeV}$ then falls. The processes $e^+e^- \rightarrow Z h^0 \gamma$ and $e^+e^- \rightarrow Z H^0 \gamma$ share the same trend overall, but the asymmetry in $e^+e^- \rightarrow Z H^0 \gamma$ increases slightly in higher CM energies due to the kinematics of the final state particles. The asymmetry at $\sqrt{s}=1\text{ TeV}$ is obtained to be close to -0.490 and -0.456 for $e^+e^- \rightarrow Z h^0 \gamma$ and $e^+e^- \rightarrow Z H^0 \gamma$, respectively. The asymmetry is calculated for $s_{\beta-\alpha}=0.98$, but it should be noted that the distributions do not change with $s_{\beta-\alpha}$. The dependence on $s_{\beta-\alpha}$ in cross-sections and couplings is dropped in the asymmetry distributions. The asymmetry is dependent on the masses of the Higgses in the final state. In higher $m_{H^0}$ values, asymmetry distribution of $e^+e^- \rightarrow Z H^0 \gamma$ moves to the right with the same trend. Various asymmetry distributions for the SM were presented in ref. \cite{}, the same outcome is obtained; the number of events collected in the backward direction is higher than the forward direction. Therefore, the asymmetry has to be negative. However, the numbers are positive in ref. \cite{}, we assume there is a sign error in the calculation. The difference between the asymmetry distributions of $e^+e^- \rightarrow Z h^0 \gamma$ and $e^+e^- \rightarrow Z H^0 \gamma$ are calculated due to phenomenological curiosity. The difference in asymmetry between two processes might be used to test whether both processes share similar Feynman diagrams topology. The asymmetry difference as a percentage is defined as follows: \begin{equation} \Delta A_{FB}= 100 \times\frac{(A_{\text{FB}}^{e^+e^- \rightarrow Z h^0 \gamma}-A_{\text{FB}}^{e^+e^- \rightarrow Z H^0 \gamma})}{A_{\text{FB}}^{e^+e^- \rightarrow Z h^0 \gamma}}. \end{equation} The percentage of difference in asymmetry between two processes is plotted in figure \ref{fig6} (right) as a function of the CM energy with various $m_{H^0}$ masses. It can be seen that the difference is steady for $m_{H^0}=125\text{ GeV}$ in $0.5-1$ TeV range. Since the asymmetry for both of the processes has the same tendency as a function of the CM energy, and moves to the right with increasing $m_{H^0}$, the difference rises at high CM energies with increasing $m_{H^0}$ values. It should be noted that the cross-section lowers dramatically with higher $m_{H^0}$. The difference in the asymmetry is around 3.6\% at the FCC-ee, and it quickly reaches to 6\% at the ILC energies for $m_{H^0}=150\text{ GeV}$. The figure shows that the ILC is more advantageous for measuring the asymmetry difference in various $m_{H^0}$ scenarios. \begin{widetext} \onecolumngrid \begin{figure}[htbp] \centering \includegraphics[width=0.48\textwidth]{plot6a_cba01.pdf}\hfill \includegraphics[width=0.48\textwidth]{plot6c.pdf} \caption{\label{fig6} The asymmetry is plotted with $p_{T,\text{cut}}^\gamma=10\text{ GeV}$, and $t_\beta=10$ for both processes. $s_{\beta-\alpha}=0.98$ is considered for each processes, but it has no impact on the asymmetry. (left): The asymmetry distributions with and $m_{H^0}=150\text{ GeV}$. (right): The asymmetry difference between two processes as a percentage, ($100\times(A_{\text{FB}}^{e^+e^- \rightarrow Z h^0 \gamma}-A_{\text{FB}}^{e^+e^- \rightarrow Z H^0 \gamma})/A_{\text{FB}}^{e^+e^- \rightarrow Z h^0 \gamma}$) with various $m_{H^0}$ masses. } \end{figure} \end{widetext} \section{Conclusion} \label{sec5} In this study, we focused on computing the scattering processes $e^+e^-\rightarrow Z H_i \gamma$ (i=1,..,3) at the tree level. They were investigated for polarized and unpolarized incoming beams in the context of 2HDM. The results were obtained for a scenario that is favored by recent experimental outcomes and also widely accepted. When $s_{\beta-\alpha}=1$, the couplings of the $h^0$ boson become the same with the SM Higgs boson, therefore one of the Higgs bosons ($h^0$) becomes indistinguishable from the SM Higgs boson (H); this limit is called the exact alignment limit. The cross-section of $e^+e^- \rightarrow Z H^0 \gamma$ diminished completely. No conclusion can be drawn about the 2HDM or the underlying scalar sector from $e^+e^- \rightarrow Z h^0 \gamma$ because it has the same result as the SM. The production rates of both processes were explored by letting $s_{\beta-\alpha}$ deviated from this limit. Even a small deviation raised the cross-section of process $e^+e^- \rightarrow Z H^0 \gamma$, while the cross-section of $e^+e^- \rightarrow Z h^0 \gamma$ decreased by more than $4\%$. Since the cross-section distribution of $e^+e^- \rightarrow Z h^0 \gamma$ was a mere function of $s_{\beta-\alpha}$, measuring the cross-section of $e^+e^- \rightarrow Z h^0 \gamma$ is useful to extract the parameter $s_{\beta-\alpha}$. Thus, the couplings $c_{[h^0, H^0]A^0Z}$ and $c_{[h^0, H^0]ZZ}$ can be determined. The scattering process $e^+e^- \rightarrow Z H^0 \gamma$ is a function of the same couplings, so its cross-section can be obtained. It could be argued that the Yukawa couplings too take part in the calculation, but they are very small to make a noticeable contribution. Therefore, their contribution could be ignored. Polarizing the incoming beams changes the contribution of various Feynman diagrams given in figure \ref{fig1}, and the cross-section can get higher values. We compared the polarization configurations, and obtained that the cross-section was enhanced up to a factor of 1.8 with $P_{e^-}=-0.80$ and $P_{e^+}=+0.60$ beams. In conclusion, the polarization of the incoming beams increased the production cross-section and the number of events in the collider. The forward-backward asymmetry of the Higgses and the photon was presented for both processes. We obtained that the parameter $s_{\beta-\alpha}$ did not have an impact on the asymmetry distributions. Therefore, it can be deduced that the asymmetry distributions alone can not indicate the underlying model in nature. Among three proposals, the FCC-ee was the one with a higher cross-section for the $e^+e^- \rightarrow \hzg$. The CEPC had low CM energy, but it still had the potential to study the process $e^+e^- \rightarrow Z h^0 \gamma$, but not $e^+e^- \rightarrow Z H^0 \gamma$. The ILC shared the same remarks with the FCC-ee. The leading motivations of these future colliders are to measure the properties of the Higgs boson and obtain hints of the extended Higgs sector. This study presented the potential of exploring these processes in future lepton colliders. If these proposed experiments can find deviations from the SM or solid experimental proof of the existence of new physics, a new era will be opened in the history of mankind, and our understanding of nature will change completely. \acknowledgments The numerical calculations reported in this paper were partially performed at TUBITAK ULAKBIM, High Performance and Grid Computing Center (\textsc{TRUBA} resources) and the computing resource of \textsc{FENCLUSTER} (Faculty of Science, Ege University). Ege University supports this work, project number 17-FEN-054.

\section{Introduction} \label{} The translocation of biomolecules through nanopores has been an active research field for more than two decades now, implying both experimental and theoretical interesting aspects. Translocation consists in a biomolecule crossing a membrane through a hole of nanometric size (the nanopore) from the cis- to the trans- side. This process can either be natural (unbiased) or forced. Two main methods have emerged in the latter case. In the first, the translocating polymer is driven by an electrophoretic field by applying a potential bias between cis and trans sides. In the second, driving consists of a mechanical forces applied directly to one end of the polymer by using optical or magnetic tweezers. Kasianowicz {\em et al.}~\cite{Kasianowicz1996} were first to successfully perform a forced DNA translocation through a biological nanopore. This experiment paved the way for the use of translocation to characterize single objects such as single or double stranded DNA, proteins, or, at larger scales, even cells~\cite{Dekker2007, Keyser, Meller2003, Movileanu2008, Palyulin2014,Wanunu2012, Liu2013}. High impact applications of translocation in biotechnology and medical diagnostics are expected, with clear emphasis on quicker and cheaper methods for DNA sequencing~\cite{Branton2008,Clarke2009,Merchant2010,Schneider2012}. Current limitations to the use of DNA translocation as a sequencing tool are both spatial and temporal. More precisely, in the case of common biological and artificial nano-pores, several nucleotides are present simultaneously within the nanopore during the translocation. This hinders the possibility of single-base sequence resolution~\cite{Schneider2010}. Also, in most experiments a base spends about 1~$\mu$s within the pore, while current measurement resolution times would require a significant slowing down of the process, with a typical occupation time of about 1~ms~\cite{Branton2008}. Although attaching an optical (or magnetic) bead at the extremity of a polymer is substantially less straightforward compared to direct electrophoresis, this method is an effective candidate for a better control over the translocation process. In particular, it has been shown that it can reduce the mean translocation time by one order of magnitude~\cite{Kantor2004}. Another potentially valuable strategy to reduce the translocation speed is to use {\em narrow} nano-pores thus adding supplementary friction to the system. Both the pulled translocation~\cite{Keyser2006, Peng2009, VanDorp2009, Sischka2010, Bulushev2014, Bulushev2015, Bulushev2016} and the effects of the nanopore size~\cite{Yameen2009} have been investigated in experiments. Only few theoretical or numerical studies, however, have been carried out for both cases~\cite{Kantor2004, Grosberg2006, Huopaniemi2007, Panja2008, deHaan2010, deHaan2015, Menais2016,2017arXiv170809184S}. Furthermore, contrary to the case of a field driven translocation that has been extensively investigated~\cite{Sakaue2007,Sakaue2010,Saito2011,Saito2012,Ambjrnsson2005,2Ambjrnsson2005,Rowghanian2011,Dubbeldam2012,Ikonen2012,2Ikonen2012,Ikonen2013}, the main physical understanding of the translocation process of pulled polymers through nanopores remains controversial, only early results are available for high forces~\cite{2017arXiv170809184S}. This is mainly based on the methods of statistical mechanics, looking for universal mechanisms controlled by general parameters. In particular, there are discrepancies about the correct values for the critical exponents $\alpha$ and $\delta$ governing the translocation time, $\tau \propto N^{\alpha}F^{-\delta}$, as function of the polymer length $N$ and the applied force $F$~\cite{Milchev2011}. \begin{figure*}[t] \centering \includegraphics[width=0.4\textwidth,angle=90]{figure1.png} \caption{Polymer during a translocation process through the nanopore aperture of the membrane. Membrane beads are in green. Left: Linear polymer grains are depicted in red. Right: Structured polymer, the backbone is in red (sugar beads) and white (phosphate beads) while the lateral grains in purple are bases bonded to the sugar beads (red) from the backbone.} \label{FigModel} \end{figure*} In this paper, we report numerical simulations of a simple model for linear and structured polymers, translocating through a nanometric pore carved in a structured immobile membrane. We have considered toy numerical models which allow for a complete statistical treatment of the translocation processe. The {\em linear} polymer is a classic Rouse-like chain formed by connected beads. The {\em structured} polymer is obtained by the same linear chain (representing the polymer backbone), with additional beads regularly grafted to the backbone and mimicking single-stranded-DNA bases, at the coarse-grained level. The membrane is constituted by immobile interaction centers arranged on a hexagonal lattice reminiscent of the structure of graphene. Based on extended sets of data, we provide general insight on the effect of both polymer structure and nano-pore sizes. We propose simple scaling arguments based on the waiting times inside the pore to confirm the predicted value for $\alpha = 2$~\cite{Huopaniemi2007, Panja2008} for intermediate range forces and expand recent results~\cite{2017arXiv170809184S} to this regime. Furthermore, we demonstrate the existence of a threshold force for small pore size and we relate the underestimation of the predicted value $\delta=1$~\cite{Huopaniemi2007, Panja2008} at high forces by numerical simulations to the bond stretching~\cite{Luo2009}. \begin{figure*}[t] \centering \resizebox{0.3\textwidth}{!}{\includegraphics[trim={2cm 0 2cm 0 0}]{fig2a.pdf}}\resizebox{0.3\textwidth}{!}{\includegraphics[trim={2cm 0 2cm 0 0}]{fig2b.pdf}}\resizebox{0.3\textwidth}{!}{\includegraphics[trim={2cm 0 2cm 0 0}]{fig2c.pdf}}\rotatebox{90}{\resizebox{0.33\textwidth}{!}{\hspace{-0.8cm}\includegraphics[height=0.07\textwidth]{fig2leg.pdf}} } \caption{Representation of the various pore sizes studied with the polymer beads represented inside the pore. Backbone beads are depicted in blue and lateral bases in purple with their respective LJ radius. The membrane beads are depicted in black except the first layer of the nanopore depicted in red with their respective LJ radius for clarity. (Left) Large pore size for a simple linear polymer. (Middle) The same pore size corresponds to a small pore size for the structured polymer. (Right) Large pore size for the structured polymer.} \label{FigPlotofpores} \end{figure*} \section{Models and simulation details} \label{} In view of a complete statistical analysis of the translocation time and its dependence on the polymer structure and nanopore size, we have considered a coarse-grained model for both the polymer and the membrane. Our aim is to define a model simple enough to produce very extended ensembles of translocation events in order to completely characterize the statistical properties of the translocation time. However, we aim at keeping the essential properties of the DNA translocation which means to consider also a structured polymer model with additional bases attached to the backbone. We give below a few details about the models and the molecular dynamics simulation procedures. We refer the reader to the Supplemental Material (SM) \footnotemark[1] \footnotetext[1]{See Supplemental Material at [URL will be inserted by publisher] for supporting text and figures} and Ref.~\cite{Menais2016} for additional details. \subsection{The linear and structured polymer models} A minimalistic model for single stranded DNA was considered incorporating three bead types~\cite{Menais2016}. The alternating phosphate (P) and sugar (S) beads form the linear polymer backbone (Fig.~\ref{FigModel}, Top). Lateral beads, modeling the DNA bases (B), are attached to the S-beads of the polymer backbone to form the structured polymer (Fig.~\ref{FigModel}, Bottom). Note that no distinction is made at this level between the four bases (A-C-T-G). The total number of chain beads is $N$, one supplementary bead represents the pulling bead. For the structured polymer a lateral grain is attached every even chain bead ($N/2$ lateral beads). Steric and binding interactions only have been incorporated in the model. Steric interactions between any two beads, $i$ and $j$, are modeled by a truncated and shifted Lennard-Jones (LJ) potential: \begin{equation} U_{LJ}(r_{ij})= 4\epsilon \left[\left(\frac{\sigma_{ij}}{r_{ij}}\right)^{12}- \left(\frac{\sigma_{ij}}{r_{ij}}\right)^{6}\right] + \epsilon \label{Eq:LJ} \end{equation} with $U_{LJ}(r_{ij})=0$ for $r_{ij}\le 2^{1/6} \sigma_{ij}$. Here $r_{ij}$ is the distance between $i$ and $j$, $\epsilon_{ij}=\epsilon=1$, and $\sigma_{ij} = (\sigma_i + \sigma_j)/2$ with $\sigma_i$ the radius of bead $i$. This potential is therefore purely repulsive and prevents the beads from interpenetrating at very short distances. The sizes of the polymer backbone beads are $\sigma_{S}=\sigma_{P}=a=0.3$~nm. The size of the lateral bases in the structured polymer are slightly larger, $\sigma_{B}=3/2 a=0.45$~nm. Bonds between adjacent beads (both pertaining or grafted to the backbone) are modeled by the Finitely Extensible Nonlinear Elastic (FENE) anharmonic potential: \begin{equation} U_{FENE}(r_{ij})= -15\epsilon \left(\frac{R_{ij}}{\sigma_{ij}}\right)^2 \ln\left[1-\left(\frac{r_{ij}}{R_{ij}}\right)^2\right] \label{} \end{equation} with the the bond length value, $R_{ij}=3/2\sigma_{ij}$. Hydrogen bonding, base stacking and backbone bending are, therefore, not taken into account and, as a consequence, we cannot tackle issues involving the effect of stiffness, or the helicoidal structure building of the DNA double helix~\cite{Linak2011}. Note however that, since we focus our study on short single stranded DNA translocation without secondary structures formation like hairpins, these omitted degrees of freedom are expected to be irrelevant in the present context. To further reduce computational costs allowing for an accurate an statistical sampling of the translocation process, we disregard the presence of (explicit) solvent and neglect hydrodynamic interactions. Only Rouse dynamics of polymers is thus considered. We have tested the consistency of the considered polymer models with the exact predictions for the free polymer chain (see Sect.~II of the SM \footnotemark[1]). More in details, we have verified that the diffusion constant scales as the inverse of the polymer length, $D\propto N^{-1}$, due to the absence of hydrodynamic interactions. The friction of the polymer also scales as the length of the polymer, and we have also verified that the fluctuation-dissipation relation relating the two variables is fulfilled. Finally, the polymer radius of gyration (similar to the end-to-end distance) scales as $N^\nu$. The Flory exponent $\nu \simeq0.58$ is close to $\nu = 0.588$ expected for a Rouse-like linear polymer, while a higher $\nu \simeq 0.71$ has been measured for the structured polymer. These latter result can be rationalized in terms of an effective larger persistence length due to the presence of the side beads, and thus of stronger finite size effects overestimating the exponent $\nu$. For translocation purposes a constant force is exerted at the incoming end of the polymer as performed experimentally with optical or magnetic tweezers~\cite{Bell2012, HernndezAinsa2013, Peng2009, Trepagnier2007,VanDorp2009}. \subsection{The membrane} To remove effects originating from a non-negligible thickness of the nano-pore, we consider a perfectly 2-dimensional membrane formed by carbon (C) atoms arranged on a honeycomb lattice (see Fig.~\ref{FigPlotofpores}) reminiscent of the graphene structure. In our model, we also preserve a realistic ratio between the typical length scales associated to graphene and DNA, respectively. This amounts to a membrane lattice constant $b=a/2=0.15$~nm. Periodic boundary conditions are used in the plane parallel to the membrane, while free boundaries are employed in the direction orthogonal to the pore. We note that the size of the membrane is chosen sufficiently large to accomodate a fully stretched polymer while avoiding self-interactions with its own image. The membrane grains are immobile, and interact with the polymer beads via Eq.~(\ref{Eq:LJ}) with $\sigma_C=1$, ensuring that polymer beads cannot cross the membrane outside the nano-pore. In Fig.~\ref{FigModel} we show typical snapshots (generated by VMD~\cite{Humphrey1996}) of both polymer models pulled through the nanopore carved in the membrane. Details of the color code used are given in the caption of the figure. Figure~\ref{FigPlotofpores} represents the different sizes of polymer grains (blue and pink) with respect to the pore through which their translocation is studied (red). \subsection{Molecular Dynamics Simulations} Concerning the time integration, an efficient parallelised algorithm is used by the LAMMPS~\cite{Plimpton1995} software to solve the Langevin equation of motion \begin{equation} \large{m_n \frac{\partial^2 \textbf{r}_n}{\partial t^2} = -\frac{\partial U_{n}}{\partial \textbf{r}_n} - \nu_n \textbf{v}_n + \textbf{g}_n} \label{} \end{equation} with $m_n$ and $\textbf{r}_n$ being the mass and the position of the $n^{th}$ grain, $t$ the time, $U_{n}$ the sum of all the potentials applied to the $n^{th}$ grain, $\nu_n$ the friction coefficient of the $n^{th}$ grain and $\textbf{g}_n$ is the white thermal noise, of average $\left<\textbf{g}_n\right> = 0$, and variance $\left<\textbf{g}_n^2\right>=6\nu_nk_B T$. The temperature $T$ of the polymer beads is set to $3/2\text{ }\epsilon$. In the following both the mass and the friction coefficient are considered independent of the beads $m_n=m$ and $\nu_n=\nu$. Eq.~\ref{} is integrated numerically with a time step $\delta t=5 \text{ } 10^{-3}$ ($26$ fs in SI units). Translocation simulations started with an equilibrated polymer with one end attached at the entrance of the nanopore. For each forces, a set of 1000 translocations were performed for statistical analysis. We have used absorbing boundary conditions (i.e. if the first monomer does not cross the membrane and wanders in the cis side, the simulation is aborted). \section{Results and discussion} \subsection{Linear polymer} We considered as a reference polymer, the general case of Rouse-like linear polymer in order to be able to compare with the case of the structured polymer. The translocations have been performed through a large pore by pulling one end of the polymer (see top of Figure~\ref{FigModel}). The translocation of polymer forced by electrostatic interactions imposed by a voltage difference between the cis and trans side of the membrane have been extensively studied in the literature. However, the case of end polymer pulling while experimentally relevent~\cite{Keyser2006, Peng2009, VanDorp2009, Sischka2010, Bulushev2014, Bulushev2015, Bulushev2016} has been poorly considered by numerical simulations~\cite{Kantor2004, Huopaniemi2007, Panja2008} until very recently~\cite{2017arXiv170809184S}. Theory for biased polymer translocation aims at finding so called critical exponents, $\alpha$ and $\delta$, linking the mean translocation time to namely the polymerisation index and the bias force: $\tau \propto N^\alpha / F^\delta$. As formerly discussed, many theories and simulations investigate the case of a gradient bias force applied in the center of the nanopore~\cite{Sakaue2007,Sakaue2010,Saito2011,Saito2012,Ambjrnsson2005,2Ambjrnsson2005,Rowghanian2011,Dubbeldam2012,Ikonen2012,2Ikonen2012,Ikonen2013}. In this study, we are interested in the case of a pulling force at one end of the translocating polymer. Kantor and Kardar were the first to tackle the issue of a translocation driven by a pulling force and predicted $\alpha=2$ and $\delta=1$ scaling exponents~\cite{Kantor2004}. Their prediction is backed by 2D MD simulations which lead to a lower critical exponent $\alpha=1.875$~\cite{Kantor2004}. Such values were later confirmed in the literature~\cite{Huopaniemi2007} and are in agreement with this present work (see Figure~\ref{FigLinpolTranslocationtime}). \begin{figure}[t] \begin{center} \includegraphics[width=0.5\textwidth]{fig3a.pdf} \includegraphics[width=0.5\textwidth]{fig3b.pdf} \caption{Mean translocation time $\tau$ (top) and its rescaled value $\tau^*$ (bottom) as function of the bias pulling force (top) and the reduced force $Fa/k_BT$ (bottom). We investigate four polymerisation degrees ($N=16,32,64,128$ respectively in pink, blue, red and black) over three orders of magnitude for the bias force (from 0.1 to 100). The rescaled mean translocation time is as $\tau^*=\tau*F/N^{1.78}$. The rescaling of the translocation time clearly shows two regimes corresponding to a diffusive regime translocation and a force driven one.} \label{FigLinpolTranslocationtime} \end{center} \end{figure} We find that the mean translocation time shows two regimes as expected from the literature~\cite{Kantor2004,Huopaniemi2007}. In both regimes, $\alpha = 1.78$, only the value of $\delta$ differs. For low pulling forces, we have a diffusive regime ($\delta=0$) and a force driven regime when the force is increased ($\delta=1$). For very high forces we observe a decrease in the critical exponent (increasing behaviour of the rescaled translocation time). The underestimation of the critical exponent $\delta \simeq 0.8 < 1$ at high forces is due to bond strengths and/or friction coefficients set too low. Those findings were also observed when the bias force originates from a difference of potential~\cite{Luo2009}. We analyzed deterministic polymers at zero temperature for various frictions and bond strengths (see Supplementary Information Section II). Our findings show the importance of such parameters in the value of the critical exponent $\delta$. Furthermore, in the precise case of a pulled polymer, the crowding in the trans side may not be responsible for this decrease in $\delta$ as suggested by one of the last reviews on the topic~\cite{Palyulin2014} (the crowding in the trans side does not occur for a polymer under traction see Figure~\ref{FigModel} and SM Section II \footnotemark[1]). Those results are in agreement with recent results for field driven translocations that suggest the trans side crowding is irrelevant and force tension (directly connected to the parameters fore mentioned) alone can explain the decrease of the critical exponent $\delta$~\cite{2017arXiv170706663S}. Looking at the translocation time distribution (see SM Section III \footnotemark[1]), we find that they are in agreement with a first passage probability density function proposed by Ling and Ling~\cite{Ling2013} for a translocation biased by electrophoresis and confirmed for the case of a pulling force~\cite{deHaan2015}. The fitting parameters are the diffusion coefficient of the membrane along the polymer and the mean translocation velocity. An accurate estimation of the diffusion coefficient is impossible due to a large range of validity for the parameter. However, one can estimate the mean translocation velocity which coincides with the average speed deduced from the mean translocation time divided by the chain length. \begin{figure}[t] \begin{center} \includegraphics[width=0.5\textwidth]{fig4.pdf} \caption{Average waiting time depending on the translocation coordinate at a given force ($F=34$) for 4 polymer lengths ($N=16, 32, 64, 128$, respectively in pink, blue, red and black). The average waiting time curve presents the same pattern composed of 3 different regimes for the 4 polymer lengths studied: a linear regime dominated by the pulling force, then a plateau dominated by diffusion and finally a collapse of the waiting time due to tail retraction from the cis side.} \label{FigLinpolwaitingtime} \end{center} \end{figure} We pushed the analysis further by looking at the average waiting time for the translocation coordinate at different polymerisation degrees. The translocation coordinate corresponds to the number $n$ of monomers already translocated to the trans side. The average waiting time is the average time that the $n^{th}$ monomer needs to translocate (including possible returns if $n$ is not equal to 1 or $N$). The sum of all the average waiting time over the polymer length gives the average mean translocation time. We find waiting times curves in agreement with the general shape already published in the literature~\cite{Huopaniemi2007} (see Figure~\ref{FigLinpolwaitingtime}). In particular, the average waiting time as function of the translocation coordinate behaves linearly at the beginning of the translocation and for a given pulling force. Thus, at the beginning of the process, the translocation is entirely driven by the pulling force whose influence linearly decreases as it is spread on the $n$ monomers already translocated. The force is felt only by the trans side and the linear regime is present as long as the diffusion is negligible. When the diffusion force from the cis side compares to the pulling force, a plateau regime is reached. This second regime ends sharply with the retraction of the polymer tail from the cis side. This collapse or retraction of the polymer tail occurs over a constant value of 4-5 monomers. \begin{figure}[t] \begin{center} \includegraphics[width=0.5\textwidth]{fig5.pdf} \caption{Mean translocation time as function of the polymer length $N$ at a given pulling force $F=34$. A second order polynomial fits the numerical results with better accuracy than a simple power law with critical exponent $\alpha = 1.85 < 2$ provided one takes into account the effect of the polymer tail retraction ($N_e = N-4$).} \label{FigScaleresult} \end{center} \end{figure} From those observations, we develop a scaling analysis of the translocation of pulled polymers. From Figure~\ref{FigLinpolwaitingtime}, the waiting time $w(n) = a(F) + b(F) n$ is linear for the force driven regime ($n<\beta N$) with a constant term $a(F)$ and a slope $b(F)$. The second unbiased regime corresponds to a quasi-constant waiting time although slightly decreasing when $(N-n) \gg 1$ and sharply decreasing on 4-5 monomers for the third regime of polymer tail retraction. Thus, we obtain the following mean translocation time: \begin{eqnarray} \tau(F,N) & = & \int_0^{\beta N} dn\; [a(F)+b(F)\;n] \nonumber \\ & + &\int_{\beta N}^{N} dn\; [a(F) + b(F)\;\beta N]. \label{EqIntegral} \end{eqnarray} Numerical evidence shows that $a(F)$ and $b(F)$ are inversely proportional to $F$, and $\beta N$ (or at least the arithmetical prefactor) remains constant on a large range of parameters (see calculation in appendix and SM Section IV \footnotemark[1]), hence suggesting the following scaling for the translocation time: \begin{eqnarray} \tau(F,N) = \Gamma \frac{N^2}{F} + \frac{{\cal O}(N)}{F} \label{EqTauformula} \end{eqnarray} where $\Gamma$ is a constant. A second order polynomial fit of the translocation time as function of the polymer length using $N_e = N-4$ to avoid the 4-5 tail retracting monomers was adjusted to the numerical results (See Figure~\ref{FigScaleresult}). The agreement obtained is clearly better than a simple power law scaling with exponent $\alpha = 1.85$. The next to leading behavior term in (\ref{EqTauformula}) is similar to what was found by Ikonen et al.~\cite{Ikonen2013} in the case of gradient biased translocation with $\tau \propto N^{1+\nu} + O(N)$. $N^{1+\nu}$ is the asymptotic value and the $O(N)$ term is due to pore friction. For intermediate pulling forces, friction from the trans side and the pore are the main contributors to explain the scaling of the mean translocation time. The idea that friction from the trans side can have an important role was first mentioned in the case of a semi flexible polymer for an electric field driven translocation~\cite{Sarabadani2017}. For a very high pulling force recent results~\cite{2017arXiv170809184S} show that one has to take into account both friction from the trans side and from the cis side due to tension propagation. Those results stand as long as the force is high enough so that the polymer would always be fully stretched if pulled in the bulk. Based on work conducted on polymers in the bulk~\cite{Sakaue2012}, the minimum force should satisfy $F> 3k_BTN/a$ which would correspond to unit-less forces greater than 300 in most simulations and is questionably satisfied in ref~\cite{2017arXiv170809184S}. Let us predict an intermediate forces scaling based on our simulations and theoretical work from Sarabadani et al.~\cite{2017arXiv170809184S} that was conducted for high pulling forces and yields: \begin{eqnarray} \tau(F,N) = \frac{1}{F} \left[\tilde{\eta}_PN + \frac{A_\nu}{1+\nu}N^{1+\nu} + \frac{1}{2}N^2 \right] \label{EqTauformulaforhighpulling} \end{eqnarray} With $\tilde{\eta}_P$ the pore friction coefficient and $A_\nu$ the prefactor in the scaling of the end to end distance of the polymer. The three terms of eq.~\ref{EqTauformulaforhighpulling} are the respective friction contributions of the polymer with the pore, the cis side and the trans side. In our polynomial fit using eq.~\ref{EqTauformula} the prefactor of the $N^2$ contribution is in agreement with the prefactor of eq.~\ref{EqTauformulaforhighpulling} and pore friction contribution remains relevant. It seems that the contribution of the cis side friction is weak when the force is intermediate. Our reasoning is the following: in the case of a high pulling force, eq.~\ref{EqTauformulaforhighpulling} stands but if $F<3Nk_BT/a$, as soon as there is blob formation, the pore prevents the trumpet formation, the tension in the cis side vanishes and so does it's influence on the mean translocation time. The mean translocation time scaling becomes: \begin{eqnarray} \tau(F,N) = \frac{1}{F} \left[\tilde{\eta}_PN + \frac{A_\nu}{1+\nu}\left(\frac{aF}{3k_BT}\right)^{1+\nu} + \frac{1}{2}N^2 \right] \label{EqTauformulaintermediate} \end{eqnarray} For $N>aF/3k_BT$ ; for smaller chains, one recovers the high forces limit scaling. When the waiting time behaves linearly, it is clearly a $N^2$ contribution to the mean translocation time which indicates trans side friction is predominant. The contribution of the cis side friction is mainly seen in ref~\citep{2017arXiv170809184S} at the begining of the translocation when the tension propagation towards the cis side is maximum. This contribution with a mean waiting time increasing slower than linearly can also be seen in the early stages of the translocation on Figure~\ref{FigLinpolwaitingtime}. The fact that waiting times are independent of the chain length at the begining of the translocation process is further indication that the cis side friction contribution is constant and becomes negligible with the increase of the chain length in the intermediate pulling forces range. \subsection{Structured polymer translocating through large and narrow pores} The linear polymer discussed previously will now serve as a reference for the translocation of structured polymers through narrow and large pores (see narrow and large pore sizes on Figure~\ref{FigPlotofpores} middle and bottom respectively). The general behaviour of the translocation timescale observed for the linear polymer remains for the structured polymerwhen translocating throug the large pore. The difference is a $3/2$ rescaling factor of the timescale due to the larger number of beads present in the structured polymer. The proportion of beads between the structured and linear polymer is $3/2$ due to a side bead corresponding to the bases attached every two beads on the backbone. Furthermore, for the structured polymer, the crossover between unbiased and biased regimes is not as clear as previously observed for the linear polymer. Although, this does not seem to affect the critical exponents $\alpha = 1.76$ and $\delta \simeq 1$. \begin{figure}[t] \begin{center} \includegraphics[width=0.5\textwidth]{fig6.pdf} \caption{Mean translocation time for polymer chain length of $N=32$ for a large and a narrow pore (reproduced from our previous work~\cite{Menais2016}). At high pulling forces, their behaviour is similar with $\delta<1$ due to bonds elongation. When the bias force is decreased, a sharp increase in the translocation for the narrow pore compared to the large one occurs. This difference can be up to one order of magnitude and when the bias force is too low, the process does not even occur (hence the investigation of only 2 orders of magnitude in force range for the narrow pore). No general $\alpha$ and $\delta$ value can be established in the moderate forces regime.} \label{FigComppores} \end{center} \end{figure} The case of the structured polymer translocating through a narrow pore differs at low forces as we have already mentioned in previous work~\cite{Menais2016}. After a look at the mean translocation time (see Figure~\ref{FigComppores}), there seems to be almost no difference at high pulling forces. However when decreasing the pulling force, the mean translocation time increases more significantly for a narrow pore compared to the previous large pore case and no diffusive regime is observed. This increase can reach up to one order of magnitude and may be easily explained by an increased interaction (or friction) between the polymer and the pore. Those results confirm those already observed for a simple polymer and a narrow pore~\cite{Huopaniemi2007}. Polymer structure interactions with the pore can be observed through the waiting times on Figure~\ref{FigWaitingtimeStructured}. Although a slight effect of structure is observed for the large pore the case of the small pore clearly demonstrates that the translocation process is highly slowed down for (odd) grains linked with a lateral grain. This clear effect disappears with increasing forces. The shift that can be observed at the origin of waiting times for the structured polymer in Figure~\ref{FigWaitingtimeStructured} suggest that indeed the term in $O(N)$ for the translocation comes from frictions with the pore. \begin{figure}[t] \begin{center} \includegraphics[width=0.5\textwidth]{fig7.pdf} \caption{Average waiting time depending on the translocation coordinate at given force ($F=12$) and polymer chain length ($N=32$) for our linear polymer (black) and our structured polymer in the case of a large (blue) and small (red) pore. Polymer structure interactions with the small pore increases strongly the waiting time for odd translocation coordinates. This effect occurs also slightly in the case of a large pore.} \label{FigWaitingtimeStructured} \end{center} \end{figure} The critical exponent $\delta = 1.72$ observed for the low force regime is also strongly different than the usual $\delta \simeq 1$. Our simulations could not probe the very low forces as the translocation success probability is too small in this regime (see Figure ~\ref{FigTranslocproba}). Instead of a power law dependence, an exponential decrease of the translocation time as suggested in~\cite{Bacci2012} may also be expected as function of the force for the narrow pore in case of thermally activated translocation process. In fact, we will show in the following that there exists a threshold force below which the translocation process is strongly reduced. \begin{figure}[t] \begin{center} \includegraphics[width=0.5\textwidth]{fig8.pdf} \caption{Mean translocation velocity for the three studied cases at a given chain length ($N=32$) in function of the reduced force ($Fa/K_BT$). For large pores we have a linear dependency with a zero ordinate at the origin, the factor $3/2$ difference in slope between the linear and structured polymer originates from the lateral grains friction contribution. For the translocation of the structured polymer through a narrow pore, the slope is similar to the same polymer through the large pore but with a negative ordinate at the origin. implying a necessary threshold reduced force to initiate the translocation process.} \label{FigTranslocvelocities} \end{center} \end{figure} The strong increase of the translocation time for narrow pore may be interpreted by a strong friction of the polymer with the narrow pore. This effect is prevalent for small forces but almost negligible for high bias forces where the energy brought by the force is sufficient to circumvent this friction. In order to test this hypothesis, we considered the mean translocation velocity for our three polymer models. This velocity is obtained by dividing the mean translocation time by the polymer chain length (see Figure~\ref{FigTranslocvelocities}). For large pores, the translocation velocity ($v_{lp}$) scales linearly with the pulling force: \begin{center} \begin{eqnarray} v_{lp} = \xi F \label{EqFrcitiontransloclarge} \end{eqnarray} \end{center} Unsurprisingly we found that in the case of the structured polymer, the translocation velocity is divided by $3/2$ given the supplementary lateral grains friction. In the case of the structured polymer translocation through a narrow pore, the mean translocation velocity has the same slope as the same polymer through the large pore but shifted by a positive threshold force. Thus, for the narrow pore mean translocation velocity ($v_{np}$), eq.~\ref{EqFrcitiontransloclarge} becomes: \begin{center} \begin{eqnarray} v_{np} = \xi (F-F^*) \label{EqFrcitiontranslocnarrow} \end{eqnarray} \end{center} The rescaled threshold force $F^*a/k_B T$ is comprised between 1.5 and 2. The same value for the threshold force necessary to start the translocation process is observed when we consider the probability of successful translocation (defined as the fraction of translocation trial ended in the trans side) (see Figure~\ref{FigTranslocproba}). \begin{figure}[t] \begin{center} \includegraphics[width=0.5\textwidth]{fig9.pdf} \caption{Successful translocation probability of the structured polymer in function of the reduced force ($Fa/K_BT$). For both pores, the translocation is a success at a given pulling force. In the case of the narrow pore, one needs to add a threshold reduced force (between 1 and 2) to allow the translocation process to take place at a same success rate as in the large pore case.} \label{FigTranslocproba} \end{center} \end{figure} The friction in the case of a large pore is almost entirely imputable to the solvent leading to a viscous friction whereas the contribution of a narrow pore is closer to the combination of both a viscous friction (independent of the pore size) and a solid-solid friction implying a threshold force necessary to translocate the polymer which is pore size dependent. Structure effects and force thresholds have already been spotted in the literature but the mechanism behind those results are very different. The closest to our setup (and providing statistical confidence) is a work focused on RNA hairpins~\cite{Vocks2009}, the translocation is also biased by pulling but at constant speed instead of constant force. The authors determine a threshold force above which the structure is destroyed (i.e. the hairpin unzips), thus enabling the translocation to occur. Unfortunately, the use of a constant pulling speed prevents the identification of a solid-solid friction behavior. Structure and threshold forces were also observed in the case of knotted proteins~\cite{Szymczak2016} but in this case, the translocation becomes impossible at large pulling forces because the knot is tightened and does not slide along the protein anymore. To our knowledge, solid-solid friction behaviour induced by polymer structure-pore interactions is an original result in the polymer translocation field. \section{Conclusions} In this paper we have analyzed the average translocation time of a simple linear polymer through a large pore. This reference is then used to analyze the effect of a structured polymer on the translocation process in particular in case of large and narrow pores. We proposed a scaling analysis of characteristic velocities implied in the process as well as the waiting time as function of the translocation coordinate. A force driven and a diffusive-like regime were established and the sum over the corresponding waiting times of the $N$ monomers leads to the critical exponent $\alpha = 2$ with finite size effects. This leads us to a confirmation of the already found $\alpha=2$ behaviour with finite size effects that may explain the underestimation of the effective critical exponent $\alpha$ observed in previous and present numerical simulations. The translocation of the structured polymer through a large nanopore showed small difference with the linear polymer behaviour except for a slowing down of the translocation process by a factor $3/2$ corresponding to the supplementary friction coming from the lateral grains. This interesting results is in agreement with the linearity of all the interactions (given that we do not take into account hydrodynamics interactions). Concerning the translocation of the structured polymer through a narrow nanopore, we have shown that at high force regime, the scaling behaviour of the mean translocation time is the same compared to a large pore. As the force is reduced, no relevant value for $\alpha$ can be established. A sharp increase in the value of the critical exponent $\delta$ occurs linked to a strong decrease in the successful translocation probability at a finite threshold force. This threshold force is also apparent in the linear behaviour of the mean translocation velocity. The existence of this threshold force without modification in the slope of the mean translocation speed exhibits a solid-solid friction behavior for structure-pore interactions. \section*{Appendix} Here we develop the calculations including a velocity scale analysis using the waiting time function. We assume that over the translocation, diffusion is slow and the polymer is always quasi-equilibrated in the cis side. Thus, we consider separately the trans side fully stretched and the cis side undergoing diffusion. This separation is consistant with the polymer conformation on both side as shown on Figure~\ref{FigModel}.a. We compare two velocity scales corresponding to the cis and trans sides. Concerning the cis side, the diffusion velocity originates from the Brownian motion through energy equipartition of the $N-n$ monomers present in the cis side: \begin{equation} v_{cis} \propto \frac{\sqrt{\nu(N-n)k_BT}}{\nu(N-n)} = {\sqrt{\frac{k_BT}{\nu(N-n)}}}. \end{equation} For the trans side, two velocity scales exist: the same Brownian velocity which is relevant only for very small bias forces, i.e. in the quasi-unbiased regime (see Figure~\ref{FigLinpolTranslocationtime}) and the velocity originating from the pulling force: \begin{equation} v_{trans} \propto \frac{F}{\nu n}. \end{equation} The transition between the two first regimes observed for the waiting time occurs when both velocities equalize $v_{trans} \approx v_{cis}$. The crossover between the first regime where the transolcation is imposed by the pulling force and the second regime of quasi-unbiased translocation occurs for $n$ monomers translocated at a force $F(n)$: \begin{equation} F(n) \propto \sqrt{\frac{\nu k_B T n^2}{N-n}}. \label{eqforce} \end{equation} Equation~\ref{eqforce} may be simply inverted to get the monomer $n^*(F,N)$ at which the crossover between the force driven and diffusive regime occurs for a given pulling force $F$ and polymer length $N$: \begin{equation} n^*(F,N)=\frac{F^2}{2C}\sqrt {1+\frac{4NC}{F^2}}-\frac{F^2}{2C} \label{EqBoundary} \end{equation} where $C$ is a constant. For large pulling forces, when $F \gg \sqrt{CN}$, $n^* \rightarrow N$, meaning that the force driven regime always predominates. For low forces, $n^* \rightarrow 1$ and the unbiased translocation regime is recovered on the whole translocation. The fraction of the polymer corresponding to the crossover, $\beta(F,N)=n^*/N$ is only function of the reduced parameter $A=2CN/F^2$: \begin{equation} \beta(F,N) = \beta(A)= \frac{\sqrt{1+2A}-1}{A}. \end{equation} As previously stated, the mean translocation time is the sum over all the translocation coordinates of the average waiting times $w(n)$. It may be reduced to the integration over the two separate regimes if one neglects the tail retraction. From Figure~\ref{FigLinpolwaitingtime}, the waiting time $w(n)$ is linear for the force driven regime with a constant term $a(F)$ and a slope $b(F)$. The second diffusive regime corresponds to a quasi-constant waiting time although slightly decreasing when $(N-n) \gg 1$ and sharply decreasing on 4-5 monomers for the third regime of polymer tail retraction. Thus, we obtain equation \ref{EqIntegral} mentioned in the main text. Numerical evidence show that $a(F)$ and $b(F)$ are inversely proportional to $F$\footnotemark[1], hence: \begin{center} \begin{eqnarray} \tau(F,N) = \frac{a'}{F} N + \left(\beta(A)-\frac{\beta(A)^2}{2}\right) \frac{b'N^2}{F} \label{EqTauformulaappendix} \end{eqnarray} \end{center} with $a'$ and $b'$ constants. For a reasonable couple of $N$ and $F$ values ($A < 1$), the pre-factor in front of the term scaling as $N^2/F$ remains a constant as shown in SM \footnotemark[1]. We then have equation \ref{EqTauformula} valid provided the applied force is large enough to avoid highly diffusive regimes (i.e. $1\ll \sqrt{2CN} < F$). The expected scaling goes from $\tau \propto N^{2+\nu}$ for the unbiased case to $\tau \propto N^{2}/F$ in the force driven case far from finite size effects. \section*{Acknowledgments} We thank L\'eo Brunswick for fruitful discussions. Grand Équipement National de Calcul Intensif (GENCI) for the allocation of computing resources INAC-SYMMES is part of the Arcane Labex program, funded by the French National Research Agency (ARCANE project no. ANR-12-LABX-003). \bibliographystyle{apsrev4-1}

\section{Introduction} One of the most important and popular completely integrable nonlinear PDE is the nonlinear Schr\"{o}\-din\-ger (NLS) equation \cite{ZaSh*74a,FaTa,ZMNP,GVYa*08}: \begin{equation}\label{Introduction1} {\rm i}q_t+ q_{xx}+2|q|^2 q(x,t)=0. \end{equation} Here $q(x,t)$ is a complex valued function, tending fast enough to zero as $\left|x\right|\rightarrow \infty$ \cite{GVYa*08, AblPriTru}. It has been derived as a governing equation describing processes and phenomena in such diverse fields as deep water waves, plasma physics and nonlinear fibre optics. For instance, in optics, NLS models wave propagation in Kerr media, where the nonlinearity is proportional to the intensity of the field. The NLS equation \eqref{Introduction1} appeared at the early stage of the development of the inverse scattering method (ISM) and the theory of solitons \cite{AblPriTru, DraJoh, FaTa,GVYa*08, ZMNP}, and exhibits all remarkable properties of PDEs and systems of PDEs, integrable by the ISM: it allows soliton solutions, infinite set of integrals of motion, multi-Hamiltonian formulation and so on \cite{GVYa*08,FaTa}. The key tool in studying integrability by the ISM is the existence of Lax representation of the nonlinear evolutionary equation (NLEE), that is, the NLEE can be represented as a compatibility condition of two linear operators \cite{ZMNP,FaTa,ZaSh*74a,GVYa*08,IP2}. The scattering problem for \eqref{Introduction1} is given by the Zakharov-Shabat system (related to $sl(2,{\Bbb C})$ algebra): \begin{align}\label{(ZS) System} L\chi\equiv & \left({\rm i}\frac{d}{dx}+q(x,t)-\lambda\sigma_3\right)\chi\left(x,t,\lambda\right)=0,\nonumber \\ q(x,t)= & \left( \begin{array}{cc} 0 & q^+ \\ q^- & 0\\ \end{array} \right), \qquad \sigma_3=\left( \begin{array}{cc} 1 & 0\\ 0 & -1\\ \end{array} \right), \end{align} The first integrable multi-component generalisation of the scalar NLS equation (\ref{Introduction1}) is the Manakov vector NLS equation: \begin{eqnarray}\label{Examples of MNLS-type equations4} {\rm i}q_{1,t}+q_{1,xx}+2 (|q_1|^2+|q_2|^2)q_1(x,t)=0;\nonumber\\ {\rm i}q_{2,t}+q_{2,xx}+2 (|q_1|^2+|q_2|^2)q_2(x,t)=0. \end{eqnarray} Here, $ q(x,t)$ is a $2$-component vector-valued function. It is associated with a scattering problem of Zakharov-Shabat type related to the algebra $sl(3,{\Bbb C})$. It was proposed by S. V. Manakov \cite{Manakov} as an asymptotic model for the propagation of the electric field in a waveguide. Subsequently, this system was derived as a key model for light propagation in optical fibres \cite{Menyuk,AckMil}. The Manakov model \eqref{Examples of MNLS-type equations4} can be generalised to $n$-component vectors (see \cite{ZMNP}) \begin{equation}\label{} i{\bf v}_t+{\bf v}_{xx}+2 ({\bf v}^\dag , {\bf v}) {\bf v}=0,\qquad {\bf v}= {\bf v}(x,t). \end{equation} Here ${\bf v}$ is an $n $-component complex-valued vector and $(\cdot ,\cdot) $ is the standard scalar product. It is again integrable by the ISM \cite{AblPriTru,ZMNP,FaTa,GVYa*08}. The classical ZS system can be generalised to a matrix (multi-component) form in a number of ways. One of the standard ways for doing this is by considering Lax operators taking values in a simple Lie algebra $\mathfrak{g}$: \begin{align}\label{Matrix generalisations of NLS Hierarchies1} L=\mathrm{i} \partial_x + Q -\lambda \mathrm{J}. \end{align} Here, $\mathrm{J}$ is a constant element of simple Lie algebra $\mathfrak{g}$ \cite{Helg, Gilmore}. Generalizations of the NLS equation to symmetric spaces related to Lie algebra $\mathfrak{g}$ are proposed in \cite{ForKu*83,AthFor,For}. This includes coupled NLS systems with Lax pairs related the symmetric spaces {\bf A.III}, {\bf C.I}, {\bf D.III} and {\bf BD.I} types in the Cartan classification \cite{Helg,loos}. Among this class, one can get as a special case the Manakov vector NLS equation and the one studied by Kulish and Sklyanin \cite{KulSkl}. All these generalizations are solvable by the ISM \cite{KulSkl,ForKu*83,GVYa*08}. When the rank $r$ of the underlying simple Lie algebra ${\frak g}$ grows, the corresponding generic NLEE (or systems) will contain as many independent complex-valued functions as the number of all roots of ${\frak g}$ \cite{DriSok,58,vgrn}. They are solvable for any $r $ but their possible applications to physics for large $r $ do not seem realistic. However one still may extract new integrable and physically useful NLEE by imposing reductions on $L(t) $, i.e. algebraic restrictions on $Q(x,t) $ which diminish the number of independent functions in them and the number of equations \cite{2}. Of course, such restrictions must be compatible with the dynamics of the NLEE \cite{58,vgrn}. Recently, in \cite{AblMus} was proposed a nonlocal integrable equation of nonlinear Schr\"odinger type \begin{equation}\label{AblMus} {\rm i}q_t+ q_{xx}+V(x,t) q(x,t)=0, \qquad V(x,t)=2q(x,t)q^*(-x,t). \end{equation} with ${\cal PT}$-symmetry, due to the invariance of the so-called self-induced potential $V(x,t)$ under the combined action of parity and time reversal symmetry. In the same paper, the 1-soliton solution for this model is derived and it was shown that it develops singularities in finite time. Soon after this, nonlocal ${\cal PT}$-symmetric generalisations are found for the Ablowitz-Ladik model in \cite{AblMus1}. All these models are integrable by the Inverse Scattering Method (ISM) \cite{AblMus2,GeSa,zm1,ZaSh*74a}. Such nonlocal reductions of the NLS equation and its multi-component generalisations are of particular interest in regards to applications in ${\cal PT}$-symmetric optics, especially in developing of theory of electromagnetic waves in artificial heterogenic media \cite{Konotop, Barash, Barash1, Nature}. For an up-to-date review, see for example \cite{UFN}. Historically the first pseudo-hermitian hamiltonian with real spectrum is the ${\cal PT}$-symmetric one in \cite{Bender1}. Pseudo-hermiticity here means that the Hamiltonian ${\cal H}$ commutes with the operators of spatial reflection ${\cal P}$ and time reversal ${\cal T}$: ${\cal P}{\cal T}{\cal H}={\cal H}{\cal P}{\cal T}$ \cite{Ali1, Ali2}. The action of these operators is defined as follows: ${\cal P}: x\to -x$ and ${\cal T}: t\to -t$ \cite{Bender2, Fring2}. Supposing that the wave function is a scalar, this leads to the following action of the operator of spatial reflection on the space of states: ${\cal P}\psi(x,t)=\psi(-x,t)$ and ${\cal T}\psi(x,t)=\psi^*(x,-t)$. As a result, the Hamiltonian and the wave function are ${\cal PT}$-symmetric, if ${\cal H}(x,t)={\cal H}^*(-x,-t)$ and $\psi(x,t)=\psi^*(-x,-t)$ \cite{Fring1}. Here we used also, that the parity operator ${\cal P}$ is linear and unitary while the time reversal operator ${\cal T}$ is anti-linear and anti-unitary. The aim of this paper is study nonlocal reductions and to derive the corresponding soliton solutions for multi-component NLS models related to symmetric spaces of {\bf A.III}-type and {\bf BD.I}-type. This will be done based on examples of the vector NLS equations related to symmetric spaces of ${\bf A.III}\simeq SU(s+p)/S(U(s)\otimes U(p))$ type and the Kulish-Sklyanin model related to symmetric spaces of {\bf BD.I}$\simeq SO(2r+1)/SO(2)\otimes SO(2r-1)$ type. The structure of this paper is as follows: In Section 2 we outline the form of the Lax operators and the general form of the NLEEs as well as the nonlocal symmetries (involutions) of interest. In Section 3 we present the direct scattering problem: the Jost solutions, the scattering matrix and the minimal set of scattering data and the fundamental analytic solutions (FAS). Furthermore, In Section 4, based on appropriate modification of the Zakharov-Shabat dressing method, we derive 1- and 2-soliton solution of the corresponding NLEE equation related to ${\bf A.III}$ and ${\bf BD.I}$ symmetric spaces and study the effect on nonlocal reductions on them. \section{Preliminaries}\label{} \subsection{Lax pair and general form of the equations}\label{} Let us start with the generic Lax pair for the MNLS equations on symmetric spaces \cite{G2005}, it can be represented by the following form: \begin{subequations}\label{Lax Representation g=so(n)} \begin{align} L\chi(x,t,\lambda)& \equiv \rm i \partial_{x} \chi+(Q(x,t)-\lambda \mathrm{J})\chi (x,t,\lambda) =0,\quad U(x,t,\lambda)=Q(x,t)-\lambda \mathrm{J},\label{Lax Representation g=so(n)a} \\ M\chi(x,t,\lambda)& \equiv \rm i \partial_{t} \chi +(V_{0}(x,t)+\lambda V_{1}(x,t)- \lambda^{2} \mathrm{J})\chi(x,t,\lambda)=0, \label{Lax Representation g=so(n)b} \\ V_{1}(x,t)& = Q(x,t), \quad V_{0}(x,t)= \mathrm{i} \text{ad}^{-1}_{\mathrm{J}}\frac{\mathrm{d} Q}{\mathrm{d} x} +\frac{1}{2}[\text{ad}^{-1}_{\mathrm{J}} Q,Q(x,t)].\label{Lax Representation g=so(n)c} \end{align} \end{subequations} Here $\chi (x,t,\lambda)$ are the eigenfunction of the Lax operators, $U(x,t,\lambda)$ and $V(x,t,\lambda)$ are chosen to take values in a simple Lie algebra $\mathfrak{g}$ of rank $r$, while the eigenfunctions $\chi(x,t,\lambda)$ belong to the corresponding Lie group $\mathfrak{G}$. Here also $\mathrm{J}$ is a constant element of the Cartan subalgebra $\mathfrak{h}\subset \mathfrak{g}$ (which can be always chosen to be represented by a diagonal matrix) and $\lambda\in \mathbb{C}$ is a spectral parameter. On the ${\bf A.III} \simeq SU(p+s)/S(U(p)\times U(s))$ symmetric spaces the potential of $L(\lambda)$ \eqref{Lax Representation g=so(n)a} reads: \begin{align} Q(x,t)= & \left(\begin{array}{cc} 0 & \mathbf{q}^+ \\ \mathbf{q}^- & 0 \\ \end{array} \right), \hspace{0.3in} \mathrm{J}=\frac{2}{s+p} \left( \begin{array}{cc} p \mathds{1}_s & 0 \\ 0 & -s\mathds{1}_p \\ \end{array} \right),\label{Lax representation1d} \end{align} where $\hm{q}^+(x,t)$ and $(\hm{q}^-)^T(x,t)$ are $s \times p$ matrix-valued functions belonging to the simple Lie algebra $\mathfrak{g}$, $\mathds{1}_s$ and $\mathds{1}_p$ are the $s\times s$ and $p\times p$ identity matrices, respectively, $s+p=n$. The NLEE can be written as a compatibility condition \begin{equation}\label{Generalizations of the AKNS Approach1} [L(\lambda),M(\lambda)]=0, \end{equation} of the two Lax operators \eqref{Lax Representation g=so(n)a}-\eqref{Lax Representation g=so(n)b}. In particular, if $L(\lambda)$ and $M(\lambda)$ are related to $ {\bf A.III}\simeq SU(p+1)/S(U(1)\otimes U(p))$ - symmetric spaces \cite{ForKu*83}, the explicit parametrisation \eqref{Lax representation1d} of the ${\bf A.III}$ symmetric spaces will give the system \begin{eqnarray}\label{Examples of MNLS-type equations4a} {\rm i}\mathbf{q}^+_{t}+\mathbf{q}^+_{xx}+2 \mathbf{q}^+\mathbf{q}^-\mathbf{q}^+(x,t)&=&0,\\ -{\rm i}\mathbf{q}^-_{t}+\mathbf{q}^-_{xx}+2 \mathbf{q}^-\mathbf{q}^+\mathbf{q}^-(x,t)&=&0. \end{eqnarray} The particular choice $s=1$ and $p=2$ (assuming also the standard involution ${\bf q}^- = ({\bf q}^+)^*$ for the NLS type of models) corresponds to the well-know Manakov system \cite{Manakov}. Its generalisations for $n$-dimensional vectors ${\bf q}^{\pm}$ is known as vector nonlinear Schr\"odinger equation \cite{GVYa*08}. Another class of the multi-component NLS (MNLS) equations is known as Kulish and Sklyanin (KS) models and are related to the symmetric spaces of {\bf BD.I}$\simeq SO(2r+1)/SO(2)\otimes SO(2r-1)$ type \cite{GKKV*08, G37-52, G2017, GG2010, G2009}. The generic NLEEs of this class can be written as: \begin{align}\label{Some Algebraic Properties for Cartan Weyl Basis5} \mathrm{i}\vec{\mathbf{q}}_{t}^{+}+\vec{\mathbf{q}}_{xx}^{+}+ 2\left(\vec{\mathbf{q}}^{+},\vec{\mathbf{q}}^{-}\right) \vec{\mathbf{q}}^{+}- \left(\vec{\mathbf{q}}^{+},s_{0}\vec{\mathbf{q}}^{+} \right)s_{0}\vec{\mathbf{q}}^{-}= &0,\nonumber \\ \mathrm{i}\vec{\mathbf{q}}_{t}^{-}-\vec{\mathbf{q}}_{xx}^{-}- 2\left(\vec{\mathbf{q}}^{+},\vec{\mathbf{q}}^{-}\right) \vec{\mathbf{q}}^{-}+ \left(\vec{\mathbf{q}}^{-},s_{0}\vec{\mathbf{q}}^{-} \right)s_{0}\vec{\mathbf{q}}^{+}=&0, \end{align} which is associated with ${\mathfrak{g}\simeq so(2r+1,{\Bbb C})}$ linear system \eqref{Lax Representation g=so(n)a} where \begin{align}\label{Some Algebraic Properties for Cartan Weyl Basis6} Q(x,t)= \left( \begin{array}{ccc} 0 & \mathbf{(q^{+})^{T}} & 0 \\ \mathbf{q^{-}} &{\bf 0} & s_{0}\mathbf{q^{+}} \\ 0& \mathbf{(q^{-})^{T}}s_{0} & 0 \\ \end{array} \right), \qquad \mathrm{J} =\mbox{diag}\, (1, {\bf 0},-1). \end{align} Here $\mathbf{q^{\pm}}$ are $2r-1$-component vectors, and $s_{0}$ is the matrix that defines the orthogonal algebra ${\bf B}_r\simeq so(2r+1)$: \begin{align}\label{Lax Representation g=so(n)3} X\in so(2r+1) \ \ \text{iff} \ \ X+S_{0}X^{T}S_{0}^{-1}=0, \quad S_{0}=\sum_{s=1}^{2r+1}(-1)^{s+1}E_{s,2r+2-s}^{(2r+1)}= \left( \begin{array}{ccc} 0 &0 & 1 \\ 0 &-s_{0} & 0 \\ 1& 0 & 0 \\ \end{array} \right), \end{align} where $E_{s,p}^{(2r+1)}$ is a $2r+1 \times 2r+1$ matrix with matrix elements given by $(E_{s,p}^{(2r+1)})_{ij}= \delta_{si}\delta_{pj}$. \subsection{Symmetries of the Lax operator and reductions}\label{Symmetries of the Lax operator and reductions} A systematic way of description and classification of the class of admissible reductions for a given Lax pair is introduced by A. Mikhailov in \cite{2}. This is done by the so-called reduction group. The reduction group is a finite group which preserves the Lax representation, i.e. it ensures that the reduction constraints are automatically compatible with the evolution \cite{58, g2003, vgrn, TV}. The reduction constrains diminish the number of independent functions and the number of equations. Here we will restrict ourselves with $\mathbb{Z}_2$ reductions of the following two types: \begin{align}\label{The Local Involutions2m} \mbox{(A)} &&\qquad U(x,t,\lambda) = B (U(x,t,\lambda^{\ast}))^{\dag}B^{-1}, \\ \mbox{(B)} &&\qquad -U(x,t,\lambda) =B (U(-x,t,-\lambda^{\ast}))^{\dag}B^{-1},\label{The Local Involutions2m1} \end{align} where $B$ is constant block-diagonal matrix of the form $B = \left( \begin{array}{cc} \mathbf{B_+} &0 \\ 0& \mathbf{B_-} \\ \end{array}\right)$ for ${\bf A.III}$ symmetric spaces. For symmetric spaces of type ${\bf BD.I}$ it can be written in the form: \begin{align}\label{The Local Involutions3mKS} B = \left( \begin{array}{ccc} \mathbf{B_+ } &0 &0 \\ 0& \mathbf{B_\pm}&0 \\ 0 &0 &\mathbf{B_- } \\ \end{array}\right). \end{align} Here we assume also that the blocks $\mathbf{B_+}$ and $\mathbf{B_-}$ are nonsingular matrices. Note also, that the involution of type (B) in (\ref{The Local Involutions2m1}) results in nonlocal reduction conditions. Recalling the standard block-structure on ${\bf A.III}$ symmetric spaces (\ref{Lax representation1d}) one can write explicitly the reductions conditions for the matrix blocks: \begin{align}\label{The Local Involutions4m} \mbox{(A)} &&\qquad\mathbf{q}^{-}(x,t)=\mathbf{B_-}(\mathbf{q}^{+}(x,t))^{\dag}(\mathbf{B_+})^{-1},\quad \mathbf{q}^{+}(x,t)=\mathbf{B_+}(\mathbf{q}^{-}(x,t))^{\dag}(\mathbf{B_-})^{-1},\\ \mbox{(B)} &&\qquad\label{The Nonlocal Involution4m} \mathbf{q}^{-}(x,t)=-\mathbf{B_-}(\mathbf{q}^{+}(-x,t))^{\dag}(\mathbf{B_+})^{-1},\quad \mathbf{q}^{+}(x,t)=-\mathbf{B_+}(\mathbf{q}^{-}(-x,t))^{\dag}(\mathbf{B_-})^{-1}. \end{align} Using the the local and nonlocal involutions \eqref{The Local Involutions4m} and \eqref{The Nonlocal Involution4m} respectively, one can write the resulting Kulish-Sklyanin models. The local involution\eqref{The Local Involutions4m} will give the following equations: \begin{equation}\label{The Local and Nonlocal (KS) type Models VNLS Equation1} \mathrm{i} \mathbf{q}_{t}^{+}+\mathbf{q}_{xx}^{+}+2\left(\mathbf{q}^{+}(x,t),(\mathbf{q}^{+}(x,t))^{\ast}\right)\mathbf{q}^{+}-\left(\mathbf{q}^{+}, s_{0}\mathbf{q}^{+}\right)s_{0}(\mathbf{q}^{+}(x,t))^{\ast}=0, \end{equation} while the nonlocal involution \eqref{The Nonlocal Involution4m} will result in: \begin{equation}\label{The Local and Nonlocal (KS) type Models VNLS Equation2} \mathrm{i} \mathbf{q}_{t}^{+}+\mathbf{q}_{xx}^{+}+2\left(\mathbf{q}^{+}(x,t),(\mathbf{q}^{+}(-x,t))^{\ast}\right)\mathbf{q}^{+}-\left(\mathbf{q}^{+}, s_{0}\mathbf{q}^{+}\right)s_{0}(\mathbf{q}^{+}(-x,t))^{\ast}=0. \end{equation} \noindent {\bf Example A. [Manakov model]} {\it If ${\frak g}\simeq sl(3,{\Bbb C})$, $p=1$ and $s=2$, then $\hm{q}^+(x,t)$ and $\hm{q}^-(x,t)$ are 2-component vector functions. In addition, If we apply a (local) reduction of type (A) with $B={\Bbb I}$, this will reproduce the standard Manakov VNLS equation: \begin{equation}\label{The Nonlocal MNLS Equation1} \begin{array}{cc} -{\rm i}q_{1,t}+ q_{1,xx}+2 \ (|q_{1}(x,t)|^2+ |q_{2}(x,t)|^2) \ q_{1}(x,t)=0, \vspace{0.1in}\\ -{\rm i}q_{2,t}+ q_{2,xx}+2 \ (|q_{1}(x,t)|^2+ |q_{2}(x,t)|^2) \ q_{2}(x,t)=0. \end{array} \end{equation} Taking an involution of type (B) with $\mathbf{B}_+=\mp1$ and $\mathbf{B}_-=\text{diag}(\pm1, \pm1)$, then we obtain the following nonlocal reduction of Manakov model:} \begin{equation}\label{The Nonlocal MNLS Equation3} \begin{array}{cc} -{\rm i}q_{1,t}+ q_{1,xx}+2 \ (q_{1}(x,t) \ q_{1}^{\ast}(-x,t)+ q_{2}(x,t) \ q_{2}^{\ast}(-x,t)) \ q_{1}(x,t)=0, \vspace{0.1in}\\ -{\rm i}q_{2,t}+ q_{2,xx}+2 \ (q_{1}(x,t) \ q_{1}^{\ast}(-x,t)+ q_{2}(x,t) \ q_{2}^{\ast}(-x,t)) \ q_{2}(x,t)=0. \end{array} \end{equation} \noindent {\bf Example B. [Kulish-Sklyanin model]} {\it In the simplest case of Lax operators related to $SO(5)/SO(2)\otimes SO(3)$ (${\rm rank}\, {\frak g}=2$), we can set ${\bf q}^+=(q_{12}^+,q_{13}^+,q_{14}^+)$ and ${\bf q}^-=(q_{12}^-,q_{13}^-,q_{14}^-)$. After assuming also the standard involution of type (A) with ${B}={\Bbb I}$, the compatibility condition (\ref{Generalizations of the AKNS Approach1}) will lead to the following 3-component NLS system: \begin{align}\label{Some Algebraic Properties for Cartan Weyl Basis7} \mathrm{i}(q_{t}^{+})_{12}&+(q_{xx}^{+})_{12}+ 2\left(\left|q_{12}\right|^{2}+2\left|q_{13}\right|^{2}\right) q_{12} +2(q^{+}_{14})^{*}(q^{+}_{13})^{2}=0 ,\nonumber \\ \mathrm{i}(q_{t}^{+})_{13}&+(q_{xx}^{+})_{13}+ 2\left(\left|q_{12}\right|^{2}+\left|q_{13}\right|^{2} +\left|q_{14}\right|^{2}\right) q_{13} +2(q^{+}_{13})^{*}q^{+}_{14}q^{+}_{12}=0 ,\nonumber \\ \mathrm{i}(q_{t}^{+})_{14}&+(q_{xx}^{+})_{14}+ 2\left(\left|q_{14}\right|^{2}+2\left|q_{13}\right|^{2}\right) q_{14} +2(q^{+}_{12})^{*}(q^{+}_{13})^{2}=0. \end{align} This appears to be a model describing ${\cal F}=1$ spinor Bose-Einstein condensates in one-dimensional approximation \cite{KulSkl,G2009,AGGK,GGK2005c}. } \section{Direct Scattering Transform for $L(\lambda)$}\label{} \subsection{Jost solutions and scattering matrix} The starting point here are the so-called Jost solutions, which are determined by their asymptotics for $\left|x\right|\rightarrow \infty$: \begin{equation}\label{Jost Functions2} \lim_{x\rightarrow \infty}\mathrm{Re\,}^{{\rm i}\lambda \mathrm{J} x}\psi(x,t,\lambda)=\mathds{1}_{s+p},\qquad \lim_{x\rightarrow -\infty}\mathrm{Re\,}^{{\rm i}\lambda \mathrm{J} x}\phi(x,t,\lambda)=\mathds{1}_{s+p}, \qquad \lambda \in \mathbb{R}. \end{equation} Along with these functions, one can also use "normalised to unity" Jost solutions \begin{equation}\label{Jost Functions6} \xi(x,t,\lambda)=\psi(x,t,\lambda)\mathrm{Re\,}^{({\rm i}\lambda \mathrm{J} x)},\qquad \varphi(x,t,\lambda)=\phi(x,t,\lambda)\mathrm{Re\,}^{({\rm i}\lambda \mathrm{J} x)}, \end{equation} satisfying the following equation \begin{equation}\label{Jost Functions8} {\rm i}\frac{\mathrm{d} \xi}{\mathrm{d} x}+Q(x,t)\xi(x,t,\lambda)-\lambda\left[\mathrm{J},\xi(x,t,\lambda)\right]=0, \end{equation} provided that \eqref{Jost Functions2} satisfy \eqref{Lax Representation g=so(n)a}. On the continuous spectrum of $L(\lambda)$ the two Jost solutions $\psi(x,t,\lambda)$ and $\phi(x,t,\lambda)$ are related via scattering matrix $T(\lambda)$ \begin{equation}\label{Scattering Data4} \phi(x,t,\lambda)=\psi(x,t,\lambda)T(\lambda), \qquad \lambda\in\mathbb{R}. \end{equation} The scattering matrix $T(\lambda)$ belongs to the Lie group ${\frak G}$ corresponding the Lie algebra /symmetric space of $L(\lambda)$. For symmetric spaces of type ${\bf A.III}$ in the Cartan classification, it has the following block structure: \begin{equation}\label{Scattering Data6} T(\lambda)=\left(\begin{array}{cc} \hm{a}^+(\lambda) & -\hm{b}^-(\lambda) \\ \hm{b}^+(\lambda) & \hm{a}^-(\lambda) \end{array}\right), \end{equation} where, $\hm{a}^+(\lambda)$ and $\hm{a}^-(\lambda)$ are square matrices, while $\hm{b}^+(\lambda)$ and $\hm{b}^-(\lambda)$ are rectangular matrices. The block structure of the inverse of $T(\lambda)$ can be written as: \begin{align}\label{Scattering Data25} \hat{T}(\lambda)=\left(\begin{array}{cc} \hm{c}^-(\lambda) & \hm{d}^-(\lambda) \\ -\hm{d}^+(\lambda) & \hm{c}^+(\lambda) \end{array}\right), \qquad \hat{T}(\lambda)= T^{-1}(\lambda), \end{align} where \begin{align}\label{Scattering Data27} \begin{array}{cc} \hm{c}^\pm(\lambda)=\hat{\hm{a}}^\mp(\lambda)(\mathds{1}+\rho^{\pm}\rho^{\mp})^{-1}= (\mathds{1}+\tau^{\mp}\tau^{\pm})^{-1}\hat{\hm{a}}^{\mp}(\lambda), \\ \hspace{0.8in} \hm{d}^\pm(\lambda)=\hat{\hm{a}}^\mp(\lambda) \rho^{\pm}(\lambda)(\mathds{1}+\rho^{\mp}\rho^{\pm})^{-1}= (\mathds{1}+\tau^{\mp}\tau^{\pm})^{-1}\tau^{\mp}(\lambda)\hat{\hm{a}}^{\pm}(\lambda). \end{array} \end{align} Here $\rho^{\pm}(\lambda)$ and $\tau^{\pm}(\lambda)$ are the reflection and transmission coefficients respectively: \begin{align}\label{Scattering Data28} \rho^{\pm}(\lambda)=\mathbf{b}^{\pm}\hat{\mathbf{a}}^{\pm}(\lambda) =\hat{\mathbf{c}}^{\pm}\mathbf{d}^{\pm}(\lambda),\qquad \tau^{\pm}(\lambda)=\hat{\mathbf{a}}^{\pm}\mathbf{b}^{\mp}(\lambda) =\mathbf{d}^{\mp}\hat{\mathbf{c}}^{\pm}(\lambda). \end{align} For symmetric spaces of ${\bf BD.I}$ type, the block structure of $T(t,\lambda)$ and its inverse take the form: \begin{align}\label{The Direct Scattering Transform for $L$1} T(t,\lambda)=\left( \begin{array}{ccc} m_{1}^{+} & -\vec{\mathbf{B}}^{- T} &c_{(1)}^{-} \\ \vec{\mathbf{ b}}^{+} & \mathbf{T}_{22} &-s_{0}\vec{\mathbf{b}}^{-} \\ c_{(1)}^{+} & \vec{\mathbf{B}}^{+ T}s_0 & m_{1}^{-} \\ \end{array} \right), \qquad \hat{T}(t,\lambda)=\left( \begin{array}{ccc} m_{1}^{-} & \vec{\mathbf{b}}^{- T} &c_{(1)}^{-} \\ -\vec{\bf{B}}^{+} & \hat{\mathbf{T}}_{22} &s_{0}\vec{\bf{B}}^{-} \\ c_{(1)}^{+} & -\vec{\mathbf{b}}^{+ T}s_0 & m_{1}^{+} \\ \end{array} \right). \end{align} Here, $\vec{\bf{B}}^{\pm}(t,\lambda)$ and $\vec{\bf{b}}^{\pm}(t,\lambda)$ are $2r-1$-component vectors and $c_{(1)}^{\pm}(\lambda)$ and $m_{1}^{\pm}(\lambda)$ are scalar functions. The reflection coefficients $\vec{\rho}^\pm (\lambda)$, the transmission coefficients $\vec{\tau}^\pm (\lambda)$ and the functions $c^\pm_{(1)} (\lambda)$ are determined by the generalised Gauss decomposition of $T(t,\lambda)$: \begin{align}\label{The Direct Scattering Transform for $L$1a} T(t,\lambda)= T^{-}_{\mathrm{J}}(t,\lambda)D^{+}_{\mathrm{J}}(t,\lambda)\hat{S}^{+}_{\mathrm{J}}(t,\lambda)= T^{+}_{\mathrm{J}}(t,\lambda)D^{-}_{\mathrm{J}}(t,\lambda)\hat{S}^{-}_{\mathrm{J}}(t,\lambda). \end{align} Here, $S^{\pm}_{\mathrm{J}}$ and $T^{\pm}_{\mathrm{J}}$ are upper- and lower- block-triangular matrices, which can be written in the form: \begin{align}\label{The Direct Scattering Transform for $L$3} S^{\pm}_{\mathrm{J}}=\mathrm{Re\,}^{(\pm(\vec{\tau}^{\pm}(\lambda,t).\vec{E}^{\pm}_{1}))}, \qquad T^{\pm}_{\mathrm{J}}=\mathrm{Re\,}^{(\mp(\vec{\rho}^{\mp}(\lambda,t).\vec{E}^{\pm}_{1}))}, \end{align} where \begin{align}\label{The Direct Scattering Transform for $L$4} \vec{\tau}^{\pm}(\lambda,t)=\vec{\mathbf{B}}^{\mp}/m^{\pm}_{1}, \quad \vec{\rho}^{\pm}(\lambda,t)=\vec{\mathbf{b}}^{\pm}/m^{\pm}_{1}, \quad c_{(1)}^\pm=\frac{m_{1}^{\pm}}{2}(\vec{\rho}^{\pm,T}, s_{0}\vec{\rho}^{\pm})= \frac{m_{1}^{\mp}}{2}(\vec{\tau}^{\mp,T}, s_{0}\vec{\tau}^{\mp}), \end{align} and $D^{\pm}_{\mathrm{J}}(t,\lambda)$ is the block-diagonal factor in \eqref{The Direct Scattering Transform for $L$1a}: \begin{align}\label{The Direct Scattering Transform for $L$5} D^{+}_{\mathrm{J}}=\left( \begin{array}{ccc} m^{+}_{1} & 0 & 0 \\ 0 & \mathbf{m}_{2}^{+} & 0 \\ 0 & 0 & 1/m_{1}^{+}\\ \end{array} \right), \qquad D^{-}_{\mathrm{J}}=\left( \begin{array}{ccc} 1/m_{1}^{-} & 0 & 0 \\ 0 & \mathbf{m}_{2}^{-} & 0 \\ 0 & 0 & m_{1}^{-} \\ \end{array} \right). \end{align} Here, $m_k^\pm(t,\lambda)$ are the upper/lower rank $k$ principal minors of the scattering matrix $T(t,\lambda)$ \eqref{The Direct Scattering Transform for $L$1a} and \begin{align}\label{The Direct Scattering Transform for $L$6} \mathbf{m}_{2}^{+}=\mathbf{T}_{22}+\frac{\vec{b}^{+}\vec{B}^{-,T}}{m^{+}_{1}}, \qquad \mathbf{m}_{2}^{-}=\mathbf{T}_{22}+\frac{s_{0}\vec{B}^{-}\vec{b}^{+,T}s_{0}}{m^{-}_{1}}. \end{align} If the potential matrix $Q(x,t)$ satisfies the NLEE \eqref{Generalizations of the AKNS Approach1} then the associated scattering matrix $T(t,\lambda)$ evolves in time linearly, i.e. it satisfies the equation \begin{equation}\label{} {\rm i}\ \frac{{\rm d}T}{{\rm d}t}+[{\rm f}(\lambda),T(t,\lambda)]=0. \end{equation} Here $f(\lambda)$ is the dispersion law of the NLEE \eqref{Generalizations of the AKNS Approach1}. For the NLS type of equations, we have $f(\lambda)=-\lambda^2 J$. \subsection{Fundamental Analytic Solutions (FAS)} In the this section, we will briefly outline the construction of the Fundamental Analytic Solutions (FAS) $\chi(x,,t, \lambda)$ of the generalised Zakharov-Shabat system \eqref{Lax Representation g=so(n)} \cite{G2005, GVYa*08}. The FAS can be directly obtained from Jost solutions of \eqref{Lax Representation g=so(n)}: \begin{align}\label{The Direct Scattering Transform for $L$2} \chi^{\pm}(x,t,\lambda)=\phi (x,t,\lambda) S ^{\pm}_{\mathrm{J}}(t,\lambda)=\psi(x,t,\lambda)T^{\mp}_{\mathrm{J}}(t,\lambda) D^{\pm}_{\mathrm{J}}(t,\lambda), \end{align} by using the generalised Gauss decomposition \eqref{The Direct Scattering Transform for $L$1a} of the scattering matrix $T(t,\lambda)$. On the real axis (i.e. on the continuous spectrum of $L(\lambda)$), the two FAS are linearly dependent: \begin{align} \chi^{+}(x,t,\lambda)=&\chi^{-}(x,t,\lambda) G_0(t,\lambda),\qquad \lambda\in \mathbb{R},\label{The Zakharov-Shabat Dressing Method Local1} \end{align} where the sewing function $G_0(t,\lambda)$ can be expressed in terms of the generalised Gauss factors $S ^{\pm}_{\mathrm{J}}(t,\lambda)$: \begin{equation}\label{} G_{0,J}(t,\lambda)=\hat{S}_J^{-}(t,\lambda)S_J^{+}(t,\lambda)|_{t=0}. \end{equation} The independent matrix elements of $G_{0}(t,\lambda)$, together with the discrete spectrum of $L(\lambda)$ form up the minimal set of scattering data of $L$. Based on the completeness relations of the associated square solution and Wronskian type of relation, one can recover the potential $Q(x,t)$ out of the minimal set of scattering data \cite{G*86}. We end up this section with the remark that, although the general form of the Gauss decomposition \eqref{The Direct Scattering Transform for $L$1a} holds true for any symmetric space, for symmetric spaces of type ${\bf A.III}$ one can simplify the form of the matrix block by slightly modifying the Gauss decomposition of $T(\lambda)$ \footnote{Decomposition of type \eqref{the FAS16} are known as LU decomposition of $T(\lambda)$, while a decomposition of type \eqref{The Direct Scattering Transform for $L$1a} is known as LDU decomposition of $T(\lambda)$.}: \begin{align}\label{the FAS16} T(\lambda)= & \mathbf{T}_J^{-}(\lambda) \hat{\mathbf{S}}_J^{+}(\lambda) =\mathbf{T}^{+}_J(\lambda) \hat{\mathbf{S}}_J^{-}(\lambda). \end{align} Now, $\mathbf{S}^{\pm}(\lambda)$ and $\mathbf{T}^{\pm}(\lambda)$ are the block-triangular matrices: \begin{equation}\label{the FAS15} \begin{array}{cc} \mathbf{S}_J^{+}(\lambda)=\left(\begin{array}{cc} \mathds{1}_s, & \hm{d}^-(\lambda) \\ 0, & \hm{c}^+(\lambda) \end{array}\right),\qquad % \mathbf{T}_J^{-}(\lambda)=\left(\begin{array}{cc} \hm{a}^+(\lambda), & 0 \\ \hm{b}^+(\lambda), & \mathds{1}_p \end{array}\right), \\ \\ \mathbf{S}_J^{-}(\lambda)=\left(\begin{array}{cc} \hm{c}^-(\lambda), & 0 \\ -\hm{d}^+(\lambda), & \mathds{1}_p \end{array}\right),\qquad % \mathbf{T}_J^{+}(\lambda)=\left(\begin{array}{cc} \mathds{1}_s, & -\hm{b}^-(\lambda) \\ 0, & \hm{a}^-(\lambda) \end{array}\right). \end{array} \end{equation} Using the Gauss decomposition \eqref{the FAS16} one can write explicitly the sewing function $G_{0,J}(t,\lambda)$ \eqref{} in terms of the blocks of $T(t,\lambda)$ and it inverse: \begin{align} G_{0}(\lambda)= \hat{D}^{-}(\lambda)(\mathds{1}+K^{-}(\lambda)),\qquad \hat{G}_{0}(\lambda)= \hat{D}^{+}(\lambda)(\mathds{1}-K^{+}(\lambda)).\label{the FAS18} \end{align} Here the block-diagonal factors $D^{\pm}(\lambda)$ and block-off-diagonal $K^{\pm}(\lambda)$ factors are expressed in terms of the blocks ${\bf a}^\pm (\lambda)$, ${\bf b}^\pm (\lambda)$ of $T(\lambda)$ \eqref{Scattering Data6} and in terms of the blocks ${\bf c}^\pm (\lambda)$, ${\bf d}^\pm (\lambda)$ \eqref{Scattering Data25} of its inverse $\hat{T}(\lambda)$, respectively; \begin{align}\label{the FAS19} D^{-}(\lambda)= &\left( \begin{array}{cc} \mathbf{c}^{-}(\lambda) & 0 \\ 0 & \mathbf{a}^{-}(\lambda) \\ \end{array} \right),\qquad K^{-}(\lambda)=\left( \begin{array}{cc} 0 & \mathbf{d}^{-}(\lambda)\\ \mathbf{b}^{+}(\lambda) &0 \\ \end{array} \right), \\ D^{+}(\lambda)= &\left( \begin{array}{cc} \mathbf{a}^{+}(\lambda) & 0 \\ 0 & \mathbf{c}^{+}(\lambda) \\ \end{array} \right),\qquad K^{+}(\lambda)=\left( \begin{array}{cc} 0 & \mathbf{b}^{-}(\lambda)\\ \mathbf{d}^{+}(\lambda) &0 \\ \end{array} \right). \end{align} The superscripts $``\pm''$ in the expressions $D^{\pm}(\lambda)$ above mean analyticity for $\lambda \in {\Bbb C}_\pm$. \section{Dressing method and soliton solutions}\label{} \subsection{The Zakharov-Shabat Dressing Method}\label{} The main idea of the Zakharov-Shabat dressing method is to start with a regular solution $\eta^{\pm}_{0}(x,t,\lambda)$ of \eqref{Lax Representation g=so(n)a} and to perform a "dressing" by adding singularities at two prescribed points $\lambda_k^\pm$ \cite{RI, GKKV*08, 66, 67, G2005, ZMNP, Sh, DokLeb}. The new singular solution will have the form: \begin{subequations}\label{The Dressing Method: The Singular SolutionsLocal2} \begin{align} \eta^{\pm}(x,t,\lambda)= & u_{k} (x,t,\lambda)\eta^{\pm}_{0}(x,t,\lambda)w_{k,\pm}^{-1}(\lambda).\label{The Dressing Method: The Singular SolutionsLocal2a} \end{align} \end{subequations} Here, the matrices $w_{k,\pm}(\lambda)$ can be given by: \begin{align}\label{The Dressing Method: The Singular SolutionsLocal20} w_{k,+}(\lambda)=&\mathds{1},\qquad w_{k,-}(\lambda) = \left( \begin{array}{cc} u_{11,k}^{+} & 0 \\ 0 & u_{22,k}^{-} \\ \end{array} \right), \end{align} where the limits of the dressing factor can be obtained by: \begin{align}\label{The Dressing Method: The Singular SolutionsLocal19} \lim_{x \rightarrow \infty}u_{k}(x,t,\lambda)=\left( \begin{array}{cc} u_{11,k}^{+} & 0 \\ 0 & \mathds{1}\\ \end{array} \right), \qquad u_{11,k}^{+} = \mathds{1}+ (c_{k}(\lambda)-1)P_{11, k}^{+}, \\ \lim_{x \rightarrow -\infty}u_{k}(x,t,\lambda)=\left( \begin{array}{cc} \mathds{1}& 0 \\ 0 & u_{22,k}^{-}\\ \end{array} \right), \qquad u_{22,k}^{-} = \mathds{1}+ (c_{k}(\lambda)-1)P_{22, k}^{-}. \end{align} For Lax operators relatred to ${\bf A.III}$ symmetric spaces, the dressing factors have the form: \begin{equation} u_{k} (x,t,\lambda)= \mathds{1} + (c_{k}(\lambda)-1)P_{k}(x,t), \qquad c_{k}(\lambda)= \frac{\lambda-\lambda_{k}^{+}}{\lambda-\lambda^{-}_{k}}. \label{The Dressing Method: The Singular SolutionsLocal2b} \end{equation} Here $u_{k}(x,t,\lambda)$ is the dressing factor, $P_{k}(x,t)$ is a projector of rank $1$ has the form: \begin{align}\label{The Dressing Method: The Singular SolutionsLocal3} P_{k}(x,t) = \frac{|n_k(x,t)\rangle \langle m_k(x,t)|}{\langle m_k(x,t)|n_k(x,t)\rangle}, \end{align} where $|n_k(x,t)\rangle =\chi^{+}_{0}(x,t,\lambda^{+}_{1})|n_{0,1}\rangle$ is a column vector and $\langle m_{1}(x,t)|= \langle m_{0,1}| \hat{\chi}^{-}_{0}(x,t,\lambda^{-}_{1})$ is a row vector and both of them are eigenvector. The projectors $P_{k}(x,t)$ automatically satisfy the condition $P_{k}^{2}(x,t)=P_{k}(x,t)$. The functions $\eta^{\pm}(x,t,\lambda)$ satisfies the linear system: \begin{equation}\label{The Zakharov-Shabat Dressing Method Local12} \left({\rm i}\frac{\mathrm{d} \eta^\pm}{\mathrm{d} x}+Q(x,t)\eta^\pm(x,t,\lambda)- \lambda\left[\mathrm{J}, \eta^\pm(x,t,\lambda)\right]\right)=0. \end{equation} where $\eta^{\pm}(x,t,\lambda)$ are related to the ZS system with an unknown potential $Q(x,t)$, which can be found later, while $\eta_{0}^{\pm}(x,t,\lambda)$ is related to the ZS system with a known potential $Q_{0}(x,t)$: \begin{equation}\label{The Dressing Method: The Singular SolutionsLocal11} {\rm i}\frac{\mathrm{d} \eta_{0}^\pm}{\mathrm{d} x}+Q_{0}(x,t)\eta_{0}^\pm(x,t,\lambda)-(\lambda)\left[\mathrm{J}, \eta_{0}^\pm(x,t,\lambda)\right]=0. \end{equation} Now, from \eqref{The Dressing Method: The Singular SolutionsLocal2a} and \eqref{The Dressing Method: The Singular SolutionsLocal11}, we find that the dressing factor $u_k(x,t,\lambda)$ satisfies the following equation: \begin{equation}\label{The Dressing Method: The Singular SolutionsLocal12} {\rm i}\frac{\mathrm{d} u_k}{\mathrm{d} x}+Q(x,t)u_k(x,t,\lambda)-u_k(x,t,\lambda)Q_{0}(x,t)- \lambda \left[\mathrm{J}, u_k(x,t,\lambda)\right]=0, \end{equation} since the anzats for the dressing factor $u_k(x,t,\lambda)$ in \eqref{The Dressing Method: The Singular SolutionsLocal2b} and \eqref{The Dressing Method: The Singular SolutionsLocal12} are compatible with respect to $\lambda$, then there are two conditions can be applied to the left-hand side of \eqref{The Dressing Method: The Singular SolutionsLocal12} which are the limit for $\lambda\rightarrow\infty$ and the residue at $\lambda=\lambda^{-}_{k}$ and both of them are vanish. The first condition leads to the potential: \begin{align}\label{The Dressing Method: The Singular SolutionsLocal13} Q(x,t)-Q_{0}(x,t)=-(\lambda^{+}_{k}-\lambda^{-}_{k})[\mathrm{J},P_k(x,t)], \end{align} and the second condition gives the following nonlinear equation for $P_k(x,t)$: \begin{align} {\rm i}\frac{\mathrm{d} P_{k}}{\mathrm{d} x}& + Q_{0}(x,t)P_{k}(x,t)-P_{k}(x,t)Q_{0}(x,t)-\lambda^{+}_{k}\mathrm{J} P_{k}(x,t)+ \lambda^{-}_{k} P_{k}(x,t) \mathrm{J} \\ &+ (\lambda^{+}_{k}-\lambda^{-}_{k})P_{k}(x,t) \mathrm{J} P_{k}(x) =0. \label{The Dressing Method: The Singular SolutionsLocal14} \end{align} In addition, we have the normalization condition $\lim_{\lambda\to \infty} u_k (x,t,\lambda)= \mathds{1}$. In the case of ${\bf BD.I}$ symmetric spaces the dressing factor $u_k(x,t,\lambda)$ can be written in the form: \begin{align} u_{k} (x,t,\lambda)= & \mathds{1} + (c_{k}(\lambda)-1)P_{k}(x,t)+ \left(\frac{1}{c_{k}(\lambda)}-1\right)\overline{P}_{k}(x,t), \qquad \overline{P}_{k}=S_{0}P^{T}_{k}S_{0}^{-1}, \label{One Soliton } \end{align} where $P_{k}(x,t)$ and $\overline{P}_{k}(x,t)$ are mutually orthogonal projectors with rank $1$ \cite{vgrn}. In a similar fashion, one can write the dressed potential $Q(x,t)$ as: \begin{align}\label{One Soliton } Q(x,t)= Q_{0}(x,t)-(\lambda^{+}_{k}-\lambda^{-}_{k}) [\mathrm{J},P_k(x,t)-\overline{P}_{k}(x,t)]. \end{align} \subsection{One Soliton Solutions}\label{} \begin{example} If we take ${\frak g}\simeq sl(3,{\Bbb C})$ and a dressing factor in the form \eqref{The Dressing Method: The Singular SolutionsLocal2b} satisfying the nonlocal reduction conditions \eqref{The Local Involutions2m1}, this will imply the following involution on the dressing factor: \begin{align}\label{The Nonlocal Involution } B \ u_{1}(-x,t,-\lambda^{\ast})^{\dagger}\ B^{-1}=& \ u^{-1}_{1}(x,t,\lambda), \end{align} where $B$ is constant block-diagonal matrix. As a result, the projector $P_1$ must satisfy: \begin{align}\label{The Nonlocal Involution } P_{1}(x,t)= & BP_{1}^{\dagger}(-x,t)B^{-1},\qquad (-\lambda^{\pm})^{\ast}=\lambda^{\mp}. \end{align} That means the projector $P_{1}(x,t)$ and $c_{1}(\lambda)$ become: \begin{align}\label{The Nonlocal Involution } P_{1}(x,t)= & \frac{|n_1(x,t)\rangle \langle n_1^{\ast}(-x,t)|B}{\langle n_1^{\ast}(-x,t)|B | n_1(x,t)\rangle}, \qquad \langle m_1(x,t)|=(B | n_1(-x,t)\rangle)^{\dagger}, \\ c_{1}(\lambda)= & \frac{\lambda-\lambda_{1}^{+}}{\lambda+(\lambda^{+}_{1})^{\ast}}. \end{align} So, from \eqref{The Dressing Method: The Singular SolutionsLocal13}, the components of the one-soliton solution can be written as: \begin{subequations}\label{The Nonlocal Involution } \begin{align}\label{The Nonlocal Involution } q_{1j}(x,t)= &-2(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast}) \ \frac {n_{0,1}^{1} (n_{0,1}^{j})^{\ast} \mathrm{Re\,}^{-\mathrm{i}( \tilde{M}_{1}^{+}(x,t)+ (\tilde{M}_{1}^{+})^{*}(-x,t))}}{2 R_{0,1} \rm {cosh} (2\nu_{1}x+2 \tilde{\mu_{1}}t+\xi_{0,1})}, \qquad j=2,3, \nonumber \\ q_{j1}(x,t)= & \ -2(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast}) \ \frac {n_{0,1}^{j} (n_{0,1}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i}(\tilde{M}_{1}^{+}(x,t)+ (\tilde{M}_{1}^{+})^{*} (-x,t))}}{2 R_{0,1} \rm {cosh} (2\nu_{1}x+2\tilde{\mu_{1}}t+ \xi_{0,1})}, \quad \qquad j=2,3. \end{align} Here, \begin{align} \tilde{M}_{1}^{\pm}(x,t)= & \lambda_{1}^{\pm}x + 2 \ \rm f_{0, 1}^{\pm} t, \qquad (\tilde{M}_{1}^{+}(-x,t))^{*} = (-\lambda_{1}^{+})^{*}(-x) + 2 \ \rm (f_{0}(-\lambda_{1}^{+})^{\ast}) t, \label{The Nonlocal Involution } \\ \nu_{1}= & \rm i ((-\lambda^{+}_{1})^{*}-\lambda^{+}_{1})/2, \qquad \tilde{\mu_{1}}= \rm i (\rm f(-\lambda^{+}_{1})^{*}-\rm f(\lambda^{+}_{1})), \label{The Nonlocal Involution } \\ R_{0,1}= & \sqrt{(n_{0,1}^{1})^{\ast}n_{0,1}^1 ( (n_{0,1}^{2})^{\ast}n_{0,1}^2 + (n_{0,1}^{3})^{\ast}n_{0,1}^3)}, \quad \xi_{0,1}=\frac{1}{2}\rm {ln} \frac{ (n_{0,1}^{1})^{\ast}n_{0,1}^1}{(n_{0,1}^{2})^{\ast}n_{0,1}^2+ (n_{0,1}^{3})^{\ast}n_{0,1}^3}. \label{The Nonlocal Involution } \end{align} \end{subequations} \end{example} \begin{example} If we take again ${\frak g}\simeq sl(3,{\Bbb C})$ and the involution automorphism to be the Weyl reflection with respect to the second simple root ${\rm e}_2-{\rm e}_3$ of $sl(3,{\Bbb C})$ and impose a reduction of type (B) in \eqref{The Local Involutions2m1}. This will correspond to another block matrix $B = \left( \begin{array}{cc} \mathbf{B_+} &0 \\ 0& \mathbf{B_-} \\ \end{array}\right)$ where $ \mathbf{B_-}$ is a block off-diagonal matrix. The blocks can be written as \begin{align}\label{another nonlocal involution} \mathbf{B_+}=\mp1, \qquad \text{and} \qquad \mathbf{B_-}= \left(\begin{array}{cc} 0&\pm1 \\ \pm1&0\\ \end{array}\right). \end{align} The reductions conditions for the matrix blocks can be written as: \begin{align}\label{another nonlocal involution1} \mbox{(C)} &&\qquad \mathbf{q}^{-}(x,t)=-\mathbf{B_-}(\mathbf{q}^{+}(-x,t))^{\dag}(\mathbf{B_+})^{-1},\quad \mathbf{q}^{+}(x,t)=-\mathbf{B_+}(\mathbf{q}^{-}(-x,t))^{\dag}(\mathbf{B_-})^{-1}. \end{align} Now, if we apply the involution of type (C) with using \eqref{another nonlocal involution}, then the standard Manakov VNLS equation reproduced as: \begin{equation}\label{another nonlocal involution2} \begin{array}{cc} -{\rm i}q_{1,t}+ q_{1,xx}+2 \ (q_{1}(x,t) \ q_{2}^{\ast}(-x,t)+ q_{2}(x,t) \ q_{1}^{\ast}(-x,t)) \ q_{1}(x,t)=0, \vspace{0.1in}\\ -{\rm i}q_{2,t}+ q_{2,xx}+2 \ (q_{1}(x,t) \ q_{2}^{\ast}(-x,t)+ q_{2}(x,t) \ q_{1}^{\ast}(-x,t)) \ q_{2}(x,t)=0. \end{array} \end{equation} Also, we can write one-soliton solution of the Manakov VNLS equation for this type of involution as the dressing factor \eqref{The Dressing Method: The Singular SolutionsLocal2b} is satisfied by the involution \eqref{The Nonlocal Involution } with \eqref{The Nonlocal Involution }. So, from \eqref{The Dressing Method: The Singular SolutionsLocal13}, the components of the one-soliton solution can be written as: \begin{subequations}\label{another nonlocal involution3} \begin{align}\label{another nonlocal involution3a} q_{12}(x,t)= &-2(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast}) \ n_{0,1}^{1} \mathrm{Re\,}^{-\mathrm{i} \tilde{M}_{1}^{+}(x,t)} R_{1}^{-1}(x,t)(n_{0,1}^{3})^{\ast} \mathrm{Re\,}^{-\mathrm{i}\tilde{M}_{1}^{+,\ast}(-x,t)}, \nonumber \\ q_{13}(x,t)= &-2(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast}) \ n_{0,1}^{1} \mathrm{Re\,}^{-\mathrm{i} \tilde{M}_{1}^{+}(x,t)} R_{1}^{-1}(x,t)(n_{0,1}^{2})^{\ast} \mathrm{Re\,}^{-\mathrm{i}\tilde{M}_{1}^{+,\ast}(-x,t)}, \nonumber \\ q_{21}(x,t)= & \ -2(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast}) \ n_{0,1}^{2} \mathrm{Re\,}^{\mathrm{i} \tilde{M}_{1}^{+}(x,t)} R_{1}^{-1}(x,t) (n_{0,1}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} \tilde{M}_{1}^{+,\ast}(-x,t))}, \nonumber \\ q_{31}(x,t)= & \ -2(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast}) \ n_{0,1}^{3} \mathrm{Re\,}^{\mathrm{i} \tilde{M}_{1}^{+}(x,t)} R_{1}^{-1}(x,t) (n_{0,1}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} \tilde{M}_{1}^{+,\ast}(-x,t))}. \end{align} Here, \begin{align} \tilde{M}_{1}^{\pm}(x,t)= & \lambda_{1}^{\pm}x + 2 \ \rm f_{0, 1}^{\pm} t, \qquad (\tilde{M}_{1}^{+}(-x,t))^{*} = (-\lambda_{1}^{+})^{*}(-x) + 2 \ \rm (f_{0}(-\lambda_{1}^{+})^{\ast}) t, \label{another nonlocal involution3b} \\ R_{1}(x,t)= &2 R_{0,1} \rm {cosh} (2\nu_{1}x+2\tilde{\mu_{1}}t+ \xi_{0,1}), \label{another nonlocal involution3c} \\ \nu_{1}= &\rm i ((-\lambda^{+}_{1})^{*}-\lambda^{+}_{1})/2, \qquad \tilde{\mu_{1}}= \rm i (\rm f(-\lambda^{+}_{1})^{*}-\rm f(\lambda^{+}_{1})), \label{another nonlocal involution3c1} \\ R_{0,1}= & \sqrt{(n_{0,1}^{1})^{\ast}n_{0,1}^1 ( (n_{0,1}^{3})^{\ast}n_{0,1}^2 + (n_{0,1}^{2})^{\ast}n_{0,1}^3)}, \qquad \xi_{0,1}=\frac{1}{2}\rm {ln} \frac{ (n_{0,1}^{1})^{\ast}n_{0,1}^1}{(n_{0,1}^{3})^{\ast}n_{0,1}^2+ (n_{0,1}^{2})^{\ast}n_{0,1}^3}. \label{another nonlocal involution3d} \end{align} \end{subequations} \end{example} \begin{example} If we take ${\frak g}\simeq so(5,{\Bbb C})$, then the dressing factor will take the form \eqref{One Soliton }. If we impose the nonlocal reduction conditions \eqref{The Local Involutions2m1}: \begin{align}\label{The nonlocal Involution } B \ u_{1}(-x,t,-\lambda^{\ast})^{\dagger}\ B^{-1}=& \ u^{-1}_{1}(x,t,\lambda), \end{align} this will lead to the following reduction condition: \begin{align}\label{The nonlocal Involution } P_{1}(x,t)= & BP_{1}^{\dagger}(-x,t)B^{-1},\qquad (-\lambda^{\pm})^{\ast}=\lambda^{\mp}. \end{align} That means the projector $P_{1}(x,t)$ and $c_{1}(\lambda)$ become: \begin{align} P_{1}(x,t)= & \frac{|n_1(x,t)\rangle \langle n_1^{\ast}(-x,t)|B}{\langle n_1^{\ast}(-x,t)|B | n_1(x,t)\rangle}, \qquad \langle m_1(x,t)|=(B | n_1(-x,t)\rangle)^{\dagger},\label{The nonlocal Involution } \\ c_{1}(\lambda)= & \frac{\lambda-\lambda_{1}^{+}}{\lambda+(\lambda^{+}_{1})^{\ast}}.\label{The nonlocal Involution } \end{align} Thus, taking the trivial solution $Q_{0}(x,t)=0$, the components of one soliton solution can be written as: \begin{subequations}\label{The nonlocal Involution nonlocal Involution } \begin{align} q_{12}^{+}(x,t)= & q_{45}^{+}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{1} (n_{0,1}^{2})^{\ast} \mathrm{Re\,}^{-\mathrm{i} M_{1}^{+}(x,t)} -n_{0,1}^{4}(n_{0,1}^{5})^{\ast} \mathrm{Re\,}^{-\mathrm{i} (M_{1}^{+}(-x,t))^{\ast}}), \nonumber \\ q_{13}^{+}(x,t)= &- q_{35}^{+}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{1} (n_{0,1}^{3})^{\ast} \mathrm{Re\,}^{-\mathrm{i} M_{1}^{+}(x,t)} +n_{0,1}^{3}(n_{0,1}^{5})^{\ast} \mathrm{Re\,}^{-\mathrm{i} (M_{1}^{+}(-x,t))^{\ast}}), \nonumber \\ q_{14}^{+}(x,t)= & q_{25}^{+}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{1} (n_{0,1}^{4})^{\ast} \mathrm{Re\,}^{-\mathrm{i} M_{1}^{+}(x,t)} -n_{0,1}^{2} (n_{0,1}^{5})^{\ast} \mathrm{Re\,}^{-\mathrm{i} (M_{1}^{+}(-x,t))^{\ast}}), \nonumber \\ q_{21}^{-}(x,t)= & q_{54}^{-}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{2} (n_{0,1}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} (M_{1}^{-}(-x,t))^{\ast}}-n_{0,1}^{5}(n_{0,1}^{4})^{\ast} \mathrm{Re\,}^{\mathrm{i} M_{1}^{+}(x,t)}), \nonumber \\ q_{31}^{-}(x,t)= & - q_{53}^{-}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{3} (n_{0,1}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} (M_{1}^{-}(-x,t))^{\ast}} +n_{0,1}^{5}(n_{0,1}^{3})^{\ast} \mathrm{Re\,}^{\mathrm{i} M_{1}^{+}(x,t)}), \nonumber \\ q_{41}^{-}(x,t)= & q_{52}^{-}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{4} (n_{0,1}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} (M_{1}^{+}(-x,t))^{\ast}}-n_{0,1}^{5}(n_{01}^{2})^{\ast} \mathrm{Re\,}^{\mathrm{i} M_{1}^{+}(x,t)}).\label{The nonlocal Involution } \end{align} Here, \begin{align} (M_{1}^{+}(-x,t))^{\ast}= &(-\lambda_{1}^{+})^{\ast}(-x) + 2 \ (\rm f(-\lambda^{+}_{1})^{*}) t, \quad R_{1}(x,t)=2 R_{0,1} \rm {cosh} (2\nu_{1}(x)+2\tilde{\mu_{1}}t+ \xi_{0,1})+S, \label{The nonlocal Involution } \\ \nu_{1}= & \rm i ((-\lambda^{+}_{1})^{*}-\lambda^{+}_{1})/2, \qquad \tilde{\mu_{1}}= \rm i (\rm f(-\lambda^{+}_{1})^{*}-\rm f(\lambda^{+}_{1})),\label{The nonlocal Involution } \\ S= &(n_{0,1}^{2})^{*} n_{0,1}^{2}+ (n_{0,1}^{3})^{*} n_{0,1}^{3}+ (n_{0,1}^{4})^{*} n_{0,1}^{4},\label{The nonlocal Involution } \\ R_{0,1}= & \sqrt{ (n_{0,1}^1)^{*} n_{0,1}^1 + (n_{0,1}^5)^{*} n_{0,1}^5 }, \qquad \xi_{0,1}=\frac{1}{2}\rm {ln} \frac{ (n_{0,1}^1)^{*} n_{0,1}^1 }{ (n_{0,1}^5)^{*} n_{0,1}^5 }. \label{The nonlocal Involution } \end{align} \end{subequations} \end{example} \begin{example} If we take again ${\frak g}\simeq so(5,{\Bbb C})$ and another $\mathbb{Z}_2$ reduction of type (B) \eqref{The Local Involutions2m1} with a block matrix $B$ containing an off-diagonal block $\mathbf{B_\pm}$: $\mathbf{B_+}=\mathbf{B_-}=\mp1$ and $\mathbf{B_\pm}= \text{off-diag}(\pm1,\ldots,\pm1)$. In this case, the reductions conditions for the matrix blocks can be written as: \begin{align}\label{another nonlocal involution1a} \mathbf{q}^{-}(x,t)=-\mathbf{B_-}(\mathbf{q}^{+}(-x,t))^{\dag}(\mathbf{B_+})^{-1},\quad \mathbf{q}^{+}(x,t)=-\mathbf{B_+}(\mathbf{q}^{-}(-x,t))^{\dag}(\mathbf{B_-})^{-1}. \end{align} Now, recall example A from Section 2, if we apply the involution of type (B) with using $\mathbf{B_+}=\mathbf{B_-}=\mp1$ and $\mathbf{B_\pm}= \text{off-diag}(\pm1,\ldots,\pm1)$, then the 3-component NLS system in \eqref{Some Algebraic Properties for Cartan Weyl Basis7} can be rewritten as: \begin{align}\label{another nonlocal involutionso(5)} \mathrm{i}(q_{t}^{+})_{12}&+(q_{xx}^{+})_{12}+ 2\left(q_{12}(x,t)q_{14}^{\ast}(-x,t)+2 q_{13}(x,t)q_{13}^{\ast}(-x,t)\right) q_{12} +2 q^{+,\ast}_{12}(-x,t)(q^{+}_{13})^{2}=0 ,\nonumber \\ \mathrm{i}(q_{t}^{+})_{13}&+(q_{xx}^{+})_{13}+ 2\left(q_{12}(x,t)q_{14}^{\ast}(-x,t)+ q_{13}(x,t)q_{13}^{\ast}(-x,t)+ q_{14}(x,t)q_{12}^{\ast}(-x,t) \right) q_{13}\nonumber \\ &+2 q^{+,\ast}_{13}(-x,t)q^{+}_{14}q^{+}_{12}=0 ,\nonumber \\ \mathrm{i}(q_{t}^{+})_{14}&+(q_{xx}^{+})_{14}+ 2\left(q_{14}(x,t)q_{12}^{\ast}(-x,t)+2 q_{13}(x,t)q_{13}^{\ast}(-x,t) \right) q_{14} +2 q^{+,\ast}_{14}(-x,t)(q^{+}_{13})^{2}=0. \end{align} Also, we can write one-soliton solution of this system for this type of involution as the dressing factor \eqref{One Soliton } is satisfied by the involution \eqref{The nonlocal Involution } with \eqref{The nonlocal Involution }. So, from \eqref{One Soliton }, the components of the one-soliton solution can be written as: \begin{subequations}\label{The nonlocal Involution } \begin{align} q_{12}^{+}(x,t)= & q_{45}^{+}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{1} (n_{0,1}^{4})^{\ast} \mathrm{Re\,}^{-\mathrm{i} M_{1}^{+}(x,t)} -n_{0,1}^{4}(n_{0,1}^{5})^{\ast} \mathrm{Re\,}^{-\mathrm{i} (M_{1}^{+}(-x,t))^{\ast}}), \nonumber \\ q_{13}^{+}(x,t)= &- q_{35}^{+}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{1} (n_{0,1}^{3})^{\ast} \mathrm{Re\,}^{-\mathrm{i} M_{1}^{+}(x,t)} +n_{0,1}^{3}(n_{0,1}^{5})^{\ast} \mathrm{Re\,}^{-\mathrm{i} (M_{1}^{+}(-x,t))^{\ast}}), \nonumber \\ q_{14}^{+}(x,t)= & q_{25}^{+}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{1} (n_{0,1}^{2})^{\ast} \mathrm{Re\,}^{-\mathrm{i} M_{1}^{+}(x,t)} -n_{0,1}^{2} (n_{0,1}^{5})^{\ast} \mathrm{Re\,}^{-\mathrm{i} (M_{1}^{+}(-x,t))^{\ast}}), \nonumber \\ q_{21}^{-}(x,t)= & q_{54}^{-}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{2} (n_{0,1}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} (M_{1}^{-}(-x,t))^{\ast}}-n_{0,1}^{5}(n_{0,1}^{2})^{\ast} \mathrm{Re\,}^{\mathrm{i} M_{1}^{+}(x,t)}), \nonumber \\ q_{31}^{-}(x,t)= & - q_{53}^{-}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{3} (n_{0,1}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} (M_{1}^{-}(-x,t))^{\ast}} +n_{0,1}^{5}(n_{0,1}^{3})^{\ast} \mathrm{Re\,}^{\mathrm{i} M_{1}^{+}(x,t)}), \nonumber \\ q_{41}^{-}(x,t)= & q_{52}^{-}(x,t)= -(\lambda^{+}_{1}+(\lambda^{+}_{1})^{\ast})R_1^{-1}(x,t)(n_{0,1}^{4} (n_{0,1}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} (M_{1}^{+}(-x,t))^{\ast}}-n_{0,1}^{5}(n_{01}^{4})^{\ast} \mathrm{Re\,}^{\mathrm{i} M_{1}^{+}(x,t)}).\label{The nonlocal Involution } \end{align} Here, \begin{align} (M_{1}^{+}(-x,t))^{\ast}= &(-\lambda_{1}^{+})^{\ast}(-x) + 2 \ (\rm f(-\lambda^{+}_{1})^{*}) t, \quad R_{1}(x,t)=2 R_{0,1} \rm {cosh} (2\nu_{1}(x)+2\tilde{\mu_{1}}t+ \xi_{0,1})+S, \label{The nonlocal Involution } \\ \nu_{1}= & \rm i ((-\lambda^{+}_{1})^{*}-\lambda^{+}_{1})/2, \qquad \tilde{\mu_{1}}= \rm i (\rm f(-\lambda^{+}_{1})^{*}-\rm f(\lambda^{+}_{1})),\label{The nonlocal Involution } \\ S= &(n_{0,1}^{4})^{*} n_{0,1}^{2}+ (n_{0,1}^{3})^{*} n_{0,1}^{3}+ (n_{0,1}^{2})^{*} n_{0,1}^{4},\label{The nonlocal Involution } \\ R_{0,1}= & \sqrt{ (n_{0,1}^1)^{*} n_{0,1}^1 + (n_{0,1}^5)^{*} n_{0,1}^5 }, \qquad \xi_{0,1}=\frac{1}{2}\rm {ln} \frac{ (n_{0,1}^1)^{*} n_{0,1}^1 }{ (n_{0,1}^5)^{*} n_{0,1}^5 }. \label{The nonlocal Involution } \end{align} \end{subequations} \end{example} \subsection{Two Soliton Solution}\label{} \begin{example} For ${\frak g}\simeq sl(3,{\Bbb C})$, in order to get the 2-soliton solution, the dressing factor c \eqref{The Dressing Method: The Singular SolutionsLocal2} can be rewritten as: \begin{subequations}\label{involution of two Soliton Solution} \begin{align} \eta^{+}(x,t,\lambda)= & u_{1,2} (x,t,\lambda)\eta^{+}_{0}(x,t,\lambda) (\omega ^{\pm}_{1,2}(\lambda))^{-1} , \label{involution of two Soliton Solutiona} \\ u_{1,2}(x,t,\lambda) = & \mathds{1} + (c_{1}(\lambda)-1)P_{1}(x,t)+ (c_{2}(\lambda)-1)P_{2}(x,t).\label{involution of two Soliton Solutionb} \end{align} \end{subequations} Thus, the singular solution (two soliton solution) with singularities located at $\lambda_{1}^{\pm}$ and $\lambda_{2}^{\pm}$ can be obtained as follows: \begin{align}\label{involution of two Soliton Solution2} Q(x,t)=& Q_{0}(x,t)-\sum^{2}_{k=1}(\lambda^{+}_{k}-\lambda^{-}_{k})[\mathrm{J},P_k(x,t)]. \end{align} Assuming a canonical reduction of type (B) \eqref{The Local Involutions2m1} with ${\bf B}={\Bbb I}$, from \eqref{involution of two Soliton Solution2} the components of the two-soliton solution can be written as: \begin{subequations}\label{involution of two Soliton Solution6} \begin{align} q_{1j}(x,t)= &-2\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast}) \ \frac {n_{0,k}^{k} (n_{0,k}^{j})^{\ast} \mathrm{Re\,}^{-\mathrm{i}( \tilde{M}_{k}^{+}(x,t)+ (\tilde{M}_{k}^{+})^{*}(-x,t))}}{2 R_{0,k} \rm {cosh} (2\nu_{k}x+2 \tilde{\mu_{k}}t+\xi_{0,k})}, \qquad j=2,3, \nonumber \\ q_{j1}(x,t)= & \- 2\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast}) \ \frac {n_{0,k}^{j} (n_{0,k}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i}(\tilde{M}_{k}^{+}(x,t)+ (\tilde{M}_{k}^{+})^{*} (-x,t))}}{2 R_{0,k} \rm {cosh} (2\nu_{k}x+2\tilde{\mu_{k}}t+ \xi_{0,k})}, \quad \qquad j=2,3. \label{involution of two Soliton Solution6a} \end{align} Here, \begin{align} \tilde{M}_{k}^{\pm}(x,t)= & \lambda_{k}^{\pm}x + 2 \ \rm f_{0, k}^{\pm} t, \qquad (\tilde{M}_{k}^{+}(-x,t))^{*} = (-\lambda_{k}^{+})^{*}(-x) + 2 \ (\rm f(-\lambda^{+}_{1})^{*}) t, \label{involution of two Soliton Solution6b} \\ \nu_{k}= & \rm i ((-\lambda^{+}_{k})^{*}-\lambda^{+}_{k})/2, \qquad \tilde{\mu_{k}}= \rm i (\rm f(-\lambda^{+}_{k})^{*}-\rm f(\lambda^{+}_{k})), \label{involution of two Soliton Solution6c} \\ R_{0,k}= & \sqrt{(n_{0,k}^{1})^{\ast}n_{0,k}^1 ( (n_{0,k}^{2})^{\ast}n_{0,k}^2 + (n_{0,k}^{3})^{\ast}n_{0,k}^3)}, \quad \xi_{0,k}=\frac{1}{2}\rm {ln} \frac{ (n_{0,k}^{1})^{\ast}n_{0,k}^1}{(n_{0,k}^{2})^{\ast}n_{0,k}^2+ (n_{0,k}^{3})^{\ast}n_{0,k}^3}. \label{involution of two Soliton Solution6d} \end{align} \end{subequations} \end{example} \begin{remark} Alternatively, in order to obtain 2-soliton solution, one can apply dressing method again on the 1-soliton solution used as a seed solution: \begin{align}\label{involution of two Soliton Solution1} \eta^{+}(x,t,\lambda)= & u_{2}(x,t,\lambda)\eta^{+}_{1}(x,t,\lambda)(\omega ^{\pm}_{2}(\lambda))^{-1}, \qquad u_{2} (x,t,\lambda)=\mathds{1} + (c_{2}(\lambda)-1)P_{2}(x,t). \end{align} \end{remark} \begin{example} In the case of the involution of Example 2, the 2-soliton dressing factor \eqref{involution of two Soliton Solutiona} is automatically compatible with the involution. Therefore, from \eqref{involution of two Soliton Solution2}, the components of the two-soliton solution can be written as: \begin{subequations}\label{two another nonlocal involution sl(3)} \begin{align} q_{12}(x,t)= &-2\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast}) \ n_{0,k}^{1} \mathrm{Re\,}^{-\mathrm{i} \tilde{M}_{k}^{+}(x,t)} R_{k}^{-1}(x,t)(n_{0,k}^{3})^{\ast} \mathrm{Re\,}^{-\mathrm{i}\tilde{M}_{k}^{+,\ast}(-x,t)}, \nonumber \\ q_{13}(x,t)= &-2\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast}) \ n_{0,k}^{1} \mathrm{Re\,}^{-\mathrm{i} \tilde{M}_{k}^{+}(x,t)} R_{k}^{-1}(x,t)(n_{0,k}^{2})^{\ast} \mathrm{Re\,}^{-\mathrm{i}\tilde{M}_{k}^{+,\ast}(-x,t)}, \nonumber \\ q_{21}(x,t)= & \ -2\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast}) \ n_{0,k}^{2} \mathrm{Re\,}^{\mathrm{i} \tilde{M}_{k}^{+}(x,t)} R_{k}^{-1}(x,t) (n_{0,k}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} \tilde{M}_{k}^{+,\ast}(-x,t))}, \nonumber \\ q_{31}(x,t)= & \ -2\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast}) \ n_{0,k}^{3} \mathrm{Re\,}^{\mathrm{i} \tilde{M}_{k}^{+}(x,t)} R_{k}^{-1}(x,t) (n_{0,k}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} \tilde{M}_{k}^{+,\ast}(-x,t))}. \label{two another nonlocal involution sl(3)a} \end{align} Here, \begin{align} \tilde{M}_{k}^{\pm}(x,t)= & \lambda_{k}^{\pm}x + 2 \ \rm f_{0, k}^{\pm} t, \qquad (\tilde{M}_{k}^{+}(-x,t))^{*} = (-\lambda_{k}^{+})^{*}(-x) + 2 \ (\rm f(-\lambda^{+}_{1})^{*}) t, \label{two another nonlocal involution sl(3)b} \\ R_{1}(x,t)= &2 R_{0,1} \rm {cosh} (2\nu_{1}x+2\tilde{\mu_{1}}t+ \xi_{0,1}),\label{two another nonlocal involution sl(3)c} \\ \nu_{k}= & \rm i ((-\lambda^{+}_{k})^{*}-\lambda^{+}_{k})/2, \qquad \tilde{\mu_{k}}= \rm i (\rm f(-\lambda^{+}_{k})^{*}-\rm f(\lambda^{+}_{k})), \label{two another nonlocal involution sl(3)c1} \\ R_{0,k}= & \sqrt{(n_{0,k}^{1})^{\ast}n_{0,k}^1 ( (n_{0,k}^{3})^{\ast}n_{0,k}^2 + (n_{0,k}^{2})^{\ast}n_{0,k}^3)}, \qquad \xi_{0,k}=\frac{1}{2}\rm {ln} \frac{ (n_{0,k}^{1})^{\ast}n_{0,k}^1}{(n_{0,k}^{3})^{\ast}n_{0,k}^2+ (n_{0,k}^{2})^{\ast}n_{0,k}^3}. \label{two another nonlocal involution sl(3)d} \end{align} \end{subequations} \end{example} \begin{example}If we take ${\frak g}\simeq so(5,{\Bbb C})$, then the 2-soliton dressing factors $u(x,t,\lambda)$ can be considered with two more poles in the following form: \begin{align}\label{Two Soliton } u_{1,2} (x,t,\lambda)= & \mathds{1} + (c_{1}(\lambda)-1)P_{1}(x,t)+ \left(\frac{1}{c_{1}(\lambda)}-1\right)\overline{P}_{1}(x,t),\nonumber \\ & + (c_{2}(\lambda)-1)P_{2}(x,t)+ \left(\frac{1}{c_{2}(\lambda)}-1\right)\overline{P}_{2}(x,t), \\ c_{k}(\lambda)= \frac{\lambda-\lambda_{k}^{+}}{\lambda-\lambda^{-}_{k}} & , \quad P_{k}(x,t) = \frac{|n_k(x,t)\rangle \langle m_k(x,t)|}{\langle m_k(x,t)|n_k(x,t)\rangle}, \quad \overline{P}_{k}=S_{0}P^{T}_{k}S_{0}^{-1}, \quad k=1,2.\label{Two Soliton } \end{align} Thus, the new potential $Q(x,t)$ corresponding to a given trivial solution $Q_{0}(x,t)=0$ can be obtained by the following form: \begin{align}\label{Two Soliton } Q(x,t)= -\sum_{k=1}^{2}(\lambda^{+}_{k}-\lambda^{-}_{k}) [\mathrm{J},P_k(x,t)-\overline{P}_{k}(x,t)]. \end{align} The components of new potential $Q(x,t)$ can be obtained by: \begin{subequations}\label{Two Soliton } \begin{align} q_{12}^{+}(x,t)= q_{45}^{+}(x,t)= & - \sum^{2}_{k=1}(\lambda^{+}_{k}-\lambda^{-}_{k}) R_k^{-1}(x,t)(n_{0,k}^{1}m_{0,k}^{2} \mathrm{Re\,}^{-\mathrm{i} M_{k}^{+}(x,t)} +n_{0,k}^{4}m_{0,k}^{5} \mathrm{Re\,}^{-\mathrm{i} M_{k}^{-}(x,t)}), \nonumber \\ q_{13}^{+}(x,t)= -q_{35}^{+}(x,t) = & -\sum^{2}_{k=1}(\lambda^{+}_{k}-\lambda^{-}_{k}) R_k^{-1}(x,t)(n_{0,k}^{1}m_{0,k}^{3} \mathrm{Re\,}^{-\mathrm{i} M_{k}^{+}(x,t)} -n_{0,k}^{3}m_{0,k}^{5} \mathrm{Re\,}^{-\mathrm{i} M_{k}^{-}(x,t)}), \nonumber \\ q_{14}^{+}(x,t)= q_{25}^{+}(x,t) =& -\sum^{2}_{k=1}(\lambda^{+}_{k}-\lambda^{-}_{k}) R_k^{-1}(x,t) (n_{0,k}^{1}m_{0,k}^{4} \mathrm{Re\,}^{-\mathrm{i} M_{k}^{+}(x,t)} +n_{0,k}^{2}m_{0,k}^{5} \mathrm{Re\,}^{-\mathrm{i} M_{k}^{-}(x,t)}), \nonumber \\ q_{21}^{-}(x,t)= q_{54}^{-}(x,t) = & -\sum^{2}_{k=1}(\lambda^{+}_{k}-\lambda^{-}_{k}) R_k^{-1}(x,t)(-n_{0,k}^{2}m_{0,k}^{1} \mathrm{Re\,}^{\mathrm{i} M_{k}^{-}(x,t)}-n_{0,k}^{5}m_{0,k}^{4} \mathrm{Re\,}^{\mathrm{i} M_{k}^{+}(x,t)}), \nonumber \\ q_{31}^{-}(x,t)= -q_{53}^{-}(x,t) = & -\sum^{2}_{k=1}(\lambda^{+}_{k}-\lambda^{-}_{k}) R_k^{-1}(x,t)(-n_{0,k}^{3}m_{0,k}^{1} \mathrm{Re\,}^{\mathrm{i} M_{k}^{-}(x,t)} +n_{0,k}^{5}m_{0,k}^{3} \mathrm{Re\,}^{\mathrm{i} M_{k}^{+}(x,t)}), \nonumber \\ q_{41}^{-}(x,t)= q_{52}^{-}(x,t) =& -\sum^{2}_{k=1}(\lambda^{+}_{k}-\lambda^{-}_{k}) R_k^{-1}(x,t) (-n_{0,k}^{4}m_{0,k}^{1} \mathrm{Re\,}^{\mathrm{i} M_{k}^{-}(x,t)}-n_{0,k}^{5}m_{0,k}^{2} \mathrm{Re\,}^{\mathrm{i} M_{k}^{+}(x,t)}).\label{Two Soliton } \end{align} Here \begin{align} M_{k}^{\pm}(x,t)= &\lambda_{k}^{\pm}x + 2 f_{0, k}^{\pm} t,\qquad R_{k}(x,t)=2 R_{0,k} \rm {cosh} (2\nu_{k}x+2\tilde{\mu_{k}}t+ \xi_{0,k})+S,\label{Two Soliton } \\ \nu_{k}= &\rm i (\lambda^{-}_{k}-\lambda^{+}_{k})/2, \quad \tilde{\mu_{k}}= \rm i (\rm f(\lambda^{-}_{k})-\rm f(\lambda^{+}_{k})), \quad S= m_{0,k}^{2}n_{0,k}^{2}+ m_{0,k}^{3}n_{0,k}^{3}+ m_{0,k}^{4}n_{0,k}^{4},\label{Two Soliton } \\ R_{0,k}= & \sqrt{ m_{0,k}^1 n_{0,k}^1 + m_{0,k}^5 n_{0,k}^5 }, \qquad \xi_{0,k}=\frac{1}{2}\rm {ln} \frac{ m_{0,k}^1 n_{0,k}^1 }{ m_{0,k}^5 n_{0,k}^5 }. \label{Two Soliton } \end{align} \end{subequations} \end{example} \begin{example} Again in the case when ${\frak g}\simeq so(5,{\Bbb C})$, if we impose a nonlocal involution of the form \eqref{The Local Involutions2m1} then the dressing factor in \eqref{Two Soliton } will satisfy the following reduction condition: \begin{align}\label{involution of two Soliton } B \ \prod _{k=1}^{2} u_{k}(-x,t,-\lambda^{\ast})^{\dagger}\ B^{-1}=& \ \prod _{k=1}^{2} u^{-1}_{k}(x,t,\lambda), \end{align} where $B$ is constant block-diagonal matrix. This involution is satisfied if: \begin{align}\label{involution of two Soliton } P_{k}(x,t)= & BP_{k}^{\dagger}(-x,t)B^{-1},\qquad (-\lambda^{\pm})^{\ast}=\lambda^{\mp}. \end{align} That means the projector $P_{k}(x,t)$ and $c_{k}(\lambda)$ become: \begin{align} P_{k}(x,t)= & \frac{|n_k(x,t)\rangle \langle n_k^{\ast}(-x,t)|B}{\langle n_k^{\ast}(-x,t)|B | n_k(x,t)\rangle}, \quad \langle m_k(x,t)|=(B | n_k(-x,t)\rangle)^{\dagger},\quad c_{k}(\lambda)= & \frac{\lambda-\lambda_{k}^{+}}{\lambda+(\lambda^{+}_{k})^{\ast}}.\label{involution of two Soliton } \end{align} So, from \eqref{One Soliton }, it follows that \begin{align}\label{involution of two Soliton } Q(x,t)=& Q_{0}(x,t)-(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast}) [\mathrm{J},BP_{k}^{\dagger}(-x,t)B^{-1}]. \end{align} Thus, the components of two soliton solution can be written as: \begin{subequations}\label{involution of two Soliton } \begin{align} q_{12}^{+}(x,t)= & q_{45}^{+}(x,t)= -\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast})R_k^{-1}(x,t)(n_{0,k}^{1} (n_{0,k}^{2})^{\ast} \mathrm{Re\,}^{-\mathrm{i} M_{k}^{+}(x,t)} -n_{0,k}^{4}(n_{0,k}^{5})^{\ast} \mathrm{Re\,}^{-\mathrm{i} (M_{k}^{+}(-x,t))^{\ast}}), \nonumber \\ q_{13}^{+}(x,t)= &- q_{35}^{+}(x,t)= -\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast})R_k^{-1}(x,t)(n_{0,k}^{1} (n_{0,k}^{3})^{\ast} \mathrm{Re\,}^{-\mathrm{i} M_{k}^{+}(x,t)} +n_{0,k}^{3}(n_{0,k}^{5})^{\ast} \mathrm{Re\,}^{-\mathrm{i} (M_{k}^{+}(-x,t))^{\ast}}), \nonumber \\ q_{14}^{+}(x,t)= & q_{25}^{+}(x,t)= -\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast})R_k^{-1}(x,t)(n_{0,k}^{1} (n_{0,k}^{4})^{\ast} \mathrm{Re\,}^{-\mathrm{i} M_{k}^{+}(x,t)} -n_{0,k}^{2} (n_{0,k}^{5})^{\ast} \mathrm{Re\,}^{-\mathrm{i} (M_{k}^{+}(-x,t))^{\ast}}), \nonumber \\ q_{21}^{-}(x,t)= & q_{54}^{-}(x,t)= -\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast})R_k^{-1}(x,t)(n_{0,k}^{2} (n_{0,k}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} (M_{k}^{-}(-x,t))^{\ast}}-n_{0,k}^{5}(n_{0,k}^{4})^{\ast} \mathrm{Re\,}^{\mathrm{i} M_{k}^{+}(x,t)}), \nonumber \\ q_{31}^{-}(x,t)= & - q_{53}^{-}(x,t)= -\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast})R_k^{-1}(x,t)(n_{0,k}^{3} (n_{0,k}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} (M_{k}^{-}(-x,t))^{\ast}} +n_{0,k}^{5}(n_{0,k}^{3})^{\ast} \mathrm{Re\,}^{\mathrm{i} M_{k}^{+}(x,t)}), \nonumber \\ q_{41}^{-}(x,t)= & q_{52}^{-}(x,t)= -\sum^{2}_{k=1}(\lambda^{+}_{k}+(\lambda^{+}_{k})^{\ast})R_k^{-1}(x,t)(n_{0,k}^{4} (n_{0,k}^{1})^{\ast} \mathrm{Re\,}^{\mathrm{i} (M_{k}^{+}(-x,t))^{\ast}}-n_{0,k}^{5}(n_{0,k}^{2})^{\ast} \mathrm{Re\,}^{\mathrm{i} M_{k}^{+}(x,t)}).\label{involution of two Soliton } \end{align} Here, \begin{align} (M_{k}^{+}(-x,t))^{\ast}= &(-\lambda_{k}^{+})^{\ast}(-x) + 2 \ (\rm f(-\lambda^{+}_{1})^{*}) t, \quad R_{k}(x,t)=2 R_{0,k} \rm {cosh} (2\nu_{k}(x)+2\tilde{\mu_{k}}t+ \xi_{0,k})+S, \label{involution of two Soliton } \\ \nu_{k}= & \rm i ((-\lambda^{+}_{k})^{*}-\lambda^{+}_{k})/2, \qquad \tilde{\mu_{k}}= \rm i (\rm f(-\lambda^{+}_{k})^{*}-\rm f(\lambda^{+}_{k})),\label{involution of two Soliton } \\ S= &(n_{0,k}^{2})^{*} n_{0,k}^{2}+ (n_{0,k}^{3})^{*} n_{0,k}^{3}+ (n_{0,k}^{4})^{*} n_{0,k}^{4},\label{involution of two Soliton } \\ R_{0,k}= & \sqrt{ (n_{0,k}^1)^{*} n_{0,k}^1 + (n_{0,k}^5)^{*} n_{0,k}^5 }, \qquad \xi_{0,k}=\frac{1}{2}\rm {ln} \frac{ (n_{0,k}^1)^{*} n_{0,k}^1 }{ (n_{0,k}^5)^{*} n_{0,k}^5 }. \label{involution of two Soliton } \end{align} \end{subequations} \end{example} \section{Conclusions}\label{sec:6} We have studied here multi-component generalisations of NLS models on {\bf A.III} and {\bf BD.I} symmetric spaces. Our study is based on two types of main examples: the Manakov vector NLS equation, related to symmetric spaces of {\bf A.III} type and Kulish-Sklyanin models, related to symmetric spaces of {\bf BD.I} type. Firstly, we formulated the direct scattering problem for both models. This includes: the construction of the Jost solutions, the scattering matrix and the minimal set of scattering data. Based on the Gauss decomposition of the scattering matrix, we have also constructed the fundamental analytic solutions (FAS). It turns out that the spectral properties of the Lax operator depend crucially on the choice of representation of the underlying Lie algebra or symmetric space while the minimal set of scattering data is provided by the same set of functions \cite{GG2010}. Finally, we have presented a modification of the dressing method and obtained 1- and 2-solitons for both models with nonlocal reductions. Depending on the positions of the discrete eigenvalues $\lambda_j^\pm$ in the spectral plane, there are two regimes for the 2-soliton solution: if two of the discrete eigenvalues are in the upper half of the complex plane, while the other two are in the lower half, then the nonlocal involution will preserve their number inside each of the contours, so it will result in a Riemann-Hilbert problem with balanced number of singularities and therefore the corresponding 2-soliton solutions will be regular for all $t$. Otherwise, the 2-soliton solution will develop again a singularity in finite time as in the case, studied in \cite{AblMus,AblMus1,AblMus2}. The results of this paper can be extended in several directions: \begin{itemize} \item To construct gauge covariant formulation of the multi-component NLS hierarchies on symmetric spaces, including the Wronskian relations, the ``squared solutions'' and their completeness relations, the descriptions of the class of NLEEs associated to a given scattering problem, the generating (recursion) operator and it spectral decomposition, the description of the infinite set of integrals of motion, the hierarchy of Hamiltonian structures and the $r$-matrix formulation. \item To study the gauge equivalent systems of multi-component ferromagnetics on symmetric spaces \cite{66,GGMV2011a,GGMV2011b}. \item To study different types of reductions of multi-component integrable systems on symmetric spaces \cite{gc1,gc,Metin}. \item To study the associated Darboux transformations and their generalizations for NLS equations over Hermitian symmetric spaces and to obtain multi-soliton solutions via such generalizations. This includes also rational solutions \cite{DokLeb}. \item To extend our results for the case of non-vanishing boundary conditions (a non-trivial background) \cite{AblMus3,Li1}. The considerations required in this case are more complicated and will be discussed it elsewhere. \item To study other types of integrable hierarchies on symmetric spaces, e.g. quadratic bundles (related to the derivative NLS equation, the Kaup-Newell equation or Gerdjikov-Ivanov equation) \cite{GGI2016,GGI2017,GVYa*08,Rei,ReiSTSH,78} or rational bundles \cite{GGMV2011a,GGMV2011b}. \end{itemize} \section*{Acknowledgements} The authors have the pleasure to thank Prof. Mark Ablowitz and Prof. Vladimir Gerdjikov for numerous useful discussions.

\section{Introduction} \label{} Observations of the redshifted 21-cm signal from neutral Hydrogen is the most promising method for revealing astrophysical processes occurring during the Epoch of Reionization (EoR) and the Cosmic Dawn (CD), and has great potential at independently constraining the cosmological parameters~\citep[see e.g.][for reviews]{Furlanetto06,Morales10}. Several experiments are currently underway aiming at statistically detecting the 21-cm signal from the Epoch of Reionization (e.g. LOFAR~\footnote{Low Frequency Array, }, MWA~\footnote{Murchison Widefield Array, }, PAPER~\footnote{Precision Array to Probe EoR, }), already achieving increasingly attractive upper limits on the 21-cm signal power spectra~\citep{Patil17,Beardsley16,Ali15}, and paving the way for the second generation experiments such as the SKA~\footnote{Square Kilometre Array, } and HERA~\footnote{Hydrogen Epoch of Reionization Array, } which will be capable, with their order of magnitude improvement in sensitivity, of robust power spectra characterization and for the first time directly image the large scale neutral hydrogen structures from EoR and CD. A major obstacle in achieving this exciting goal is that the cosmological signal is considerably weaker than the astrophysical foregrounds. The foregrounds must be accurately and precisely removed from the observed data as any error at this stage has the ability to strongly affect the 21-cm signal extraction. While the brightest extragalactic sources can be modeled and removed after direction dependent calibration~\citep[e.g.,][]{Yatawatta13}, the remaining foregrounds, composed of extragalactic emission below the confusion noise level and diffuse and partly polarised galactic emission, are still approximately 3 to 4 orders of magnitude brighter than the 21-cm signal. They are nevertheless expected to be spectrally smooth while the 21-cm signal is anticipated to be uncorrelated on frequency scales on the order of MHz or larger. This important difference is the main characteristic exploited by the many techniques that have been proposed to model and remove the foreground emission, including parametric fits~\citep[e.g.,][]{Jelic08,Bonaldi15} and non-parametric methods~\citep[e.g.,][]{Harker09,Chapman13}. The assumption made here of a smooth foreground signal is however strongly affected by the limitations and constraints of the observational setup. Many additional contaminants have been identified related to the reality of radio interferometry, and observation in the low frequency domain. The chromatic (i.e., wavelength dependent) response of the instrument manifests itself as a frequency dependence of both the synthesized beam, also called the Point Spread Function (PSF), and the Primary Beam (PB) of a receiver station, producing chromatic side lobes from sources inside the field of view (FoV)~\citep{Vedantham12,Hazelton13} and outside it~\citep{Thyagarajan15,Mort17, Bharat17}. Calibration errors and mis-subtraction of sources due to imperfect sky modeling will also contribute to additional side lobe noise~\citep{Datta10,Morales12,Trott12,Ewall17,Barry16,Patil16}. The rapid phase and sometime amplitudes modifications of radio waves caused by small scale structures in the ionosphere also produce scintillation noise~\citep{Koopmans10, Vedantham16}. These different mechanisms will all add spectral structure to the otherwise smooth astrophysical foregrounds, and are well-known as ``mode-mixing" effects in the literature. Both simulations and analytic calculations have demonstrated that these mode-mixing contaminants are essentially localized inside a wedge-like region in the two-dimensional angular ($k_{\perp}$) versus line-of-sight ($k_{\parallel}$) power spectra (see Fig.~\ref{}). This peculiar shape is explained by the fact that larger baselines (higher $k_{\perp}$) change length more rapidly as a function of frequency than smaller baselines, causing increasingly faster spectral fluctuations, and thus producing power into proportionally higher $k_{\parallel}$ modes. Mitigating those additional foreground contaminants has proven to be extremely difficult. Increasing the degrees of freedom of a parametric fit would considerably increase the fitting error and might also suppress the 21-cm signal at the lower-value $k$ modes. Non-parametric methods are in theory not limited to smooth models but modeling an increasingly more complex foreground often means increasing the numbers of components (without a clear understanding about what they include), which risks the leakage of 21-cm signal into the reconstructed foreground model and vice versa. In~\cite{Patil17}, six to eight components of the Generalized Morphological Component Analysis (GMCA;~\cite{Chapman13}) were necessary to model, even imperfectly, the foreground contaminants, reaching limits where it is increasingly more difficult to assess and be confident about the accuracy of the foreground removal process. We note that the GMCA is not based on a statistical framework but simply separates the signal in the least number of morphological components. This makes it hard to build in a-priori knowledge about the signal in any kind of signal separation. Ideally, we would like to consistently account for every single mode-mixing contaminant that have been identified so far. Recently,~\cite{Ghosh17} has demonstrated that estimating the 21-cm power spectrum using a maximum likelihood inversion of the spherical-wave visibility equation can considerably reduce the chromatic effects due to the frequency dependence of the PSF, effectively recovering a PSF-deconvolved sky.~\cite{Vedantham12} also proposed a new imaging technique in the attempt of decreasing visibilities gridding artifacts. Convolving the visibilities with a `frequency independent' window function makes it easier to strongly attenuate the frequency dependent response to the side-lobes of the primary antenna pattern and RFI sources, which are mostly located on the ground~\citep{Ghosh11b}. Improving the primary beam characterization~\citep{Thyagarajan16}, and using calibration scheme which enforce smooth gain solution in frequency~\citep{Barry16, Yatawatta16}, also contribute to reducing the mode-mixing. Nevertheless, most of the improvements are done with the purpose of limiting the leakage of foreground contaminants outside the foreground wedge, and any foreground removal strategy will still be required to properly handle mode-mixing contaminants inside the wedge. \begin{figure} \includegraphics[trim={2mm 2mm 0 0},clip]{materiel/wedge+equi-k.pdf} \centering \caption{\label{} Schematic representation of the 2D power spectra (inspired by a similar Figure in~\protect\cite{Barry16}), illustrating the foreground wedge and the EoR window. Instrumental chromaticity and imperfect calibration and sky model will produce foreground mode-mixing contaminants which are mainly concentrated inside the primary beam field of view line (dashed line) and can leak up to the horizon line. Only modes above this line are theoretically free of foreground contaminants. Lines of equidistant $k = \sqrt{k_{\perp}^2 + k_{\parallel}}$ are over-plotted in gray.} \end{figure} An alternative, which has been increasingly popular, is to try to avoid as much as possible the foregrounds, and only probe a triangular-shaped region in $k$-space where the 21-cm signal is dominant. Because most of the instrumental chromatic effects are confined inside the wedge, there exists in theory an ``EoR window'' (see Fig.~\ref{}) within which one could perform statistical analyses of the 21-cm signal without significantly being affected by foreground contaminants. ~\cite{Liu14a,Liu14b} proposed a mathematical formalism describing the wedge, allowing one to maximize the extent of the accessible EoR window. Several methods have also been developed to estimate the covariance of the foregrounds~\citep{Dillon15,Murray17} which can then be included in a power spectra estimator~\citep{Trott16b}. These foreground avoidance or suppression methods have the disadvantage, however, of considerably reducing the sensitivity of the instruments, because they reduce the numbers of modes that can be probed~\citep{Furlanetto16}. ~\cite{Pober14} have estimated the impact of avoiding the foreground wedge region to be a factor $\sim$ 3 for PAPER or HERA, and even a factor $\sim$ 6 for LOFAR. It is thus not a viable alternative for experiments such as LOFAR-EoR, most sensible at $k \leq 0.3 \ \rm{h\, cMpc^ {-1}}$ with a peak sensibility at $k \sim 0.1 \ \rm{h\, cMpc^{-1}}$, and for which very little foreground-free modes are available (see Fig.~\ref{}). Additionally, ignoring the wedge can also introduce a bias in the recovered 21-cm signal power spectra~\citep{Jensen16} and it is also much harder to probe the redshift space distortion effects of the 21-cm signal if the foreground cleaning in the wedge region is discarded~\citep{Jonathan15}. Considering that for a successful foreground removal strategy all the foreground contaminants need to be accounted for, and that ad-hoc modeling is not an option for most of them, we propose a novel non-parametric method based on Gaussian Process Regression (GPR). In this framework, the different components of the problem, including the astrophysical smooth foregrounds, mid-scale fluctuations associated with mode-mixing, the noise, and a basic 21-cm signal model, are modeled with Gaussian Process (GP), allowing for a clean separation of their contributions, and a precise estimation of there uncertainty. GPR is extensively used in machine learning applications and has been successfully used in astronomy, for example to model blazar broadband flares~\citep{Karamanavis16}, inferring stellar rotation periods~\citep{Hojjati13}, or modeling instrumental systematics~\citep{Aigrain16}. It provides flexibility and avoids having to specify an arbitrary functional form for the variations we seek to model. Implemented in a Bayesian framework, it enables us to incorporate relevant physical information in the form of covariance structure priors (spectral and possible spatial) on the various components. We introduce the foreground modeling and removal method in Section~\ref{}. To demonstrate the ability of the technique we perform simulations including realistic astrophysical foreground models, mid-scale frequency fluctuations, and the simulated 21-cm signal. We introduce the simulation pipeline in Section~\ref{}, before presenting the results in Section~\ref{}. Finally, we summarize the main conclusions in Section~\ref{}. \section{Formalism} \label{} In this section, we first introduce the Gaussian Process Regression (GPR) formalism and then proceed to describe the application of this technique to foreground modeling and removal in 21-cm signal observations. \subsection{Gaussian Process} A Gaussian Process (GP) is a probability distribution over functions~\citep{Rasmussen05,Gelman14}. It constitutes the generalization of the Gaussian distribution of random variables or vectors, into the space of functions. A GP $f \sim \mathcal{GP}\left(m, \kappa\right)$ is fully defined by its mean $m$ and covariance function $\kappa$ (also called ``kernel") so that any set of points $\mathbf{x}$ in some continuous input space is associated with normally distributed random variables $\mathbf{f} = f(\mathbf{x})$, with mean $m(\mathbf{x})$ and where the value of $\kappa$ specifies the covariance between the function values at any two points. The Gaussian Process is the joint distribution of all those random variables which all share the desired covariance properties, \begin{equation} f(\mathbf{x}) \sim \mathcal{N}\left(m(\mathbf{x}), K(\mathbf{x}, \mathbf{x})\right). \end{equation} with $K(\mathbf{x}, \mathbf{x})$ an $n \times n$ covariance matrix with element $(p, q)$ corresponding to $\kappa(x_p, x_q)$. In Gaussian Process Regression, we seek a function $f(\mathbf{x})$ that would model our noisy observation $\mathbf{d} = f(\mathbf{x}) + \mathbf{n}$, where $\mathbf{n}$ is a Gaussian distributed noise with variance $\sigma_n^2$, observed at the data points $\mathbf{x}$. Given a Gaussian Process prior $\mathcal{GP}\left(m, \kappa\right)$, the joint density distribution of the observations $\mathbf{d}$ and the predicted function values $ \mathbf{f'} = f(\mathbf{x'})$ at a set of points $\mathbf{x'}$ is, \begin{equation} \left[ \begin{array}{c} \mathbf{d} \\ \mathbf{f'} \end{array}\right] \sim \mathcal{N}\left( \left[\begin{array}{c} m(\mathbf{x}) \\ m(\mathbf{x'}) \end{array}\right], \left[ \begin{array}{cc} K(\mathbf{x}, \mathbf{x}) + \sigma_n^2 I & K(\mathbf{x},\mathbf{x'}) \\ K(\mathbf{x'},\mathbf{x}) & K(\mathbf{x'},\mathbf{x'}) \end{array} \right] \right). \end{equation} where $I$ is the identity matrix. Conditioning the joint prior distribution on the observations, we obtain the joint posterior distribution of our model at data points $\mathbf{x}'$, \begin{equation} \label{} \mathbf{f'}|\mathbf{x}, \mathbf{d}, \mathbf{x'} \sim \mathcal{N}\left(\mathrm{E} ( \mathbf{f'}), \mathrm{cov}(\mathbf{f'})\right), \end{equation} where $\mathrm{E}(.)$ and $\mathrm{cov}(.)$ are the standard notations for the mean and covariance respectively, and with, \begin{align} \label{} \nonumber \mathrm{E}(\mathbf{f'}) &= m(\mathbf{x}') + K(\mathbf{x}', \mathbf{x})\left[K( \mathbf{x}, \mathbf{x}) + \sigma_n^2 I\right]^{-1} (\mathbf{d} - m(\mathbf{x}'))\\ \mathrm{cov}(\mathbf{f'}) &= K(\mathbf{x}', \mathbf{x}') - K(\mathbf{x}', \mathbf{x})\left[K(\mathbf{x}, \mathbf{x}) + \sigma_n^2 I\right]^{-1}K(\mathbf{x}, \mathbf{x}'). \end{align} The function values $\mathbf{f'}$ can then be sampled from the joint posterior distribution by evaluating the mean and covariance matrix above, the mean being the maximum a-posterior solution. Gaussian Process Regression can be seen as a fitting method in which we assign prior information on the function values of the model in the form of a covariance function. The results are marginalized over all functions drawn from the probability distribution function (PDF) in Eq.~\ref{}, unlike parametric modeling where the model family is fixed and one only marginalizes over the parameters. While we assume here a data model with Gaussian noise, GP could be used in theory as priors associated with other likelihood functions, such as a Poisson likelihood~\citep{Diggle98} or a Student-t likelihood~\citep{Neal97}. Even with current Gaussian data model, the predictive mean of the posterior PDF (Eq.~\ref{}) is not required to be Gaussian distributed over the data points $\mathbf{x}$, enabling one to model non-Gaussian variation. \subsection{Covariance functions} The covariance function $\kappa$ determines the structure that the GP will be able to model. A common class of covariance functions is the Matern class~\citep{Stein99}. It is defined by, \begin{equation} \label{} \kappa_{\mathrm{Matern}}(x_p, x_q) = \frac{2^{1 - \eta}}{\Gamma(\eta)}\left( \frac{ \sqrt{2\eta}r}{l}\right)^{\eta} K_{\eta}\left(\frac{\sqrt{2\eta}r}{l}\right), \end{equation} where $r = |x_q - x_p|$ and $K_{\eta}$ is the modified Bessel function of the second kind. Functions obtained with this class of kernel are at least $\eta$-times differentiable. The kernel is also parametrized by the `hyper-parameter' $l$, which is the characteristic coherence-scale. It denotes the distance in the input space after which the function values change significantly and thus defines the `smoothness' of the function. Special cases of this class are obtained by setting $\eta$ to $\infty$, in which case we obtain a Gaussian kernel, and by setting $\eta = 1 / 2$, in which case we obtain an exponential kernel. Throughout the paper we use the functional form in Eq.~\ref{} because of its flexibility. Importantly, if the observation we seek to model is composed of multiple additive sources, a GP model kernel can be the addition of their covariance functions. It is then possible to separate the contribution of the different terms. We show in Appendix~\ref{} that GPR can be formulated as a linear regression problem where one models the data $\mathbf{d}$ as $\mathbf{d} = \mathbf{H} \mathbf{f} + \mathbf{n}$, where $\mathbf{f}$ are the weights of the basis functions and $\mathbf{n}$ is the noise contribution. In general this is an ill-posed problem and one needs to set additional prior or constraints on $ \mathbf{f}$. Usually, in GPR the constraint is statistical and set in the form of covariance matrix which can be modeled as a sum of covariance functions corresponding to the signals from the EoR, foregrounds and noise. \subsection{Covariance function optimization} Model selection in the context of GPR is a two-fold process. The first choice is that of the type of covariance function that could model the data, and the second is that of optimizing the `hyper-parameters' of this covariance function. Both can be done in a Bayesian framework, selecting the model that maximizes the marginal-likelihood, also called the evidence. This is the integral of the likelihood times the prior \begin{equation} p(\mathbf{d}|\mathbf{x}, \theta) = \int{p(\mathbf{d}|\mathbf{f} ,\mathbf{x}, \theta)p( \mathbf{f}|\mathbf{x}, \theta)d\mathbf{f}}, \end{equation} with $\theta$ being the hyper-parameters of the covariance function $\kappa$. Under the assumption of Gaussianity, we can integrate over $\mathbf{f}$ analytically, yielding the log-marginal-likelihood~(LML), \begin{equation} \label{} \mathrm{log}\,p(\mathbf{d} |\mathbf{x}, \theta) = -\frac{1}{2} \mathbf{d}^\intercal (K + \sigma_n^2 I)^{-1} \mathbf{d} - \frac{1}{2} \mathrm{log}\,|K + \sigma_n^2 I| - \frac{n}{2} \mathrm{log}\,2\pi \end{equation} where we have used the shorthand $K \equiv K(\mathbf{x}, \mathbf{x})$ and with $n$ the number of sampled points. The posterior probability density of the hyper-parameters is then found by applying Bayes' theorem: \begin{equation} \label{} \mathrm{log}\,p(\theta| \mathbf{d}, \mathbf{x}) \propto \mathrm{log}\,p(\mathbf{d} |\mathbf{x}, \theta) + \mathrm{log}\,p (\theta). \end{equation} We may then either select the model that maximizes Eq.~\ref{} (maximum likelihood estimate), or incorporate prior information on the hyper-parameters and maximize Eq.~\ref{} (maximum a-posteriori estimate). The marginal likelihood does not only favor the models that fit best the data, overly complex models are also disfavored~\citep{Rasmussen05}. Selecting the values of $\theta$ that maximizes the LML is a non-linear optimization problem. Because the covariance function is defined analytically, it is trivial to compute the partial derivatives of the marginal likelihood with respect to the hyper-parameters, which allow the use of efficient gradient-based optimization algorithm. \begin{figure} \includegraphics[trim={1mm 3mm 0 0},clip]{materiel/eor_21cmfast_covariance.pdf} \centering \caption{\label{} Exponential covariance functions for different values of the coherence-scale $l$ (gray lines), compared to the covariance of a simulated 21-cm EoR signal at different redshift (colored lines).} \end{figure} \subsection{GPR for 21-cm signal detection} \label{} In the context of 21-cm signal detection, we are interested in modeling our data $\mathbf{d}$ observed at frequencies $\mathbf{\nu}$ by a foreground, a 21-cm and a noise signal $\mathbf{n}$: \begin{equation} \mathbf{d} = f_{\mathrm{fg}}(\mathbf{\nu}) + f_{\mathrm{21}}(\mathbf{\nu}) + \mathbf{n}. \end{equation} To separate the foreground signal from the 21-cm signal, we can exploit their different frequency behavior: the 21-cm signal is expected to be uncorrelated on scales of a few MHz, while the foregrounds are expected to be smooth on that scale. The covariance function of our GP model can then be composed of a foreground covariance function $K_{\mathrm{fg}}$ and a 21-cm signal covariance function $K_{\mathrm{21}}$, \begin{equation} K = K_{\mathrm{fg}} + K_{\mathrm{21}}. \end{equation} The aim behind including explicitly a 21-cm signal component is not so much to model it but to isolate its covariance contribution from the covariance of the foregrounds. A complete model is also necessary to insure accurate estimation of the error covariance matrix. We can now write the joint probability density distribution of the observations $\mathbf{d}$ and the function values $ \mathbf{f}_{\mathrm{fg}}$ of the foreground model $f_{\mathrm{fg}}$ at the same frequencies $\nu$: \begin{equation} \left[ \begin{array}{c} \mathbf{d} \\ \mathbf{f_{\mathrm{fg}}} \end{array}\right] \sim \mathcal{N}\left( \left[\begin{array}{c} 0 \\ 0 \end{array}\right], \left[ \begin{array}{cc} (K_{\mathrm{fg}} + K_{\mathrm{21}}) + \sigma_n^2 I & K_{\mathrm{fg}} \\ K_{\mathrm{fg}} & K_{\mathrm{fg}} \end{array} \right] \right). \end{equation} Here again we use the shorthand $K \equiv K(\nu, \nu)$. We note that we use a GP prior with a zero mean function, which is common practice in Gaussian Process Regression~\citep{Rasmussen05,Gelman14} and allows the foregrounds to be fully defined by its covariance function. We tested the algorithm with a zero mean function and a polynomial parametric mean function and found the former to be a better choice for our application. The selection of a covariance function for the 21-cm signal can be done by comparison to a range of 21-cm signal simulations. In Fig.~\ref{}, we show the covariance as a function of frequency difference $\Delta\nu$ of a 21-cm signal, calculated with 21cmFAST~\citep{Mesinger11} when compared to the Matern $\eta = 1 / 2$ covariance functions for various values of the frequency coherence-scale $l$. For this particular set of simulations, the 21-cm signal can be well modeled using an exponential ($\eta = 1 /2$) kernel with a frequency coherence-scale ranging between 0.3 MHz and 1.2 MHz depending on the reionization stage. The foregrounds need to be modeled by a smoother function. The Gaussian kernel ($\eta = \infty$) yields very smooth models which might be unrealistic for modeling physical processes and a better alternative may be a Matern kernel with $\eta = 5/2$ or $\eta = 3/2$. Ultimately, the choice of the foreground covariance function is driven by the data in a Bayesian sense, by selecting the one that maximizes the evidence. Because the 21-cm signal is faint compared to the foregrounds and the noise, finding the correct hyper-parameters of the 21-cm signal would be close to impossible if this were done on each spatial line of sight individually. We therefore first optimize the LML for the full set of visibilities, assuming the frequency coherence scale is spatially invariant. This determines the covariance matrix structure that we then use to model the data for each spatial line of sight separately. This way we find that it is possible to perform much deeper modeling and reach the level of the 21-cm signal. After GPR, we retrieve the foregrounds part of the model: \begin{align} \label{} E(\mathbf{f}_{\mathrm{fg}}) &= K_{\mathrm{fg}}\left[K + \sigma_n^2 I\right]^{-1} \mathbf{d}\\ \label{} \mathrm{cov}(\mathbf{f}_{\mathrm{fg}}) &= K_{\mathrm{fg}} - K_{\mathrm{fg}}\left[K + \sigma_n^2 I\right]^{-1}K_{\mathrm{fg}}. \end{align} We are interested in estimating the residual after foregrounds are subtracted, \begin{equation} \label{} \mathbf{d}_{\mathrm{res}} =\mathbf{d} - E(\mathbf{f}_{\mathrm{fg}}). \end{equation} \section{Simulation} \label{} In this section, we describe the simulated astrophysical diffuse foregrounds, 21-cm EoR signal, instrumental mode-mixing contaminants and noise that are used to test the performance of the GPR foregrounds. Bright unresolved sources are not included in the simulation, assuming they can be properly modeled and subtracted from the data. \subsection{21-cm EoR signal} \label{subsec_simu_eor} We use the semi-analytic code 21cmFAST~\citep{Mesinger07,Mesinger11} to simulate 21-cm signal corresponding to the field-of-view of one LOFAR-HBA station beam. The code treats physical processes with approximate methods, and it is therefore computationally much less expensive than full radiative transfer simulations. The semi-analytic codes generally agree well with hydrodynamical simulations for comoving scales $> 1 {\rm Mpc}$ . We use the same 21-cm signal simulation as described in~\cite{Chapman12} and further used in~\cite{Ghosh15} and~\cite{Ghosh17}, which was initialized with $1800^{3}$ dark matter particles at z~=~300. The velocity fields were calculated on a grid of $450^{3}$ which was used to perturb the initial conditions and the simulation boxes of the 21-cm brightness temperature fluctuations. A minimum virial mass of $10^{9} \, M_{\odot}$ was defined for the halos contributing to ionizing photons. Once the evolved density, velocity and ionization fields have been obtained, 21cmFAST computes the $\delta T_{\rmn{b}}$ fluctuations at each redshift. For further details of the simulation, we refer the reader to~\citet{Chapman12}. Figure~\ref{} shows that to first order the 21-cm signal can be approximated and modeled by a Gaussian Process with an exponential covariance function, and that the frequency coherence-scale is a function of redshift i.e. of the stage of reionization. The coherence scale of fluctuations in frequency of the mode-mixing contaminants and of the 21-cm signal can affect the GPR method. To test this, we also generate 21-cm signal via a GP with an exponential kernel for which we vary the frequency coherence-scale $l_{\mathrm{21}}$ between 0.3 and 1.2 MHz. This range should cover a wide range of possible 21-cm signal models during the EoR. \subsection{Astrophysical diffuse foregrounds} \label{subsec_simu_mix} We use the foreground simulation from~\citet{Jelic08,Jelic10}. The Galactic foregrounds have three main contributions: \begin{enumerate}[label=\roman*),leftmargin=0.4cm,itemsep=0.2cm] \item The largest contribution (70\% around $100 - 200$ MHz) comes from the Galactic diffuse synchrotron emission (GDSE) due to the interaction of cosmic ray electrons with the galactic magnetic field. \item The next contribution is coming from synchrotron emission from extended sources, mostly supernova remnants (SNRs). \item The final component is the free-free radio emission from diffuse ionized gas which contributes roughly $1\%$ to the total Galactic foreground emission. \end{enumerate} The individual Galactic foreground components are modeled as Gaussian random fields. The GDSE is modeled as a power law as a function of frequency with a spectral index of $-2.55 \pm 0.1$~\citep{Shaver99} and $-2.15$ for the free-free emission. We have not included polarization of the foregrounds in our simulation. We also assume that point sources brighter than 0.1 mJy can be identified and accurately removed from the maps and therefore these sources are not included in the current diffuse foreground simulation~\citep{Jelic08}. Unresolved extragalactic sources were added to the simulation based on radio source counts at 151 MHz~\citep{Jackson05}. The simulated radio galaxies are clustered using a random walk algorithm. \begin{figure} \includegraphics[trim={1mm 2.5mm 0 0},clip]{materiel/signal_components.pdf} \caption{\label{} The different components of the simulated signal. The astrophysical diffuse emission (top-left panel), instrumental mode-mixing contaminants (top-right panel), 21-cm signal (bottom-left panel) and noise component (bottom-right panel) of a randomly selected visibility from the simulated cube is plotted as a function of frequency.} \end{figure} \subsection{Instrumental mode-mixing contaminants} The source of mode-mixing contaminants are manifold (Section~\ref{}). In essence, they are due to the combination of the instrument chromaticity and imperfect calibration. In the present paper, we will not attempt to simulate those effects, and we defer that to a future publication. Instead, we will simulate them using a Gaussian Process. This treatment is motivated by the analysis of LOFAR data which shows that these medium-scale fluctuations can be well modeled by a GP with a Matern covariance function, $\eta = 3/2$ and a coherence-scale $l_{\mathrm{mix}} \sim 2\ \mathrm{MHz}$\footnote{A more detailed description of LOFAR-HBA mode-mixing modeling will be given in a forthcoming publication. We refer the reader to~\cite{Patil16} and ~\cite{Patil17} for a recent analysis of this contaminants.}. In Section~\ref{} we will test GPR against others methods to generate mode-mixing contaminants using random polynomials and Matern kernel with different hyper-parameters for the different baselines. The mode-mixing are usually confined to a wedge-like structure in k space~\citep{Datta10,Morales12} (see Fig.~\ref{}). In the present publication, we do not simulate the $k_{\perp}$ dependence of the wedge and also defer this to future work. In fact, current assessments of the mode-mixing contaminants in LOFAR data tend to favor a baseline independent `brick' effect observed in~\cite{Ewall17}, which probably comes mainly from transferring the gain errors from longer to shorter baselines~\citep{Barry16, Patil16}. For the purpose of testing the impact of the `brick' extent, we simulate instrumental mode-mixing contaminants with frequency coherence-scale $l_{\mathrm{mix}}$ varying between 1 MHz and 8 MHz. \subsection{Noise} \label{subsec_simu_noise} In order to obtain realistic simulations of the noise, we first compute weights maps $W(u, v, \nu)$ which reflect the baseline distribution in the gridded uv plane. A noise visibility cube is created by filling it with random Gaussian noise for the real and imaginary part of the visibility separately with a noise standard deviation, \begin{equation} \sigma(u, v, \nu) = \frac{1}{\sqrt{W(u, v, \nu)}} \frac{\rm{SEFD(\nu)}}{ \sqrt{2 \, \Delta \nu\, \Delta t}}, \end{equation} where $\Delta \nu$ and $\Delta t$ are the frequency bandwidth and integration time, respectively, and the SEFD is the system equivalent flux density. We note that the SEFD is generally frequency dependent and varies across the sky. The SEFD depends largely on the sky temperature ($\rm{T_{sky}} \propto \nu^{-2.55}$) of the total sky brightness and the effective area of the LOFAR array ($\rm{A_{eff}}$). Here, we assume a constant SEFD $\sim 4000\, \mathrm{Jy}$~\citep{vanHaarlem13} over the simulated band-width, and assume a LOFAR-HBA data set of about 100 nights of 12 hour long observations. \subsection{Simulation cube} The simulation spans a frequency range of $132 - 148$ MHz with a spectral resolution of 0.2 MHz, i.e. a bandwidth of 16 MHz and 80 sub-bands, from which 12 MHz are used for power-spectra calculation centered around a redshift of $z \sim 9.1$. The maps cover a field of view of 6 degrees with a pixel size of 1.17 arc-minute. The mean value of the brightness temperature is subtracted to mimic a typical interferometric observation. The intrinsic foreground, mode-mixing, 21-cm signal and noise, respectively, of the simulation are converted into visibilities via a Fourier transform and added together to create an observation cube: \begin{equation} \begin{split} V_{\mathrm{obs}}(\mathbfit{u}, \nu) =\ & V_{\mathrm{sky}}(\mathbfit{u}, \nu) + V_{\mathrm{mix}}(\mathbfit{u}, \nu) + V_{\mathrm{21}}(\mathbfit{u}, \nu) + V_{\mathrm{n}}(\mathbfit{u}, \nu), \end{split} \end{equation} where $\mathbfit{u} = (u, v)$ is the vector representing the coordinates in wavelength in the uv plane and $\nu$ is the observing frequency. We restrict our analysis to the baseline range $50 - 250\,\lambda$ currently used by LOFAR~\citep{Patil17}. An example of these components are shown in Figure~\ref{} as a function of frequency. The distinct frequency-correlation is the characteristic exploited in the GPR method to separate these signals. We note that the signal separation (in this case foreground) method could be applied equally well to visibilities, image pixels, or spherical harmonics coefficients~\citep{Ghosh17}. The simulation cube is parametrized by four main parameters: \begin{description} [labelwidth=1.5cm,labelsep=1em,leftmargin=!,font=\normalfont,itemsep=0.2cm] \item[$\sigma_{\mathrm{21}} / \sigma_{\mathrm{n}}$] The ratio between the standard deviation of the 21-cm signal cube and the standard deviation of the noise cube, for the $50 - 250\,\lambda$ baselines range. This allows to test different reionization scenario while keeping the same noise level, and vice-versa. \item[$l_{\mathrm{21}}$] The frequency coherence-scale of the exponential covariance kernel in the case when a Gaussian Process is used to simulate the 21-cm signal. This parameter is ignored when 21cmFAST is used instead. \item[$\sigma_{\mathrm{mix}} / \sigma_{\mathrm{n}}$] The ratio between the standard deviation of the instrumental mode-mixing contaminants cube and the standard deviation of the noise cube, for the $50 - 250\,\lambda$ baselines range. \item[$l_{\mathrm{mix}}$] The frequency coherence-scale of the Matern covariance kernel used to simulate the instrumental mode-mixing contaminants. \end{description} \section{Results} \label{} In the following section, the Gaussian Process Regression (GPR) procedure described in Section~\ref{} is applied to the simulated datasets described in Section~\ref{}, in order to model and remove the foreground components, and subsequently compute the power spectrum of the 21-cm signal. Specifically, we apply the method on simulated cubes which reproduce the level of noise, mode-mixing contaminants, and foregrounds diffuse emission that we currently or theoretically can achieve with LOFAR, and subsequently explore various values of simulation parameters. \subsection{Recovering the 21-cm signal power spectra} \subsubsection{Foregrounds modeling and removal} The simulated foregrounds cube is composed of a frequency smooth sky signal and less smooth mode-mixing contaminants. We build this property into our GP covariance function by decomposing our foregrounds covariance into two separate parts, \begin{equation} K_{\mathrm{fg}} = K_{\mathrm{sky}} + K_{\mathrm{mix}} \end{equation} with `sky' denoting the intrinsic sky and `mix' denoting the mode-mixing contaminants. We use a Matern covariance function for all components of our data GP model. A Matern kernel has three hyper-parameters, $l$, $\sigma$ and $\eta$. The function becomes especially simple when $\eta$ is half integer~\citep{Rasmussen05}, which is why only discrete values of $\eta$ are used, $\eta \in (1 / 2, 3 / 2, 5 / 2, 7 / 2)$, choosing the best value based on the log-marginal-likelihood. This reduces the numbers of hyper-parameters to be optimized to six (two for each of the intrinsic sky, mode-mixing and 21-cm components of the GP model). We use the python package \textsc{GPy}\footnote{} to do the optimization using the full set of visibilities. This is done in two steps. We first use a uniform prior on the hyper-parameters and test different values of $\eta$, selecting the model that maximizes the evidence. A final run is then done with a more restricted range for the hyper-parameters. The foreground subtracted visibility is then obtained by computing the residual: \begin{equation} \label{} V_{\mathrm{res}}(\mathbfit{u}, \nu) = V_{\mathrm{obs}}(\mathbfit{u}, \nu) - V^{\mathrm{rec}}_{\mathrm{fg}}(\mathbfit{u}, \nu), \end{equation} where $V^{\mathrm{rec}}_{\mathrm{fg}}(\mathbfit{u}, \nu)$ is the maximum a-posteriori GPR foregrounds model. We recollect that for this particular set of simulations, the 21-cm signal was modeled using an exponential ($\eta = 1 /2$) kernel with a frequency coherence-scale ranging between 0.3 MHz and 1.2 MHz. For foregrounds we choose a Matern kernel with $\eta = 5/2$ or $\eta = 3/2$. Ultimately, the choice of the foreground covariance function is driven by the data in a Bayesian sense, by selecting the one that maximizes the evidence. Because the 21-cm signal is faint compared to the foregrounds and the noise, we therefore first optimize the LML for the full set of visibilities, assuming the frequency coherence scale is spatially invariant. In this way, we determine the covariance matrix structure that we then use to model the data for each spatial line of sight separately. In GPR, we retrieve the foregrounds part of the model first using Eqn.~\ref{} and the residuals were subsequently calculated using Eqn.~\ref{}. \subsubsection{Power spectrum estimation} Next, we determine the power spectra to quantify the scale dependent second moment of the signal by taking the Fourier transform of the various visibility cubes $V(\mathbfit{u}, \nu)$ in the frequency direction. We define the cylindrically averaged power spectrum as~\citep{Aaron12}: \begin{equation} P(k_{\perp}, k_{\parallel}) = \frac{X^2 Y}{\Omega_{\mathrm{PB}} B} \left<\left| \hat{V} (\mathbfit{u}, \tau)\right|^2\right>, \end{equation} where $\hat{V}(\mathbfit{u}, \tau)$ is the Fourier transform in the frequency direction, $B$ is the frequency bandwidth, $\Omega_{\mathrm{PB}}$ is the primary beam field of view, X and Y are conversion factors from angle and frequency to comoving distance, and $<..>$ denote the averaging over baselines. The Fourier modes are in units of inverse comoving distance and are given by~\citep{Morales06,Trott12}: \begin{align} &k_{\perp} = \frac{2 \pi |\mathbfit{u}|}{D_M(z)},\\ &k_{\parallel} = \frac{2 \pi H_0 \nu_{21} E(z)}{c(1+z)^2} \tau,\\ &k = \sqrt{k_{\perp}^2 + k_{\parallel}^2}, \end{align} where $D_M(z)$ is the transverse co-moving distance, $H_0$ is the Hubble constant, $\nu_ {21}$ is the frequency of the hyperfine transition, and $E(z)$ is the dimensionless Hubble parameter~\citep{Hogg10}. Finally, we average the power spectrum in spherical shells and define the spherically averaged dimensionless power spectrum as, \begin{equation} \Delta^2({k}) = \frac{k^3}{2 \pi^2} P(k). \end{equation} The recovered 21-cm signal power spectrum is obtained by subtracting the noise bias from the residual power spectra, derived from the residuals in Eq.~\ref{}. In general, the noise bias can be estimated with reasonable accuracy from the Stokes V image cube (circularly polarized sky), or by taking the difference between Stokes I data separated by a small frequency or time interval. The sky is only weakly circularly polarized and the Stokes V image cube is expected to provide a good estimator of the thermal noise. In our simulation the noise bias is estimated using the same noise cube used to generate the simulation cube. This ensures that the variance in the recovered 21-cm signal that we estimate are inherent to GPR and not due to thermal noise sampling variance limitations. \subsection{Application on the reference simulation} Our reference simulation is representative of the capability of LOFAR-HBA based on current observation of the noise and the level of mode-mixing errors. Specifically, the foregrounds data cube is composed of diffuse emission foreground and instrumental mode-mixing contaminants simulated using a Matern $\eta_{\mathrm{mix}} = 3 / 2$ covariance function with frequency coherence-scale of $l_{\mathrm{mix}} = 2~\mathrm{MHz}$ and a variance $(\sigma_{\mathrm{mix}} / \sigma_{\mathrm{n}})^2 =\,2$. The 21-cm signal is simulated from 21cmFAST with a variance $(\sigma_{\mathrm{21}} / \sigma_{\mathrm{n}})^2 = 0.007$. The noise realization corresponds to 1200 hours of LOFAR-HBA observations and a SEFD = 4000\,K. The input parameters of the reference simulation are summarized in Table~\ref{}. \subsubsection{Power spectrum results} \label{} We generate a total of 200 simulations, each with different noise and instrumental mode-mixing contaminants realizations, but with exactly the same astrophysical foregrounds and 21-cm signal. The power spectra of the different components are shown in Fig.~\ref{}. The top panel shows the spherically averaged power spectra. The intrinsic foregrounds are orders of magnitude brighter than the 21-cm signal on large scales (small $k$), but drop below the 21-cm signal at $k > 0.3 \ \rm{h\, cMpc^{-1}}$. While the mode-mixing component is only a small percent of the total power, it occupies a wider range of $k$ modes. This is better understood when looking at the cylindrically averaged power spectra as a function of $k_{\parallel}$ (bottom panel in Fig.~\ref{}); while most of the power of the intrinsic foregrounds is concentrated at low $k_{\parallel}$, the mode-mixing components still dominate the 21-cm signal at large $k_{\parallel}$, due to their smaller coherence in the frequency direction. This illustrates the importance of adding mode-mixing to any foreground removal strategy. We note that the $k$ mode at which the foreground power steeply decreases depends on the maximum baseline considered for the analysis. For this baseline configuration, we also note that a characterization of the power-spectra is theoretically possible for $k \le 0.3 \ \mathrm{h\, cMpc^{-1}}$ assuming perfect foreground removal and considering only the thermal noise uncertainty on the 21-cm signals (see also Fig.~\ref{}). The initial GPR runs with uniform priors on all hyper-parameters reveal that, in about 40\% of the cases, the 21-cm coherence-scale hyper-parameter $l_{\mathrm{21}}$ converges to the prior higher bound. A more informative prior can be used to solve this issue and better constrain $l_{\mathrm{21}}$. Figure~\ref{} shows that the simulated 21-cm signal coherence-scales range between about 0.3 and 1.2 MHz. A gamma distribution prior, thus honoring the positivity of the hyper-parameter, can then be used instead of the uniform prior with a variance broad enough such that it includes all probable values. The probability density function of the gamma distribution $\Gamma(\alpha, \beta)$, parametrized by the shape $\alpha$ and rate $\beta$, is defined as, \begin{equation} P(x| \alpha, \beta) = \frac{\beta^{\alpha} x^{\alpha - 1} e^{-\beta x}}{\gamma (\alpha)}, \end{equation} where $\gamma(\alpha)$ is the gamma function. For the hyper-parameter $l_{\mathrm{21}}$, we use the $\Gamma(3.6, 4.2)$ prior which is characterized by an expectation value of 0.85, a median value of 0.77, a 16th percentile value of 0.42 and an 84th percentile value of 1.29. To test the impact of this prior on the recovery of the 21-cm signal, we perform simulations similar to the ones described above but with the 21-cm signal simulated from a GP for which we know the true value of $l_{\mathrm{21}}$. We then compare the input value of $l_{\mathrm{21}}$ and the value estimated from the GPR. This shows that in case of a uniform prior, the values of $l_{\mathrm{21}}$ are not well estimated while, using a $\Gamma(3.6,4.2)$ prior, the estimated values of $l_{\mathrm{21}}$ are significantly less biased and have an uncertainty of $\sim 0.2$. We found that using this prior is only necessary because the reference simulation is characterized by a low signal-to-noise of the 21-cm signal and a low frequency coherence-scale of the mode-mixing component. The gamma prior helps in better separating the contributions from the mode-mixing and 21-cm signal. The initial GPR runs are also used to set the values of $\eta$ of the Matern covariance function for the different GP components. We find that the evidence is maximized using $\eta_{\mathrm{sky}} = 5 / 2$, $\eta_{\mathrm{mix}} = 3 / 2$ and, $\eta_{\mathrm{21}} = 1 / 2$. \begin{figure} \includegraphics{materiel/ps_fiducial_full_1col.pdf} \vspace{-0.15in} \caption{\label{} Detection of the EoR signal with the reference simulation. The top panel shows the spherically averaged power spectra. The central and bottom panels show the cylindrically averaged power spectra, averaged over $k_{\parallel}$ and over $k_{\perp}$ respectively. The simulated observed signal (dark blue) is composed of intrinsic astrophysical foregrounds (dotted dark blue), instrumental mode-mixing contaminants (dashed light blue), noise (green) and a simulated 21-cm signal (dashed gray). Using our GPR method to model and remove the foregrounds from the simulated cube, the 21-cm signal (orange) is well recovered with limited bias. The orange filled region represents the standard deviation of the recovered 21-cm signal over 200 simulated cubes.} \end{figure} Having found the most probable settings of GP model and hyper-parameters priors, we perform a final GPR on each of the simulated cubes. Figure~\ref{} shows the power spectra of the recovered 21-cm signal compared to the input cosmological signal power spectra. The orange filled region represents the standard deviation of the recovered signal over the 200 simulated cubes, the line corresponding to the mean. This provides an estimate respectively of the variance and the bias of the method. The bias is overall limited but is more pronounced at low $k$ modes. It is maximum at $k = 0.073 \ \rm{h\, cMpc^{-1}}$ where we have a bias equal to 86\% of the uncertainty. The variance is almost always similar or below, on the $k$ modes probed, the thermal noise limit. We however find it to be 30\% greater at $k = 0.18 \ \rm{h\, cMpc^{-1}}$. We recall that the noise bias is estimated using the same noise cube used in the simulated cube. Hence, the variance that we estimate is inherent to GPR and does not include thermal noise sampling variance. Investigating the cylindrically averaged power spectra reveals that most of the bias of the current implementation of the GPR method is introduced because of the one-dimensional fit to the data in the frequency direction. The power spectra as a function of $k_ {\parallel}$ (bottom panel of Figure~\ref{}) show an excellent correspondence between the input and recovered signal with small uncertainty. On the contrary, the power spectra as a function of $k_{\perp}$ (central panel of Figure~\ref{}) show a much larger bias and uncertainty. The method is capable of retaining the correct variance in the frequency direction but not so well in the baseline direction. This is explained by the fact that the regression is currently only done in the frequency direction and assumes that the frequency coherence-scale of the different components is spatially invariant. In Section~\ref{} we explore various improvements to the method that may be implemented to reduce the bias and uncertainty. Nevertheless, current results already demonstrate that the approach is able to achieve a reliable first measurement of the 21-cm signal and an initial characterization of its power spectra in 1200 hours of LOFAR observations. \subsubsection{Estimating the model hyper-parameter uncertainties} The maximum a-posteriori (MAP) solution of the model hyper-parameters is evaluated through an optimization algorithm, using the analytically defined likelihood function (Eq.~\ref{}). However, to fully sample the posterior distribution of the hyper-parameter, characterize its topology, and analyse the correlations between parameters, we resort to Monte Carlo Markov Chain (MCMC). An MCMC method samples the posterior probability distribution of the model parameters given the observed data. We use an ensemble sampler algorithm based on the affine-invariant sampling algorithm~\citep{Goodman10}, as implemented in the \textsc{emcee} python package\footnote{}. Figure~\ref{} shows the resulting posterior probability distribution of the GP model hyper-parameters. We find that the input values are always inside the 68\% confidence interval. The hyper-parameters of the mode-mixing covariance function are very well constrained. The confidence interval on the 21-cm signal kernel hyper-parameters are relatively larger, because in this particular simulation the 21-cm signal is an order of magnitude fainter than the noise. The parameter estimates and confidence intervals are summarized in Table~\ref{}, along with their input values and associated priors. We note that for this setup the 21-cm signal has no input $l_{\mathrm{21}}$ because it was simulated using 21cmFAST. \begin{figure} \includegraphics[trim={0 4mm 0 0}]{materiel/post_prob_main_simu.pdf} \centering \caption{\label{} Posterior probability distributions of the GP model hyper-parameters for the reference simulation. We show here the coherence-scale and strength of the EoR covariance function ($l_{\mathrm{21}}$ in MHz and $\sigma_{\mathrm{21}}$), and the coherence-scale and strength of the mode-mixing foreground kernel ($l_{ \mathrm{mix}}$ in MHz and $\sigma_{\mathrm{mix}}$). The input parameters of the simulation are marked in blue. The orange contours show the 68\%and 95\% confidence interval. We note that the PDFs are all narrower than their priors.} \end{figure} \renewcommand{\arraystretch}{1.4} \begin{table} \centering \caption{Summary of the input parameters of the reference simulation and estimate on the median and confidence interval of there respective GP model hyper-parameters obtained using an MCMC method. The input intrinsic sky is simulated using astrophysical foreground simulation from~\protect\cite{Jelic08} while the 21-cm signal is simulated from 21cmFAST~\protect\citep{Mesinger11}.} \begin{threeparttable} \label{} \begin{tabularx}{0.45 \textwidth}{XXXX} \toprule & Input & Prior & Estimate \\ \midrule $\sigma_{\mathrm{sky}} / \sigma_{\mathrm{n}}$ & --& $\mathcal{U}(30, 45)$ & $37.4 \substack{+0.4\\ -0.4}$ \\ $l_{\mathrm{sky}}$ [MHz] & -- & $\mathcal{U}(60, 100)$ & $80.1 \substack{+1.2\\ -1.2}$ \\ \midrule $\sigma_{\mathrm{mix}} / \sigma_{\mathrm{n}}$ & 1.478 & $\mathcal{U}(1, 2)$ & $1.47 \substack{+0.01\\ -0.01}$ \\ $l_{\mathrm{mix}}$ [MHz] & 2 & $\mathcal{U}(1.5, 2.5)$ & $2.01 \substack{+0.02\\ -0.02}$ \\ \midrule $\sigma_{\mathrm{21}} / \sigma_{\mathrm{n}}$ & 0.083 & $\mathcal{U} (0.002, 0.25)$ & $0.11 \substack{+0.03\\ -0.04}$\\ $l_{\mathrm{21}}$ [MHz] & -- & $\Gamma(3.6, 4.2)$ & $0.90 \substack{+0.05\\-0.04}$ \\ \bottomrule \end{tabularx} \end{threeparttable} \end{table} \begin{figure} \includegraphics[trim={2mm 4mm 0 0},clip] {materiel/ps3d_fg_err_gpr_vs_gmca.pdf} \caption{\label{} Spherically averaged power spectra of the foreground modeling error using the GPR and GMCA method. With GPR (blue line) the foreground error is at the level of the 21-cm signal (dashed black line) and is close to the thermal noise uncertainty (plain black line) which is the inherent statistical error level we could achieve, while the foreground error with GMCA is at the level of the noise (dotted black line).} \end{figure} \begin{figure*} \includegraphics{materiel/ratio_all_+th_zero_1row.pdf} \caption{\label{} Fractional bias of the recovered 21-cm signal ($r_{\mathrm{rec}}$) with varying coherence-scale of the 21-cm signal ($l_{\mathrm{21}}$, left panel), coherence-scale of the mode-mixing contaminants ($l_{\mathrm{mix}}$, central panel) and strength of the 21-cm signal ($\sigma_{\mathrm{21}} / \sigma_{\mathrm{n}}$, right panel), for different k ranges. We show the mean (plain line) and standard deviation (filled area) of the fractional bias calculated from a total of 3000 simulations, giving an estimate of the bias and uncertainty, respectively, introduced by the method. GPR performance is optimal for $l_{\mathrm{mix}} > 3.0$ MHz, $\sigma_{\mathrm{21}} \gtrsim 0.1~\sigma_{\mathrm{n}}$ and for $0.6~\mathrm{MHz} < l_{\mathrm{21}} < 1~\mathrm{MHz}$. The vertical dashed lines represent the nominal values around which we vary the parameters.} \end{figure*} \subsubsection{Comparison between GPR and GMCA} Next, we compare GPR to another well-tested foreground removal method. From the currently available algorithms, the Generalized Morphological Component Analysis (GMCA)~\citep{Bobin07,Chapman13} is the one that has demonstrated the best results~\citep{Chapman15,Ghosh17}. We use the python based toolbox \textsc{pyGMCALab}\footnote{} and run the algorithm on our simulated cubes. We model the foregrounds by the minimum numbers of components that minimize the overall fitting error. An optimal eight components are used to represent the foregrounds. We then compare the power spectra of the foreground modeling error when using GPR and GMCA. Figure~\ref{} shows that GMCA has difficulty to correctly model the complex mode-mixing contaminants and does not reach a level of modeling error better than the noise for $k \leq 0.3 \ \rm{h\, cMpc^{-1}}$. Using GPR, we improve these results by an order of magnitude, and this allows us to achieve an error in the foreground power spectra that is at or below the 21-cm signal power spectrum. We also note that this level is similar to the thermal noise uncertainty which is the ultimate error level we can achieve. \subsection{Performance of the GPR method} \subsubsection{Exploring the input parameter space} The efficiency of a foreground removal algorithm depends on the characteristics of the foregrounds and of the 21-cm signal. To explore the performance of GPR in terms of bias and variance, we explore the input parameters of the simulated cube, varying one parameter at a time. As a quality criterion, we use the fractional bias of the recovered spherically averaged 21-cm signal power spectra, \begin{equation} r_{\mathrm{rec}}(k) = \frac{\Delta_{\mathrm{rec}}^2(k) - \Delta_{\mathrm{21}}^2 (k)}{\Delta_{\mathrm{21}}^2(k)}. \end{equation} where $\Delta_{\mathrm{rec}}^2(k)$ is the GPR recovered power spectrum, and $\Delta_{\mathrm{21}}^2(k)$ is the power spectrum of the input 21-cm signal. For these tests we build simulation cubes with central parameters $\sigma_{ \mathrm{mix}} = 1.478~\sigma_{\mathrm{n}}$, $l_{\mathrm{mix}} = 3~ \mathrm{MHz}$, $\sigma_{\mathrm{21}} = 0.12~\sigma_{\mathrm{n}}$ and $l_{\mathrm{21}} = 0.75~\mathrm{MHz}$ around which we vary the parameters. We use a Gaussian Process with an exponential covariance function (see Section~\ref{subsec_simu_eor}) to generate 21-cm signals such that we can control the frequency correlation of the signal (i.e. $l_{\mathrm{21}}$). A total of 3000 simulations with different realizations of the noise, 21-cm signal, and mode-mixing contaminants are generated. We determine the relative difference between recovered and input power spectra for different $k$ bins and compute its mean and standard deviation\footnote{We note that the distribution of $r_{\mathrm{rec}}$ is actually not Gaussian, being the ratio of two distributions, but the mean and standard deviation were found to be appropriate enough to characterize this distribution.} over the full set of simulated cubes (Fig.~\ref{}). This provide us with an estimate of the fractional bias and uncertainty introduced by the method. We also compare the later to the minimal uncertainty due to thermal noise. By varying the strength of the 21-cm signal, we find that the bias is limited (below 35\%) for the full range of the investigated values and falls below 20\% for $\sigma_{\mathrm{21}} \geq 0.12~\sigma_{\mathrm{n}}$. The uncertainty and bias increase with lower S/N as expected, and we find it to be significantly higher than the thermal noise uncertainty for $\sigma_{\mathrm{21}} \lesssim 0.1~\sigma_{\mathrm{n}}$. Varying the frequency coherence-scale of the mode-mixing contaminants, we also find limited bias and a small increase of the uncertainty at low $l_{\mathrm{mix}}$. As $l_{\mathrm{mix}}$ approaches that of $l_{\mathrm{21}}$, it becomes increasingly more difficult to statistically differentiate the two signals. This is the reason why the uncertainty increases for values $l_{\mathrm{mix}} < 3$ MHz. A decrease in the value of $l_{\mathrm{mix}}$ also corresponds to increasing the extent of the foreground wedge (or `brick'), and equivalently reducing the EoR window. Varying the frequency coherence-scale of the 21-cm signal, we find that some bias is introduced at small and large $l_{\mathrm{21}}$, related to the use of a Gamma prior to this GP hyper-parameters. Overall, GPR is limited in situation of very low S/N and/or when the foregrounds start to mix with the 21-cm signal. In most situations it performs relatively well, with limited bias and uncertainty level on par with the thermal noise uncertainty. \begin{figure} \includegraphics[trim={1mm 1.5mm 0 0},clip] {materiel/detection_confidence_interval.pdf} \caption{\label{} Detection confidence interval for the reference simulation as a function of the S/N of the 21-cm signal $\sigma_{\mathrm{21}} / \sigma_{\mathrm{n}}$. The measured confidence levels (blue points) are fitted using an inverse function (blue line). The dashed gray lines show the 95\% and 99.7\% confidence level.} \end{figure} \subsubsection{Detection confidence level} We define the detection confidence level as the probability that the model is preferred (i.e. the evidence is maximal) if it contains a 21-cm signal component compared to one that does not. In GPR, the evidence as a function of the hyper-parameters $\theta$ is analytically defined (Eq.~\ref{}) and can be efficiently estimated for the optimal values of $\theta$. We note that comparing this maximum evidence for two different covariance structures parameterized by different numbers of hyper-parameters does not usually provide definitive answer on which kernel is the most suitable to model the data, especially if the difference of the evidences is small~\citep{Fischer16,Rasmussen05}. Nevertheless, this criterion is fast to compute and can still provide informative approximation on the confidence level of the detection. To determine it as a function of S/N of the 21-cm signal, we generate new reference simulations, varying now the input 21-cm signal strength $\sigma_{\mathrm{21}}$. We use Eq.~\ref{} to compute the evidence for the optimal values of the hyper-parameters $\theta$. In Figure~\ref{}, we show the detection confidence level as a function of the input 21-cm signal $\sigma_{\mathrm{21}} /\sigma_{\mathrm{n}}$, calculated using a total of 3000 simulations. A 95\% and 99.7\% detection confidence level is observed for $\sigma_{ \mathrm{21}} \gtrsim 0.09\ \sigma_{\mathrm{n}}$ and $\sigma_{ \mathrm{21}} \gtrsim\ 0.12 \sigma_{\mathrm{n}}$ respectively, rapidly increasing with signal to noise. The above calculation is obtained using the expression of the evidence from Eq.~\ref{} which is a function of the hyper-parameters $\theta$. A more robust way to compare the models is to estimate the evidence values integrated over the hyper-parameters and take their ratio, also called the Bayes factor. This is generally much more computationally expensive, and we only perform this test, as a confirmation of the above results, for a limited number of cases. We compute the evidence with an implementation of the nested sampling algorithm of~\cite{Mukherjee06}. For $\sigma_{\mathrm{21}} = 0.083\,\sigma_{ \mathrm{n}}$ (i.e., the reference simulation), we obtain Bayes factors ranging between 3.8 and 19 corresponding to a `substantial' to `strong' strength of evidence according to the scale of~\cite{Harold61}. For $\sigma_{\mathrm{21}} = 0.12\,\sigma_{\mathrm{n}}$, we obtain Bayes factors ranging between 5.2 and 55 corresponding to a `substantial' to `very strong' strength of evidence. Finally, for $\sigma_{\mathrm{21}} = 0.2\,\sigma_{ \mathrm{n}}$, we obtain a Bayes factors ranging between 328 and $1.9 \times 10^4$ corresponding to a `decisive' strength of evidence. \begin{figure*} \includegraphics{materiel/ps2d_rnd_fg_med.pdf} \includegraphics{materiel/ratio_rnd_poly_sum_rnd_rbf.pdf} \caption{\label{} Fractional bias of the recovered 21-cm signal ($r_{\mathrm{rec}}$) for mode-mixing contaminants generated using random polynomials with maximum order $o_{\mathrm{max}}$ (left), GP with random coherence-scale selected in the range $l^{\mathrm{max}}_{\mathrm{mix}} - l^{\mathrm{min}}_{\mathrm{mix}}$ with $l^{\mathrm{min}}_{\mathrm{mix}} = 3$ (middle), and GP with decreasing coherence-scale as function of baseline length with $l^{\mathrm{50}}_{\mathrm{mix}} = 6\,\mathrm{MHz}$ (right). The top panel shows the 2D cylindrically averaged power spectra of the different mode-mixing contaminants for the most extreme tested scenarios (the axis are in log scale). The bottom panel show the mean (plain line) and standard deviation (filled area) of the fractional bias calculated from a total of 1000 simulations.} \end{figure*} We note that these estimates are only for a single frequency bandwidth of 12MHz, and that usually several redshift bins are combined which will increase the overall confidence level on the detection of the 21-cm signal. \subsection{Testing different methods of simulating mode-mixing} \label{} In this sub-section, we test the versatility of GPR against alternative form of mode-mixing contaminants. In previous simulations, we used a Matern kernel with fixed coherence-scale. We now perform similar simulation with three others methods to generate the instrumental mode-mixing components. The simulation cubes are generated with parameters $\sigma_{ \mathrm{mix}} = 1.478~\sigma_{\mathrm{n}}$, $\sigma_{\mathrm{21}} = 0.12~\sigma_{\mathrm{n}}$ and $l_{\mathrm{21}} = 0.75~\mathrm{MHz}$. \subsubsection{Random polynomial} We generate mode-mixing visibilities using polynomial functions of random order taken in the range $3 - o_ {\mathrm{max}}$ and random coefficients. Applying GPR to this simulation shows that this component is best modeled (i.e. the evidence is maximized) using a Matern covariance function with $\nu = \infty$ (equivalent to a Gaussian covariance kernel). The results of this test are shown in the left panel of Fig.~\ref{}. The measured bias is minimal for all tested cases. \subsubsection{Random coherence-scale} We now generate mode-mixing visibilities using a Matern kernel and randomly selected coherence-scale $l_{\mathrm{mix}}$ in the range $3 - l^{\mathrm{max}}_{ \mathrm{mix}}\,\mathrm{MHz}$ for each different visibilities modes $ \mathbfit{u}$. For this test we set $\nu = \infty$. Running GPR on this simulation shows that the mode-mixing component is best modeled by a Rational Quadratic covariance function which is defined as: \begin{equation} \label{} \kappa_{\mathrm{RQ}}(x_p, x_q) = \left(1 + \frac{|x_q - x_p|^2}{2 \alpha l}\right)^{-\alpha}, \end{equation} and can be seen as an infinite sum of Gaussian covariance functions with different characteristic coherence-scales~\citep{Rasmussen05}. The results of this test is shown in the middle panel of Fig.~\ref{}. We again find limited bias which is also independent of the range of coherence-scales. \subsubsection{Decreasing coherence-scale} A wedge-like feature can be simulated by generating mode-mixing visibilities with decreasing coherence-scale as a function of baseline. For this test, we use a Matern kernel with $\nu = \infty$, and a coherence-scale that is linearly decreasing as a function of baseline with the coherence-scale of the 50 lambda baselines $l^{\mathrm{50}}_{\mathrm{mix}} = 6\,\mathrm{MHz}$, and the coherence-scale of the 250 lambda baselines $l^{\mathrm{250}}_{\mathrm{mix}}$ taken in the range $3 - 5.5\, \mathrm{MHz}$. The result of this test is shown in the right panel of Fig.~\ref{}. It shows that an increase of the bias with increasing range of coherence-scales. The maximum bias is nevertheless limited to about 30\%. In the future, we will implement the ability of GPR to perform a fit of the hyper-parameters with different coherence-scale for different baselines ranges, which should reduce further this bias. \section{Discussion and Conclusion} \label{} In this paper, we have introduced a novel signal separation method for Epoch of Reionization and Cosmic Dawn experiments. The method uses Gaussian Process Regression (GPR) to model various mixed components of the observed signal, including the spectrally smooth sky, mode-mixing associated with the instrument chromaticity and imperfect calibration, and a 21-cm signal model. Including covariance functions for each of these components in the GPR ensures a relatively unbiased separation of their contribution and accurate uncertainty estimation, even in very low signal to noise observations. In building the GP model, we make use of prior information about the different components of the signal. This makes the method very useful in the initial diagnostic and analysis stage of the data processing as it allows one to get a better insight into the data in terms of potential contaminants (i.e. mode-mixing but also the ionosphere). Additionally, GPR is flexible, and the GP model can be easily adapted to integrate new systematics. Cable reflexion for example could be easily modeled in this framework, adding a periodic covariance function component to the model. GPR is shown to accurately model the foreground contaminants including instrumental mode-mixing which have proven to be an Achilles heel of current foreground removal algorithms. When applied to simulation datasets, equivalent to LOFAR 1200 hours of observations and based on its current assessment of noise and systematic errors, GPR limits biasing the 21-cm signal, and recovers the input power spectrum well across the whole $k$ range $0.07 - 0.3 \ \rm{h\, cMpc^{-1}}$. When compared to GMCA, we find that GPR decreases the uncertainty on the recovered 21-cm signal power spectra by an order of magnitude, in the presence of mode-mixing. Exploring the performance of GPR using a range of different foregrounds and EoR signal, we find an optimal recovery for $l_{\mathrm{mix}} \geq 3$ MHz, $\sigma_{\mathrm{21}} \gtrsim 0.12~\sigma_{\mathrm{n}}$, with fractional bias below 20\% and with at least a 3 sigma confidence level on the detection. Outside this range, the detectability of the signal is still adequate, but with larger bias and larger uncertainty. These values hold for a single frequency bandwidth of 12 MHZ, and combining several redshift bins will improve the confidence limit on the detection. They partially depend as well on the observation configuration, such as uv-coverage, the lower and upper baseline limit, the FoV, and so they are most representatives for LOFAR-HBA in 1200 hours of observations. The fundamental improvement of GPR resides in its complete statistical description of all components contributing to the observed signal. In its current implementation\footnote{The code implementing the algorithm described in this paper is freely available at \url{}}, we use a generic model for the 21-cm signal and mode-mixing components which only make use of our prior knowledge on the frequency dependence of the signals. While this treatment may be sufficient for a detection of the 21-cm signal and its characterization with LOFAR, an improved model may be built for future experiments with e.g. the more sensitive SKA. The mode-mixing model for example can be improved by integrating the $k_{\perp}$ dependency of the foreground wedge and folding into the model the analytic work describing the effect on the signal of the instrumental chromaticity, calibration errors and sky-model incompleteness. Exploiting the isotropic nature of the 21-cm signal and its evolution at different redshift-bins will also ensure a more sensitive and accurate modeling. Finally, in the course of determining the physical 21-cm signal parameters from the 21-cm signal power spectra using for example an MCMC sampler~\citep{Greig15,Kern17}, the GPR bias could be determined and integrated at each MCMC steps. \section*{Acknowledgements} FGM and LVEK would like to acknowledge support from a SKA-NL Roadmap grant from the Dutch ministry of OCW. AG acknowledge Postdoctoral Fellowship from the South African Square Kilometer Array, South Africa (SKA-SA) for financial support. We would also like to thanks the anonymous referee for the careful review that improved the quality of this paper. FGM thanks M. Mevius, A. Offringa, M. Sardarabadi, and S. Zaroubi for helpful discussions. \bibliographystyle{mnras}

\section{Introduction} \label{intro} The discovery of the Kondo effect \cite{hewson} in semiconducting quantum dots (QDs),\cite{gold,cro,gold2,wiel,grobis,kreti,ama,hubel,keller} has spurred the study of electronic transport through QDs. Later molecular QDs,\cite{liang,kuba,yu,leuen,parks,roch,scott,parks2,NatelsonReview,serge,vincent} and QDs in carbon nanotubes,\cite{jari,maka,anders} were also studied. Molecular QDs in general have large asymmetric coupling to source and drain leads. Semiconducting QDs are characterized by the enormous possibilities for tuning the different parameters. In all these systems, several physical effects are generically observed when the system is cooled at cryogenic temperatures due to the large Coulomb repulsion in these nanoscopic QDs, such as Coulomb blockade and the Kondo effect, which implies a resonance at the Fermi energy in the spectral density of the dot state, that leads to an anomalous peak in the differential conductance $G(V)=dI/dV$ at zero bias voltage $V$, where $I$ is the current through the QD. These physical effects are usually well described by an impurity Anderson model (IAM). While intense research has been devoted to the Kondo effect, also important effects of the interactions take place at finite $V$ that are unexpected for independent electrons. In particular, Coulomb blockade peaks were shown to present a strong width renormalization in the situation of large tunnel asymmetries between source and drain electrodes.\cite{haug} Specifically, \"{o}nemann \textit{et al.} started from an equilibrium situation in which the dot level $E_{d}>0$. Then they applied a bias voltage, which shifts the chemical potential $\mu _{L}$ of one of the leads (we call it the left lead) in $eV>0$ and also $E_{d}$ by about $eV/2$ due to capacitance effects. A peak in $G(V)$ is observed when $\mu _{L}\sim E_{d}$. When instead a bias voltage such that $eV<0$ is applied, another peak in $G(V)$ was observed (when $E_{d}\sim \mu _{R}$, the chemical potential of the right lead). In a noninteracting picture, one would expect that both peaks have nearly the same width and the same height.\cite{park,capac} However, the latter turns out to be about two times wider and with a maximum nearly five times smaller.\cite{haug} The reason is that the QD has a much stronger coupling to the right lead, and then for $eV>0$ the QD is empty, while for $eV<0$ it is in an intermediate valence situation, and one knows that the half width at half maximum $\Delta _{\text{CT}}$ of the spectral density of the charge-transfer (CT) peak in the IAM increases with the occupancy.\cite{capac,pru,logan} An explanation of the experiment in the framework of the SU(2) IAM, including the effects of capacitance and tunneling asymmetries and the Kondo effect is provided in Ref. \onlinecite{capac}. Several tunable systems of the kind discussed above in which there is an orbital, dot or valley degeneracy, in addition to the spin one, are described by the SU(4) IAM, like those of carbon nanotubes,\cite{jari,maka,anders} a nanoscale Si transistor,\cite{tetta} an As dopant in a Si nanostructure,\cite{lans} and systems of two QDs where the occupation of one or the other QD plays the role of the orbital degeneracy.\cite{hubel,keller} In the latter case, the tunneling coupling of both dots are in general different, reducing the symmetry to SU(2), but tuning other parameters the SU(4) symmetry can be recovered as an emergent one at low temperatures.\cite{restor,nishi}. Crossovers \cite{tetta,buss,desint,buss2,thermo,nishi2,oks,lopes} and abrupt transitions \cite{klee} between SU(2) and SU(4) symmetry were also studied. In the noninteracting case, for a flat very wide conduction band as we assume, the spectral density of the localized level for each of the N components in the SU($N$) impurity Anderson model is a Lorentzian with half width at half maximum $\Delta$ [defined by Eq. \ref{delta} in terms of the microscopic parameters of the model]. Thus in this case, $\Delta _{\text{CT}}=\Delta$ independently of on site energy $E_d$. As we show in this work, the effect of correlations on $\Delta _{\text{CT}}$ increases with N and one expects a width $\Delta _{\text{CT}}=N\Delta $ for the SU($N$) model in the Kondo regime $-E_d, E_d+U \gg \Delta$, where $U$ is the Coulomb repulsion. Therefore if the experiments like those of K\"{o}nemann \textit{et al.} are performed for any of the above realizations of the SU(4) IAM, this effect would be more evident. Furthermore for the experiments with the setup of two QDs in which each QD is connected to its own pair of conducting leads and the tunneling matrix elements and the voltages at the four leads can be controlled independently with high precision,\cite{ama,keller} the spectral density can be read out directly form the differential conductance through one of the QD under appropriate conditions.\cite{oks} In this work we use a 1/N expansion based on variational wave functions,\cite{varma,gunn} to show that one expects in general that $\Delta_{\text{CT}} \sim N \Delta$ for $-E_d \gg N \Delta$ in the SU($N$) IAM. We also calculate $\Delta_{\text{CT}}$ as a function of $E_d$ for N=4 using dynamical density-matrix renormalization group (DDMRG) and the noncrossing-approximation (NCA). In this case $\Delta_{\text{CT}}$ changes rather abruptly from $4 \Delta$ to $\Delta$ as the effective $E_d$ (including renormalization due to effects of the hybridization \cite{hald}) changes sign from negative to positive. We also discuss the possible experimental relevance of these results. The paper is organized as follows. In Sec. \ref{model}, we describe the SU($N$) IAM and the methods used. In Sec. \ref{vari} we discuss the CT peak for general $N$ at zero temperature in the Kondo regime $-E_{d}\gg N\Delta $ using a variational approximation. Sec. \ref{su4} contains NCA and DDMRG results for $N$=4. In Sec. \ref{occ} we show the results for the occupancy as a function of on-site energy level and compare them with the noninteracting case and with exact Bethe ansatz results. In Sec. \ref{exp} we discuss the relevance of our results to possible experimental realizations. Sec. \ref{sum} contains a summary and a discussion. \section{Model and methods} \label{model} We consider the one-level SU($N$) IAM with infinite on-site repulsion $U$. The impurity states involve a singlet configuration $|s\rangle $ together with a degenerate configuration $|m\rangle $, $m=1$ to N, corresponding to one additional electron (or hole) in the \textquotedblleft impurity\textquotedblright , which can be a QD, a system of two QDs, a part of a carbon nanotube, or an atom with degenerate levels as discussed in the introduction. When discussing transport experiments, for simplicity we assume that the \textquotedblleft impurity\textquotedblright\ is connected to a left ($L$) and a right ($R$) conducting lead. An extension to the case of the system of two QDs in which both are connected to a pair of independent leads \cite{ama,hubel,keller} is straightforward.\cite{oks} The Hamiltonian reads: \begin{eqnarray} H &=&\sum_{m}E_{d}|m\rangle \langle m|+\sum_{\nu km}\epsilon _{\nu k}c_{\nu km}^{\dagger }c_{\nu km} \notag \\ &&+\sum_{\nu km}(V_{k}^{\nu }|m\rangle \langle s|c_{\nu km}+\mathrm{H.c}.), \label{ham} \end{eqnarray} where the constraint $|s\rangle \langle s|+\sum_{m}|m\rangle \langle m|=1$ is imposed (this is equivalent to the assumption of infinite Coulomb repulsion $U$). Here $c_{\nu km}^{\dagger }$ create conduction states in the lead $\nu $ with projection $m$ and wave vector $k$. The tunnel couplings of the quantum dot to the leads and the total resonant level widths are \begin{eqnarray} \Delta _{\nu }&=&\pi \sum_{k}|V_{k}^{\nu}|^{2}\delta (\omega -\epsilon _{\nu k}), \notag \\ \Delta &=&\Delta _{L}+\Delta _{R} \label{delta} \end{eqnarray} taken in general independent of energy $\omega $.\cite{note} For the discussion of the impurity spectral density $\rho _{m}(\omega )$, only $\Delta $ is relevant To obtain $\rho _{m}(\omega )$ we use three different methods. The variational one, based on Refs. \onlinecite{varma,gunn} is the simplest one and is described in detail in the next Section. It is limited to the Kondo regime $-E_{d}\gg N\Delta $ and zero temperature. We also use NCA and DDMRG. The NCA is equivalent to a sum of an infinite series of diagrams in perturbations in the hybridization.\cite{hewson,bickers,kroha} In the Kondo regime, it is known to reproduce correctly the relevant energy scale $T_{K}$ and its dependence of parameters. An advantage for our purpose over the numerical-renormalization group in which finite-energy features are artificially broadened due to the logarithmic discretization of the conducting band,\cite{zitko,loig} NCA correctly describes these features. For instance, the NCA works satisfactorily in cases in which the conduction density of states is not smooth,\cite{kroha} including in particular a step in the conduction band.\cite{ds} Furthermore, it has a natural extension to non-equilibrium conditions,\cite{wingreen} and it is specially suitable for describing satellite peaks of the Kondo resonance, as those observed in Ce systems,\cite{reinert,ehm} or away from zero bias voltage in non-equilibrium transport due to phonons \cite{fon,sate} or magnetic and orbital excitations.\cite{restor,NFL,st} Alternatives to NCA for non-equilibrium problems are renormalized perturbation theory (but limited to small $\omega $, $V$ and temperature $T$)\cite{hbo,ng,com3} or the equation of motion method,\cite{rome1,rapha} although it does nor reproduce correctly the functional dependence of $T_{K}$ on $E_{d}$.\cite{rapha,rome2} However, the NCA has important limitations out of the Kondo regime. In particular for moderate positive $E_{d}$, the impurity self energy has an unphysical positive imaginary part (this means that the Green function violates causality) and the impurity spectral density presents a spurious peak at the Fermi energy. As a consequence, the NCA results in this region of $E_{d}$ are unreliable. For this reason, we also use the DDMRG in its correction-vector-method approach. Since it was introduced by Kuehner and White,\cite{kuhne99a} the correction vector has shown to be a reliable way to do dynamical calculations with DMRG in different low-dimensional strongly correlated models. Different strategies were done to include the correction vector in the target states of a standard DDMRG.\cite{hovel00,freun93a,jecke02} We choose a recently presented version introduced by Nocera and Alvarez,\cite{alva} which based on a Krylov-space approach, has been shown to be more accurate and efficient than conjugate gradient.\cite{jecke02} However, the two major difficulties of the correction-vector method persist. First, the need to be computed in small frequency intervals which is unavoidable but with parallelization strategies becomes affordable. Second the artificial broadening $\eta$ that is necessarily introduced in the calculation (calculations cannot be done at $\eta=0$). Therefore one is computing $G(\omega+i\eta)$, and the resulting impurity density spectral function $\rho^{(\eta)}(\omega) = - \frac{1}{\pi} {\text{Im}} G(\omega+i\eta)$ can be visualized as the real spectral density $\rho(\omega)$ convoluted by the Lorentzian $\rho_d(\omega) =(\eta/\pi)/((\omega-\omega_0)^2+\eta^2)$ of width $\eta$. Since we are interested in the real spectral density for $\eta \rightarrow 0^+$, this compels to a deconvolution of the spectrum, and many proposals have been presented to do this difficult task successfully,\cite{Gebhard2003,Ulbricht2010,Weichselbaum2009,Nishimoto2004,Raas2004,Raas2005,Paech1014} and resolve structures with small width in the real spectrum. Fortunately, as it was mentioned above, we are interested in the line shapes of the CT peak for the IAM which as it has been very well described in Ref. \onlinecite{Raas2004}, the deconvolution becomes easiest, particularly for large $U$. The authors have also shown that the noninteracting case $U=0$ is well described by the DDMRG.\cite{Raas2004} Assuming that the original CT peak is a Lorentzian, then the data obtained with DDMRG will be the convolution of two Lorentzians of half width at half maximum $\Delta_{\text{CT}}$ and $\eta$ respectively. This convolution will result in a new Lorentzian whose half width is the sum of the two original ones, so that the half width at half maximum of the peak obtained by fitting the DDMRG calculations will be $\eta_{\text{eff}}=\Delta_{\text{CT}}+\eta$. The strategy works fine while the CT peak does not merge with the Kondo peak, obtaining reliable results in the range $E_d \ge 10\Delta$. When both peaks overlap, we still can distinguish the left part of the CT (at lower frequencies) and we fit just the left half part of CT peak using the DDMRG data. Calculations were done with up to 1000 states keeping the truncation error less than $10^{-8}$, assuring a numerical error in DDMRG data much smaller than the size of the symbols in Fig. \ref{eta}. \vspace{0.5cm} \begin{figure}[h] \includegraphics[width=8.cm]{CVDMRG.eps} \caption{(Color online) DDMRG correction vector data of the CT at $\eta=10\Delta$. Open circles (Crosses) for $N_s=40$ ($N_s=80$), for $E_d=40\Delta$ (see Fig.~\ref{anchos}). In legends we indicate the $\eta_{\text{eff}}$ obtained from a fit with a Lorentzian $\propto K/((\omega-\omega_{up})^2+\eta_{\text{eff}}^2)$ (thin line) for each size.} \label{eta} \end{figure} In order to visualize the procedure, we show in Fig.~\ref{eta} the DDMRG correction vector results, for $E_d=40\Delta$ with $\eta=10\Delta$, and for two sizes ($N_s=40$, 80). As it is apparent in the figure, there are no significant finite-size effects. Then we fit the peaks with Lorentzians to obtain the half width $\eta_{\text{eff}}$ and the position $E_{d}^{\text{eff}}$ for each lattice size, allowing us to take the estimation $\eta_{\text{eff}}=(10.9\pm0.1)\Delta$. Finally we can calculate the values presented in Fig.~\ref{anchos} doing the subtraction of the width obtained with the $\eta=10\Delta$ used in the correction vector. We derive in our example $\Delta_{\text{CT}}=(0.9\pm0.1)\Delta$. \section{Width and weight of the charge-transfer peak for large negative on-site energy} \label{vari} In this Section we calculate $\Delta _{\text{CT}}$ well inside the Kondo regime -E_{d}\gg N\Delta $, using a variational procedure.\cite{varma,gunn} The ground-state wave function of the Hamiltonian Eq. (\ref{ham}) is approximated as \begin{equation} |g\rangle =A|F\rangle +\sum_{qm}B_{q}d_{m}^{\dagger }c_{qm}|F\rangle , \label{gvar} \end{equation where $|F\rangle =\Pi _{q}^{\prime}c_{qm}^{\dagger }|0\rangle \bigotimes |s\rangle$ is the state with the filled Fermi see in the band and the impurity in the singlet state, the subscript $q $ refers to both lead and wave vector index [$\nu $ and $k$ in Eq. (\re {ham})] , the prime over the product symbol means that only $q$ for which \epsilon _{q}<\epsilon _{F}$ are included, where $\epsilon _{F}$ is the Fermi energy. We denote $d_{m}^{\dagger }=|m\rangle \langle s|$. The variational parameters $A$ and $B_{q}$ are determined minimizing the energy. The $B_{q}$ are independent of $m$ as a consequence of SU($N$)\ symmetry. Normalization of $|g\rangle $ implies \begin{equation} |A|^{2}+N\sum_{q}|B_{q}|^{2}=1. \label{norm} \end{equation In the following we take the origin of energies at the Fermi energy \epsilon _{F}=0$. We define the energy gain due to hybridization as \begin{equation} T_{K}=E-E_{0}, \label{tkvar1} \end{equation} where $E$ is the ground-state energy and $E_{0}=E_{d}+\langle F|H|F\rangle $, the ground-state energy for $V_{q}=0$. Minimizing the energy one obtains \begin{equation} B_{q}=\frac{-\bar{V}_{q}A}{T_{K}-\epsilon _{q}}, \label{bq} \end{equation and the equation for the ground-state energy, which can be written in the form \begin{equation} T_{K}-E_{d}-N\sum_{q}\frac{|V_{q}|^{2}}{T_{K}-\epsilon _{q}}=0. \label{evar} \end{equation The sum can be evaluated assuming $\Delta =\Delta _{L}+\Delta _{R}=\pi \sum_{q}|V_{q}|^{2}\delta (\omega -\epsilon _{q})$ independent of $\omega $ in the range $-D<\omega <D$ and gives $(\Delta /\pi )\ln [(D+T_{K})/T_{K}]$. In the Kondo limit $-E_{d}\gg N\Delta $, one can neglect $T_{K}$ in comparison with $D$ and $|E_{d}|$ obtaining \begin{equation} T_{K}=D\exp \left( \frac{\pi E_{d}}{N\Delta }\right) , \label{tkvar} \end{equation which has the correct exponential dependence on $E_{d}$ (although the correct prefactor is smaller.\cite{hewson,fili}) At zero temperature, the impurity spectral density is \begin{eqnarray} \rho _{m}(\omega ) &=&\rho _{m}^{d}(\omega )+\rho _{m}^{c}(\omega ), \notag \\ \rho _{m}^{d}(\omega ) &=&\sum_{e}|\langle e|d_{m}|g\rangle |^{2}\delta (\omega +E_{e}-E), \notag \\ \rho _{m}^{c}(\omega ) &=&\sum_{e}|\langle e|d_{m}^{\dagger}|g\rangle |^{2}\delta (\omega -E_{e}+E), \label{rhov} \end{eqnarray where $E_{e}$ is the energy of the excited state $|e\rangle $. Since we are interested in $\omega $ near $E_{d}$ and in this Section $E_{d}$ is well below the Fermi energy, we can neglect the creation part $\rho _{m}^{c}(\omega )$ of the spectral density. Using Eq. (\ref{gvar}) we can write \begin{eqnarray} \rho _{m}^{d}(\omega ) &=&-\frac{1}{\pi }\sum_{q}|B_{q}|^{2}\text{Im G_{qm,qm}(E-\omega ), \notag \\ G_{qm,qm}(z) &=&\langle qm|\hat{G}|qm\rangle ,\text{ }|qm\rangle =c_{qm}|F\rangle , \notag \\ \hat{G}(z) &=&\frac{1}{z+i\eta -H}, \label{gop} \end{eqnarray where $\eta \rightarrow 0^{+}$ is a positive infinitesimal. Defining the operators $\hat{G}^{0}(z)=z+i\eta -H_{0}$, where $H_{0}=H-\hat{ }$, and $\hat{V}=\sum_{qm}(V_{q}d_{m}^{\dagger}c_{qm}+\mathrm{H.c}.)$ is the hybridization part of $H$, one has the Dyson equation \begin{equation} \hat{G}=\hat{G}^{0}+\hat{G}^{0}\hat{V}\hat{G}. \label{dyson} \end{equation Taking matrix elements of both members Eq. (\ref{dyson}) between states |qm\rangle $ and/or $|q^{\prime}m^{\prime}qm\rangle =d_{m^{\prime}}^{\dagger }c_{q^{\prime}m^{\prime}}|qm\rangle $, we obtain after some algebra \begin{eqnarray} &&G_{qm,qm}(z)= \notag \\ &&=\frac{G_{qm,qm}^{0}(z)}{1-G_{qm,qm}^{0}(z)\sum_{q^{\prime }m^{\prime }}|V_{q^{\prime }}|^{2}G_{q^{\prime }m^{\prime }qm,q^{\prime }m^{\prime }qm}^{0}(z)}, \notag \\ &&G_{qm,qm}^{0}(z)=\frac{1}{z+i\eta +\epsilon _{q} -E_{0}+E_{d}}, \notag \\ &&G_{q^{\prime }m^{\prime }qm,q^{\prime }m^{\prime }qm}^{0}(z)=\frac{1} z+i\eta +\epsilon _{q}+\epsilon _{q^{\prime }}-E_{0}}. \label{gqm} \end{eqnarray The sum over $m^{\prime }$ above is just a factor $N$ due to SU($N$) symmetry. As we shall see, this factor enters into the width of the charge-transfer (CT) peak $\Delta _{\text{CT}}$ The sum over $q^{\prime }$ can be evaluated using the density of conduction states [as we have done to obtain Eq. (\ref{tkvar})], leading to \begin{eqnarray} \text{Im}G_{qm,qm}(E-\omega ) &=&\frac{-N\Delta }{\left( \omega +T_{K}-\epsilon _{q}-E_{d}-\Lambda \right) ^{2}+\left( N\Delta \right) ^{2}}, \notag \\ \Lambda &=&\frac{N\Delta }{\pi }\ln \left\vert \frac{D+\omega +T_{K}-\epsilon _{q}}{\omega +T_{K}-\epsilon _{q}}\right\vert , \label{imgqm} \end{eqnarray} Inserting this in the first Eq. (\ref{gop}) and using Eqs. (\ref{norm}) and (\ref{bq}) one obtains the desired spectral density. In the limit $-E_{d}\gg N\Delta $, $T_{K}$ becomes exponentially small [see Eq. (\ref{tkvar})]. Furthermore, from Eq. (\ref{bq}), one realizes that the values of $\epsilon _{q}$ which lead to the larger values of $B_{q}\sim A/T_{K}$ are of the order of $T_{K}$. From Eqs. (\ref{norm}) and (\ref{bq}) one sees that for $T_{K}\rightarrow 0$, also $A\rightarrow 0$ and \sum_{q}|B_{q}|^{2}\rightarrow 1/N$. Thus in this limit T_{K}\rightarrow 0$, for any function $F(\epsilon _{q})$, we can use \begin{equation} \sum_{q}|B_{q}|^{2}F(\epsilon _{q})\rightarrow \frac{1}{N}F(0) \label{lim} \end{equation Furthermore since $\Lambda $ has a week logarithmic dependence and we are interested in $\omega \sim E_{d}$, we can evaluate it at $E_{d}$. Then Eqs. \ref{imgqm}), the first (\ref{gop}) and (\ref{lim}) lead to \begin{eqnarray} \rho _{m}(\omega ) &\simeq &\frac{\Delta /\pi }{\left( \omega -E_{d}^{\text eff}}\right) ^{2}+\left( N\Delta \right) ^{2}}, \label{rovar} \\ E_{d}^{\text{eff}} &=&E_{d}+\frac{N\Delta }{\pi }\ln \left\vert \frac{D+E_{d }{E_{d}}\right\vert , \label{edeff} \end{eqnarray well inside the Kondo limit $-E_{d}\gg N\Delta .$ This simple result reflects three important physical effects of the correlations on the CT peak: i) The position is shifted upwards. The origin of this shift was explained by Haldane for the SU(2) case,\cite{hald} and extensions to the SU(4) case and the SU(2) $\leftrightarrow $SU(4) crossover were discussed.\cite{fili,restor} Depending on details different but similar estimations for the shift were given. This point is discussed in more detail in Section \ref{t0} (see Fig. \ref{shift}). ii) The half width at half maximum of the peak is $\Delta _{\text{CT}}=N\Delta $. This is a factor $N$ with respect to the noninteracting case. iii) the total weight of the peak is $1/N$. The maximum intensity is thus reduced in a factor $1/N^{2}$ compared to the noninteracting case. \section{Width of the charge-transfer peak in the SU(4) IAM} \label{su4} In this section we show our results for $\Delta_{\text{CT}}$ as a function of $E_d$ in the SU(4) case using NCA and DDMRG. For the NCA we used a constant density of unperturbed conduction states extending from $-D$ to $D$ with $D=100 \Delta$. The assumptions made in DDMRG imply a semielliptical density of conduction states.\cite{note,proe} \subsection{Temperature $T \rightarrow 0$} \label{t0} \begin{figure}[h] \includegraphics[width=8.cm]{Nvar.eps} \caption{(Color online) Impurity spectral density calculated with the NCA (full lines) and variational wave function (dashed lines) for $E_d= - 15 \Delta$. } \label{nca} \end{figure} We discuss first the results at low temperature ($T = T_{K}/10$). Here $T_{K}$ was chosen such that $G(T_{K})=G_{0}/2$, where $G_{0}$ is the conductance of the system at $T=0$ in the extreme Kondo limit in which the total occupancy is 1 (the maximum possible conductance of the system). Using Friedel sum rule \cite{lang,yoshi} one has \begin{equation} G_0 = N \frac{e^2}{h} \sin ^{2}\left( \frac{\pi}{N} \right). \label{g0} \end{equation} In Fig.~\ref{nca} we show the spectral density $\rho_m(\omega)$ obtained with the NCA in the Kondo regime ($-E_{d}\gg N\Delta $). It shows two peaks. The broad CT peak (which is the focus of this work) at $E_{d}^{\text{eff}}$ (slightly larger than $E_{d}$) and the very narrow Kondo peak at the Fermi energy. Instead, for positive $E_{d}$ (not shown) only the CT peak is present. We also show in the figure the result for $\rho _{m}^{d}(\omega )$ obtained using Eq. (\ref{gop}), (\ref{norm}) and (\ref{bq}). One can see that there is a very good agreement between the NCA and variational calculations for the CT peak. We have verified that the CT peak in the empty orbital regime (not shown) looks identical to the noninteracting spectral density, as expected from the discussion of the previous section. Instead, in the Kondo regime it is nearly four times broader and its maximum is sixteen times smaller. We have repeated these calculations for several values of $E_d$. In the empty orbital regime, the peak has been fit with a Lorentzian with three parameters: the width identified with $\Delta _{\text{CT}}$, the position $E_{d}^{\text{eff}}$ and its weight $w$, which is the integral in energy of the Lorentzian. In the Kondo regime we used a similar fit with two Lorentzians, one for the CT peak and one for the Kondo peak. In the intermediate valence (IV) regime $|E_{d}^{\text{eff}}| \sim \Delta$, this procedure failed because in addition to the fact that the CT peak merges with a Kondo peak that becomes increasingly wider, the NCA fails and gives a self energy with a positive imaginary part. Thus, in this region, the DDMRG results are particularly useful. \begin{figure}[h] \vspace{1.cm} \includegraphics[width=8.cm]{Deff.eps} \caption{(Color online) Half width at half maximum of the charge transfer peak as a function of the effective on site energy using NCA (green circles) and DDMRG (blue squares). Dashed lines is a fit of the filled symbols using Eq. (\ref{dfit}).} \label{anchos} \end{figure} In Fig.~\ref{anchos} we show the resulting $\Delta _{\text{CT}}$ as a function of $E_{d}^{\text{eff}}$ using both techniques. The filled symbols denote the most reliable result between both. Due to the fact that the band is flat in the NCA and semielliptical in DDMRG (with a total width $2D=200 \Delta$), the resonant level width $2\Delta_{\text{DDMRG}}(\omega)$ is energy dependent in DDMRG and smaller than the corresponding width $2\Delta$ within NCA, except for $\omega=0$ where both widths are equal.\cite{note} As a consequence for large $|E_{d}|$, a smaller $\Delta _{\text{CT}}$ is expected in the DDMRG calculation, as can be seen in the figure. In spite of the lack of NCA results near $E_{d}^{\text{eff}}=0$, the results of both techniques indicate a rather sudden crossover from $\Delta _{\text{CT}} \sim 4 \Delta$ to $\Delta _{\text{CT}} = \Delta$ at a slightly positive $E_{d}^{\text{eff}}$. In addition, there is an unexpected increase in $\Delta _{\text{CT}}$ for small negative $E_{d}^{\text{eff}}$. This might be due to the fact that the CT peak is deformed as it merges with the Kondo peak in this region. For later use (Section \ref{exp}), we also show in the figure a fit with the following function, which is an extension of that used in the SU(2) case \begin{equation} \frac{\Delta _{\text{CT}}}{\Delta }=a-b\tanh \left( \frac{E_{d}^{\text{eff }-E_{\text{IV}}}{c\Delta }\right) . \label{dfit} \end{equation For general $N$ one expects $a+b\simeq N$, $a-b\simeq 1$, so that this expression interpolates between the Kondo regime ($\Delta _{\text{CT}}\simeq N\Delta $ for $-E_{d}\gg N\Delta $) and the empty orbital one ($\Delta _{\text{CT}}=\Delta $ for $E_{d}\gg \Delta $) passing through the intermediate valence (IV) regime. From the fit we obtain $a=2.61$, $b=1.68$, $c=2.90$, and $E_{\text{IV}}=1.58\Delta $ for the effective level in the intermediate valence regime. \begin{figure}[t] \includegraphics[width=8.cm]{hald.eps} \caption{(Color online) Shift of the position of the CT peak as a function of on-site energy calculated with NCA for different temperatures $T$.} \label{shift} \end{figure} In Fig.~\ref{shift} we show the shift $\delta = E_{d}^{\text{eff}}- E_d$ as a function of $E_d$. Approximate expressions of this shift were given based on functional renormalization group.\cite{hald,fili,restor} Generalizing Haldane's treatment to the SU($N$) case \cite{restor} one obtains \begin{equation} \delta=\frac{N-1}{\pi}\Delta\ln \left(\frac{D}{C}\right), \label{shifte} \end{equation} where $C$ is a low-energy cutoff. Haldane in his study of the SU(2) model used $C=\Delta$, while Filipone {\it et al.} have taken $C=|E_d|/\alpha$ with $\alpha$ of the order of 1 in the SU(4) case.\cite{fili} The NCA results in Fig.~\ref{shift} seem to agree with a cutoff $C$ of the order of the maximum between $\Delta$ and $|E_d|$. In the present paper the shift $\delta$ is larger than in similar calculations of Ref. \onlinecite{restor} because of the larger value of $D$ chosen ($D=100$ here and $D=10$ in Ref. \onlinecite{restor}). \subsection{Finite temperatures} \label{fint} As the temperature increases, the above mentioned shortcomings of the NCA tend to disappear. For $T= 2 \Delta$ the Kondo peak has disappeared (although some weak structure persists near $\omega=0$ for small negative $E_{d}^{\text{eff}}$). Thus we have fit the CT peak using one Lorentzian. The results for $\Delta _{\text{CT}}$ as a function of $E_{d}^{\text{eff}}$ are plotted in Fig.~\ref{anchost}. For $T= 2 \Delta$ still an upturn for slight negative $E_{d}^{\text{eff}}$ is present, as for $T$ near to $0$ (see Fig. \ref{anchos}). For $T= 2 \Delta$ this structure is greatly reduced and the fit using Eq. (\ref{dfit}) improves. The parameters of the fit are for $T= 2 \Delta$: $a=2.81$, $b=1.74$, $c=2.51$, and $E_{\text{IV}}=-0.99\Delta $. For $T= 4 \Delta$: $a=2.82$, $b=1.73$, $c=6.26$, and $E_{\text{IV}}=-1.64\Delta $. Note that between $T= 2 \Delta$ and $T= 4 \Delta$, a broadening of crossover region between $\Delta _{\text{CT}} \sim 4 \Delta$ and $\Delta _{\text{CT}} \sim \Delta$ by a factor larger than two takes place. In addition, between very low $T$ and $T= 2 \Delta$, this region moves from slightly positive to slightly negative $E_{d}^{\text{eff}}$. \begin{figure}[t] \includegraphics[width=8.cm]{DeffT.eps} \caption{(Color online) Half width at half maximum of the charge transfer peak as a function of the effective on site energy using NCA at two different temperatures $T$. Dashed lines are fits using Eq. (\ref{dfit}).} \label{anchost} \end{figure} \section{Occupancy for SU(2) and SU(4)} \label{occ} To illustrate some of the effects of correlations, we show in Fig. \ref{ocup} the total occupancy $n= \sum_m \langle n_m \rangle$ as a function of the energy level calculated from the integral of the pesudofermion density of states in the NCA \cite{bickers}. This procedure is superior to the integral of the spectral density of the dot level, and is free from the shortcomings of this density for positive $E_d$, like a spurious peak at the Fermi energy. This technical point is discussed in Ref. \onlinecite{costi}. For $N$=2 we include in the figure results obtained using the Bethe ansatz (BA) as described in Ref. \onlinecite{ba}. These results were shifted to the left by a constant energy $C$ to compensate by the Haldane shift and possible uncertainties in the position of $E_d$ in the BA treatment. We have used $C=1.67$. We also include in the figure the noninteracting results. For $N=2$ one can observe a very good agreement between NCA and BA results for all $E_d$. For positive $E_d$ both results agree also with the $U=0$ (noninteracting) case. However, for negative $E_d$, the effects of correlations become apparent. In particular for $U=0$ and large negative $E_d$, $n=N$ for the SU($N$) model, while for $U, -E_d \rightarrow \infty$, $n=1$. This is due to the decreasing weight of the charge-transfer peak, discussed in Section \ref{vari} [see also Eq. (\ref{w}) below]. The deviations form the noninteracting case are stronger in the SU(4) case. They are significant even for $E_d$. This is due in part to the larger Haldane shift in the SU(4) case. The fact that the increase in $n$ with decreasing $E_d$ is smoother than for the N=2 case, is due to the larger width of the charge-transfer peaks in the region of negative $E_d$. \begin{figure}[t] \includegraphics[width=8.cm]{n.eps} \caption{(Color online) Total occupancy as a function of on-site energy level calculated with NCA for the SU($N$) IAM with $N$=2 and 4. Dashed blue line corresponds to the Bethe ansatz (BA) result for $N$=2 (see text). Dotted lines are the noninteracting results.} \label{ocup} \end{figure} \section{Possible experiments} \label{exp} \subsection{Two dots independently connected to its own pair of leads} \label{2dots} In the setup with two dots studied in Refs. \onlinecite{ama,hubel,keller}, the dot $i=1,2$ is connected to left and right ($\nu=L, R$) leads by couplings $\Delta_{\nu i}$ and all applied voltages and couplings can be controlled with high accuracy, For large enough repulsion between the dots,\cite{nishi2} if $\Delta_{L 1}+\Delta_{R 1}=\Delta_{L 2}+\Delta_{R 2}$, and equal energy levels $E_{d 2}=E_{d 1}$, the system is in the SU(4) regime.\cite{keller} Even if the equality is not satisfied exactly, the SU(4) symmetry can be restored at low energies adjusting the difference $E_{d 2} - E_{d 1}$.\cite{restor,nishi}. In this conditions, if in addition the couplings of one dot are very asymmetric (say $\Delta_{L 1} \ll \Delta_{R 1}$), moving the bias voltage $V_{L 1}$ of the $L 1$ lead, the resulting differential current through dot 1, $dI_1/dV_{L 1}$ just maps the spectral density of this dot (as in scanning tunneling spectroscopy).\cite{oks} In this case, a marked asymmetry in the $dI_1/dV_{L 1}$ response for positive and negative $E_{d i}$ should be measured, as in Fig. (\ref{anchos}). Previous non-equilibrium calculations show that if $\Delta_{L 1} < \Delta_{R 1}/9$, the resulting $dI_1/dV_{L 1}$ is very similar to the equilibrium spectral density for coupling $\Delta_{L 1} + \Delta_{R 1}$ and chemical potential corresponding to the right lead of dot 1. \subsection{Conductance through an SU(4) "impurity" including capacitance effects} \label{haugsu4} Here we discuss an experiment like that of K\"{o}nemann \textit{et al. }\cite{haug} introduced in Section \ref{intro}, in which asymmetries with the sign of bias voltage $V$ were observed in the widths and intensities of conductance peaks, but here we assume the QD or impurity described by the SU(4) IAM instead of the SU(2) one discussed previously.\cite{haug,capac} Specifically, the interacting system, which we call impurity is coupled to two SU(4) interacting leads (left $L$ and right $R$) with chemical potentials $\mu _{\nu }$ and tunnel couplings $\Delta _{\nu }$ ($\nu =L$, $R$), as described by the Hamiltonian Eq. (\ref{ham}). We take the sign of the bias voltage in such a way that $\mu _{L}-\mu_{R}=eV $. Note that the interchange of right and left leads or electrons by holes is equivalent to a change of sign of $V$. In this section, we take the origin of energies at $\mu _{R}=0$. The capacitance effects modify the energy necessary to add an electron to the dot with the lever arm parameter $\alpha $ (which depends on the source, drain and gate capacitances).\cite{haug,park} Usually $E_{d}=E_{d}^{0}+\alpha eV$ is assumed, where $\alpha \sim 1/2$, so that $E_{d}$ is displaced in approximately half the magnitude of $\mu _{L}-\mu_{R}$. Since the experimentally accessible quantity is $E_{d}^{\text{eff}}$ and not $E_{d}$, we assume \begin{equation} E_{d}^{\text{eff}}=E_{d}^{0}+\alpha eV, \label{Ed} \end{equation} where $E_{d}^{0}$ is the effective dot level (the position of the CT peak) when $V=0$ and $\alpha$ with $ \le \alpha \le 1$ describes how the effective level is modified by the bias voltage. A scheme of how $E_d$ is modified applying a bias voltage is presented in Fig. \ref{scheme} \begin{figure}[h] \includegraphics[width=8.cm]{sch.eps} \caption{(Color online) Scheme of the Fermi level of both leads and the displacement of the dot level $E-d$ due to capacitance effects. The top (bottom) figures represent different situations for positive (negative) $E_{d}^{0}$. When the Fermi level of one of the leads coincides with $E_d$ a peak in the conductance is observed. Assuming that the right lead is more strongly coupled to the dot, the system can be in Kondo, intermediate valence (IV) or empty orbital regime (EOR) as indicated.} \label{scheme} \end{figure} The current through the impurity is given by \cite{meir} \begin{eqnarray} I &=&C\int d\omega \rho _{m}(\omega ,V,E_{d})[f_{L}(\omega )-f_{R}(\omega )], \notag \\ C &=&\frac{4N\pi e\Delta _{L}\Delta _{R}}{h\Delta }, \label{i} \end{eqnarray} where $N=4$ in this section, $f_{\nu }(\omega )=f(\omega -\mu _{\nu })$ is the Fermi distribution in each lead, with $f(\omega )=1/(e^{\omega /kT}+1)$, and $\rho _{m}(\omega ,V,E_{d})$ is the non-equilibrium spectral density of the impurity level with symmetry $m$, which depends on the voltage $V$ explicitly and also implicitly through the voltage dependence of $E_{d}$ [Eq. (\ref{Ed})]. The conductance is $G=dI/dV.$ As shown before,\cite{capac} to observe the asymmetry in the peaks in $G(V)$ one needs asymmetric coupling to the leads. Therefore we assume $\Delta _{R}\gg \Delta _{L}$. It has been shown before \cite{oks} that for $\Delta _{L} \lesssim \Delta _{R}/9$, the spectral density at the dot practically coincides with the corresponding one with the dot at equilibrium with the right lead at chemical potential $\mu _{R}$. This allows us to avoid a cumbersome non-equilibrium calculation. To simplify the calculation further, we extend a phenomenological approach used before \cite{capac} to the present case, and assume that the impurity spectral density near the CT peak can be approximated by \begin{eqnarray} \rho _{m}(\omega ) &=&\frac{(w \Delta _{\text{CT}})/\pi }{(\omega -E_{d}^{\text{eff}})^{2}+\Delta _{\text{CT}}^{2}}, \label{rf} \\ w &=&1-\sum_{m^{\prime }\neq m}\langle n_{m^{\prime }}\rangle \label{w} \end{eqnarray where $\langle n_{m}\rangle =\int d\omega \rho _{m}(\omega )f_{R}(\omega )$ is the occupation of the dot for symmetry $m$. The weight $w$ is the probability that the site of the impurity is not occupied by electrons of symmetry different than $m$. This weight appears naturally in the atomic limit $ V_q \rightarrow 0$ and was used in Hubbard-III \cite{hub3} and similar \cite{aaa,dyn} approximations to Hubbard and intermediate-valence models. Integrating Eq. (\ref{rf}), using Eq. (\ref{w}) and SU($N$) symmetry so that $\langle n_{m}\rangle $ is independent of $m$, one obtains a linear equation for $\langle n_{m}\rangle $. Solving it we obtain in general \begin{eqnarray} \langle n_{m}\rangle &=&\frac{1-2\Psi _{R}}{(N+1)-2(N-1)\Psi _{R}}, \notag \\ \Psi _{\nu } &=&\frac{1}{\pi }\text{Im}\psi (\chi _{v}), \notag \\ \chi _{v} &=&\frac{1}{2}+\frac{\Delta _{\text{CT}}+i(E_{d}^{\text{eff}}-\mu _{\nu })}{2\pi T}, \label{nm} \end{eqnarray where $\psi (x)$ is the digamma function. \begin{figure}[h] \vspace{1.cm} \includegraphics[width=8.cm]{P.eps} \caption{(Color online) Weight of the CT peak obtained from the fit of the NCA CT peak (circles and squares) and from Eqs. (\ref{w}) and (\ref{nm}) (dashed lines) with $N=4$ as a function of the effective impurity level for several temperatures $T$. Blue dashed line corresponds to $T=0$.} \label{weight} \end{figure} In Fig.~\ref{weight} we compare the weight obtained fitting the NCA results with those of the phenomenological approach given here. With the exception of the region in which both $T$ and $E_{d}^{\text{eff}}$ are near zero, we see that Eq. \ref{rf} reproduces well the NCA result. In particular for $|E_{d}^{\text{eff}}| \gg \Delta$, Eq. (\ref{rf}) reproduces accurately the spectral density at the CT peak. Replacing Eq. (\ref{rf}) in Eq. (\ref{i}) gives the current \begin{equation} I=C [1-(N-1)\langle n_{m}\rangle ](\Psi _{R}-\Psi _{L}), \label{if} \end{equation} Differentiating this expression with respect to the voltage, using Eqs. (\ref{dfit}), (\ref{Ed}), and (\ref{nm}), we obtain for $N=4$ \begin{eqnarray} G(V) &=&C [ \left( 1-3\langle n_{m}\rangle \right) \left( \Psi _{R}^{\prime }-\Psi _{L}^{\prime }\right) \notag \\ &&+\frac{12\Psi _{R}^{\prime }}{\left( 5-6\Psi _{R}\right) ^{2}}(\Psi _{R}-\Psi _{L})] , \notag \\ \Psi _{L}^{\prime } &=&\frac{1}{2\pi ^{2}T}\text{Im}\left\{ \psi ^{\prime }(\chi _{L})\left( i(\alpha -1)-\xi \right) \right\} , \notag \\ \Psi _{R}^{\prime } &=&\frac{1}{2\pi ^{2}T}\text{Im}\left\{ \psi ^{\prime }(\chi _{R})\left( i\alpha -\xi \right) \right\} , \notag \\ \xi &=&\frac{b\alpha }{c}\text{Sech}\left[ \frac{E_{d}^{\text{eff}} -E_{\text{IV}}}{c\Delta }\right] ^{2}, \end{eqnarray} and $\psi ^{\prime }(x)$ is the derivative of the digamma function. The weak point in this approach is the use of the fit Eq. (\ref{dfit}), which as can be seen in Figs. \ref{anchos} and \ref{anchost} fails near the intermediate valence regime ($E_{d}^{\text{eff}} \sim 0$). However, this equation is accurate enough in the Kondo ($-E_{d}^{\text{eff}} \gg N \Delta$) and empty orbital ($E_{d}^{\text{eff}} \gg \Delta$) regimes. \begin{figure}[h] \vspace{1.cm} \includegraphics[width=8.cm]{Ed+15.eps} \caption{(Color online) Differential conductance for an SU(4) impurity as a function of bias voltage for asymmetric couplings $\Delta_R \gg \Delta_L$, capacitance ratio $\alpha=1/2$, $E_{d}^0=15 \Delta$ and several temperatures $T$. The constant $C^{\prime}_{1} = eC/(2 \pi \Delta)$} \label{haugp} \end{figure} In Fig.~\ref{haugp} we show the resulting conductance for $E_{d}^0$ slightly smaller than $E_{d}^{\text{eff}}$ of the dotted line in Fig.~\ref{nca}. At $V=0$, $E_{d}^{\text{eff}}$ lies $15 \Delta$ at higher energy than $\mu_L=\mu_R=0$. as $V$ increases, a peak develops which reaches its maximum at $eV= 30 \Delta$, for which $\mu_L=E_{d}^{\text{eff}}=eV$. This configuration is similar to a scanning tunneling microscope (STM), for which the tip is very weakly coupled to the impurity. However, due to capacitance effects, the peak is broader by a factor $1/(1-\alpha)$ (two in the figure),\cite{capac} and the intensity at the maximum is near $(1-\alpha)eC/(\pi \Delta)$ as it can be seen in the figure. This peak, which corresponds to the empty orbital regime, is accurately reproduced by our phenomenological approach. When a negative voltage is applied, $E_{d}^{\text{eff}}$ decreases until for $eV=-30 \Delta$ it reaches the value of $\mu_R=0$. In this case, the system is in the intermediate-valence regime and for $T=0$ our approach is not accurate because of the uncertainty in the width and weight of the CT in this region. From the spectral density, one would expect a peak nearly two times wider and a total weight reduced by a factor 2/5, but also the strong dependence of $\Delta_{\text{CT}}$ on $V$ (through $E_{d}^{\text{eff}}$) affects the shape. As expected, increasing the temperature has the effect of broadening both peaks, with more dramatic effects on the narrower one. If the system is instead prepared starting from the Kondo regime $E_{d}^0=-15 \Delta$ (Fig.~\ref{haugn}), at low temperatures one has the well known Kondo peak near zero bias.\cite{capac} Applying a positive $V$ one reaches again the intermediate-valence regime for $eV= 30 \Delta$ for which $E_{d}^{\text{eff}}=\mu_R=0$. Applying instead a negative voltage $eV \sim -30 \Delta$, $\mu_L=eV$ again coincides with $E_{d}^{\text{eff}}$ and a peak corresponding to a configuration similar to an STM one, but with a width increased by a factor $1/(1-\alpha)$ with respect to the spectral density, is obtained. Due to the increase of the width by a factor $N$ and a decrease in the total weight of the peak by a factor $\sim 1/N$, the intensity at this relative maximum in $G(V)$ is near $(1-\alpha)eC/(N^2\Delta)$, as seen in the figure. This peak corresponds to the Kondo regime and is well reproduced by our approach. Then we can anticipate that if an experiment with symmetric values of $E_{d}^0$ (with the same $|E_{d}^0|$ can be achieved, applying a gate voltage to the less coupled lead with the sign such that $|E_{d}^0|$ increases, in the case of negative $E_{d}^0$ (Kondo regime) the peak will be four times wider and nearly sixteen times less intense than for positive $E_{d}^0$ (empty orbital regime). \begin{figure}[h] \vspace{1.cm} \includegraphics[width=8.cm]{Ed-15.eps} \caption{(Color online) Same as Fig. \ref{haugp} for $E_{d}^0=-15 \Delta$. The constant $C^{\prime}_{2} = eC/(32 \pi \Delta)$} \label{haugn} \end{figure} \section{Summary and discussion} \label{sum} We have calculated the width of the charge transfer peak in the infinite-$U$ SU(4) Anderson model as a function of the impurity level $E_d$ and for the general SU($N$) case in the Kondo regime $-E_d \gg N \Delta$, where $\Delta$ is half the resonant level width. While it is clear that in the empty orbital regime $E_d \gg \Delta$, the half width at half maximum of the charge transfer peak $\Delta_{\text{CT}}= \Delta$, we obtain that in the Kondo regime $\Delta_{\text{CT}}=N\Delta$ extending previous results for the SU(2) case.\cite{pru,logan} The weight of the peak in the Kondo regime is reduced by a factor $\sim 1/N$ due correlation effects, so that the maximum intensity is reduced by a factor $1/N^2$. The variational calculation presented in Section \ref{vari} provide a simple physical picture of the non-trivial effect of correlations in broadening the charge transfer peak in the Kondo regime. The variational calculation of Section \ref{vari} suggest a simple interpretation of the width $\Delta_{\text{CT}}$. For a peak below the Fermi level $\epsilon_F=0$, the shape of the spectral density is defined by the spectral distribution of the state $|v>$ (not an eigenstate) obtained after annihilating one $d_m$ electron [see the Eq. (\ref{rhov}) for $\rho _{m}^{d}(\omega )$]. The hybridization term mixes this state with other states in which a $d_m$ electron is created again. The number of ways $M$ of adding this electron gives the number of bands that hybridize with $|v>$ and the width would be $\Delta_{\text{CT}}=M \Delta$. For peaks above the Fermi energy the same argument would apply interchanging creation and annihilation. To check this idea and investigate the effects of finite $U$, we have run the DDMRG code for $E_d=-30$, and $U=20$. We obtain peaks at $E_d+U$ and $E_d+2U$ of width $\sim 3.2 \Delta$ consistent with the expected $M=3$. However, there is also an intense central peak which overlaps with the charge-transfer peak and renders it dangerous to draw definite conclusions. To analyze general values of $E_d$, it is more useful to discuss it in terms of the position of the charge-transfer peak $E_{d}^{\text{eff}}$ which incorporates a shift in the effective $E_d$ discussed first by Haldane for the SU(2) case and more recently for the SU(4) case.\cite{fili,restor}. Our results for SU(4) symmetry suggest a rather abrupt decrease of $\Delta_{\text{CT}}$ as $E_{d}^{\text{eff}}$ increases from $\sim -\Delta$ to $\sim \Delta$, but our analysis at low temperature for small $|E_{d}^{\text{eff}}|$ is complicated by the merge of the charge-transfer and Kondo peaks and limitations of our approaches. The strong contrast between the charge-transfer peaks in the Kondo and empty-orbital regimes should be observable in experiment. This is discussed in Section \ref{exp}. One possibility is a setup of two quantum dots,\cite{ama,hubel,keller} which can be tuned to prepare the system in a similar way as in scanning tunneling spectroscopies, with one quantum dot very weakly coupled to a lead to which a bias voltage is applied. Another possibility is to extend to an SU(4) system the experiment of K\"{o}nemann \textit{et al.}, in which strong asymmetries were observed in the width and magnitude of conductance peaks as a function of gate voltage for a quantum dot described by the SU(2) impurity Anderson model.\cite{haug} These experiments, in which also capacitance effects play a role,\cite{capac} would display stronger asymmetries in the SU(4) case. We hope that these results will encourage experimental work along these lines and be useful for the interpretation of Coulomb blockade peaks in conductance experiments. On the theoretical side, a deeper analysis of the change in $\Delta_{\text{CT}}$ at zero temperature as $E_{d}^{\text{eff}}$ crosses zero would be useful. \section*{Acknowledgments} J. F. and A. A. A. are sponsored by PIP 112-201101-00832 of CONICET and PICT 2013-1045 of the ANPCyT. F. L. and C. G. acknowledge support from CONICET, through Project PIP No. 112-201201-00389CO

\section{Introduction} In the Alvarez-Macovski\cite{Alvarez1976} method for extracting x-ray spectral data, the attenuation coefficient is expanded as a linear combination of functions of energy multiplied by basis set coefficients that depend on the composition of the material at points within the object. The line integrals of the basis set coefficients are then computed by inverting the transformation from the measurements with multiple spectra. Summarizing the measurements by the vector $\mathbf{L}$ and the line integrals of the basis set coefficients as the vector $\mathbf{A}$, we need to show that the transformation $\mathbf{L(A)}$ is invertible. This question is examined in this paper. In particular, the transformation for a noise-free system using two measurement spectra with a two function basis set, \emph{i.e.} a dual energy system, is studied. Invertibility is of more than mathematical interest since spectra that are near non-invertible can lead to a large multiplication of noise in the output images. In the past, dual energy systems used measurements with two different x-ray tube voltages or a fixed tube voltage and filtering with two materials both with energy integrating detectors. Experience showed that these were invertible. However, the introduction of photon counting detectors into clinical systems\cite{taguchi2013vision} leads to potential invertibility problems due to the imperfections in these detectors\cite{Overdick2008} such as pileup, K-radiation escape, charge sharing, and others. My recent paper\cite{alvarez2017noninvertibility} showed that a large pileup factor can lead to non-invertibility for a three dimension A-vector with three bin pulse height analysis and that this non-invertibility leads to large increases in noise. I discussed invertibility in other papers. In early work\cite{Alvarez1982,Lehmann1986}, I showed that a general, two dimensional inversion theorem\cite{Fulks1978} could be applied to the dual energy transformation. The theorem requires that the transformation be invertible on a closed contour in the first quadrant of the A-vector plane and have a non-zero Jacobian determinant inside the contour. I showed that the transformation is invertible on the contour if the two spectra have different maximum energies. The proof was for ordinary energy integrating detectors as well as photon counting detectors without pileup. Later, I showed that the transformation is invertible on the contour for photon counting detectors with pileup\cite{alvarez2017DEInvertibilityProof}. Recently, Levine\cite{levine2017nonuniqueness} described a different approach to determining invertibility. He studied a system with energy spectra with only three discrete energies, no scatter and a perfect detector. He found that, with appropriately chosen spectral values, the transformation is not one-to-one and is therefore not invertible. The spectra had the same maximum energy so my previous results were not applicable. My papers did not address the proof that the Jacobian determinant is non-zero within the contour. In this paper, I derive analytical formulas are derived to test for the non-zero condition for a set of measurement spectra and detector. The formulas are applied to several systems including (1) the set proposed by Levine\cite{levine2017nonuniqueness} that are not invertible as well as (2) similar spectra that are invertible and (3) invertible voltage switched x-ray tube spectra that have long been used in dual energy imaging. The results are verified by testing the errors with an iterative algorithm that inverts the transformation. \section{Methods} \subsection{The transformation\label{}} This section reviews the Alvarez-Macovski method\cite{Alvarez1976}, introduces notation, and specifies the transformation that is required to be inverted for the method. For biological materials, we can approximate the x-ray attenuation coefficient $\mu(\mathbf{r},E)$ with a two function basis set\cite{alvarez2013dimensionality} \begin{equation} \mu(\mathbf{r},E)=a_{1}(\mathbf{r})f_{1}(E)+a_{2}(\mathbf{r})f_{2}(E).\label{} \end{equation} \noindent{}In this equation, the coefficients $a_{i}(\mathbf{r})$ are functions only of the material composition at a position $\mathbf{r}$ within the object and the functions $f_{i}(E)$ depend only on the x-ray energy $E$. If there is a high atomic number contrast agent, we can extend the basis set to higher dimensions. Neglecting scatter, the expected value of the number of transmitted photons $\lambda_{k}$ with an effective measurement spectrum $S_{k}(E)$ is \begin{equation} \lambda_{k}=\int S_{k}(E)e^{-\int_{\mathcal{L}}\mu\left(\mathbf{r},E\right)d\mathbf{r}}dE\label{} \end{equation} \noindent{}where the line integral in the exponent is on a line $\mathcal{L}$ from the x-ray source to the detector. The effective measurement spectra include the effects of the source spectrum and the detector response. Using the attenuation coefficient decomposition, Eq. \ref{}, the line integral is \begin{equation} \int_{\mathcal{L}}\mu\left(\mathbf{r},E\right)d\mathbf{r}=A_{1}f_{1}(E)+A_{2}f_{2}(E).\label{} \end{equation} \noindent{}where $A_{i}=\int_{\mathcal{L}}a_{i}\left(\mathbf{r}\right)d\mathbf{r},\ i=1\ldots2$ are the line integrals of the basis set coefficients. If the $A_{i}$ are summarized as the components of the A-vector, $\mathbf{A}$, and the basis functions at energy $E$ by the vector $\mathbf{f(E)}$, the expected values in Eq. \ref{} are \begin{equation} \lambda_{k}\left(\mathbf{A}\right)=\int S_{k}(E)e^{-\mathbf{A\cdot f(E)}}dE.\label{} \end{equation} This paper assumes noise free data so the measurements are the expected values. These measurements are summarized by a vector, $\mathbf{N}$, whose components are the expected value of the photon counts with each effective spectrum. Since the body transmission is exponential in $\mathbf{A}$, we can approximately linearize the measurements by taking logarithms. The results is the log measurement vector \begin{equation} \mathbf{L(A)}=-\log(\mathbf{\frac{N(A)}{\mathbf{\boldsymbol{N}}_{0}}})\label{} \end{equation} \noindent{}where $\mathbf{\boldsymbol{N}_{0}}$ is the expected value of the measurements with no object in the beam and the division means that corresponding members of the vectors are divided. Eq. \ref{} defines a transformation from the log measurements $\mathbf{L(A)}$ to the A-vector, $\mathbf{A}$. \subsection{The Invertibility Theorem} The following theorem (Fulks\cite{Fulks1978} page 284) gives conditions for any two dimension transformation to be invertible: \begin{quotation} Let F be a continuously differentiable mapping defined on an open region D in E2, with range R in E2, and let its Jacobian be never zero in D. Suppose further that C is a simple closed curve that, together with its interior (recall the Jordan curve theorem), lies in D, and that $\mathbf{F}$ is one-to-one on C. Then the image $\Gamma$ of C is a simple closed curve that, together with its interior, lies in R. Furthermore, $\mathbf{F}$ is one-to-one on the closed region consisting of C and its interior, so that the inverse transformation can be defined on the closed region consisting of $\Gamma$ and its interior. \end{quotation} \noindent{}In Fulks' notation, $E2$ is a two dimensional Euclidean space, so this theorem is applicable to the transformation $\mathbf{L(A)}$ for the two-measurement, two-basis function case. I paraphrase the theorem as \begin{quotation} If the Jacobian of a continuously differentiable two dimensional mapping is nonzero throughout an open region D and if the mapping is one to one on a simple closed curve C which lies in D, then the mapping is one to one and invertible on C and its interior. \end{quotation} \noindent{}Fulks uses the term ``Jacobian'' to refer to the determinant of the Jacobian matrix \[ J=\det\left[\frac{\partial L_{i}}{\partial A_{j}}\right]. \] I will assume that the converse of the theorem is also true. That is, if the Jacobian determinant is zero in a region, then the transformation is not invertible in that region. This is reasonable given the one-dimension case where a zero derivative implies the transformation is not one-to-one in a region around the zero value and is therefore not invertible. I discussed the application of this theorem to $\mathbf{L(A)}$ in previous papers\cite{Alvarez1982,Lehmann1986,alvarez2017DEInvertibilityProof}. The proofs in these papers used the contour $C$, shown in Fig. \ref{}, which consists of the $A_{1}$ and $A_{2}$ axes and a section of a circle centered on the origin and joining the end points of the segments on the axes. In the previous papers, I proved that $\mathbf{L(A)}$ is invertible on the $A_{1}$ and $A_{2}$ axes under very general conditions since objects whose A-vectors are on these axes are composed of only one material. The previous invertibility proofs on the circular portion of $C$ are not applicable to some of the spectra in this paper. The proof assumes that the maximum energies in the spectra are different but the spectra discussed below have the same three energies but with different weights. The previous argument was that, due to beam hardening, as the radius of the contour becomes larger the transmitted spectrum approaches a peak at the maximum energy. If the maximum energies of the measurement spectra are different, then the transformation will be invertible on the circle. However, if the maximum energies are the same, the transformation becomes non-invertible for large object thicknesses by the No-Linearize theorem (see Section 6.4 of my dissertation\cite{Alvarez} and Chapter 44 of my book\cite{AlvarezBlogBook}). However, the transformation may be invertible for moderate thicknesses if we can guarantee that the circular portion of the contour with a smaller radius is in a region where the Jacobian is non-zero. \begin{figure} \includegraphics[scale=0.6]{AlvarezFig1_1}\centering{} \protect\caption{Closed contour used in proof of invertibility. The contour consists of the two axes, drawn slightly offset here to make them visible, and a quadrant of a circle centered at the origin joining the endpoints of the line segments along the axes.\label{}} \end{figure} \subsection{Interpretation of the M Matrix and its derivative} For our case, the Jacobian matrix is the matrix $\mathbf{M}$ with coefficients \begin{equation} M_{ij}=\frac{\partial L_{i}}{\partial A_{j}}.\label{} \end{equation} Since $L_{i}=-\log\left(\frac{\lambda_{i}}{\lambda_{i,0}}\right)$, we can rewrite the elements of $\mathbf{M}$ as \begin{equation} M_{ij}=-\frac{1}{\lambda_{i}}\frac{\partial\lambda_{i}}{\partial A_{j}}.\label{} \end{equation} Substituting the expected value of measurements Eq. \ref{} in this equation \begin{equation} M_{ij}=\frac{\int f_{j}(E)S_{i}(E)e^{-\mathbf{A\cdot f(E)}}dE}{\int S_{i}(E)e^{-\mathbf{A\cdot f(E)}}dE}.\label{} \end{equation} \noindent{}Defining the normalized spectrum \[ \hat{s_{i}}(E)=\frac{S_{i}(E)e^{-\mathbf{A\cdot f(E)}}}{\int S_{i}(E)e^{-\mathbf{A\cdot f(E)}}dE} \] \begin{equation} M_{ij}=\int f_{j}(E)\hat{s}_{i}(E)dE=\left\langle f_{j}(E)\right\rangle _{i}.\label{} \end{equation} \noindent{}That is, $M_{ij}$ is the effective value of the $f_{j}(E)$ basis function in the transmitted measurement spectrum $S_{i}(E)e^{-\mathbf{A\cdot f(E)}}$. The derivative of $M_{ij}$ with respect to each of the A-vector components can be computed by differentiating Eq. \ref{} using the quotient rule \begin{equation} \begin{array}{ccc} \frac{\partial M_{ij}}{\partial A_{k}} & = & \frac{-\lambda_{i}\left[\int f_{j}(E)f_{k}(E)S_{i}(E)e^{-\mathbf{A\cdot f(E)}}dE\right]+\left[\int f_{j}(E)S_{i}(E)e^{-\mathbf{A\cdot f(E)}}dE\right]\left[\int f_{k}(E)S_{i}(E)e^{-\mathbf{A\cdot f(E)}}dE\right]}{\lambda_{i}^{2}}\\ & = & -\left[\left\langle f_{j}(E)f_{k}(E)\right\rangle _{i}-\left\langle f_{j}(E)\right\rangle _{i}\left\langle f_{k}(E)\right\rangle _{i}\right]\\ & = & M_{ij}M_{ik}-\left\langle f_{j}(E)f_{k}(E)\right\rangle _{i}. \end{array}\label{} \end{equation} \noindent{}In this equation, $\lambda_{i}$ is given by Eq. \ref{} and $\left\langle \right\rangle _{i}$ is defined in Eq. \ref{}. The effective values of $\mathbf{M}$ depend on the transmitted spectrum and are, for most spectra, a function of $\mathbf{A}$. If the spectra are single delta functions at energy $E_{1},\ E_{2}$ such as from an isotope source or a synchrotron source, then \begin{equation} \left\langle f_{j}(E)f_{k}(E)\right\rangle _{i,delta\ function}=M_{ij}M_{ik}\label{eq:<mimj>delta} \end{equation} and $\nicefrac{\partial M_{ij}}{\partial A_{k}}=0$. With single delta function spectra, there is no beam hardening and the $M_{ij}$ are constants for all values of $\mathbf{A}$. Notice that Eq. \ref{} has the same form as a covariance. \subsection{Analytical formula for endpoints of zero determinant line\label{}} Since $\mathbf{M}$ is a $2\times2$ matrix, its determinant is \begin{equation} \det\left(\mathbf{M}\right)=M_{11}M_{22}-M_{12}M_{21}\label{} \end{equation} \noindent{}To apply the inversion theorem, we need to determine whether the determinant is zero in the first quadrant of the $\mathbf{A}$ plane. The simulations in Sec. \ref{} show that the product terms $M_{ij}M_{pq}(\mathbf{A})$ are approximately planar and do not intersect if the transformation is invertible. If the transformation is not invertible, the $M_{11}M_{22}(\mathbf{A})$ and $M_{12}M_{21}(\mathbf{A})$ surfaces do intersect and the curve of intersection is approximately a straight line. One approach to develop analytical formulas for invertibility is to determine the line by points of zero determinant on the $A_{1}$ and $A_{2}$ axes. the intersections of the the product terms on the $A_{1}$ and $A_{2}$ axes. If the determinant is zero in the first quadrant so the system is invertible, both of the intersections will be not physically feasible, that is, the product terms at the intersection are less than or equal to zero. Using a linear approximation to the determinant along the $A_{k},\ k=1,2$ axes \begin{equation} \begin{array}{ccc} \det\left(\mathbf{M}\right)(A_{k}) & \approx & \det\left(\mathbf{M}\right)(\mathbf{0})+\left\{ \frac{\partial}{\partial A_{k}}\left[M_{11}M_{22}\right](\mathbf{0})-\frac{\partial}{\partial A_{k}}\left[M_{12}M_{21}\right](\mathbf{0})\right\} A_{k}.\end{array}\label{} \end{equation} \noindent{}Setting the right hand side equal to zero and solving for $A_{k}$ \begin{equation} A_{k,0}=-\frac{\det\left(\mathbf{M}\right)(\mathbf{0})}{\frac{\partial}{\partial A_{k}}\left[M_{11}M_{22}\right](\mathbf{0})-\frac{\partial}{\partial A_{k}}\left[M_{12}M_{21}\right](\mathbf{0})}.\label{} \end{equation} We can use the product rule to evaluate the derivatives in the denominator \begin{equation} \begin{array}{ccc} \frac{\partial}{\partial A_{k}}\left[M_{ij}M_{pq}\right] & = & M_{ij}\frac{\partial M_{pq}}{\partial A_{k}}+M_{pq}\frac{\partial M_{ij}}{\partial A_{k}}\end{array}\label{} \end{equation} \subsection{Analytical condition for invertibility\label{}} Using a linear approximation, the value of the diagonal product term, $\left[M_{11}M_{22}\right]$, at $A_{k,0}$ is \begin{equation} \left[M_{11}M_{22}\right]_{lin}(A_{k,0})\approx M_{11}M_{22}\left(\mathbf{0}\right)+\frac{\partial}{\partial A_{1}}\left[M_{11}M_{22}\right](\mathbf{0})A_{k,0}.\label{} \end{equation} \noindent{}Note that the diagonal and off diagonal product terms are equal at the $\det(\mathbf{M}$)=0 points. In Sec. \ref{}, I show numerically that the linear approximation to $\left[M_{11}M_{22}\right]_{lin}(A_{k,0})$ is positive for the non-invertible spectrum and negative for at least one of the $A_{k}$ axes for the invertible spectra. We can derive an analytical formula for $\left[M_{11}M_{22}\right]_{lin}(A_{k,0})$ by substituting $A_{k,0}$ from Eq. \ref{} into Eq. \ref{} and simplifying:\begin{Large} \begin{equation} \begin{array}{cc} \left[M_{11}M_{22}\right]_{lin}(A_{k,0})=\ldots\\ \\ \frac{M_{11}M_{22}M_{12}M_{21}\left[\frac{\left\langle f_{2}f_{k}\right\rangle _{1}}{M_{12}}-\frac{\left\langle f_{1}f_{k}\right\rangle _{1}}{M_{11}}+\frac{\left\langle f_{1}f_{k}\right\rangle _{2}}{M_{21}}-\frac{\left\langle f_{2}f_{k}\right\rangle _{2}}{M_{22}}\right]}{\det\left(\mathbf{M}\right)(\mathbf{0})(M_{1k}+M_{2k})+M_{21}\left\langle f_{2}f_{k}\right\rangle _{1}-M_{22}\left\langle f_{1}f_{k}\right\rangle _{1}+M_{12}\left\langle f_{1}f_{k}\right\rangle _{2}-M_{11}\left\langle f_{2}f_{k}\right\rangle _{2}} \end{array}\label{} \end{equation} \end{Large} The simulations also show that the denominator of \emph{Eq. \ref{}} is positive for the three spectra tested. Since we are only interested in whether the product value is negative or zero and the simulations show the denominator is always positive, we need consider the sign of the numerator of \emph{Eq. \ref{} \begin{equation} Numerator=M_{11}M_{22}M_{12}M_{21}\left[\frac{\left\langle f_{2}f_{k}\right\rangle _{1}}{M_{12}}-\frac{\left\langle f_{1}f_{k}\right\rangle _{1}}{M_{11}}+\frac{\left\langle f_{1}f_{k}\right\rangle _{2}}{M_{21}}-\frac{\left\langle f_{2}f_{k}\right\rangle _{2}}{M_{22}}\right].\label{} \end{equation} \end{Large} \noindent{}The elements of $\mathbf{M}$ are effective values of the basis functions and are always positive so the sign is determined by the expression in brackets of Eq. \ref{}. Notice that for delta function spectra the first and second and the third and fourth terms of the numerator cancel it is equal to zero. Similarly, the denominator is also zero for delta functions so the product term is indeterminate $\frac{0}{0}$ and the system will be invertible if the Jacobian determinant at zero thickness is not equal to zero. \subsection{Simulations of det(M(A))\label{}} In the following sections, I will study by computer simulation whether the Jacobian determinant is zero inside $C$ for several spectra: (1) a set that are known to be non-invertible for two specific A-vectors, (2) a similar set that are better conditioned, and (3) x-ray tube spectra with different voltages. \subsubsection{3 energy non-invertible} Levine\cite{levine2017nonuniqueness} described a set of spectra that he found were not one-to-one. In this section, we analyze these spectra using the methods described in the previous sections. Fig. \ref{} (a) shows a three dimension plot of the diagonal and off-diagonal $\mathbf{M}$ elements product terms. Since the determinant is $\det(M)=M_{11}M_{22}-M_{12}M_{21}$, it is zero if the surfaces intersect. Each term is plotted as a separate color-coded surface as shown in the legend of the figure. Part (b) of the figure shows the spectra. From the figure, it is clear that the product term surfaces do intersect on approximately a straight line. \begin{figure}[H] \centering{}(a)\includegraphics[scale=0.6]{AlvarezFig2_1}(b)\includegraphics[scale=0.5]{AlvarezFig2_2} \protect\caption{Three dimension plot (a) of the products of the diagonal and off-diagonal elements of the determinant $\det(M\left(\mathbf{A}\right))=M_{11}M_{22}-M_{12}M_{21}$ as a function of $\mathbf{A}$ for the Levine non-invertible three energy spectrum. The determinant is zero if the surfaces intersect. Each product term is plotted as a separate color-coded surface as shown in the legend. Part (b) is a plot of the spectra as a function of x-ray energy. \label{}} \end{figure} Fig. \ref{} shows views of the diagonal and off-diagonal $\mathbf{M}$ elements product data projected onto the A-plane. Part (a) is a view of the diagonal and off-diagonal determinant terms in Fig. \ref{} projected along the Z-axis. Part (b) is an image where the grayscale is proportional to $\nicefrac{1}{\left|det(M(A))\right|}$. The bright line in the image shows the zero values. A straight line was fit to the line of maxima and is plotted on the image in yellow. The close fit of the straight line shows that the surfaces are close to planes so their intersection is a line. Fig. \ref{} also shows the ambiguous A-vectors found by Levine plotted as the green crosses. Notice that the points do not fall on the line of zero determinant values. \begin{figure}[H] \centering{}(a)\includegraphics[scale=0.5]{AlvarezFig3_1} (b)\includegraphics{AlvarezFig3_2} \protect\caption{Projection of the terms of $det(M(A))$ onto the A-plane. Part (a) is a view of the diagonal and off-diagonal determinant terms in Fig. \ref{} projected along the Z-axis. Part (b) is an image where the grayscale is proportional to $\nicefrac{1}{\left|det(M(A))\right|}$. The bright line in the image shows the zero values. A straight line was fit to the line of maxima and is plotted on the image in yellow. The close fit of the straight line shows that the surfaces are close to planes so their intersection is a line. Also shown is the Levine ambiguous A-vectors plotted as the green crosses. \label{}} \end{figure} Fig. \ref{} shows one dimensional plots of the diagonal and off-diagonal determinant terms on the $A_{bone}$ and $A_{water}$ axes. The curved lines are the actual surfaces with a circle at their intersection where $\det(M)$ is zero. Also shown are straight line fits to the curved lines using the slope at the origin. The slope was computed with Eq. \ref{}. The intersection of the fit lines is the diamond shape. \begin{figure}[H] \centering{}\includegraphics[scale=0.6]{AlvarezFig4_1} \protect\caption{Plot of the diagonal and off-diagonal surfaces along the $A_{1}$ and $A_{2}$ axes. The curved lines are the actual surfaces with a circle at their intersection where $\det(M)$ is zero. Also shown are straight line fits to the curved lines using the slope at the origin. The slope was computed with Eq. \ref{}. The intersection of the fit lines is the diamond shape. \label{}} \end{figure} Once the zero determinant points are found on the $A_{k}$ axes, we can use them to determine the zero determinant region as the line joining the intersection points. This is the yellow line in part (b) of Fig. \ref{}. The yellow line and the region of large inverse determinant coincide. \subsubsection{3 energy invertible} Fig. \ref{} is a three dimension plot of the products of the diagonal and off-diagonal elements of $\mathbf{\det(M})$ as a function of $\mathbf{A}$ for well-conditioned three energy spectra. Part (b) is a plot of the spectra. Each product term is plotted as a separate color-coded surface as shown in the legend. Notice that the surfaces do not intersect so the determinant is not equal to zero in the first quadrant of the A-plane. \begin{figure}[H] \centering{}(a)\includegraphics[scale=0.6]{AlvarezFig5_1}(b)\includegraphics[scale=0.5]{AlvarezFig5_2} \protect\caption{Three dimension plot of the products of the diagonal and off-diagonal elements of $\mathbf{\det(M})$ as a function of $\mathbf{A}$ for well-conditioned three energy spectra. Each product term is plotted as a separate color-coded surface as shown in the legend. Notice that the surfaces do not intersect so the determinant is not equal to zero in the first quadrant of the A-plane. Part (b) is a plot of the spectra. \label{}} \end{figure} Fig. \ref{} shows plots of the diagonal and off-diagonal surfaces along the $A_{1}$ and $A_{2}$ axes for invertible 3 energy spectra. Notice that the actual product terms, shown by the curved lines, do not intersect. Also shown are the straight lines fit to the derivative at the origin. Their intersection, shown by the diamond icon, occurs outside the region of interest and the product term values are equal to zero, which is mathematically not possible. \begin{figure}[H] \centering{}\includegraphics[scale=0.8]{AlvarezFig6_1} \protect\caption{Plots of the diagonal and off-diagonal surfaces along the $A_{1}$ and $A_{2}$ axes for invertible 3 energy spectra. Notice that the actual product terms, shown by the curved lines, do not intersect. Also shown are the straight lines fit to the derivative at the origin. Their intersection, shown by the diamond icon, occurs outside the region of interest and the product term values at the intersection are equal to zero, which is physically not possible. \label{}} \end{figure} \subsubsection{Voltage switched x-ray tube spectra} Fig. \ref{} shows three dimension plot of the diagonal and off-diagonal product terms of $\mathbf{\det(M})$ for voltage switched (80 and 120 kV) x-ray tube spectra. The surfaces do not intersect so the determinant is not equal to zero. \begin{figure}[H] \centering{}(a) \includegraphics[scale=0.6]{AlvarezFig7_1} (b) \includegraphics[scale=0.5]{AlvarezFig7_2} \protect\caption{Three dimension plot of the products of the diagonal and off-diagonal elements of $\mathbf{\det(M})$ as a function of $\mathbf{A}$ for voltage switched (80 and 120 kV) x-ray tube spectra. Notice that the surfaces do not intersect so the determinant is not equal to zero in the first quadrant of the A-plane. Part (b) is a plot of the spectra. \label{}.} \end{figure} Fig. \ref{} shows plots of the diagonal and off-diagonal surfaces along the $A_{1}$ and $A_{2}$ axes for the voltage switched spectra. The actual product terms, shown by the curved lines, do not intersect. Also shown are the straight lines fit to the derivative at the origin. At their intersection, shown by the diamond icon, the product term values are negative, which is not physically possible. \begin{figure}[H] \centering{}\includegraphics[scale=0.8]{AlvarezFig8_1} \protect\caption{Plots of the diagonal and off-diagonal surfaces along the $A_{1}$ and $A_{2}$ axes for the voltage switched spectra. The actual product terms, shown by the curved lines, do not intersect. Also shown are the straight lines fit to the derivative at the origin. At their intersection, shown by the diamond icon, the product term values are negative, which is not physically possible. \label{}} \end{figure} \subsubsection{$\mathbf{L}$ data for non-invertible spectra} Fig. \ref{} shows the $\mathbf{L}$ data for non-invertible 3 energy spectra. Part (a) shows images where the gray scale is proportional to the $\mathbf{L}$ components. The images show the line where the determinant is zero and several lines parallel to it. Part (b) has plots of the $\mathbf{L}$ data on the lines. The values on the central point of the zero determinant line are subtracted for ease of display. Notice that the $\mathbf{L}$ data are not constant on the $det(\mathbf{M(A)})=0$ line. \begin{figure}[h] \centering{}(a) \includegraphics[scale=0.6]{AlvarezFig9_1} (b) \includegraphics[scale=0.6]{AlvarezFig9_2} \protect\caption{$\mathbf{L}$ data for non-invertible 3 energy spectra. Part (a) shows images where the gray scale is proportional to the $\mathbf{L}$ components. The images show the line where the determinant is zero and several lines parallel to it. Part (b) has plots of the $\mathbf{L}$ data on the lines. The values on the central point of the zero determinant line are subtracted for ease of display. Notice that the $\mathbf{L}$ data are not constant on the $det(M(A))=0$ line. \label{}} \end{figure} \subsection{Invert transform with iterative algorithm} The effect of the zero values of the Jacobian determinant on the invertibility of $\mathbf{L(A)}$ was tested with an iterative inverse transform algorithm. The algorithm iteratively found the value of $A$ that minimized $\left|\mathbf{L(A)-L_{input}}\right|^{2}$. It was implemented using the Matlab $fminsearch$ function. The norm of the errors of the iterative algorithm for $\mathbf{A}$ values on the green line in Fig. \ref{ were computed. They were plotted as a function of distance along the line in Fig. \ref{. The $\det(M)$ was zero on the yellow line in Fig. \ref{ and the large errors in Fig. \ref{ are near the point of intersection of the green and yellow lines. \begin{figure} \centering{}(a) \includegraphics[scale=1.1]{AlvarezFig10_1}\includegraphics[scale=0.6]{AlvarezFig10_2} \protect\caption{Effect of $det(\mathbf{M(A)})$ on errors with iterative inverse transform algorithm. The norm of the errors of the iterative algorithm for $\mathbf{A}$ values on the green line in panel (a) were computed. They were plotted as a function of distance along the line in panel (b). The $\det(M)$ was zero on the yellow line in panel (a) and the large errors in (b) are near the point of intersection of the green and yellow lines.\label{}} \end{figure} \subsubsection{Iterative algorithm with Levine ambiguous points} The iterative algorithm was also used to test the ambiguous $\mathbf{A}$ vectors noted by Levine\cite{levine2017nonuniqueness}, $[2.5,\ 1]$ and $[4.09,\ 0.13]$. Using the $\mathbf{L}$ vector computed for the first $\mathbf{A}$ as input to the iterative algorithm, the solution depended on the initial point of the iteration. If the point was near one of the two values, the algorithm converged to that value. If the point was far from both, for example the origin, the algorithm converged to $[2.5,\ 1]$. \subsection{\label{ of determinant terms with invertible and non-invertible spectra} As discussed in Sec. \ref{}, the analytical formula for signs of values of the product terms at the intersections of the line of zero Jacobian values determines whether the spectra are invertible. The formula was tested with the simulation data described in Sec. \ref{} and the results for the product terms for the three spectra tested are shown in Table \ref{}. \begin{table}[H] \protect\caption{Determinant product terms at intersections of zero Jacobian line with $A_{1}$ and $A_{2}$ axes. \label{}} \centering{}% \begin{tabular}{|l|c|c|} \hline & \textbf{$A_{1}$} & \textbf{$A_{2}$}\tabularnewline \hline \textbf{3 Energy Non-invertible} & 0.055661 & 0.063935\tabularnewline \hline \textbf{3 Energy Good Condition} & 0.0013775 & -0.0011822\tabularnewline \hline \textbf{Tube Voltage Switched} & -0.077947 & -0.084209\tabularnewline \hline \end{tabular} \end{table} The theoretical results indicated that the sign of the product terms is determined by the numerator of Eq. \ref{}. Table \ref{} shows the numerator and denominator of the product terms equation for the tested spectra. \begin{table}[H] \protect\caption{Numerator and denominator of Eq. \ref{} for the product terms of the tested spectra. \label{}} \centering{}% \begin{tabular}{|l|c|c|} \hline Numerator $M_{11}M_{22}$ & \textbf{$A_{1}$} & \textbf{$A_{2}$}\tabularnewline \hline \textbf{3 Energy Non-invertible} & 0.00035307 & 8.1111e-05\tabularnewline \hline \textbf{3 Energy Good Condition} & 2.8091e-05 & -4.3262e-06\tabularnewline \hline \textbf{Tube Voltage Switched} & -0.0055665 & -0.0010313\tabularnewline \hline \end{tabular} $\ $ $\ $ \centering{}% \begin{tabular}{|l|c|c|} \hline Denominator $M_{11}M_{22}$ & \textbf{$A_{1}$} & \textbf{$A_{2}$}\tabularnewline \hline \textbf{3 Energy Non-invertible} & 0.0063432 & 0.0012686\tabularnewline \hline \textbf{3 Energy Good Condition} & 0.020393 & 0.0036596\tabularnewline \hline \textbf{Tube Voltage Switched} & 0.071413 & 0.012247\tabularnewline \hline \end{tabular} \end{table} \section{Discussion} The invertibility of spectral measurements is an important issue. Fig. \ref{} shows that the errors with the iterative algorithm show a sharp increase near the line of non-invertibility. My previous results with a three dimension system \cite{alvarez2017noninvertibility} showed that the noise increased by a huge factor near the plane of non-invertibility. X-ray system designers need tools to determine whether their system may become non-invertible in some part of the A-vector operating range because the system may become unusable near the region of non-invertibility. In the results shown here, the non-invertible region is nearly a straight line in the two dimension A-space. In my previous paper\cite{alvarez2017noninvertibility} the non-invertible region for a three dimension basis set is a plane. The non-invertibility regions for both cases may be explained by the observation that the elements of the $\mathbf{M}$ matrix are nearly linear as a function of $\mathbf{A}$. The results in Figures \ref{}, \ref{}, and \ref{} show that the terms in the determinant of $\mathbf{M}$ are also close to linear so their intersection, which results in a zero determinant value, is close to a straight line in two dimensions or a plane in three dimensions. Sections \ref{} and \ref{} derive analytical formulas for the line of non-invertibility in the two dimension A-space using a linear approximation to the terms in the determinant of $\mathbf{M}$. The line is determined by its intersection with the A axes. For the non-invertible Levine spectrum, Fig. \ref{} shows that the actual and linear approximation intersection points are physically feasible. That is, at positive values and with positive values for the determinant terms at the intersection points. For the invertible three spike spectra, Fig. \ref{} shows that the actual determinant terms do not intersect. The linear approximation does intersect but at the intersection points the determinant product terms are negative which is not physically feasible. For the two voltage x-ray tube spectra, Fig. \ref{} shows that the actual determinant terms again do not intersect and, with the linear approximation, the determinant terms are again negative at the points of intersection. Table \ref{} gives the values for the product terms at the intersections with the axes for the three spectra tested. As indicated by the theory, they are positive for the non-invertible spectra and at least one is negative for the three energy good condition and voltage switched spectra, which are both invertible. Table \ref{} shows that the denominator of Eq. \ref{} is always positive so the sign of the result for the product term is determined by the numerator and in particular by the expression in brackets in Eq. \ref{}. Therefore, for the cases tested, the linear approximation gives useful results and the formulas give insight into the conditions leading to non-invertibility. There may, of course, be cases where the linear approximation does not give accurate results. The approach taken was a trade-off between accuracy and simplicity of the formulas. Simple formulas allow us to gain physical insight into the invertibility conditions at the expense of lower accuracy. Levine\cite{levine2017nonuniqueness} introduces an alternate approach to the dual energy inversion problem to that described here. The formulation focuses on the exponent of the transmission equation, Eq. \ref{}. It puts the logarithm of the spectral density in the exponent and combines them with $\mathbf{A\cdot f(E)}$. The combination is used to derive A-space surfaces specified by the measurement spectra. The $\mathbf{A}$ vectors that invert the equations are at the intersection of the surfaces. With a particular set of spectra, Levine found two A-vector points where the surfaces intersect. The transformation is therefore not one-to-one and is not invertible. Interestingly, the two ambiguous points are not on the line of zero determinant. The results in Fig. \ref{} show that the $\mathbf{L}$ data on the line of non-invertibility are not constant. Instead, the $\mathbf{L}$ data vary but the two product terms of the determinant balance out resulting in a zero determinant value on this line. Research is on-going into more accurate theoretical equations for invertibility and ways to apply the formulation described in this paper to clinical systems design. \section{Conclusion} An approach to studying the invertibility of spectral measurements is presented and applied to invertible and non-invertible spectra. Non-invertible dual energy spectra have zero Jacobian determinant in the first quadrant of the $\mathbf{A}$ vector plane. For the non-invertible spectra tested, the Jacobian determinant is equal to zero on a curve in the $\mathbf{A}$ vector plane that is close to a straight line. An analytical formula is derived for a linear approximation to the points where the line crosses the A-axes. If the spectra are invertible, the intersection points are not physically feasible while for the non-invertible spectra they are feasible indicating zero Jacobian determinant values in the first quadrant.

\section{Introduction} With the rise of machine learning and the tasks such as automated teaching and assessment, there is an increased interest in understanding how machine learning models can be grounded in theories of language acquisition. Additionally, with the abundance of learner data in archive and generation, we now have an avenue through which we can not only evaluate our theories of learning, but also explore whether these theories can be used to train agents for the purpose of general AI. Language Acquisition is a multidisciplinary field that includes linguistics, psychology, neuroscience, philosophy, and more recently computer science. At the intersection of language acquisition and pedagogy lie theories of educational practices for language learners, including for example, an optimal curriculum for both L1 and L2 learners. A {\it{curriculum}} is a guide that helps teachers decide what content to present and the order of which it needs to be presented. The aim of a curriculum is to provide a highly structured method of introducing concepts in order to maximise the rate of learning. The idea of a curriculum to facilitate the rate of learning has been discussed from the perspective of {\it{animal training}} \cite{skinner1958teaching,peterson2004day}, where it is defined as {\it{shaping}}. It has also been referenced in an educational framework \cite{bruner1960process} where the author introduces the idea of a {\it{spiral curriculum}}, a process by which complex information is first presented in a simplified manner and then revisited at a more difficult level later on. Similarly Vygotsky, from the view of language acquisition, introduces the idea of {\it{scaffolding}} in order to provide contextual support for more complex ideas using simplified language or visuals. All of these concepts have been discussed in different fields but reference the same underlying idea of presenting information in a structured manner in order to exploit prior knowledge. Bruner \cite{bruner1961act} argues that the role of the teacher is not to present information by rote learning but rather facilitate the learning process in order to teach students to become {\it{active learners}}: put simply, they are ``learning to learn''. There are many factors that teachers need to consider when constructing a curriculum to achieve this goal, namely the {\it{difficulty}} and {\it{appropriateness}} of content. Difficulty is measured relative to the {\it{zone of proximal development (ZDP)}}, introduced by Vygotsky, which is a representation of what a learner is capable of achieving without help, with {\it{some}} help, and of concepts that are beyond the learner's current ability. Appropriateness is a measure of whether content being presented is within the ZPD or, in the case of scaffolding, comprises material from within the ZPD. Determining difficulty and appropriateness is traditionally a very laborious and resource intensive task which entails experts conducting focus groups and analysis to decide where a particular question or topic sits in the curriculum. This method is not only inefficient, it also assumes a static curriculum for all students. To address these limitations, we propose the use of reinforcement learning (RL) in order to learn an optimal policy and curriculum for each student in the task of visual vocabulary acquisition. Through this, we also discuss the similarities between the properties and features of RL and those of language acquisition. We evaluate our models by simulating three types of student at different levels of proficiency {\it{(beginner, intermediate, and advanced)}}. We find that the system is able to identify the difference in proficiency and adapt its curriculum to reflect. Previous uses of RL in pedagogy include \cite{beck2000advisor} where it is used to teach students arithmetic, aiming to minimise the time taken to answer questions. \cite{iglesias2009learning,iglesias2003experience} teach students database design using Q-learning. Both \cite{beck2000advisor} and \cite{iglesias2009learning,iglesias2003experience} evaluate results on simulated students. \cite{martin2004agentx} use RL for maths while \cite{tetreault2006comparing} use it for physics. However, as far as we know, no previous work has been done in the space of visual lexical acquisition where the principles of RL have explicitly been related to theories of language acquisition. The importance of curriculum learning in training deep learning models and agents has also been discussed by \cite{bengio2009curriculum} where its use is shown to facilitate the generalisation as well as the rate of convergence and training of deep learning networks. \cite{hermann2017grounded} also illustrate the need for some form of curriculum to improve the rate of learning for agents in a 3D simulation. However, it is worth noting that no explicit RL is used to model curriculum by either \cite{bengio2009curriculum} or \cite{hermann2017grounded}. \section{Curriculum Q-Learning} In order to automate the process of curriculum learning for visual vocabulary acquisition, we must first identify the key components of our RL system. The agent in this task is the automated tutor that must learn what information to present to the student. The environment is the student who is interacting with the agent. We assume that the student interacting with the tutor is a learner of English who has reached a given level on the \emph{Common European Framework of Reference} (CEFR) scale. CEFR is an international standard for describing language ability, using a six point scale, from A1 for beginners, up to C2 for those who have mastered a language. The RL algorithm used by our proposed system is Q-Learning, an {\it{off-policy}} algorithm for Temporal Difference (TD) Learning. Q-Learning can be defined as follows: \begin{equation} Q(s,a) \leftarrow Q(s,a) + \alpha [r + \gamma \max_{a'}Q^\pi(s',a') - Q(s,a)] \end{equation} where $Q(s,a)$ is the Q-value of a state $s$ and action $a$ tuple. The $\alpha$ is the learning rate and $\gamma$ is the discount factor. $\gamma$ models the fact that future rewards are less valuable than immediate rewards at a given time $t$. A policy $\pi$ maps states $s$ to actions $a$. The aim of the Q-Learning algorithm is to find an optimal policy $\pi$ such that it maximises the long-term cumulative reward. The policy achieves this by acting greedily and taking the action that presents the maximum Q-value given the state such that $\max_{a\in A}Q^\pi(s_{t},a)$. In action selection, there is a trade-off between {\it{exploiting}} what you have learnt so far and {\it{exploring}} other state-action tuples. In this task we model that using $\epsilon$-greedy. This means the policy will, for most part, select the actions that provide the highest estimated future reward given the state. However, with a probability of $1-\epsilon$, an action will be selected randomly and independently from a uniform distribution. Action selection is usually drawn from a Q-Table which is a table that stores all state-action Q-values. In this task, a policy can be viewed as a curriculum as it decides what should be shown and in what order. In order to learn a curriculum for vocabulary acquisition, we incorporate two models, the {\it{CEFR level model}} and {\it{word level model}}. The CEFR level model has 6 states which are defined by the 6 CEFR levels. The actions are whether the student should progress to the next level, stay in the current level, or go back a level. The word level model has two states: active (show the word), inactive (hide the word). The actions are remain in the current state and toggle state. This architecture ensures that there is also an estimated long-term reward associated with showing a student a particular word. Modelling reward is often viewed as a challenging task in RL. For this application, a student is rewarded negatively (-1) for getting a question correct and positively (+1) for getting it incorrect. The motivation behind using these values is grounded in how we learn. As the RL model acts greedily and takes the action with the maximum reward, if we review a concept we understand, then we are not gaining in terms of knowledge by reviewing it again and thus its value should be reduced. Alternatively, if we get a question wrong, the benefit of reviewing that word is higher and thus we should increase the associated Q-value. To evaluate the students' understanding, we present a word in a form of an image. The objective for the students is to describe the image, and based on their response, the Q-Learning algorithm and thus the policy is updated. A valid response is defined by the target word associated with the image or a synonym of that target word which is automatically generated by looking at the top 10 nearest words to the target word in a pre-trained {\it{word2vec}} model \cite{mikolov2013efficient}. The use of images was motivated by it's inexpensive nature of producing evaluation material. Additionally, there are countless studies that indicate the effectiveness of images for learning \cite{verdi1997organized}. \section{Experiments} For the CEFR level model, we use a learning rate $\alpha$ of $0.1$, a discount rate $\gamma$ of $0.9$ and an $\epsilon$ value of $0.95$. The word level model uses an $\alpha$ of $0.1$, a $\gamma$ of $0.9$ and an $\epsilon$ value of 1 in order to prevent words randomly going into an inactive state. To evaluate the performance of our system, we simulated three types of students at varying levels of proficiency: {\it{beginner, intermediate and advanced}}. In this case, we modelled the student's probability of getting a question correct as a negated Gompertz distribution: \begin{equation} P(success~|~u,q) = 1 - \exp(-b~\exp(-c(l(q) - l(u)))) \end{equation} where $l(u)$ denotes the level of user $u$ calibrated to a scale of $\left[0,6\right]$. Each integer in the scale represents a corresponding CEFR level from A1 to C2 (\emph{e.g.} 0 $\rightarrow$ A1, 1 $\rightarrow$ A2, \emph{etc.}). $l(q)$ represents the level of an item $q$ (i.e. a word which must be guessed from an image) calibrated to the same scale. The parameter $b$ determines the probability of success when student and item level match. This is set to $ln(0.75)$ to model a `typical' pass rate of 75\%. The calibrated curve is shown in Appendix \ref{gompertz}. The curve is flatter at the lower end as students may be expected to be comfortable with most of the material at lower CEFR levels than their own, whereas at higher levels, their ability is more uncertain. We ran simulations where each student had 100 interactions with the system. An interaction can be defined as when a student responds to a question. \begin{figure}[!tbp] \centering \begin{minipage}[b]{0.49\textwidth} \includegraphics[width=\textwidth]{levelv2.png} \caption{CEFR levels determined by the agent for students of varying levels of proficiency over 100 interactions} \label{} \end{minipage} \hfill \begin{minipage}[b]{0.49\textwidth} \includegraphics[width=\textwidth]{rewardv3.png} \caption{Cumulative reward earned by students of varying levels of proficiency from the agent over 100 interactions} \label{} \end{minipage} \end{figure} \subsection{Results} The results from Figure \ref{} show how the agent responds to the various proficiency levels. The beginner student remains relatively constant around A1 and A2 which is reflective of the student's current level. The intermediate student continually increases in CEFR level until level 3 (B2). The advanced student, although is tested with material beneath the actual level of proficiency, eventually reaches an advanced or higher CEFR level. We can also see that the agent tutor pushes the student to what can be interpreted as the edge of their ZDP. Figure \ref{} illustrates how the cumulative reward of the students varies for students at different proficiencies. The curve experiences a downward slope as the students reach their current level of vocabulary and are now being pushed to understand material beyond their scope. \section{Discussion} We have shown through the use of simulations, that we can effectively model a personalised curriculum for vocabulary acquisition using Q-Learning. Figure \ref{} and Figure \ref{} show clear indications of varying agent behaviour for students at different levels of lexical proficiency. However, beyond that, we have set up a framework that can be used in the future to extrapolate the difficulty and appropriateness of new material. The system will serve as a test bed that will yield metrics to determine where the content fits in the curriculum. Although this is foundational work, it lays the building blocks for future pedagogically inspired RL architectures. Through this work, we have also shown that there are many similarities between the principles of RL and theories of language acquisition. Specifically, parallels can be drawn between the concept of $\epsilon$-greedy and Krashen's Input Hypothesis or the {\it{i+1}}. The Input Hypothesis states that students progress their learning by comprehending language that is slightly above their current language level. The interactions between the agent and the environment in RL is analogous to the social interaction approach to language acquisition, specifically equal importance of input and output. We also use the Q-Learning algorithm as opposed to the SARSA algorithm mainly due to the properties of Q-Learning that ensure an "optimal path" is followed i.e. the minimum number of steps to reach our goal (language fluency). However, there is scope for substantial extensions in this space. Deploying the system on-line in order to collect user data will allow us to validate and improve our existing models. Incorporating memory and spaced repetition learning in order to optimise the policy and emulate cognitive processes is also an important extension that may have a great impact on the learning output. Using deep learning models to approximate the Q-value will allow the system to capture additional signals pertinent to language acquisition. Additionally, moving towards an adaptive reward model that reflects difficulty to encourage memory retention. All of these models can also be applied to agents instead of students. As discussed previously, \cite{hermann2017grounded} indicated the need for a curriculum in order to effectively train an agent in the simulated environment. Creating a dynamic environment guided by a curriculum grounded in pedagogically inspired RL may result in improved learning rates for the agent. \subsubsection*{Acknowledgements} We thank Wenchao Chen who helped develop the back-end of our web-based platform. \newpage \bibliographystyle{unsrt}

\section{Approach Overview} \label{} In typical management and orchestration frameworks~\cite{etsi-mano}, service providers need to submit exact descriptors of their network service structure, resource demands, and expected traffic from sources to a service management and orchestration system (Fig.~\ref{}). Based on the descriptors, placement, scaling, and routing decisions are made for each network service, independently from one another. \begin{figure}[t] \centering \includegraphics[width=0.9\columnwidth]{orch2} \caption{Conventional network service life-cycle, from descriptors to running services} \label{} \end{figure} \begin{figure}[t] \centering \includegraphics[width=0.8\columnwidth]{orch-jasper2} \caption{Network service management and orchestration using JASPER} \label{} \end{figure} JASPER makes two major changes to this approach, one with respect to the description of the network services (Section~\ref{}) and another with respect to handling the scaling, placement, and routing decision processes (Section~\ref{}). \subsection{Templates instead of over-specified descriptors} \label{} Because of the limited precision and flexibility of typical descriptors, we base our approach on so-called \emph{service templates}. Using service templates, service providers are required to specify neither the \emph{exact} resource demands (e.g., memory or CPU) of service components nor the required number of instances of each component. The service template describes the components of the network service and their required interconnections on an abstract level, without deployment details. Moreover, it gives the resource demands of the network service as a function of the load: \begin{itemize} \item The required computational capacity (e.g., CPU and memory) is described for each service component as a function of the input data rate. This can be used to calculate the network node capacity required to host the service component. \item The amount of traffic leaving each service component towards other components is specified as a function of the data rate that enters the component. This can be used to calculate the link capacity required to host the traffic flowing between any two interconnected instance. \end{itemize} In addition to the service templates, service providers may include the expected traffic originating from the sources of the network service in the request to embed a service template. As the traffic is constantly changing, the current traffic needs to be monitored and fed back to the template embedding process, to keep the network service in an optimal state. In this way, depending on the location and data rate of the sources of the network service, resource requirements are calculated dynamically, based on the given functions, eliminating the risk of over- or under-estimating the resource demands. Based on the functions describing the dependency of resource requirements and outgoing data rates on incoming data rates, it is also possible to reason about possible changes to the deployment and their impact, which is a pre-requisite for effective optimization. The specific functions highly depend on the type and implementation of the service component and can be derived, for example, based on historical usage data or by automatic service profiling methods~\cite{profiling2016}. \subsection{Joint, single-step scaling, placement, and routing} \label{} As shown in Fig.~\ref{}, in typical management and orchestration frameworks~\cite{etsi-mano}, based on the description of the network service and the state of the network's resources, the requested number of instances for each service component are computed and then placed and deployed with the requested amount of resources, in an appropriate location. After path selection and instantiation of the network service, the running instances are monitored and re-scaled and re-placed based on pre-defined scaling rules if required. Deciding the number of required instances for each service component, the amount of resources allocated to each component, and the optimal paths selected for routing the network service flows are, however, highly interdependent problems, which cannot be solved optimally using such independent management and orchestration steps. Our approach, illustrated in Fig.~\ref{}, changes the way network service life-cycle is handled, by combining scaling, placement, and routing steps into a joint decision process. Depending on the location and data rate of the sources, \begin{itemize} \item each service template is scaled out into an overlay with the necessary number of instances required for each service component; \item each component instance is mapped to a network node and is allocated the required amount of resources on that node; \item the connections among component instances are mapped to flows along network links, carrying the data rate. \end{itemize} JASPER is an integrated approach in multiple dimensions: (i) scaling, placement, and routing decisions are made in a single optimization step; (ii) all services that are to be placed in the same substrate network are considered together; (iii) newly requested and already deployed services are optimized jointly. This way, a global optimum can be achieved. Modern management and orchestration systems~\cite{osmwebsite,sonatawebsite,unifywebsite} have a flexible design to incorporate innovative life-cycle management approaches. For example, SONATA's service platform~\cite{sonata-paper} has a customizable service life-cycle management plugin. The platform operator can easily modify the order of life-cycle management operations and customize different operations. Using service-specific management programs it is also possible to specify when and how scaling, placement, and routing operations are performed for each network service, making the practical implementation of JASPER possible. \section{Complexity} \label{} \begin{theorem} \label{} For an instance of the Template Embedding problem as defined in Section~\ref{}, deciding whether a solution with no violations exists is NP-complete in the strong sense\footnote{NP-complete in the strong sense means that the problem remains NP-complete even if the numbers appearing in it are constrained between polynomial bounds. Under the P$\ne$NP assumption, this precludes even the existence of a pseudo-polynomial algorithm -- i.e., an algorithm the runtime of which is polynomial if restricted to problem instances with polynomially bounded numbers.}. \end{theorem} \begin{proof} It is clear that the problem is in NP: a possible witness for the positive answer is a solution -- i.e., a set of overlays and their embedding into the substrate network -- with 0 violations. The witness has polynomial size and can be verified in polynomial time wrt.\ to the input size. To establish NP-hardness, we show a reduction from the Set Covering problem (which is known to be NP-complete in the strong sense \cite{karp1972reducibility}) to the Template Embedding problem. An input of the Set Covering problem consists of a finite set $U$, a finite family $\cal W$ of subsets of $U$ such that their union is $U$, and a number $k\in\mathbb{N}$. The aim is to decide whether there is a subset $\cal Z\subseteq W$ with cardinality at most $k$ such that the union of the sets in $\cal Z$ is still $U$. From this instance of Set Covering, an instance of the Template Embedding problem is created as follows. The substrate network consists of nodes $V=\{s_1,\ldots,s_{|U|}\}\cup\{a_1,\ldots,a_{|{\cal W}|}\}\cup\{b\}$, where each $s_i$ represents an element of $U$ and each element $a_j$ represents an element of $\cal W$. There is a link from $s_i$ to $a_j$ if and only if the element of $U$ represented by $s_i$ is a member of the set represented by $a_j$. Furthermore, there is a link from each $a_j$ to $b$. The capacities of the nodes are as follows: $\text{cap}_{\text{cpu}}(s_i)=\text{cap}_{\text{mem}}(s_i)=0$ for each $i\in[1,|U|]$, $\text{cap}_{\text{cpu}}(a_j)=0$ and $\text{cap}_{\text{mem}}(a_j)=1$ for each $j\in[1,|{\cal W}|]$, and $\text{cap}_{\text{cpu}}(b)=1$ and $\text{cap}_{\text{mem}}(b)=0$. For each link, its maximum data rate is 1, its delay is 0. There is a single template consisting of a source component $S$ and two further components $A$ and $B$, and two arcs $(S,A)$ and $(A,B)$. Component $A$ has one input and one output, its resource consumption as a function of the input data rate $\lambda$ is given by $p_A(\lambda)=0$ and $m_A(\lambda)=1$; its output data rate is given by $r_A(\lambda)=1$. Component $B$ has one input and no output, its resource consumption as a function of the input data rate $\lambda$ is given by \begin{equation*} p_B(\lambda)= \begin{cases} 1, & \text{if }\lambda\le k, \\ 2, & \text{otherwise,} \end{cases} \end{equation*} and $m_B(\lambda)=0$. In each $s_i$, there is a source corresponding to an instance of $S$ with data rate $\lambda=1$. \begin{figure}[tb] \centering \subfigure[An instance of Set Covering ($k=2$) and a solution (thick lines)]{\hspace*{0.05\columnwidth}\includegraphics[width=0.35\columnwidth]{proof_setcover}\hspace*{0.05\columnwidth}} \hspace*{0.05\columnwidth} \subfigure[The generated instance of Template Embedding and the corresponding solution]{\includegraphics[width=0.45\columnwidth]{proof_te}} \caption{An example for the proof of Theorem \ref{}} \label{} \end{figure} Suppose first that the original instance of Set Covering is solvable, i.e., there is a subset $\cal Z\subseteq W$ with cardinality at most $k$ such that the union of the sets in $\cal Z$ is $U$. In this case, the generated instance of the Template Embedding problem can also be solved without any violations, as follows (see Fig.\ \ref{} for an example). Each $s_i$ must of course host an instance of $S$. In each $a_i$ corresponding to an element of $\cal Z$, an instance of $A$ is created. Since the union of the sets in $\cal Z$ is $U$, each $s_i$ has an outgoing link to at least one $a_j$ hosting an instance of $A$, which can be selected as the target of the traffic leaving the source in $s_i$ through the link $(s_i,a_j)$. Further, a single instance of $B$ is created in node $b$ and each instance of $A$ is connected to $B$ through the $(a_j,b)$ link. Since the number of instances of $A$ is at most $k$, each emitting traffic with data rate 1, the CPU requirement of the instance of $B$ is 1, so that it fits on $b$, and hence we obtained a solution to the Template Embedding problem with no violation. Now assume that the generated instance of the Template Embedding problem is solvable without violations. Then, we can construct a solution of the original instance of Set Covering, as we show next. In a solution of the generated instance of the Template Embedding problem, each $s_i$ must host an instance of $S$ and there is no other instance of $S$. Instances of $A$ can only be hosted by $a_j$ nodes because of the memory requirement, and an instance of $B$ can only be hosted in $b$ because of the CPU requirement. We define $\cal Z$ to contain those elements of $\cal W$ for which the corresponding node $a_j$ hosts an instance of $A$. Since each source generates traffic that must be consumed by an instance of $A$ and there is a path (actually, a link) from $s_i$ to $a_j$ only if the set corresponding to $a_j$ contains the element corresponding to $s_i$, it follows that the sets in $\cal Z$ cover all elements of $U$. Moreover, since the instance of $B$ must fit on $b$ and each instance of $A$ generates traffic with data rate 1, it follows that the number of instances of $A$ is at most $k$ and hence $|{\cal Z}|\le k$, thus $\cal Z$ is a solution of the original Set Covering problem. Since all numbers in the generated instance of the Template Embedding problem are constants, this reduction shows that the Template Embedding problem is indeed NP-hard in the strong sense. \end{proof} As a consequence, we can neither expect a polynomial or even pseudo-polynomial algorithm for solving the problem exactly nor a fully polynomial-time approximation scheme, under standard assumptions of complexity theory. \section{Conclusions} \label{} We have presented JASPER, a fully automatic approach to scale, place, and route multiple virtual network services on a common substrate network. JASPER can be used for both the initial allocation of newly requested services and the adaptation of existing services to changes in the demand. Besides formally defining the problem and proving its NP-hardness, we developed two algorithms for it, an MILP-based one and a custom constructive heuristic. Empiric tests have shown how our approach finds a balance between conflicting requirements and ensures that the allocated capacity quickly follows changes in the demand. The MILP-based algorithm gives optimal or near-optimal results for relatively small substrate network graphs, making it suitable for, e.g., calculations on top of a geographically distributed network where each node represents a data center. The heuristic remains very fast for even the largest networks that were tested. Overall, the tests gave evidence to the feasibility of our approach, which makes it possible (i) for service developers to specify services at a high level of abstraction and (ii) for providers to quickly re-optimize the system state after changes. Promising future research directions include, beside further algorithmic enhancements to the presented algorithms and the development of new algorithms, the consideration of queuing incoming requests in the service components and the investigation of the effects of cyclic service templates. \section{Evaluation} \label{} We implemented the presented algorithms in the form of a C++ program. For solving the MILP, Gurobi Optimizer 7.0.1\footnote{\url{}} was used. For substrate networks, we used benchmarks for the Virtual Network Mapping Problem\footnote{\url{}} from Inf\"uhr and Raidl \cite{infuhr2013solving}. As service templates, we used examples from IETF's Service Function Chaining Use Cases \cite{draft-liu-sfc-use-cases-08}. \subsection{An example} \begin{figure}[tb] \centering \includegraphics[height=55mm]{graph} \caption{Example substrate network} \label{} \end{figure} \begin{figure*}[tb] \centering \subfigure[\label{ embedding]{\includegraphics[height=54mm]{state_milp_2}} \hfil \subfigure[\label{ of increased source data rate]{\includegraphics[height=54mm]{state_milp_3}} \hfil \subfigure[\label{ of the emergence of a second source]{\includegraphics[height=54mm]{state_milp_4}} \caption{Illustrative example (memory values not shown for better readability)} \end{figure*} First, we illustrate our approach on a small substrate network of 10 nodes and 20 arcs (see Fig.~\ref{}) in which the CPU and memory capacity of each node is both 100. In this network, a service consisting of a source (S), a firewall (FW), a deep packet inspection (DPI) component, an anti-virus (AV) component, and a parental control (PC) component is deployed. Initially, there is a single source in node 1 with a moderate data rate. As a result, our algorithm deploys all components of the service in node~1 (see Fig.~\ref{}). Subsequently, the data rate of the source increases. As a result, the resource demand of the processing components of the service increases so that they do not fit onto node~1 anymore. Our algorithm automatically re-scales the service by duplicating the DPI, AV, and PC components and automatically places the newly created instances on a nearby node, namely node~3 (see Fig.~\ref{}). Later on, a second source emerges for the same service on node~9. The algorithm automatically decides to create new processing component instances on node~9 to process as much as possible of the traffic of the new source locally. The excess traffic from the new FW instance that cannot be processed locally due to capacity constraints is routed to the existing DPI, AV, and PC instances on node~3 because node~3 still has sufficient free capacity (see Fig.~\ref{}). Already this small example shows the difficult trade-offs that template embedding involves. Next, we show that our approach is capable of handling also much more complex scenarios. \subsection{Comparison of the algorithms} \begin{figure}[tb] \centering \includegraphics[width=\columnwidth]{scenario_streaming_08_with_heur} \caption{Temporal development of the demand and the allocated capacity in a complex scenario} \label{} \end{figure} \begin{figure}[tb] \includegraphics[width=0.9\columnwidth]{latency} \caption{Total latency over all created paths for the embedded template} \label{} \end{figure} We consider a substrate network with 20\,nodes and 44\,arcs, in which multiple services are deployed. Each service is a virtual content delivery network for video streaming, consisting of a streaming server, a DPI, a video optimizer, and a cache. The number of concurrently active services varies from 0 to 4, the number of sources varies from 0 to 20. Fig.~\ref{} shows how the total data rate of the sources (as a metric of the demand) and the total CPU size of the created instances (as a metric of the allocated processing capacity) change through re-optimization after each event. An event is the emergence or disappearance of a service, the emergence or disappearance of a source, or the change of the data rate of a source. As can be seen, the allocated capacity using both the heuristic and the MILP algorithms follow the demand very closely, meaning that our algorithms are successful in scaling the service in both directions to quickly react to changes in the demand. Regarding total data rate and total latency of the overlay edges, the MILP algorithm performs better than the heuristic algorithm. For example, Fig.~\ref{} shows the total latency over all paths created for the template in this scenario\footnote{In Fig.~\ref{}, in the high-load area between event 20 and 50, some problem instances are too complex to be solved within the 60 seconds time limit we have set for the optimizer. This results in solutions with zero latency, as no paths are created.}. The reason for this difference is that in the MILP algorithm, the optimal location for all required instances can be determined at the same time. This results in shorter distances between the source and the instances. The heuristic algorithm, however, needs to create instances one by one, resulting in larger data rates traveling over larger distances in the substrate network. In this scenario, to handle the peak demand, a total of 127 instances are created using the MILP algorithms, while the heuristic algorithm creates 261 instances. \subsection{Scalability} \begin{figure}[tb] \centering \subfigure[\label{ time of the MILP algorithm]{\includegraphics[width=0.9\columnwidth]{exec_time_milp}} \subfigure[\label{ gap of the MILP algorithm]{\includegraphics[width=0.9\columnwidth]{gap_milp}} \subfigure[\label{ time of the heuristic algorithm]{\includegraphics[width=0.9\columnwidth]{exec_time_heur}} \caption{Scalability of the presented algorithms} \end{figure} Since the template embedding problem is NP-hard, it is foreseeable that the scalability of the MILP solver will be limited. In order to test this, we gradually increase the source data rate of the service from our first experiment, leading to an increasing number of instances; moreover, we also consider substrate networks of increasing size. In each case, the MILP solver is run with a time limit of 60 seconds, meaning that the solution process stops at (roughly) 60 seconds with the best solution and the best lower bound that the solver found until that time. The measurements were performed on a machine with Intel Core i5-4210U CPU @ 1.70GHz and 8GB RAM. Fig.~\ref{} shows the execution time of the MILP algorithm for different data rates and substrate network sizes, while Fig.~\ref{} shows the corresponding gap between the found solution and the lower bound. As can be seen, for a small network with 10 nodes and 20 arcs, the algorithm computes optimal results for the lower half of source data rate values, and even for larger source data rates, the optimality gap is quite low (around 20\,\%), meaning that the results are almost optimal. However, for a bigger substrate network with 20 nodes and 44 arcs, the solver reaches the time limit for much smaller source data rate and also the optimality gap is much bigger. For even bigger substrate networks, the performance of the algorithm further deteriorates, up to the point where it cannot be run anymore because of memory problems. The large sensitivity to the size of the substrate network is not surprising, given that the number of variables of the MILP is cubic in the size of the substrate network. In contrast, as shown in Fig.~\ref{}, the execution time of the heuristic algorithm remains very low even for the largest substrate networks: for 1000 nodes and 2530 arcs, the execution time is still below 20 milliseconds, rendering the heuristic practical for real-world problem sizes as well. \section{Heuristic approach} \label{} Now we present a heuristic algorithm that is not guaranteed to find an optimal solution but is much faster than the mixed integer programming approach. Moreover, it has the advantage that it does not require the functions $p_j$, $m_j$, and $r_j$ to be linear. The heuristic constructs the new solution from the existing one by means of a series of small local changes.\footnote{Also the placement of a new service is done with a series of small local changes, creating component instances one by one.} While doing so, it has to be ensured that (i) the instantiation of source components is in line with the given data sources, (ii) the data flows produced by each instance are routed to appropriate instances, and (iii) capacity constraints are satisfied as much as possible. This can be achieved by iterating through the instances of each overlay once in a topological order, possibly creating new instances on the fly if necessary. Note that this may indeed be necessary, for example, if a new data source appeared or the output data rate of a data source increased. In each step, the algorithm aims at economical use of resources, e.g., by only creating new instances if necessary, deleting unneeded instances, or preferring short paths. The heuristic is shown in Algorithm~\ref{}. It starts by checking that each service has a corresponding overlay and each overlay corresponds to a service (lines 1--5). If a new service has been started or an existing service has been stopped since the last invocation of the algorithm, the corresponding overlay is created or removed at this point. Next, the mapping of the sources and source components is checked and updated if necessary (lines 6--11): if a new source emerged, an instance of the corresponding source component is created; if the data rate of a source changed, then the output data rate of the corresponding source component instance is updated; if a source disappeared, then the corresponding source component instance is removed. Finally, to propagate the changes of the sources to the processing instances, we need to iterate over all instances and ensure that the new output data rates, which are determined by the new input data rates, are discharged correctly by outgoing flows (lines 12--24). For this purpose, it is important to consider the instances in a topological order (according to the overlay) so that when an instance is dealt with, its incoming flows have already been updated. If a change in the outgoing flows is necessary, then the \textsc{increase} or \textsc{decrease} procedures are called. \begin{algorithm}[htb] \caption{Main procedure of the heuristic algorithm} \label{} \begin{algorithmic}[1] \small \If{$\exists G_{\text{OL}}(T)$ with $T\not\in {\cal T}$} \State remove $G_{\text{OL}}(T)$ \EndIf \ForAll{$T\in {\cal T}$} \If{$\nexists G_{\text{OL}}(T)$} \State create empty overlay $G_{\text{OL}}(T)$ \EndIf \ForAll{$(v,j,\lambda)\in S(T)$} \If{$\nexists i\in I_{\text{OL}}$ with $c(i)=j$ and $P_T^{(I)}(i)=v$} \State create $i\in I_{\text{OL}}$ with $c(i)=j$ and $P_T^{(I)}(i)=v$ \EndIf \State set output data rate of $i$ to $\lambda$ \EndFor \If{$\exists i\in I_{\text{OL}}$, where $c(i)$ is a source component but $\nexists (P_T^{(I)}(i),c(i),\lambda)\in S(T)$ for any $\lambda$} \State remove $i$ \EndIf \ForAll{$i\in I_{\text{OL}}$ in topological order} \If{all input data rates of $i$ are 0} \State remove $i$ and go to next iteration \EndIf \State compute output data rates of $i$ \ForAll{output $k$ of $i$} \State $\Phi$: set of flows currently leaving output $k$ \State $\lambda$: sum of the data rates of the flows in $\Phi$ \State $\lambda'$: new data rate on output $k$ \If{$\lambda'<\lambda$} \State $\cal E$: set of edges leaving output $k$ \State \Call{decrease}{$\cal E$,$\lambda-\lambda'$} \ElsIf{$\lambda'>\lambda$} \State \Call{increase}{$i$,$k$,$\Phi$,$\lambda'-\lambda$} \EndIf \EndFor \EndFor \EndFor \end{algorithmic} \end{algorithm} The auxiliary subroutines are detailed in Algorithm~\ref{}. \textsc{decrease} first removes as many edges as possible (lines 3--6); when a further decrease is necessary but no more edges can be removed, it reduces the next flow on each link by the same factor to achieve the required reduction (lines 7--9). \textsc{increase} first checks if new instances need to be created to be consistent with the template (lines 12--16), then tries to increase the existing flows (lines 17--19). If this is not sufficient to achieve the necessary increase, it creates further instances and flows (lines 20--23). \begin{algorithm}[htb] \caption{Auxiliary methods of the heuristic} \label{} \begin{algorithmic}[1] \small \State\Comment{Decrease the flows on the edges in $\cal E$ by $\Delta\lambda$ in total} \Procedure{decrease}{$\cal E$,$\Delta\lambda$} \State sort $\cal E$ in non-decreasing order of flow data rate \ForAll{$e\in {\cal E}$ while flow data rate $\lambda(e)\le\Delta\lambda$} \State $\Delta\lambda:=\Delta\lambda-\lambda(e)$ \State remove $e$ \EndFor \If{$\Delta\lambda>0$} \State let $e$ be the next edge \State reduce flow of $e$ by a factor of $(\lambda(e)-\Delta\lambda)/\lambda(e)$ \EndIf \EndProcedure \State\Comment{Increase the flows in $\Phi$ leaving output $k$ of instance $i$ by $\Delta\lambda$ in total} \Procedure{increase}{$i$,$k$,$\Phi$,$\Delta\lambda$} \ForAll{arc $(c(i),j)$ leaving output $k$ of $c(i)$} \If{$\nexists i'\in I_{\text{OL}}$ with $c(i')=j$ and $ii'\in E_{OL}$} \State $\varphi:=\;$\Call{createInstanceAndFlow}{$j$, $i$, $\Delta\lambda$} \State $\Delta\lambda:=\Delta\lambda-(\text{data rate of } \varphi)$ \State $\Phi:=\Phi\cup\{\varphi\}$ \EndIf \EndFor \ForAll{$\varphi\in\Phi$} \State $d:=\;$\Call{incrFlow}{$\varphi$,$\Delta\lambda$} \State $\Delta\lambda:= \Delta\lambda-d$ \EndFor \While{$\Delta\lambda>0$} \State $(c(i),j)$: random arc leaving output $k$ of $c(i)$ \State $\varphi:=\;$\Call{createInstanceAndFlow}{$j$, $i$, $\Delta\lambda$} \State $\Delta\lambda:=\Delta\lambda-(\text{data rate of }\varphi)$ \EndWhile \EndProcedure \State\Comment{Create an instance of component $j$ with flow from instance $i$ of high data rate (capped at cutoff)} \Procedure{createInstanceAndFlow}{$j$,$i$,cutoff} \ForAll{$v\in V$} \State create temporary instance $i'$ of $j$ on $v$ \State $\varphi$: flow of data rate 0 from $i$ to $i'$ \State \Call{incrFlow}{$\varphi$,cutoff} \State remove $i'$ and $\varphi$ \EndFor \State create instance of $j$ on node resulting in best flow \EndProcedure \State\Comment{Increase flow data rate by at most $d$} \Procedure{incrFlow}{$\varphi$,$d$} \State $v:=\;$start node of $\varphi$ \State $v':=\;$end node of $\varphi$ \State $\beta_1:=\;$maximum flow based on $cap_{CPU}(v')$ \State $\beta_2:=\;$maximum flow based on $cap_{mem}(v')$ \State $d:=\min(d,\beta_1,\beta_2)$ \State $P$: $v\leadsto v'$ path of high bandwidth ($b$) and low latency \State increase $\varphi$ by $\min(b,d)$ along $P$ \EndProcedure \end{algorithmic} \end{algorithm} In the \textsc{createInstanceAndFlow} procedure (called by \textsc{increase} to create a new instance of a component together with a flow from an existing instance), all nodes of the substrate network are temporarily tried for hosting the new instance. The candidate that leads to the best flow is selected (lines 26--31). Finally, the \textsc{incrFlow} procedure (called by both \textsc{increase} and \textsc{createInstanceAndFlow}) increases the data rate of a flow along a new path (lines 34--40). As can be seen, we avoid computing maximum flows. This is because the running time of the best known algorithms for this purpose are worse than quadratic with respect to the size of the graph \cite{hochbaum2008pseudoflow}. Since these subroutines are run many times, the high time complexity would be problematic for large substrate networks. Instead, each run of \textsc{incrFlow} increases a flow only along one new path. For finding the path, a modified best-first-search \cite{korf1993linear} is used, which runs in linear time. It should be noted that split flows can still be created if \textsc{incrFlow} is run multiple times for a flow. When improving a flow and when selecting from multiple possible flows, the \textsc{incrFlow} and \textsc{createInstanceAndFlow} routines must strike a balance between flow data rate and the increase in overall delay of the solution. Our strategy for comparing two possible flows is to first compare their data rates and compare their latencies only if there is a tie. This strategy is used in line 31 to select the best flow. The rationale is that selecting flows with high data rate leads to a small number of instances to be created. However, we also employ a cutoff mechanism: flow data rates above the cutoff (the increase in data rate that we want to achieve) do not add more value and are hence regarded to be equal to the cutoff value. This increases the likelihood of a tie, so that the tie-breaking method of preferring lower latencies is also important. An analogous strategy is used in line 39 to compare paths: the primary criterion is to prefer paths with higher bandwidth -- up to the given cutoff $d$ -- and, in case of a tie, to prefer paths with lower latency. For finding the best path, a modified best-first-search is used, in which the nodes to be visited are stored in a priority queue, where priority is defined in accordance with the comparison relation described above. \section{Mixed integer programming approach} \label{} In this section, we provide a mixed integer programming (MIP) formulation of the problem. On one hand, this serves as a further formalization of the problem; on the other hand, under suitable assumptions (to be detailed in Section~\ref{}) an appropriate solver can be used to solve the mixed integer program, yielding an algorithm for the problem. Based on the assumption that two instances of the same component cannot be mapped to a node, instances can be identified by the corresponding component and the hosting node. This is the basis for our choice of variables, which are explained in more detail in Table~\ref{}. \begin{table}[tb] \caption{Variables} \label{} \begin{tabular}{p{8mm}p{8mm}p{60mm}} \toprule Name & Domain & Definition \\ \midrule $x_{j,v}$ & $\{0,1\}$ & 1 iff an instance of component $j{\in} {\cal C}$ is mapped to node $v{\in} V$ \\ $y_{a,v,v'}$ & $\mathbb{R}_{\ge 0}$ & If $a{\in} A_T$ is an arc from an output of $j{\in} C_T$ to an input of $j'{\in} C_T$, an instance of $j$ is mapped on $v{\in} V$, and an instance of $j'$ is mapped on $v'{\in} V$, then $y_{a,v,v'}$ is the data rate of the corresponding flow from $v$ to $v'$; otherwise it is 0 \\ $z_{a,v,v',l}$ & $\mathbb{R}_{\ge 0}$ & If $a{\in} A_T$ is an arc from an output of $j{\in} C_T$ to an input of $j'{\in} C_T$, an instance of $j$ is mapped on $v{\in} V$, and an instance of $j'$ is mapped on $v'{\in} V$, then $z_{a,v,v',l}$ is the data rate of the corresponding flow from $v$ to $v'$ that goes through link $l{\in} L$; otherwise it is 0 \\ $\Lambda_{j,v}$ & $\mathbb{R}_{\ge 0}^{|\text{In}(j)|}$ & Vector of data rates on the inputs of the instance of component $j{\in} C_T$ on node $v{\in} V$, or an all-zero vector if no such instance is mapped on $v$ \\ $\Lambda'_{j,v}$ & $\mathbb{R}_{\ge 0}^{|\text{Out}(j)|}$ & Vector of data rates on the outputs of the instance of component $j{\in} C_T$ on node $v{\in} V$, or an all-zero vector if no such instance is mapped on $v$ \\ $\varrho_{j,v}$ & $\mathbb{R}_{\ge 0}$ & CPU requirement of the instance of component $j{\in} C_T$ on node $v{\in} V$, or zero if no such instance is mapped on $v$ \\ $\mu_{j,v}$ & $\mathbb{R}_{\ge 0}$ & Memory requirement of the instance of component $j{\in} C_T$ on node $v{\in} V$, or zero if no such instance is mapped on $v$ \\ $\omega_{v,\text{cpu}}$ & $\{0,1\}$ & 1 iff the CPU capacity of node $v{\in} V$ is exceeded \\ $\omega_{v,\text{mem}}$ & $\{0,1\}$ & 1 iff the memory capacity of node $v{\in} V$ is exceeded \\ $\omega_l$ & $\{0,1\}$ & 1 iff the maximum data rate of link $l{\in} L$ is exceeded \\ $ \psi_\text{cpu} $ & $\mathbb{R}_{\ge 0}$ & Maximum CPU over-subscription over all nodes \\ $ \psi_\text{mem} $ & $\mathbb{R}_{\ge 0}$ & Maximum memory over-subscription over all nodes \\ $ \psi_\text{dr} $ & $\mathbb{R}_{\ge 0}$ & Maximum capacity over-subscription over all links\\ $\zeta_{a,v,v',l}$ & $\{0,1\}$ & 1 iff $z_{a,v,v',l}>0$ \\ $\delta_{j,v}$ & $\{0,1\}$ & 1 iff $x_{j,v}\ne x^*_{j,v}$ \\ \bottomrule \end{tabular} \end{table} We use the following notations for formalizing the constraints and objectives. ${\cal C}{=}\bigcup_{T\in {\cal T}} C_T$ denotes the set of all components, ${\cal A}{=}\bigcup_{T\in {\cal T}} A_T$ the set of all arcs, and ${\cal S}{=}\bigcup_{T\in {\cal T}} S(T)$ the set of all sources across all network services that we want to map to the network. $M$, $M_1$, and $M_2$ denote sufficiently large constants. $(\Lambda_{j,v})_k$ denotes the $k$th component of the vector $\Lambda_{j,v}$. $\underbar{0}$ denotes a zero vector of appropriate length. Information about existing instances should also be taken into account during the decision process. For this, we define $x^*_{j,v} (\forall j \in {\cal C}, v \in V)$ as a constant given as part of the problem input. If there is a previously mapped instance of component $j$ on node $v$ in the network, $x^*_{j,v}$ is 1, otherwise it is 0. \subsection{Constraints} Here we define the sets of constraints that enforce the required properties of the template embedding process. \subsubsection{Mapping consistency rules} {\footnotesize \begin{align} \forall (v,j,\lambda)\in {\cal S}: && x_{j,v} & = 1 \\ \forall (v,j,\lambda)\in {\cal S}: && \Lambda'_{j,v} & = \lambda \\ \forall j\in {\cal C}, \forall v\in V, k \in [1,|\text{In}(j)|]: && (\Lambda_{j,v})_k & \le M\cdot x_{j,v} \\ \forall j\in {\cal C}, \forall v\in V, k \in [1,|\text{Out}(j)|]: && (\Lambda'_{j,v})_k & \le M\cdot x_{j,v} \\ \forall j\in {\cal C}, \forall v\in V: && x_{j,v}-x^*_{j,v} & \le \delta_{j,v} \\ \forall j\in {\cal C}, \forall v\in V: && x^*_{j,v}-x_{j,v} & \le \delta_{j,v} \end{align} } Constraints (1) and (2) enforce that the placement respectively the output data rate of source component instances are in line with the tuples specified in $\cal S$. Constraint (3) guarantees the consistency between the variables $\Lambda_{j,v}$ and $x_{j,v}$: if $\Lambda_{j,v}$ has a positive component, then $x_{j,v}$ must be 1, i.e., only an existing component instance can process the incoming flow. Constraint (3) is analogous for the outgoing flows, represented by the $\Lambda'_{j,v}$ variables. Constraints (5) and (6) together ensure that $\delta_{j,v}=1$ if and only if $x_{j,v}\ne x^*_{j,v}$. \subsubsection{Flow and data rate rules} {\footnotesize \begin{multline} \forall j\in {\cal C}, \text{$j$ not a source component}, \forall v\in V:\\ \Lambda'_{j,v} = r_j(\Lambda_{j,v})-(1-x_{j,v})\cdot r_j(\underbar{0}) \end{multline} \begin{multline} \forall j\in {\cal C}, \forall v\in V, k \in [1,|\text{In}(j)|]:\\ (\Lambda_{j,v})_k=\sum_{a\text{ ends in input }k\text{ of }j, v'\in V} y_{a,v',v} \end{multline} \begin{multline} \forall j\in {\cal C}, \forall v\in V, k \in [1,|\text{Out}(j)|]:\\ (\Lambda'_{j,v})_k=\sum_{a\text{ starts in output }k\text{ of }j, v'\in V} y_{a,v,v'} \end{multline} \begin{multline} \forall a\in{\cal A},\forall v,v_1,v_2\in V:\\ \sum_{vv'\in L}z_{a,v_1,v_2,vv'}-\sum_{v'v\in L}z_{a,v_1,v_2,v'v} =\\ =\begin{cases} 0 & \text{if }v\ne v_1\text{ and }v\ne v_2 \\ y_{a,v_1,v_2} & \text{if }v=v_1\text{ and }v_1\ne v_2 \\ 0 & \text{if }v=v_1=v_2 \end{cases} \end{multline} \begin{align} \forall a\in{\cal A}, \forall v,v'\in V, \forall l\in L: && z_{a,v,v',l} & \le M\cdot \zeta_{a,v,v',l} \end{align} } Constraint (7) computes the data rate on the outputs of a processing component instance based on the data rates on its inputs and the $r_j$ function of the underlying component. The constraint is formulated in such a way that for $x_{j,v}=1$, $\Lambda'_{j,v} = r_j(\Lambda_{j,v})$, whereas for $x_{j,v}=0$ (in which case also $\Lambda_{j,v}=0$ because of Constraint (3)), also $\Lambda'_{j,v}=0$ so that there is no contradiction with Constraint (4). Constraint (8) computes the data rate on the inputs of a component instance as the sum of the data rates on the links ending in that input. Similarly, Constraint (9) ensures that the data rate on the outputs of a component instance is distributed on the links starting in that output. Constraint (10) is the flow conservation rule, also ensuring the right data rate of each flow, thus connecting the $z_{a,v,v',l}$ variables (flow values on individual links) and the $y_{a,v,v'}$ variables (flow data rate). Constraint (11) sets the $\zeta_{a,v,v',l}$ variables (on the basis of the $z_{a,v,v',l}$ variables), so that they can be used later on in the objective function (Section~\ref{}). \subsubsection{Calculation of resource consumption} {\footnotesize \begin{align} \forall j\in {\cal C}, \forall v\in V: && \varrho_{j,v} & = p_j(\Lambda_{j,v})-(1-x_{j,v})\cdot p_j(\underbar{0}) \\ \forall j\in {\cal C}, \forall v\in V: && \mu_{j,v} & = m_j(\Lambda_{j,v})-(1-x_{j,v})\cdot m_j(\underbar{0}) \end{align} } Constraints (12) and (13) calculate CPU respectively memory consumption of each component instance based on the $p_j$ and $m_j$ functions of the underlying component\footnote{Adding more resource types would be reflected by adding corresponding constraints here.}. The logic here is analogous to that of Constraint (7). \subsubsection{Capacity constraints} {\footnotesize \begin{align} \forall v\in V: & \quad \sum_{j\in {\cal C}}\varrho_{j,v} \le \text{cap}_{\text{cpu}}(v)+M\cdot\omega_{v,\text{cpu}} \\ \forall v\in V: & \quad \sum_{j\in {\cal C}}\varrho_{j,v} - \text{cap}_{\text{cpu}}(v) \leq \psi_\text{cpu} \\ \forall v\in V: & \quad \sum_{j\in {\cal C}}\mu_{j,v} \le \text{cap}_{\text{mem}}(v)+M\cdot\omega_{v,\text{mem}} \\ \forall v\in V: & \quad \sum_{j\in {\cal C}}\mu_{j,v} - \text{cap}_{\text{mem}}(v) \leq \psi_\text{mem} \\ \forall l\in L: & \quad \sum_{a\in {\cal A};v,v'\in V}z_{a,v,v',l} \le b(l)+M\cdot\omega_l \\ \forall l\in L: & \quad \sum_{a\in {\cal A};v,v'\in V}z_{a,v,v',l} - b(l) \leq \psi_\text{dr} \end{align} } The aim of these constraints is to set the $\omega$ and $\psi$ variables (based on the already defined $\varrho$, $\mu$ and $z$ variables), which will be used in the objective function (Section~\ref{}). Constraint (14) ensures that $\omega_{v,\text{cpu}}$ will be 1 if the CPU capacity of node $v$ is overloaded, while Constraint (15) ensures that $\psi_\text{cpu}$ will be at least as high as the amount of CPU overload of any node (the appearance of $\psi_\text{cpu}$ in the objective function will guarantee that it will be exactly the maximum amount of CPU overload and not higher than that). Constraints (16), (17) do the same for memory overloads and Constraints (18), (19) do the same for the overload of link capacity. \subsubsection{Interplay of the constraints} To illustrate the interplay of the constraints, we assume that we need to optimize the embedding shown in Fig.~\ref{}. Constraints (1) and (2) ensure that instances of the source component, i.e., S1 and S2, are embedded and their output data rates are set correctly. Constraint (9) ensures that these data rates are then handed out as flows that can only end up in instances of A. These flows are mapped to network links and instances of A are assigned input data rates using Constraints (10) and (8), respectively. That being set, Constraint (3) marks the instances A1 and A2 as embedded, and Constraint (7) sets their output data rates using the respective $r_j$ function. In a similar way, the rest of the components are instantiated and embedded in the network. Constraints (5) and (6) ensure that the $\delta_{j,v}$ variables are set correctly. Constraints (12) and (13) compute the resource consumption of each instance based on the input data rates and the corresponding $p_j$ and $m_j$ functions. Constraints (14)--(19) make sure that over-subscription of node and link capacities are captured correctly, and collect the maximum value of over-subscription for each resource type. This maximum value is used in the objective function described in Section~\ref{}, which drives the decisions based on the constraints. \subsection{Optimization objective} \label{} We formalize the optimization objective based on the goals defined in Section~\ref{} as follows: {\footnotesize \begin{multline} \text{minimize} \quad M_1\cdot\Big( \sum_{v\in V}(\omega_{v,\text{cpu}} + \omega_{v,\text{mem}}) + \sum_{l\in L}\omega_{l}\Big)+ \\ + M_2\cdot\Big(\sum_{\substack{a\in {\cal A} \\ v,v'\in V \\ l\in L}}(d(l)\cdot\zeta_{a,v,v',l})+ \sum_{\substack{j\in {\cal C} \\ v\in V}}\delta_{j,v}\Big)+ \\ + \psi_\text{cpu} + \psi_\text{mem} + \psi_\text{dr} + \sum_{\substack{j\in {\cal C} \\ v\in V}}(\varrho_{j,v} + \mu_{j,v}) + \sum_{\substack{a\in {\cal A} \\ v,v'\in V \\ l\in L}}z_{a,v,v',l} \end{multline} } By assigning sufficiently large values to $M_1$ and $M_2$, we can achieve the following goals with the given priorities (1 being the highest priority): \begin{enumerate} \item Number of capacity constraint violations over all nodes and links is minimized. \item Template arcs are mapped to network paths in such a way that their total latency is minimized. Moreover, the number of instances that need to be started/stopped is minimized. \item The maximum value for capacity constraint violations over all nodes and links is minimized. Also, overlay instances and the edges among them are created in a way that their resource consumption is minimized. \end{enumerate} The objective function is in line with the objectives defined in Section~\ref{}. The primary objective is to minimize the number of constraint violations; a sufficiently large $M_1$ ensures that a decrease in the first term of the objective function has larger impact than any change in the other terms. Moreover, the resulting solution $\sigma'$ will be Pareto-optimal with respect to the other, secondary metrics: otherwise, there would be another solution $\sigma''$ that is as good as $\sigma'$ according to each secondary metric and strictly better than $\sigma'$ in at least one secondary metric, but then, $\sigma''$ would lead to a lower overall value of the objective function. This mixed integer program can be used for initial embedding of service templates as well as for optimizing existing embeddings. However, for the initial embedding of newly requested network services, the term $\sum_{j\in {\cal C}, v\in V}\delta_{j,v}$ should be removed from the objective function because it would introduce an unwanted bias towards embeddings with fewer instances, although it is possible that having more instances can decrease the overall cost of the solution. \subsection{Solving the mixed integer program} \label{} All our constraints are linear equations and linear inequalities, and also the objective function is linear. Hence, if the functions $p_j$, $m_j$, and $r_j$ are linear for all $j\in{\cal C}$, then we obtain a mixed-integer linear program (MILP), which can be solved by appropriate solvers. For non-linear functions, a piecewise linear approximation may make it possible to use MILP solvers to obtain good (although not necessarily optimal) solutions. \section{Introduction} Network services, like video streaming and online gaming, consist of different service components, including (virtual) network functions, application servers, data bases, etc. Typically, several of these network services are hosted on top of wide-area networks, serving the continuously changing demands of their users. The need for efficient and automatic deployment, scaling, and path selection methods for the network services has led to paradigms like network softwarization, including software-defined networking (SDN) and network function virtualization (NFV). SDN and NFV provide the required control and orchestration mechanisms to drive the network services through their life-cycle. Today, network services are placed and deployed in the network based on fixed, pre-defined descriptors~\cite{etsi-mano} that contain the number of required instances for each service component and the exact resource demands. More flexibility can be achieved by specifying auto-scaling thresholds for metrics of interest. Once such a threshold is reached, the affected network services should be modified, e.g., scaled. To react to addition and removal of network services, fluctuations in the request load of a network service, or to serve new user groups in a new location, (i) the network services can be scaled out/in by adding/removing instances of service components, (ii) the placement of service components and the amount of resources allocated to them can be modified, and (iii) the network flows between the service components can be re-routed through different, more suitable paths. Given this large number of degrees of freedom for finding the best adaptation, deciding scaling, placement, and routing independently can result in sub-optimal decisions for the network and the running services. Consider a service platform provider hosting a dynamically changing set of network services, where each network service serves dynamically changing user groups that produce dynamically changing data rates. Trade-offs among the conflicting goals of network services and platform operators can be highly non-trivial, for example: \begin{itemize} \item Placing a compute-intensive service component on a node with limited resources near the \emph{source} of requests (e.g., the location of users, content servers, etc.) minimizes latency but placing it on a more powerful node further away in the network minimizes processing time. \item Letting a single instance of a data-processing component serve multiple sources minimizes compute resource consumption but using dedicated instances near the sources minimizes network load. \item Changing the current configuration to a better one will hopefully pay off in the long run but keeping the current configuration avoids reconfiguration costs. \item Fulfilling the resource requirements of one service versus the requirements of another service. \end{itemize} To deal with these challenges, we propose JASPER, a comprehensive approach for the \textbf{J}oint optimiz\textbf{A}tion of \textbf{S}caling, \textbf{P}lac\textbf{E}ment, and \textbf{R}outing of virtual network services. In JASPER, each network service is described by a \emph{service template}, containing information about the components of the network service, the interconnections between the components, and the resource requirements of the components. Both the resource requirements and the outgoing data rates of a component are specified as \emph{functions of the incoming data rates}. The input to the problem we are tackling comprises service templates, location and data rate of the \emph{sources} of each network service, and the topology and available resources of the underlying network. Our optimization approach takes care of the rest: based on the location and current data rate of the sources, in a single step, the templates are scaled by replicating service components as necessary, the placement of components on network nodes is determined, and data flows are routed along network paths. Node and link capacity constraints of the network are automatically taken into account. We optimize the solution along multiple objectives, including minimizing resource usage, minimizing latency, and minimizing deployment adaptation costs. Our main contributions are as follows: \begin{itemize} \item For the case where resource demands of service components are determined as a function of the incoming data rate to each instance, we formalize \emph{template embedding} as a joint optimization problem for scaling, placing, and routing service templates in the network. \item We prove the NP-hardness of the problem. \item We present two algorithms for solving the problem, one based on mixed integer programming, the other a custom heuristic. \item We evaluate both algorithms in detail to determine their strengths and weaknesses. \end{itemize} With the proposed approach, service providers obtain a flexible way to define network services on a high level of abstraction while service platform providers obtain powerful methods to optimize the scaling and placement of multiple services in a single step, fully automatically. The rest of the paper is organized as follows. In Section~\ref{}, we give an overview of related work. Section~\ref{} presents a high-level overview of our approach and Section~\ref{} describes the details of our model and assumptions. We discuss the complexity of template embedding in Section~\ref{} and formulate the problem as a mixed integer programming model in Section~\ref{}. We present a heuristic solution in Section~\ref{} and the evaluation results of our solutions in Section~\ref{}, before concluding the paper in Section~\ref{}. \section*{Acknowledgment} This work has been performed in the context of the SONATA project, funded by the European Commission under Grant number 671517 through the Horizon 2020 and 5G-PPP programs. This work is partially supported by the German Research Foundation (DFG) within the Collaborative Research Center ``On-The-Fly Computing'' (SFB 901). The work of Z.\ \'A.\ Mann was partially supported by the Hungarian Scientific Research Fund (Grant Nr. OTKA 108947) and the European Union's Horizon 2020 research and innovation programme under grant 731678 (RestAssured). \bibliographystyle{IEEEtran} \section{Related work} \label{} The template embedding problem is a joint, single-step optimization of scaling, placement, and routing for network services. In general, our solution can be applied in different contexts, e.g., (distributed) cloud computing and Network Function Virtualization (NFV). In this section, after an analysis of related approaches from a theoretical point of view, we give an overview of related work in the cloud computing and NFV contexts. The major difference among the existing work in these two fields is usually the abstraction level considered for the substrate network and the resulting assumptions for the model. In particular, in the cloud computing context, embedding is typically done on top of physical machines in data centers, while in the NFV context, embedding is done on top of geographically distributed points of presence. \subsection{Virtual network embedding problem} The combination of the placement and path selection sub-problems of template embedding is similar to the Virtual Network Embedding (VNE) problem. Both deal with mapping virtual nodes and virtual links of a graph into another graph and do not include the scaling step. Fischer et al.~\cite{Fischer2013} have published a survey of different approaches to VNE, including static and dynamic VNE algorithms. In contrast to static VNE solutions that consider the initial mapping process only, in this paper we also deal with optimizing and modifying already embedded templates. Some VNE solutions, for example, Houidi et al.~\cite{houidi2015exact}, can modify the mapping in reaction to node or link failures. The modifications in their work, however, are limited to recalculating the location for the embedded virtual network, i.e., migrating some of the nodes and changing the corresponding paths among them. In addition to these modifications, our approach can also modify the \emph{structure} of the graph to be embedded by adding or removing nodes and links if necessary. \subsection{Cloud computing context} The related problem in cloud environments is typically formulated as resource allocation for individual components. Scaling and placing instances of virtual machines on top of physical machines while adhering to capacity constraints are the usual problems tackled in this context~\cite{lorido2014review,mann2016interplay}. The communication among different virtual machines, however, is usually left out or considered only in a limited sense~\cite{mann2015allocation}. Even the approaches that do consider the communication among virtual machines~\cite{divakaran2015towards,ahvar2015nacer,alicherry2013optimizing,ahvar2016cacev} do not include routing decisions whereas JASPER also includes routing. Relevant to the placement sub-problem of template embedding, Bellavista et al.~\cite{Bellavista2015} focus on the technical issues of deploying flexible cloud infrastructure, including network-aware placement of multiple virtual machines in virtual data centers. Wang et al.~\cite{Wang2017} study the dynamic scaling and placement problem for network services in cloud data centers, aiming at reducing costs. These papers also do not address routing. Moreover, our approach of specifying resource consumption as a function of input data rates allows a much more realistic modeling of the resource needs of service components than the constant resource needs assumed by the existing approaches in this context. Keller et al.~\cite{Keller2014b} consider an approach similar to our template embedding problem in the context of distributed cloud computing. Our terminology is partly based on their work but there are important differences in the assumptions and the models that make our approach stronger and more flexible than their solutions. In contrast to their model, where the number of users determines the number of required instances, the deciding factor in our work is the \emph{data rate} originating from different source components. Data rate can be represented, for example, as requests or bits per second and is a more perceptible metric in practical applications and gives a more fine-grained control over the embedding process. Moreover, we do not enforce strict scaling restrictions for components as done in their work. (For example, their method needs as input the exact number of instances of a back-end server that is required behind a front-end server.) Finally, the optimization objective in their model is limited to minimizing the total number of instances for embedded templates. We use a more sophisticated multi-objective optimization approach where different metrics like CPU and memory load of network nodes, data rate on network links, and latency of embedded templates are considered. \subsection{Network function virtualization context} The placement and routing problems are also relevant in the field of Network Function Virtualization (NFV). In the NFV context, the \emph{forwarding graphs} of network services composed of multiple virtual network functions (VNFs) are mapped into the network. Herrera et al.~\cite{herrera2016resource} have published an analysis of existing solutions for placing network services as part of a survey on resource allocation in NFV. Kuo et al.~\cite{Kuo2016} consider the joint placement and routing problem, focusing on maximizing the number of admitted network service embedding requests. Ahvar et al.~\cite{Ahvar2017} propose a solution to this problem, with the assumption that the VNFs can be re-used among different flows. Their objective is to find the optimal number of VNFs for all requests and to minimize the costs for the provider. Another similar approach that considers re-using components is proposed by Bari et al.~\cite{bari2016orchestrating}. Kebbache et al.~\cite{Khebbache2017} aim at solving this problem in an efficient way that can scale with the size of the underlying infrastructure and the embedded network services. They measure the efficiency of their algorithms with respect to run time, acceptance rate, and costs. Another attempt to solve this problem in an efficient and scalable way has been made by Luizelli et al.~\cite{Luizelli2017}, focusing on minimizing resource allocation. In comparison to all these approaches, we consider a more comprehensive optimization objective, trying to minimize the delay for network services, the number of added or removed instances, resource consumption, as well as overload of resources. In our work, the exact structure of the network service does not have to be fixed in the deployment request. In a previous work~\cite{draexler2017ijnm}, we have studied another type of flexibility in the network service structure, namely, the case where the network service components are specified with a partial order and can be re-ordered if desirable for the optimization objectives. Beck et al.~\cite{Beck2015} also consider placement of network services with flexibly ordered components. JASPER is based on the assumption that the \emph{order} of traversing the service components is fixed and given, however, the number of instances for each component and the amount of resources allocated to each component can be adapted dynamically, resulting in network services with malleable structures. Several other optimization approaches~\cite{mehraghdam-netsoft16,sahhaf2015,moens2014vnf,savi2015impact} and heuristic algorithms~\cite{mijumbidesign,beck2015coordinated} have been proposed for placement, scaling, and path selection problems for network services. Our template embedding approach has two important differences compared to these solutions. First, our approach can be used for initial placement of a newly requested service as well as scaling and adapting existing embeddings. Second, in our approach, the structure of the service and mapping of the service components to network nodes and the optimal routing are determined in one single step, based on the requirements of the service and current state of network resources, searching for a global optimum. A preliminary version of this work was presented at the CCGrid 2017 conference \cite{draexler2017joint}. Compared to the conference version, this paper contains the proof of NP-hardness, more detailed explanation of the problem model and the devised algorithms, and a more detailed evaluation and discussion of the practical applicability of the proposed approach. \section{Problem model} \label{} In this section, we formalize our model and define the problem we are tackling. Our model uses three different graphs for representing (i) the generic network service structure, (ii) a concrete and deployable instantiation of the network service, and (iii) the actual network. We use different names and notations to distinguish among these graphs (Table~\ref{}). Informally, the problem we address is as follows: given a substrate network, a set of -- newly requested or already existing -- network services with their templates, and the source(s) for the services in the network along with the traffic originating from them, we want to optimally embed the network services into the network. \begin{table}[t] \centering \caption{Notations Used for Graphs in the Model} \label{} \begin{tabular}{llll} \toprule Graph & Symbol & Name & Annotations \\ \midrule \multirow{2}{*}{Template $G_\text{tmpl}$} & $j{\in} C_T$ & Component & $\text{In}(j)$, $\text{Out}(j)$, $p_j,m_j,r_j$ \\ & $a{\in} A_T$ & Arc \\ \midrule \multirow{2}{*}{Overlay $G_\text{OL}$} & $i{\in} I_\text{OL}$ & Instance & $c(i)$, $P_T^{(I)}(i)$ \\ & $e{\in} E_\text{OL}$ & Edge & $P_T^{(E)}(e)$ \\ \midrule \multirow{2}{*}{Network $G_\text{sub}$} & $v{\in} V$ & Node & $\text{cap}_{\text{cpu}}(v)$, $\text{cap}_{\text{mem}}(v)$ \\ & $l{\in} L$ & Link & $b(l)$, $d(l)$ \\ \bottomrule \end{tabular} \end{table} \subsection{Substrate network} We model the \emph{substrate network} as a directed graph $G_\text{sub}{=}(V,L)$. Each \emph{node} $v\in V$ is associated with a CPU capacity $\text{cap}_{\text{cpu}}(v)$ and a memory capacity $\text{cap}_{\text{mem}}(v)$ (this can be easily extended to other types of resources). Moreover, we assume that every node has routing capabilities and can forward traffic to its neighboring nodes.\footnote{Capacities can be 0, e.g., to represent conventional switches by 0 CPU capacity or an end device by 0 forwarding capacity.} Each \emph{link} $l\in L$ is associated with a maximum data rate $b(l)$ and a propagation delay $d(l)$. For each node $v$, we assume that the internal communications (e.g., communication inside a data center) can be done with unlimited data rate and negligible delay. \subsection{Templates} The substrate network has to host a set $\cal T$ of network services. We define the structure of each network service $T\in {\cal T}$ using a \emph{template}, which is a directed acyclic graph $G_\text{tmpl}(T){=}(C_T,A_T)$. We refer to the nodes and edges of the template graph as \emph{components} and \emph{arcs}, respectively. They define the type of components required in the network service and specify the way they should be connected to each other to deliver the desired functionality. Fig.~\ref{} shows an example template. A template component $j\in C_T$ has an ordered set of inputs, denoted as $\text{In}(j)$, and an ordered set of outputs, denoted as $\text{Out}(j)$. Its resource consumption depends on the data rates of the flows entering the component. We characterize this using a pair of functions $p_j,m_j:\mathbb{R}_{\ge 0}^{|\text{In}(j)|} \to \mathbb{R}_{\ge 0}$, where $p_j$ is the CPU load and $m_j$ is the required memory size of component $j$, depending on the data rate of the incoming flows. These functions typically account for resource consumption due to processing the input data flows as well as fixed, baseline consumption (even when idle). Similarly, data rates of the outputs of the component are determined as a function of the data rates on the inputs, specified as $r_j:\mathbb{R}_{\ge 0}^{|\text{In}(j)|} \to \mathbb{R}_{\ge 0}^{|\text{Out}(j)|}$. Fig.~\ref{} shows examples for functions $p_j,m_j,r_j$ that define the resource demands and output data rates of an example component. Each arc in $A_T$ connects an output of a component to an input of another component. \emph{Source components} are special components in the template: they have no inputs, a single output with unspecified data rate, and zero resource consumption. In the example of Fig.~\ref{}, S is a source component, whereas the others are normal processing components. \subsection{Overlays and sources} A template specifies the types of components and the connections among them as well as their resource demands depending on the load. A specific, deployable instantiation of a network service can be derived by scaling its template, i.e., creating the necessary number of instances for each component and linking the instances with each other according to the requirements of the template. Depending on data rates of the service flows and the locations in the network where the flows start, different numbers of instances for each component might be required. To model this, for each network service $T$, we define a set of \emph{sources} $S(T)$. The members of $S(T)$ are tuples of the form $(v,j,\lambda)$, where $v\in V$ is a node of the substrate network, $j\in C_T$ is a source component, and $\lambda\in\mathbb{R}_+$ is the corresponding data rate assigned to the output of this source component. Such a tuple means that an instance of source component $j$ generates a flow from node $v$ with rate $\lambda$. Sources may represent populations of users, sensors, or any other component that can generate flows to be processed by the corresponding network service. Fig.~\ref{} shows two example sources for the template of Fig.~\ref{}, located on different nodes of the substrate network. \begin{figure}[tb] \centering \subfigure[\label{ template]{\includegraphics[width=0.27\columnwidth]{template}} \hfil \subfigure[\label{ component]{\includegraphics[width=0.4\columnwidth]{component}} \subfigure[\label{ on a network]{\includegraphics[width=0.57\columnwidth]{sources}} \subfigure[\label{ overlay]{\includegraphics[width=0.3\columnwidth]{overlay}} \hfil \subfigure[\label{ embedded in the network]{\includegraphics[width=0.57\columnwidth]{mapping1}} \caption{Some examples: (a) a template, (b) a component, (c) an overlay corresponding to the template, and (d) a mapping of the overlay into a substrate network. The links of the substrate network are bi-directional.} \end{figure} An \emph{overlay} is the outcome of scaling the template based on the associated sources. An overlay $\text{OL}$ stemming from template $T$ is described by a directed acyclic graph $G_\text{OL}(T){=}(I_\text{OL},E_\text{OL})$. Each component \emph{instance} $i\in I_\text{OL}$ corresponds to a component $c(i)\in C_T$ of the underlying template. Each $i\in I_\text{OL}$ has the same characteristics (inputs, outputs, resource consumption characteristics) as $c(i)$. Moreover, if there is an edge from an output of an instance $i_1$ to an input of instance $i_2$ in the overlay, then there must be a corresponding arc from the corresponding output of $c(i_1)$ to the corresponding input of $c(i_2)$ in the template. This ensures that the edge structure of the overlay is in line with the structural requirements of the network service, represented by the arcs in the template. To be able to create the required number of instances for each component, we assume either that the components are stateless or that a state management system is in place to handle state redistribution upon adding or removing instances. In this way, requests can be freely routed to any instance of a component. Alternatively, additional details can be added to the model, for example, to make sure that the flows belonging to a certain session are routed to the right instance of stateful components that have stored the corresponding state information. Fig.~\ref{} shows an example overlay corresponding to the template in Fig.~\ref{}. The naming of the instances follows the convention that the first letter identifies the corresponding component in the template, e.g., A1 is an instance of component A. An overlay might include multiple instances of a specific template component, e.g., B1, B2, and B3 all are instances of component B. An output of an instance can be connected to the input of multiple instances of the same component, like the output of A1 is connected to the inputs of B1 and B2. In a case like that, B1 and B2 share the data rate calculated for the connection between components A and B. Similarly, outputs of multiple instances in the overlay can be connected to the input of the same instance, like the input of C1 is connected to the output of B1, B2, and B3, in which case the input data rate for C1 is the sum of the output data rates of B1, B2, and B3. \subsection{Mapping on the substrate network} Each overlay $G_\text{OL}(T)$ must be mapped to the substrate network by a feasible mapping $P_T$. We define the mapping as a pair of functions $P_T=\left(P_T^{(I)},P_T^{(E)}\right)$. $P_T^{(I)}:I_\text{OL}\to V$ maps each instance in the overlay to a node in the substrate network. We make the simplifying assumption that two instances of the same component cannot be mapped to the same node. The rationale behind this assumption is that in this case it would be more efficient to replace the two instances by a single instance and thus save the idle resource consumption of one instance.\footnote{This simplification is mostly a technicality to simplify the problem write-up and could be extended if necessary.} $P_T^{(E)}:E_\text{OL}\to {\cal F}$ maps each edge in the overlay to a flow in the substrate network; $\cal F$ is the set of possible flows in $G_\text{sub}$. We assume the flows are splittable, i.e., can be routed over multiple paths between the corresponding endpoints in the substrate network. The two functions must be compatible: if $e\in E_\text{OL}$ is an edge from an instance $i_1$ to an instance $i_2$, then $P_T^{(E)}(e)$ must be a flow with start node $P_T^{(I)}(i_1)$ and end node $P_T^{(I)}(i_2)$. Moreover, $P_T^{(I)}$ must map the instances of source components in accordance with the sources in $S(T)$, mapping an instance corresponding to source component $j$ to node $v$ if and only if $\exists (v,j,\lambda)\in S(T)$. The binding of instances of source components to sources determines the outgoing data rate of these instances. As the overlay graphs are acyclic, the data rate $\lambda(e)$ on each further overlay edge $e$ can be determined based on the input data rates and the $r_j$ functions of the underlying components, considering the instances in a topological order. The data rates, in turn, determine the resource needs of the instances. Fig.~\ref{} shows a possible mapping of the overlay of Fig.~\ref{} to an example substrate network, based on the pre-defined location of S1 and S2 in the network. Note that it is possible to map two communicating instances to the same node, like A2 and D2 in the example. In this case, the edge between them can be realized inside the node, without using any links. The flow between A2 and B3 is an example of a split flow that is routed over two different paths in the substrate network. Note that Fig.~\ref{} shows only a single overlay mapped to the substrate network for the sake of clarity. In general, JASPER can embed several overlays corresponding to different network services into a substrate network. \subsection{Objectives} \label{} The \emph{system configuration} consists of the overlays and their mapping on the substrate network. A new system configuration can be computed by an appropriate algorithm for the template embedding problem. A valid system configuration must respect all capacity constraints: for each node $v$, the total resource needs of the instances mapped to $v$ must be within its capacity concerning both CPU and memory, and for each link $l$, the sum of the flow values going through $l$ must be within its maximum data rate. However, it is also possible that some of those constraints are violated in a given system configuration: for example, a valid system configuration (i.e., one without any violations) may become invalid because the data rate of a source has increased, because of a temporary peak in resource needs, or a failure in the substrate network. Therefore, given a current system configuration $\sigma$, our primary objective is to find a new system configuration $\sigma'$, in which the \emph{number of constraint violations is minimal} (ideally, zero). For this, we assume that violating node CPU, memory, and link capacity constraints is equally undesirable. There are a number of further, secondary objectives, which can be used as tie-breaker to choose from system configurations that have the same number of constraint violations: \begin{itemize} \item Total delay of all edges across all overlays \item Number of instance addition/removal operations required to transition from $\sigma$ to $\sigma'$ \item Maximum amounts of capacity constraint violations, for each resource type (CPU, memory, link capacity) \item Total resource consumption of all instances across all overlays, for each resource type (CPU, memory, link capacity) \end{itemize} Higher values for these metrics result in higher costs for the system or in lower customer satisfaction, so our objective is to minimize these values. Therefore, our aim is to select a new system configuration $\sigma'$ from the set of system configurations with minimal number of constraint violations that is Pareto-optimal with respect to these secondary metrics. \subsection{Problem formulation summary} Our aim is to handle the scaling, placement, and routing for newly requested network services as well as already deployed network services. Taking this into account, the Template Embedding problem can be summarized as follows: \begin{itemize} \item Inputs: \begin{itemize} \item Substrate network \item Template for each network service \item Location and data rate of the sources for each network service \item For the already deployed network services: overlay and its mapping onto the substrate network \end{itemize} \item Outputs: \begin{itemize} \item For the newly requested network services: overlay and its mapping onto the substrate network \item For the already deployed network services: modified overlay and its modified mapping onto the substrate network \end{itemize} \end{itemize} Scaling is performed while creating the overlay from the template, while placement and routing are performed when the instances and edges of the overlay are mapped onto the substrate network. A further important detail concerns the relationship between different network services. The creation of the overlay from the template and its mapping onto the substrate network are defined for each network service separately; however, they share the same substrate network. The objectives defined in Sec.~\ref{} relate to the whole network including all network services, aiming for a global optimum and potentially resulting in trade-offs among the network services. A further connection among different network services may arise if they share the same component type. In this case, it is also possible that the corresponding overlay instances are realized by the same instance.

\section{Introduction} Over the past several years machine learning (ML) has become one of the most popular and powerful sets of techniques and tools used for multidisciplinary scientific research. Such popularity has been continuously growing in the past years thanks to the increasing number of methodological developments, the availability of faster hardware with strong computational capabilities such as modern GPUs and coprocessors, and finally, the great interest and investment from the private sector. The recent enthusiasm has led to a new underlying code development strategy where several tools based on ML are available as open source projects, some examples are TensorFlow~\cite{tensorflow2015-whitepaper}, scikit-learn~\cite{scikit-learn}, Keras~\cite{chollet2015}, Theano~\cite{theano} among others. Easy access to these tools has simplified the integration of new modern techniques in many research fields, in particular for those where a large amount of data is available. The dissemination of the innumerable applications and developments based on ML has given way to conferences such as ICML\footnote{\url{}}, NIPS\footnote{\url{}} and ACAT, but also to the composition of specialized working groups, e.g the IML LHC working group at CERN\footnote{\url{}} which promotes the integration of new techniques for specific requirements in experimental analysis. In the next sections we focus the discussion on ML in theoretical High Energy Physics (HEP-TH). We start by defining the categories of applications observed in that field. Then we take as a case study the recent development achieved for the determination of parton distribution functions (PDFs), where several techniques from ML are employed. Finally, we conclude by listing the most plausible development directions for the integration of ML in HEP-TH. \section{Identifying machine learning applications in HEP-TH} While classifying ML applications in experimental physics may seem simple because the experimental analysis usually requires model regression, classification and noise filtering to extract useful information from large datasets. It becomes challenging when trying to determine applications from the point of view of theoretical physics. When talking about ML techniques in theoretical High Energy Physics (HEP-TH) we may naively imagine that the usage of such tools is not ideal. In fact, we can argue that the theoretical physicists are trained to decode nature by constructing conjectures and building theoretical frameworks which remap the complexity of the measured observable in conceptual model rules. Therefore, this approach is in contrast with a model determination through a ML black box model, of which a physical interpretation is very difficult or nearly impossible to achieve. However, nowadays such misleading interpretation is disappearing and machine learning is becoming a set of useful tools that helps achieving practical results in situations where the theory has a strong computational requirement or when it requires the determination of free parameters. So, before continuing our discussion, we think it is important to define and classify how ML is translated in HEP-TH. In Figure~\ref{} we summarize one plausible graphical representation of the two branches of applications that we will discuss now in turn: {\it Level-0} and {\it Level-1}. \begin{figure} \center \includegraphics[scale=0.6]{ACATdiagrams} \caption{One possible graphical schematic representation of how machine learning is identified in HEP-TH applications.} \label{} \end{figure} \subsection{Level-0: computational techniques and tools} The first level contains machine learning in terms of computational techniques and tools. In this category we find the great majority of topics presented in the Track 3 section of the ACAT conference. These applications contain the most robust implications of computing in theoretical physics. Due to the specific nature of the problems addressed in HEP-TH these tools should be considered as part of ML applications because their development has contributed in exhaustive manner to the development of new techniques and methods in this field. A non-exhaustive list of topics involved in {\it Level-0} are summarized in the points below: \begin{itemize} \item Advanced numerical methods and applications: some examples are algorithms for Monte Carlo and Quasi Monte Carlo integration, techniques for subtraction schemes and regularization of Feynman integrals, e.g.~\cite{}, resummation techniques~\cite{}. \item Monte Carlo event generators: in this category we have the methodological developments established by MC codes such as POWHEG~\cite{}, MadGraph\_aMG5~\cite{}, Pythia~\cite{}, Herwig~\cite{}, MCFM~\cite{} and several others. The possibility to reuse events independently from the parton distribution functions thanks to reweighting and weight storage techniques such as APPLgrid~\cite{} and FastNLO~\cite{}. \item Higher order computational methods: numerical techniques for $N$-loop integrals as OneLoop~\cite{}, QCDLoop~\cite{}, LoopTools~\cite{}; parton level generators at NNLO, e.g.~DYNNLO~\cite{} and the more recent N3LO~\cite{}. \item Computer algebra techniques: some examples of algebra systems developed for the HEP-TH community as FORM~\cite{} and QGRAF~\cite{} for the translation of Feynman diagram rules into compact analytic expressions for its numerical evaluation. \end{itemize} \subsection{Level-1: applications of ML modern techniques} This second level contains the ML applications in ``sensu stricto'' i.e.~using ML modern techniques used in data sciences. This kind of application usually requires hybrid projects where experimental data and theory are included together. Nevertheless, during the last few years there has been a strong development of very successful tools based on these techniques. Some examples for this category are: \begin{itemize} \item Supervised learning, such as regression and classification: parton distribution~\cite{} and fragmentation functions~\cite{} determination, Monte Carlo tunes~\cite{}, reweighting techniques, jet discrimination through deep convolutional neural networks~\cite{}. \item Techniques for uncertainty estimation and combination: in this category we find several tools from the PDF4LHC15 recommendation~\cite{}, and some recent methods to provide a reliable MC uncertainty for simulations, e.g.~by modeling jet predictions~\cite{}. \item ``Experimental'' mathematics: applications based on machine learning optimization algorithms which may lead to the determination of multivariate densities~\cite{likas,} and sampling for integral evaluation~\cite{}. \end{itemize} In the next section we describe the recent innovative achievements obtained by the parton distribution function community in HEP-TH using ML methods. \section{Case study: the proton structure determination} The Quantum Chromodynamics (QCD) theory describes the proton structure in terms of partons, e.g.~quarks and gluons, but due to the non-perturbative regime of confinement we are unable to evaluate from QCD the momentum fraction of the proton carried by each parton. However, in order to avoid our lack of knowledge we introduce the concept of parton distribution functions (PDFs). These PDFs are then inferred from data of all relevant processes, together with the theoretical knowledge on how they affect the cross section which can be calculated approximately in perturbation theory. \subsection{The NNPDF approach} \begin{figure} \center \includegraphics[scale=0.6]{nnpdf31nnlo-10}\includegraphics[scale=0.6]{nnpdf31nnlo-1e4} \caption{The NNPDF3.1 NNLO PDFs, evaluated at $\mu^2=10$ GeV$^2$ (left) and $\mu^2=10^4$ GeV$^2$ (right) from~\cite{}.} \label{} \end{figure} The NNPDF collaboration uses ML techniques to obtain a PDF determination. In contrast to other problems in ML, where finding an accurate and fast algorithm is enough, we are not only interested in the best PDF fit, but also in obtaining an uncertainty estimate of the PDF determination. The procedure we employ is described in full details in~\cite{} and can be summarized as: \begin{itemize} \item Monte Carlo generation of artificial data. Experimental data, with central values, errors and their correlations are used to generate further artificial data, consistent with the covariance matrix provided by the experiment. \item Neural network fit to artificial data. A genetic algorithm is used to fit each artificial data set to a neural network representing the PDF (more details in~\cite{}). \item Predictions are later obtained by computing statistical estimators (such as means, quantiles and standard deviations) over the set of neural networks. Figure~\ref{} illustrates the current state of the art PDFs, NNPDF3.1~\cite{}, obtained with the NNPDF framework. \end{itemize} The points above summarize the main difficulties in PDF fits. The first consists in the inclusion of multiple experimental data which introduces indirect constraints on the PDFs. The data included in such fits are based on several datasets measured during the past decades and based on different physical processes: deep-inelastic scattering, fixed target Drell-Yan and hardronic data. This data is obtained through diverse experimental techniques and statistical analyses, therefore yielding possible inconsistencies and tensions among themselves. In order to limit these effects the NNPDF methodology propagates the uncertainty of the experimental data on the PDFs by performing the Monte Carlo artificial replica generation based on the covariance matrix provided by each experiment. The second most relevant problem consists in the choice of an unbiased functional form able to adapt and allow the propagation of data uncertainties into PDF errors while keeping under control physical requirements such as momentum sum rule conservation and positivity constraints. The NNPDF collaboration has successfully employed neural networks based on feed-forward multilayer perceptron architecture to model each single PDF flavor entering the fit procedure. The large number of parameters from these PDFs are then trained through a genetic optimization algorithm. NNPDF has released several PDF sets of global unpolarized determinations in the last years\footnote{\url{}}, together with the more recent polarized PDFs~\cite{} and Fragmentation Functions~\cite{}. One of the current most important tasks of the NNPDF collaboration is to improve the performance and quality of the optimization having in mind that in the next years the number of the new measurements from LHC will increase. \subsection{PDF4LHC15 tools for LHC Run II} Another issue concerning PDF determination consists in deciding which set of PDF should be used for the calculation of a PDF-dependent quantity. The question addressed here is how to obtain the best combined PDF uncertainty from individual PDF sets. This kind of issue is studied by the PDF4LHC working group which releases a recommendation document usually every three years. The latest one published in 2015~\cite{} aims to release combined PDF sets based on methods with clear statistical interpretation. The PDF4LHC15 prescription starts by constructing a prior Monte Carlo combined set of PDFs derived from global determinations, namely MMHT14~\cite{}, CT14~\cite{} and NNPDF3.0~\cite{}. These sets satisfy requirements such as: similar dataset, DGLAP solution and $\alpha_s$. The prior construction is obtained by converting the Hessian sets to MC replicas following the procedure described in~\cite{}. After that we apply algorithms to reduce the redundant information stored in the prior and deliver sets of PDFs with the same properties of the prior but smaller number of replicas. \begin{figure} \center \includegraphics[scale=0.6]{ACATdiagrams2} \vspace{-1cm} \caption{Graphical representation of the reduction algorithms used in the PDF4LHC15 recommendation~\cite{}.} \label{} \end{figure} \subsection*{Compression algorithm} The CMC-PDFs~\cite{} implement the compression algorithm of MC replicas. This algorithm is designed to extract a subset of replicas which preserves as much as possible the underlying statistical distribution of the prior MC PDF set. This algorithm pre-computes for the input PDF the moments (central value, variance, skewness and kurtosis), the statistical distance (Kolmogorov distance) and the correlations for each flavor in a grid of $x$ points. These estimators are then compared to subsets of replicas selected by a genetic algorithm. The procedure terminates when a tolerance value for the error function comparing these estimators is flat. From a practical point of view in the context of the PDF4LHC15 the initial $N_{\rm rep}=900$ replicas of the prior set have been reduced to 100 replicas. Further technical investigations based on clustering algorithms have also been performed in~\cite{}. \subsection*{MC2H} The MC2H algorithm~\cite{} was first introduced to convert MC PDF sets into a Hessian representation. This algorithm performs the principal component analysis (PCA) of the PDF covariance matrix in a predefined grid of $x$ nodes for all flavors at a given initial scale. The eigenvectors obtained from the PCA are the basis of the Hessian representation. This representation consists in simple linear combinations of the input MC replicas. The MC2H procedure also allows the reduction of the number of required replicas of the input PDF set because we can reject the eigenvectors associated to small eigenvalues therefore obtaining a reduced set of Hessian replicas. \subsection*{SMPDF} Starting from the MC2H algorithm we have developed the Specialized Minimal PDFs (SMPDF)~\cite{} which consists in obtaining the smallest set of Hessian replicas for a given physical process. The SMPDF algorithm performs an interactive PCA reduction on top of the PDF covariance matrix computed in a grid of $x$ points determined as the region where the PDF-process correlation is maximal. We have also provided a public web-interface for the construction of SMPDFs available and documented in ~\cite{}. \section{Outlook} The advantages and results obtained thanks to ML have been essential for several developments of this research field from PDF determination to Monte Carlo event generation. In the next months and years new applications will be achieved. It is already possible to summarize the two main directions for these developments. The first consist in the development of tools for the estimation and propagation of uncertainties. New methods to deal with uncertainty will possibly improve the determination of PDFs among other applications. This point also includes the new ideas about reweighting techniques in the context of higher order calculations. The second development branch consists in the construction of new modern neural network architectures for problem specific applications together with efficient new gradient based methods. Such tools will open the possibility to obtain better methods for multidimensional density estimation. Consequently, we will obtain better sampling algorithms relevant for improved and faster integration algorithms in Monte Carlo event generators and similar tools. In these proceedings we provided a short overview of successful applications of ML in HEP-TH but this is just the prelude of a new era where ML and HEP-TH are in synergy producing innovative and unique results. \section*{Acknowledgements} S.~C. is supported by the HICCUP ERC Consolidator grant (614577) and by the European Research Council under the European Union's Horizon 2020 research and innovation programme (grant agreement n$^{\circ}$ 740006). \section*{References}

\section{Introduction} In o-minimal structures over real closed fields, the topology on the set of definable $\mathcal C^p$-maps between definable $\mathcal C^p$-manifolds induced from the Whitney strong topology is too strong for a transversality theorem to hold (see Example \ref{example1}). In other words, transversality of definable maps is not a generic condition for the induced topology. For this reason Shiota \cite{shiota1} introduced a weaker topology for the semialgebraic structure that can also be adapted to the wider setting of o-minimal structures. We call this the definable topology. In this topology Loi \cite{loi1} proved an analogue of the Thom transversality theorem in o-minimal structures. The result of Loi was just proved for definable maps of class $\mathcal C^p$ for $p \in \mathbb N$. He also mentions in his article that it is not clear whether the theorem holds for $p = \infty$ or $p = \omega$. In this article we deal with the smooth case. We will restrict to exponential o-minimal structures for which the smooth cell decomposition holds. The reason for this is two-fold: first the infinite differentiability is not well defined in o-minimal structures in general as proved by Wilkie \cite{wilkie1} and second we use an approximation result of Fischer \cite{fischer1} and smooth definable bump functions in our proofs. Smooth definable bump functions do not exist in polynomially bounded o-minimal structures as a consequence of a result of Miller \cite{miller1}, and not every o-minimal structure admits smooth cell decomposition (see Le Gal and Rolin \cite{lr}). The first result in this article is a definable smooth version of the Thom transversality theorem (Theorem \ref{thm_transversality}). To prove the result we give a simplified definition of the definable topology on the set of smooth definable maps using definable jet spaces (Subsection 2.3). This definition is equivalent to the one used in the articles of Escribano \cite{escribano1}, Fischer \cite{fischer1}, Shiota \cite{shiota1} and Loi \cite{loi1}. The advantage of our definition is that certain properties, namely Lemmas \ref{lem_jet_map1}, \ref{lem_jet_map2}, \ref{lem_proper} and \ref{lem_restriction}, of the topology follow easily and these are not so apparent from the other definition. Our result then follows by combining the approximation result of Fischer \cite{fischer1} with the method of Loi \cite{loi1}. We next discuss the stability properties of transversality to stratifications. In the smooth case the most basic result about stability is that the set of smooth maps transverse to a closed submanifold is open in the strong topology. A more general result can be stated as follows: let $M$, $N$ be smooth manifolds and $\Sigma$ be a stratification of a closed subset in $N$. If the dimension of $M$ is big enough then the following are equivalent: (1) $\Sigma$ is Whitney $(a)$-regular; (2) the set $\{f \in \mathcal C^{\infty}(M,N) : f \pitchfork \Sigma\}$ is open in the strong topology. That (1) implies (2) is a result of Feldman \cite{Feldman}. In the o-minimal setting, an analogue of Feldman's result was included in the transversality theorem of Loi \cite{loi1}. It is easy to show that the same also holds for smooth definable maps. Then, using the fact that any stratification of a definable set admits a Whitney $(a)$-regular refinement, we show that the set of definable smooth maps non-transverse to a given finite collection of definable $\mathcal C^1$-manifolds is nowhere dense in the definable topology (Theorem \ref{cor_strong_transversality}). To our knowledge such a result has not been observed before in the smooth or complex settings and might not even be true. It is worth mentioning here that Mather \cite{Mather} has also sought such a result for generic projections. He asks whether projections transverse to Boardman manifolds contain an open and dense set; see the comment after Theorem 2 of Section 5 in his article. The converse of Feldman's result, i.e. (2) implies (1) was proved by Trotman \cite{trotman1} (see also \cite{st}). The main difficulty in proving the result of Trotman is the existence of a smooth map transverse to a stratification with a given derivative at a point. Trotman used two non-trivial results to prove the existence of such a map. First the Baire property of the Whitney strong topology and second the density of transverse maps. Since it is not clear whether the definable topology has the Baire property the arguments of Trotman do not work in the definable case. To overcome this difficulty we provide a new more constructive method (see Lemma \ref{givendf}) that avoids the use of the Baire property to show that a definable version of Trotman's result holds (Theorem \ref{thm_openness_a_regularity}). This new method might have other applications. Thoughout the paper, we work with o-minimal structures defined on the ordered field of real numbers $(\mathbb R, <, +, ., 0,1)$. By a definable set (resp. a definable map) we mean a set (resp. a map) which is definable in the given o-minimal structure. We refer the reader to Coste \cite{coste1} and van den Dries \cite{dries} for definitions and basic results on o-minimal structures. Let us fix some notations used in the article. A set $M \subset \mathbb R^m$ is called \emph{a definable $\mathcal C^p$-submanifold} ($0\leq p \leq \infty$) if $M$ is a definable set and is a $\mathcal C^p$-submanifold of $\mathbb R^m$. By a definable $\mathcal C^p$-manifold we mean a definable $\mathcal C^p$-submanifold of some Euclidean space. If $p = \infty$ it is called a definable smooth manifold. By a definable stratification of a given definable set $V \subset \mathbb R^m$ we mean a partition $\Sigma$ of $V$ into finitely many definable $\mathcal C^1$-submanifolds of $\mathbb R^m$, called strata, such that for any $X, Y$ in $\Sigma$ such that $Y \cap \overline{X} \neq \emptyset$, we have $Y\subset \overline{X}$ where $\overline{X}$ denotes the closure of $X$. We call $(X, Y)$ a \emph{pair of adjacent strata}. A stratum $X$ is said to be \emph{Whitney $(a)$-regular} over $Y$ if for every point $y \in Y \cap \overline{X}$ and for any sequence of points $\{x_n\}$ in $X$ tending to $y$ such that the sequence of tangent spaces $\{T_{x_n}X\}$ tends to $\tau \in \mathbb G_m^{\dim X}$, we have $T_y Y \subset \tau$. A stratification is called Whitney $(a)$-regular if for every pair of adjacent strata $(X,Y)$, $X$ is Whitney $(a)$-regular over $Y$. We denote by $\mathbb R$ the set of the real numbers, $\mathbb N= \{0, 1, \ldots\}$ the set of the natural numbers and $\mathbb G_m^k$ the Grassmannian of all $k$-linear subspaces of $\mathbb R^m$.\\ \noindent \textbf{Acknowledgements.} We would like to thank Professor Maria Ruas for her interest and useful discussion on the paper. We also thank Professor David Trotman for his valuable comments on the manuscript. The first author was supported by FAPESP grant 2016/14330-0 and the second author was supported by FAPESP grant 2015/12667-5. \section{Topology on definable maps} In this section we give a definition of a topology on the set of definable smooth maps using definable jet spaces. This definition is equivalent to the one used by Escribano \cite{escribano1}, Fischer \cite{fischer1}, Loi \cite{loi1} and Shiota \cite{shiota1}. \subsection{Definable tubular neighborhood} Let $\mathcal D$ be an o-minimal structure and $M \subset \mathbb R^m$ be a definable $C^p$-submanifold with ($1\leq p \leq \infty$). We denote by $\mathcal NM$ the normal bundle of $M$, i.e. is the set $\{(x, v) \in M \times \mathbb R^m \text{ such that } v \perp T_x M\}$. Then $\mathcal NM$ is a definable $\mathcal C^{p-1}$-submanifold of $\mathbb R^m \times \mathbb R^m$. Consider the definable $\mathcal C^{p-1}$-map $\varphi : \mathcal NM \to \mathbb R^m, (x, v) \mapsto (x + v)$. Following the arguments of Coste \cite{coste1} we can show that there exists a definable open neighborhood $U$ of $M\times \{0\}$ in $\mathcal NM$ such that the restriction $\varphi|_U$ is a definable $\mathcal C^{p-1}$-diffeomorphism onto an open neighborhood $T_M$ of $M$ in $\mathbb R^m$. We can assume that $U$ is of the form $$\{(x, v) \in \mathcal NM, \|v\| < \varepsilon(x)\}$$ where $\varepsilon: M \to (0,\infty)$ is a definable $\mathcal C^k$-function with $k\in \mathbb N$. Let $\pi_1: \mathcal NM \to M$ defined by $\pi_i(x, v) = x$ be the projection onto $M$. Then the map $$ \pi_M: T_M \to M, w \mapsto \pi_1(\varphi^{-1} (w))$$ is a definable $\mathcal C^{p-1}$-retraction. We call the pair $(T_M, \pi_M)$ a \emph{definable tubular neighborhood} of $M$ in $\mathbb R^m$. \subsection{Definable jet space} Let $M\subset \mathbb R^m$ and $N\subset \mathbb R^n$ be definable $\mathcal C^p$-submanifolds ($1\leq p \leq \infty$). Denote by $\mathcal D^p(M,N)$ the set of all definable $C^p$-maps between $M$ and $N$ and for an integer $k < p$ denote by $J^k(M, N)$ the $k$-jet bundle of smooth maps from $M$ to $N$. Define $$J^k_{\mathcal D}(M, N):= \{ j^k f(x) \in J^k(M, N): x\in M \text{ and } f \in \mathcal D^{p} (M, N)\}\}$$ and call it the \emph{definable $k$-jet space}. It is possible to embed $J^k_{\mathcal D}(M, N)$ into a Euclidean space as follows: let us denote by $P^k(\mathbb R^m)$ the set of all polynomials in $m$ variables with degree at most $k$ and the constant term zero. Let $A$ be the cardinality of the set $\{\alpha = (\alpha_1, \ldots, \alpha_m) \in \mathbb N^m: 1\leq |\alpha|:=\alpha_1 +\ldots +\alpha_m \leq k\}$. We may identify $P^k(\mathbb R^m)$ with $\mathbb R^A$ by $$ \sum_{1\leq \alpha \leq k} a_\alpha X^\alpha \leftrightarrow (a_\alpha)_{1\leq \alpha \leq k}.$$ Suppose that $M \subset \mathbb R^m$ and $N \subset \mathbb R^n$ are open subsets. Recall that the $k$-jet of a map $f \in \mathcal D^p(M,N)$ at a point $x \in M$ is the truncated Taylor polynomial of degree $k$. That is, $$j^kf(x)(X) = f(x) + \sum_{1\leq \alpha \leq k} \frac{\partial^\alpha f(x)}{\alpha!} (X -x)^\alpha \leftrightarrow (x, f(x), \partial^\alpha f(x))_{1\leq \alpha \leq k}$$ where $$\partial^\alpha f(x):= \frac{\partial^{|\alpha|} f(x)}{\partial x_1^{\alpha_1} \ldots \partial x_m^{\alpha_m}}.$$ Since all polynomial maps are definable, we have $J_{\mathcal D}^k(M, N) = J^k(M, N)\equiv M\times N \times \mathbb R^{n A}$. In general, for any definable submanifolds $M \subset \mathbb R^m$ and $N \subset \mathbb R^n$ , we take $(T_M, \pi_M)$ and $(T_N, \pi_N)$ to be definable tubular neighborhoods of $M$ and $N$ in $\mathbb R^m$ and $\mathbb R^n$ respectively and identify $$ J_{\mathcal D}^k(M, N) \equiv \{j^k (f \circ \pi_M) (x): f\in \mathcal D^{\infty}(M, N), x \in M\} \subset J_{\mathcal D}^k (T_M, T_N).$$ Now, consider $\pi_M$ as a map from $T_M$ to $\mathbb R^m$ and $\pi_N$ as a map from $T_N$ to $\mathbb R^n$. Then we have $$J_{\mathcal D}^k(M, N) \equiv \{j^k (\pi_N \circ g \circ \pi_M) (x): g\in \mathcal D^{\infty}(T_M, T_N), x \in M\}.$$ The composition $g\circ \pi_M(x)$ for $x\in M$ makes sense because $M\subset T_M$. Note that $J_{\mathcal D}^k(M, N)$ is a definable $\mathcal C^{p-k}$-submanifold of $T_M \times T_N \times \mathbb R^{nA}$ (see \cite{loi1}, \cite{shiota2}). \subsection{Definable topology} Let $M\subset \mathbb R^m$ and $N\subset \mathbb R^n$ be definable $\mathcal C^p$-submanifolds ($1\leq p \leq \infty$). Let $k < p$ and $U$ be a definable subset of $J_{\mathcal D}^k(M, N)$. Set $$\mathcal M(U) := \{f\in \mathcal D^p(M, N): j^k f(M) \subset U\}.$$ We can define a topology on $\mathcal D^p(M,N)$ by regarding the family $\{\mathcal M(U)\}$ as its basis. We then call this topology the \emph{$\mathcal D^k$-topology}. Since $J_{\mathcal D}^k(M, N)$ is a subset of $\mathbb R^{m +n + nA}$ we will endow $J_{\mathcal D}^k(M, N)$ with the metric induced from $\mathbb R^{m +n + nA}$. For convenience we take the metric $d(x, y) := \| x -y\|_\infty = \max \{|x_1 - y_1|, |x_2 - y_2|,\ldots\}$. In fact the choice of the metric does not affect the topologies because all metrics on a Euclidean space are equivalent. Now we define $$\mathcal U^k_\varepsilon(f) := \{g \in\mathcal D^{p}(M, N) : \forall x \in M, d(j^kf(x), j^kg(x))< \varepsilon(x)\}$$ where $f \in \mathcal D^{\infty} (M, N)$ and $\varepsilon: M \to (0,\infty)$ is a continuous definable function. Then the family $\{\mathcal U^k_\varepsilon(f)\}$ also forms a basis of the $\mathcal D^k$-topology. To see this fact let us consider the map $$ \lambda: J^k_\mathcal D(M, N) \to \mathbb R, j^k g(x) \mapsto \varepsilon(x) -d(j^k f(x), j^kg (x)).$$ Since $\lambda$ is a definable continuous map, $\lambda^{-1}(0,\infty)$ is an open definable set in $J^k_\mathcal D(M, N)$. Put $U:= \lambda^{-1}(0,\infty)$, then $\mathcal U^k_\varepsilon(f) = \mathcal M (U)$. It remains to show that for any open neighborhood $W$ of $f$ in the $\mathcal D^k$-topology there is a definable continuous function $\theta : M \to (0, \infty)$ such that $\mathcal U_\theta^k(f) \subset W$. Let $V$ be a definable open subset of $W$ such that $f\in V$. Consider the following function $$ \delta (x) := \inf\{d(j^kg(x)- j^k f(x)): j^kg(x)\in J^k_{\mathcal D}(M, N)\setminus V\}.$$ Note that $\delta$ is a positive definable function. Choosing a definable continuous function $\theta : X \to (0,\infty)$ such that $\theta < \delta$ we have $\mathcal U^k_\theta (f) \subset V$. Now suppose that $M, N$ are definable smooth manifolds. We define the $\mathcal D^\infty$-topology on $\mathcal D^\infty(M, N)$ as follows. Denote by $W^k$ the basis of the $\mathcal D^k$-topology on $\mathcal D^\infty(M, N)$ and set $$ W^\infty := \bigcup_k W^k.$$ The $\mathcal D^\infty$-topology on $\mathcal D^\infty(M, N)$ is the topology whose basis is $W^\infty$. \begin{rem} The definition of the definable topology on $\mathcal D^p(M, N)$ using vector fields as used in \cite{escribano1}, \cite{fischer1}, \cite{loi1} and \cite{shiota1} is as follows: For each $k < p$, let us fix $V_1,\ldots,V_r$ definable $\mathcal C^k$-vector fields on $M$ such that for each $x \in M$, $V_1(x),\ldots,V_r(x)$ generate the tangent space $T_xM$. For $f \in \mathcal D^{p}(M,N)$ and $\varepsilon : M \to (0,\infty)$, a continuous definable function, we define $$\mathcal B^0_{\varepsilon}(f) := \{g \in \mathcal D^{p}(M,\mathbb R) : |f(x) - g(x)| < \varepsilon(x)\}$$ and \begin{align*} \mathcal B^k_{\varepsilon}(f):=& \{g \in \mathcal D^{p}(M,\mathbb R) : |V_{i_1}\ldots V_{i_j} (f-g)(x)| < \varepsilon(x) \\ & \quad \forall x \in M, 1 \leq i_1,\ldots,i_j \leq r, 1 \leq j \leq k\} \end{align*} for $k>0$. The $\mathcal D^k$-topology on $\mathcal D^{p}(M,\mathbb R)$ is the topology whose basis is given by the sets of the form $\mathcal B^0_{\varepsilon}f$ and $\mathcal B^k_{\varepsilon}f$. The $\mathcal D^k$-topology on $\mathcal D^{p}(M,\mathbb R^n) = \mathcal D^{p}(M,\mathbb R) \times \ldots \times \mathcal D^{p}(M,\mathbb R)$ is defined to be the product topology. For $N$ a definable smooth submanifold of $\mathbb R^n$, we consider $\mathcal D^{p}(M,\mathbb R^n)$ as a subset of $\mathcal D^{p}(M,\mathbb R^n)$ and define the $\mathcal D^k$-topology on $\mathcal D^{p}(M,N)$ to be the topology induced from $\mathcal D^{p}(M,\mathbb R^n)$. If $M, N$ are definable smooth manifolds then the $\mathcal D^{\infty}$-topology on $\mathcal D^{\infty}(M,N)$ is the topology whose basis is given by the union of all open sets of $\mathcal D^k$-topologies on $\mathcal D^{\infty}(M,N)$ for $k = 0,\ldots$. Notice that the definition of the topologies above does not depend on the choices of vector fields (see \cite{escribano1}). It is not difficult to prove that the topologies defined by using jet space and the one by using vector fields are the same. A hint for this is to take $V_1, \ldots, V_n$ the orthogonal projections of the canonical vector fields of $\mathbb R^m$ onto the tangent space of $M$. \end{rem} \begin{rem} Let $M, N$ be $\mathcal C^p$-manifolds. Recall that the $\mathcal C^k$-topology ($k \leq p$) on $\mathcal C^{p}(M,N)$ is the topology generated by the basis given by the sets of the form $\{f \in \mathcal C^{p}(M,N) : j^kf(M) \subset U\}$ where $U$ is an open set in the $k$-jet bundle $J^k(M,N)$. Recall also that if $\varepsilon : M \to (0, \infty)$ is a continuous function, $f \in \mathcal C^{p}(M,N)$ and $d$ is a metric on $J^k(M,N)$, then the set $$B_{\varepsilon}f = \{g \in \mathcal C^{p}(M,N) : d(j^kf(x),j^kg(x)) < \varepsilon(x)\,\,\, \forall x \in M\}$$ is an open neighbourhood of $f$ in the $\mathcal C^k$-topology. Then, it is well known that a sequence $f_i$ converges to $f$ in the $\mathcal C^k$-topology if there exists a compact set $K$ on which $f_i$ converges uniformly to $f$ and all but finitely many $f_i$'s are equal to $f$ outside $K$ (see \cite{gg}). Now suppose that $M$ and $N$ are two definable $\mathcal C^p$ manifolds in some o-minimal structure. Consider $\mathcal D^p(M,N)$ as a subset of $\mathcal C^p(M,N)$. We remark that for the topology on $\mathcal D^p(M,N)$ induced by the $\mathcal C^k$-topology on $\mathcal C^{p}(M,N)$, transversality to submanifolds in the target is not a generic condition. This is quite easy to see in the semialgebraic setting: \begin{example}\label{example1}\rm Let $M = \mathbb R$, $N = \mathbb R^2$, $S \subset N$ be the $x$-axis and $f : M \to N$ be defined by $f(x) = (x,0)$. Clearly $f$ is semialgebraic and is non-transverse to $S$. To prove that transversality is not a generic condition for the induced topology on $\mathcal D^p(M,N)$ of semialgebraic $\mathcal C^p$-maps, we exhibit a neighbourhood of $f$ that does not contain any transverse map. For this, consider the continuous function $\varepsilon : M \to \mathbb R^+$ given by $\varepsilon(x) = e^{-x}$. Since the structure of semialgebraic sets is polynomially bounded, for any semialgebraic map $g$ such that $\|f(x)-g(x)\| < \varepsilon(x)$, there exists $a \in \mathbb R$ such that $f(x) = g(x)$ for all $x > a$. Thus no semialgebraic map in the $\varepsilon$-neighborhood of $f$ can be transverse to $S$. \end{example} It is worth noticing that a sequence $\{f_i\}$ converging to $f \in \mathcal D^p(M, N)$ in the $\mathcal D^k$-topology need not be equal outside a compact set as in the case of $\mathcal C^p(M, N)$ (with the $\mathcal C^k$-topology). Again, it is easy to realize this fact in the semialgebraic case with $k=0$. \begin{example}\rm Consider $M = \mathbb R$, $N = \mathbb R^2$ as semialgebraic $\mathcal C^1$-manifolds. Let $f, f_i: M \to N$ be semialgebraic $\mathcal C^1$-maps defined by $f(x) = (x,0)$, and $f_i(x) = (x, x^{-2i})$ outside the compact set $V = [-1, 1]$ and inside $V$ $\{f_i\}$ uniformly converging to $f$. Since the semialgebraic structure is polynomially bounded, it is not hard to show that for any semialgebraic continuous function $\varepsilon: M \to \mathbb R^+$, we have $\max\{\|(f_i (x)- f(x)\| \}< \varepsilon(x), \forall x \in M$ for $i$ big enough. This means that $\{f_i\}$ converges to $f$ in the $\mathcal D^0$-topology but the $f_i$ are not equal to $f$ outside a compact set. \end{example} \end{rem} From now on we just restrict to the $\mathcal D^\infty$-topology. It can be easily checked that Propositions 3.4, 3.5, 3.6, 3.9 and 3.10 in Chapter II of Golubitsky and Guillemin \cite{gg} still hold for the $\mathcal D^\infty$-topology in any o-minimal structure. For the sake of clarity we list some of them here and will use in the later sections. \begin{lem}\label{lem_jet_map1} Let $M, N$ be definable smooth manifolds. The map $$j^k : \mathcal D^\infty (M, N) \to \mathcal D^\infty (M, J_{\mathcal D}^k(M, N)), f \mapsto j^kf$$ is continuous in the $\mathcal D^\infty$-topology. \end{lem} \begin{proof} See Proposition 3.4 in \cite{gg}. \end{proof} \begin{lem}\label{lem_jet_map2} Let $M $, $N$ and $P $ be definable smooth manifolds. Let $f: N \to P$ be a definable smooth map. The the map $f_* : \mathcal D^\infty(M, N) \to \mathcal D^\infty(M, P)$ given by $g\mapsto f \circ g$ is a continuous map in the $\mathcal D^\infty$-topology. \end{lem} \begin{proof} See Proposition 3.5 in \cite{gg}. \end{proof} \begin{lem}\label{lem_proper} Let $M, N, P$ be smooth definable manifolds. Let $f: P \to M$ be a definable smooth map. If $f$ is proper then the map $$f^*: \mathcal D^\infty(M, N) \to \mathcal D^\infty(P, N), g\mapsto g\circ f$$ is continuous in the $\mathcal D^\infty$-topology. \end{lem} \begin{proof} See note 2, page 49, \cite{gg}. \end{proof} \begin{lem}\label{lem_restriction} Let $M \subset \mathbb R^m, N \subset \mathbb R^n$ be definable smooth manifolds. Let $(T_M,\pi_M)$ be a definable smooth tubular neighborhood of $M$ in $\mathbb R^m$. Then, the restriction map $$\mathcal D^\infty(T_M, N) \to \mathcal D^\infty(M, N), f \mapsto f|_M$$ is continuous in the $\mathcal D^\infty$-topology. \end{lem} \begin{proof} Consider the following map $$\iota^*: \mathcal D^\infty(T_M, N) \to \mathcal D^\infty(M, N), f \mapsto f\circ \iota = f|_M$$ which is induced by the inclusion map $\iota: M \to T_M$. Since $M$ is closed in $T_M$, $\iota$ is proper. By Lemma \ref{lem_proper}, $\iota^*$ is continuous. \end{proof} \begin{rem} In general if $f$ is not proper the map $f^*$ in Lemma \ref{lem_proper} is not continuous. For instance, consider an o-minimal structure that contains the graph of the exponential function. Suppose $M = N= \mathbb R$, $P = (0,1) \subset M$ and $f: P \to M$ be the inclusion map. Clearly $f$ is not a proper map. We will show that $f^*$ is not continuous. Let $\varepsilon: P \to (0,\infty)$ be a definable continuous map such that $\lim_{x\to 0} \varepsilon (x) = 0$. First, note that the set $$ U_P: = \{ g\in \mathcal D^\infty(P, N): |g (x)| < \varepsilon\}$$ is an open set in $\mathcal D^\infty (M, N)$ with the $\mathcal D^\infty$-topology. Set $U_M := (f^*)^{-1} (U_P)$. To show that $f^*$ is not continuous, it suffices to show that $U_M$ is not open. It is easy to see that $U_M$ contains the zero function and if $h\in U_M$ then $h(0) =0$. We will prove that there is no basic open neighborhood of the zero function contained in $U_M$. Indeed, if $U$ is an arbitrary basic neighborhood of the zero function, then by definition, there are $k \in \mathbb N$ and $\delta: M \to \mathbb (0, \infty)$, a definable continuous function, such that $U =\mathcal U_{\delta}^k (0)$. By Lemma \ref{lem_appro_def}, there is a definable smooth positive function $\varphi:M \to N$ such that $ \varphi \in \mathcal U_{\delta}^k (0)$. Since $\varphi(0) \neq 0$, $\varphi \not\in U_M$. This implies that $U \not \subset U_M$. \end{rem} For the rest of the paper, we restrict to an \emph{exponential o-minimal structure} (i.e. a structure containing the graph of the exponential function) which admits smooth cell decomposition. \section{Transversality Theorem} Recall that a map $f : M \to N$ between two $\mathcal C^1$-manifolds is said to be transverse to a $\mathcal C^1$-submanifold $S \subset N$ at $p \in M$ if either $f(p) \notin S$ or $D_pf(T_pM) + T_{f(p)}S = T_p N$ and in that case we write $f \pitchfork_p S$ and $f \pitchfork S$ in case $f$ is transverse at all points of $M$. If $\Sigma$ is a collection of $\mathcal C^1$-submanifolds of $N$, then $f$ is said to be transverse to $\Sigma$, denoted by $f \pitchfork \Sigma$, if $f$ is transverse to each member of $\Sigma$. This section is devoted to the proof of the following theorem. \begin{thm}\label{thm_transversality} Let $M$ and $ N $ be definable smooth manifolds. Let $\Sigma$ be a finite collection of definable $\mathcal C^1$-manifolds of $J_{\mathcal D}^k(M, N)$. Then, for any $k \in \mathbb N$ (i) the set $\mathcal A^k_{ \Sigma}:=\{ f \in \mathcal D^\infty (M, N): j^k f \pitchfork \Sigma\}$ is dense in $\mathcal D^\infty(M, N)$ with the $\mathcal D^\infty$-topology. (ii) the set $\mathcal A^k_\Sigma$ is open (and dense) in $\mathcal D^\infty(M, N)$ with the $\mathcal D^\infty$-topology if $\,\,\Sigma$ is a Whitney $(a)$-regular stratification of some definable closed subset of $J_{\mathcal D}^k(M, N)$. \end{thm} To prove Theorem \ref{thm_transversality} we need the following lemmas. \begin{lem}\label{lem_composetrans} Let $f : M \to N$ and $g : N \to P$ be smooth maps. Let $g$ be transverse to a $\mathcal C^1$-submanifold $S \subset P$ and $f$ be transverse to $g^{-1}(S)$. Then, $g \circ f$ is transverse to $S$. \end{lem} \begin{proof} If $(g\circ f)(M) \cap S = \emptyset$ then the result is trivial. The only interesting case is when there exists $x \in M$ such that $g \circ f(x) \in S$. Set $p = f(x)$. Since $g$ is transverse to $S$, $g^{-1}(S)$ is a submanifold of $N$ and \begin{align} D_pg(T_pN) + T_{g(p)}S &= T_{g(p)} P. \label{1} \end{align} Since $f$ is transverse to $g^{-1}(S)$ we have \begin{align} D_xf(T_xM) + T_{f(x)}(g^{-1}S) &= T_{f(x)}N. \label{2} \end{align} But $f(x) = p$, so (\ref{2}) becomes \begin{align} D_xf(T_xM) + T_{p}(g^{-1}S) &= T_{p}N. \label{3} \end{align} Substituting $T_pN$ from (\ref{3}) into (\ref{1}) gives \begin{align*} D_pg(D_xf(T_xM) + T_p(g^{-1}(S)) +T_{g(p)} S &= T_{g(p)} P, \intertext{and by the chain rule} D_x(g \circ f)(T_xM) + D_{f(x)} g (T_{f(x)}(g^{-1}S)) + T_{g(f(x))}S &= T_{f(g(x))}P. \end{align*} Since $D_{f(x)} g(T_{f(x)}(g^{-1}S)) \subset T_{g(f(x))}S$ we get $$D_x(g \circ f)(T_xM) + T_{g(f(x))}S = T_{f(g(x))}P.$$ This shows that $f \circ g$ is transverse to $S$ and the proof is complete. \end{proof} Next, we state Theorem 1.1 of Fischer \cite{fischer1}: \begin{lem}\label{Fischer} Let $m \geq 0$ be an integer, $U \subset \mathbb R^n$ be a definable open set and $f : U \to \mathbb R$ be a definable $C^m$-function. Then, for every definable function $\varepsilon : U \to (0,\infty)$, there is a definable smooth function $g : U \to \mathbb R$ such that $$|\partial^{\alpha}f(u) - \partial^{\alpha}g(u)| < \varepsilon(u)$$ for any $u \in U$ and $|\alpha| \leq m$. \end{lem} We then have: \begin{lem}\label{lem_appro_def} Let $U$ be an open definable subset of $\mathbb R^m$. Let $\varepsilon: U \to (0, \infty)$ be a continuous definable function. Then, for any $k \in \mathbb N$, there exists a smooth positive definable function $\varphi: U \to \mathbb R$ such that for all $\alpha \in \mathbb N^m$ with $|\alpha|\leq k$ we have $$ |\partial^\alpha \varphi (x)| < \varepsilon (x), \forall x \in U .$$ \end{lem} \begin{proof} By Lemma 1 in \cite{loi1} there is a definable positive $\mathcal C^k$-function $\psi: U \to \mathbb R$ such that \begin{equation*} |\partial^\alpha \psi (x)| < \frac{\varepsilon (x)}{2}, \end{equation*} $\forall x \in U$ and $|\alpha|\leq k .$ In particular, for $\alpha = 0$ we have: \begin{equation}\label{eq12} |\psi(x)| < \frac{\varepsilon (x)}{2}. \end{equation} Now, by Lemma \ref{Fischer} there is a smooth definable function $\varphi : U \to \mathbb R$ such that \begin{equation}\label{eq13} |\partial^\alpha \varphi (x) - \partial^\alpha \psi (x)|< \psi(x), \end{equation} $\forall x \in U$ and $|\alpha|\leq k.$ In particular, for $\alpha = 0$ we have: $$|\varphi(x) - \psi(x)| < \psi(x),$$ for all $x \in U$. Since $\psi$ is positive, this implies that $\varphi$ is a positive definable function. Then, \begin{align*} |\partial^{\alpha}\varphi(x)| & = |\partial^{\alpha}\varphi(x) - \partial^{\alpha} \psi(x) + \partial^{\alpha} \psi(x)| \\ & \leq |\partial^{\alpha}\varphi(x) - \partial^{\alpha} \psi(x)| + |\partial^{\alpha} \psi(x)| \\ & <\psi(x) + \frac{\varepsilon(x)}{2}\\ &< \varepsilon(x) \tag{by (\ref{eq12}) and (\ref{eq13})} \end{align*} $\forall x \in U$ and $|\alpha|\leq k.$ This implies that $\varphi$ is the desired map and the proof is complete. \end{proof} Our next Lemma is a definable version of the basic transversality theorem whose proof is a standard application of Sard's theorem. \begin{lem}\label{lem_density_def} Let $M, S$ and $J$ be definable smooth manifolds, and $\Phi: M \times S \to J$ be a definable smooth submersion. Let $\Sigma$ be a finite collection of definable $C^1$ submanifold of $J$. Then $$\tau (\Phi, \Sigma) = \{s \in S: \Phi_s = \Phi(., s) \pitchfork \Sigma\}$$ is a definable dense subset of $S$ with $\dim (S \setminus \tau (\Phi, \Sigma)) < \dim S$. \end{lem} \begin{proof} Similar to the proof of Lemma 3 in \cite{loi1}. \end{proof} \begin{lem}\label{lem_a_open}Let $M $ and $N$ be definable smooth manifolds. Let $\Sigma$ be a definable $\mathcal C^1$ stratification of some definable closed subset of $N$. If $\Sigma$ is Whitney $(a)$-regular then $$\mathcal A_{ \Sigma} := \{ f \in \mathcal D^\infty (M, N): f \pitchfork \Sigma\}$$ is open in $\mathcal D^\infty(M, N)$ with $\mathcal D^\infty$-topology. \end{lem} \begin{proof} Notice that open in the $\mathcal D^k$-topology implies open in the $\mathcal D^p$-topology for $p\geq k$, so it is enough to prove the case $k=1$. But, this follows from the same arguments used in Theorem 1 of Trotman \cite{trotman1} to prove a similar result in the smooth case (see also \cite{trivedi}, \cite{trotman2}). \end{proof} Finally we prove the main theorem of this section. \begin{proof}[Proof of Theorem \ref{thm_transversality}] (i) We assume that $M \subset \mathbb R^m$ and $N \subset \mathbb R^n$ are definable smooth submanifolds. First we reduce the problem to the case when $M$ is an open set of $\mathbb R^m$ and $N=\mathbb R^n$. Let $(T_M, \pi_M)$ and $(T_N, \pi_N)$ be definable tubular neighborhoods of $M$ and $N$ respectively. Let $\pi_{N*}: J_{\mathcal D}^k(M, T_N) \to J_{\mathcal D}^k(M, N)$ be the map given by $j^kf(x) \mapsto j^k (\pi_N \circ f)(x)$, where $f : M \to N \subset T_N$ is a smooth definable map. For each $S \in \Sigma$, set $S' := \pi_{N*}^{-1}(S)$. Since $\pi_{N*}$ is a definable submersion, $S'$ is a definable smooth manifold in $J_{\mathcal D}^k(M, T_N)$. Let $f\in \mathcal D^\infty(M, N)$. Consider $f$ as a map from $M$ to $\mathbb R^n$. Suppose that $g\in \mathcal D^\infty(M, \mathbb R^n)$ is sufficiently close to $f$ such that $j^kg \pitchfork S'$. We may assume that $g \in \mathcal D^\infty(M, T_N)$. Note that $j^k(\pi_N \circ g) = \pi_{N*} \circ j^k g$. Then by Lemma \ref{lem_composetrans} we have that $j^k(\pi_N \circ g) \pitchfork S$. Moreover, since the map $\mathcal D^\infty(M, T_N) \to \mathcal D^\infty(M, N)$ given by $g \mapsto \pi_N \circ g$ is continuous by Lemma \ref{lem_jet_map2}, it follows that $\pi_N\circ g$ is sufficiently close to $f$. Therefore, we can reduce the proof to the case $N= \mathbb R^n$. For each $S \in \Sigma$, set $$S'' := \{(x, y, a_\alpha)_{1\leq |\alpha|\leq k} \in J_{\mathcal D}^k(T_M, \mathbb R^n): (\pi_M(x), y, a_\alpha)_{1\leq |\alpha|\leq k} \in S\}.$$ Clearly, $S''$ is a definable smooth submanifold of $J_{\mathcal D}^k(T_M, \mathbb R^n)$. Let $f\in \mathcal D^\infty(M, \mathbb R^n)$. Set $\tilde{f} := f \circ \pi_M: T_M \to \mathbb R^n$. Suppose $g \in \mathcal D^\infty(T_M, \mathbb R^n)$ is sufficiently close to $\tilde{f}$ such that $j^kg\pitchfork S''$. By Lemma \ref{lem_restriction}, $g \circ \pi_M$ is also close to $f$, and again as a consequence of Lemma \ref{lem_composetrans}, $j^k (g \circ \pi_M) \pitchfork S$. Thus we can reduce the proof to the case when $M$ is an open set of $\mathbb R^m$. Let $f \in \mathcal D^\infty( M, \mathbb R^m)$ and $U_f$ be an open neighborhood of $f$ in the $\mathcal D^\infty$-topology. It suffices to show that there is $g \in U_f$ such that $j^kg \pitchfork \Sigma$. Since $U_f$ is open in the $\mathcal D^\infty$-topology, there are $l \in \mathbb N$ and a definable continuous function $\varepsilon: M \to (0,\infty)$ such that $\mathcal U_\varepsilon^l(f) \subset U_f$. We can choose $l$ such that $l \geq k$ (since as $l$ increases, $\mathcal U_\varepsilon^l(f)$ gets smaller). We will show that $\mathcal U_\varepsilon^l(f)$ contains an element $g$ such that $j^k g \pitchfork \Sigma$. We separate the proof into two cases as follows. \textbf{Case 1:} $M$ is an open subset of $\mathbb R^m$ and $N = \mathbb R$. Let $R = \#\{\alpha \in \mathbb N^m: |\alpha|\leq l\}$, $C= R^2 (l!)^2$. By Lemma \ref{lem_appro_def}, there is a definable smooth positive function $\varphi: M \to \mathbb R$ such that $$|\partial^\alpha \varphi(x)| < \frac{\varepsilon(x)}{C(1 +\|x\|^l)}, \text{ for all } |\alpha| \leq l.$$ Consider the following family of maps $$ F(x, t)= f_s(x) := f(x) + \sum_{|\alpha|\leq k} s_\alpha x^\alpha\varphi(x),$$ where $x \in M$ and $s = (s_\alpha)_{|\alpha|\leq k}\in I^R$ with $I = (0,1)$. By the same computation as in Theorem 2 of \cite{loi1}, it follows that $f_s \in \mathcal U_\varepsilon^l (f)$ for all $s\in I ^R$. In addition the map $$\Phi: M \times I^R \to J^k_{\mathcal D}(M, \mathbb R), (x, t) \mapsto(x, (\partial^\alpha f_s(x)))_{|\alpha|\leq k}$$ is a submersion. By Lemma \ref{lem_density_def}, there exists an $s$ such that $j^k f_s \pitchfork \Sigma$. Then, $g = f_s$ has the required property. \textbf{Case 2:} $M$ is an open subset of $\mathbb R^m$ and $N = \mathbb R^n$. The result is obtained by applying the same arguments as in the case 1 for the family $$F(x, s) := (f_1(x) + \sum_{|\alpha|\leq k} s_{1,\alpha} x^\alpha\varphi(x),\ldots, f_n(x) + \sum_{|\alpha|\leq k} s_{n,\alpha} x^\alpha\varphi(x)),$$ where $s = (s_{i, \alpha})_{1\leq n, |\alpha|\leq k} \in I^{nR}$. (ii) Consider the map $$ j^k: \mathcal D^\infty(M, N) \to \mathcal D^\infty(M, J^k_{\mathcal D} (M, N)).$$ Let $\mathcal L:= \{ f\in \mathcal D^\infty(M, J^k_{\mathcal D} (M, N)), f \pitchfork \Sigma\}$. Since $\Sigma$ is an (a)-regular stratification, $\mathcal L$ is open in $\mathcal D^\infty(M, J^k_{\mathcal D} (M, N)$ (see Lemma \ref{lem_a_open}). By Lemma \ref{lem_jet_map1}, this shows that the map $j^k$ is continuous, therefore, $\mathcal A^k_\Sigma = (j^k)^{-1} (\mathcal L)$ is open. \end{proof} \section{The set of non-transverse maps} Let $M$ and $N $ be definable smooth manifolds. Let $\Sigma$ and $\Sigma'$ be finite collections of $\mathcal C^1$-submanifolds of $N$. We write $\Sigma < \Sigma'$ if elements of $\Sigma$ are also elements of $\Sigma'$. If $\Sigma$ and $\Sigma'$ are stratifications of the same set $V \subset N$ then $\Sigma'$ is said to be a refinement of $\Sigma$ if every stratum of $\Sigma$ is a union of some strata of $\Sigma'$. Denote by $\mathcal A_{\Sigma}$ the set of maps transverse to $\Sigma$, i.e. $\mathcal A_{\Sigma} := \{f \in \mathcal D^{\infty}(M,N) : f \pitchfork \Sigma\}.$ The following observations are clear from the definitions: (1) If $\Sigma < \Sigma'$ then $\mathcal A_{\Sigma'} \subset \mathcal A_{\Sigma}$. (2) If $\Sigma'$ is a refinement of $\Sigma$ then $\mathcal A_{\Sigma'} \subset \mathcal A_{\Sigma}$. We will now prove the main theorem of this section. \begin{thm}\label{cor_strong_transversality} Let $\Sigma$ be a finite collection of definable $\mathcal C^1$-submanifolds of $J_{\mathcal D}^k(M, N)$. For $k\in \mathbb N$, define $$\mathcal A^k_{ \Sigma}:=\{ f \in \mathcal D^\infty (M, N): j^k f \pitchfork \Sigma\}.$$ Then $\mathcal D^\infty(M, N) \setminus \mathcal A^k_{ \Sigma}$ is nowhere dense in $\mathcal D^\infty(M, N)$ with the $\mathcal D^\infty$-topology. \end{thm} \begin{proof} It suffices to show that $\mathcal A^k_\Sigma$ contains a subset which is open and dense in $\mathcal D^\infty(M, N)$ with the $\mathcal D^\infty$-topology. Let $X$ be the union of all elements of $\Sigma$. We denote by $\overline{X}$ the closure of $X$ in $J_{\mathcal D}^k(M, N)$. Let $\Sigma'$ be a definable $(a)$-regular stratification of $\overline{X}$ of $X$ which is compatible with the family consisting of $\overline{X}\setminus X$ and all strata of $\Sigma$ (this is possible due to \cite{loi3}, \cite{loi2} or \cite{ntt}). Write $$ \Sigma' = \{\Sigma'_1, \Sigma'_2\}$$ where $\Sigma'_1$ is the stratification of $\overline{X}\setminus X$ and $ \Sigma'_2$ is a refinement of $\Sigma$. By the observations above, that shows $$\mathcal A^k_{\Sigma'} \subset \mathcal A^k_{\Sigma'_2} \subset \mathcal A^k_{\Sigma}.$$ Since $\Sigma'$ is a Whitney $(a)$-regular stratification of the closed set $\overline{X}$, by Theorem \ref{thm_transversality}, $\mathcal A^k_{\Sigma'}$ is open and dense in $\mathcal D^\infty(M, N)$ with the $\mathcal D^\infty$-topology. This ends the proof. \end{proof} \begin{cor}\label{cor_non_transverse} Let $\Sigma$ be a finite collection of definable $\mathcal C^1$-submanifolds of $N$. Then $\mathcal D^\infty(M, N)\setminus \mathcal A_\Sigma$ is nowhere dense in $\mathcal D^\infty(M, N)$ with the $\mathcal D^\infty$-topology. \end{cor} \begin{proof} Note that $J^0_{\mathcal D}(M, N) = M \times N$ and $j^0 f(x) = (x, f(x))$. Since the projection $\pi: M \times N \to N$ $(x, y) \mapsto y$ is a submersion, $\Sigma':= \pi^{-1}(\Sigma)$ is a collection submanifolds of $M \times N$. By Lemma \ref{lem_composetrans}, if $f \in \mathcal D^\infty (M, N)$ such that $j^0 f \pitchfork \Sigma'$ then $ f\pitchfork \Sigma$. This implies that the set $\mathcal A^0_{\Sigma'} := \{f \in \mathcal D^\infty(M, N): j^0f\pitchfork \pi^{-1}(\Sigma)\}$ is contained in $\mathcal A_\Sigma$. By Theorem \ref{cor_strong_transversality}, the complement of $\mathcal A^0_{\Sigma'}$ in $\mathcal D^\infty(M, N)$ is nowhere dense, then so is the complement of $\mathcal A_\Sigma$. \end{proof} \section{Openness implies Whitney (a)-regularity} Let $\Sigma$ be a definable stratification of a definable set $V\subset \mathbb R^n$. Suppose that $p \in V$ lies in a stratum $S$ of $\Sigma$. Consider a linear subspace $H$ of $T_p\mathbb R^n \equiv \mathbb R^n$ such that $H + T_p S= \mathbb R^n$. Let $M$ be a definable smooth manifold such that $\dim M \geq \dim H$. We say that a map $f : M \to \mathbb R^n$ \emph{has a given derivative} at a point $x \in M$ if: 1. $f(x) = p$, 2. $D_xf(T_xM) = H$. Set $\mathscr D_H := \{f \in \mathcal D^{\infty}(M,\mathbb R^n) :\exists\, x \in M, f(x) = p, D_xf(T_xM) = H\}.$ We will show that there exists a map $f \in \mathscr D_H$ transverse to $\Sigma$. Notice that the transversality theorem guarantees the existence of maps transverse to any stratification but not of transverse maps with a given derivative. For example, if the dimension of $M$ is less than the codimension of a submanifold $S$ of $\mathbb R^n$, then the only maps transverse to $S$ are those that do not intersect $S$. \begin{lem}\label{givendf} There exists an $f \in \mathscr D_H$ such that $f \pitchfork \Sigma$. \end{lem} \begin{proof} Without loss of generality we assume that $H= \mathbb R^k \times \{0\} \subset \mathbb R^n$ and $p = 0$. First we will show that there exists a definable smooth map $g : \mathbb R^k \to \mathbb R^n$ such that $g(0) = 0$, $D_0g(\mathbb R^k) = H$ and $g \pitchfork \Sigma$. Consider the map $\Psi : \mathbb R^{k}\times\mathbb R^{n-k} \to \mathbb R^n$ given by $$\Psi(x,s)= \psi_s(x) := (x_1,\ldots,x_k,s_1\|x\|^2, \ldots, s_{n-k}\|x\|^2)$$ where $x = (x_1, \ldots, x_k)$, $s = (s_1, \ldots, s_{n-k})$ and $\|x\|^2 = x_1^2+\cdots+x_k^2$. The Jacobian of this map looks like \begin{align*} \left [ \begin{array}{c | c} I_k & 0\\ \hline \begin{array}{ccc} . & . & .\\ . & . & .\\ . & . & .\\ \end{array} & \begin{array}{ccc} {\scriptstyle \|x\|^2} & \ldots & 0 \\ \vdots & \ddots & \vdots \\ 0 & \ldots & {\scriptstyle \|x\|^2}\\ \end{array} \end{array}\right ]_. \end{align*} It is then clear that $\Psi$ is a submersion on $\{\mathbb R^k \setminus \{0\}\} \times \mathbb R^{n-k}$. Put $\Psi' := \Psi|_{\{\mathbb R^k \setminus \{0\}\} \times \mathbb R^{n-k}}$. By Lemma \ref{lem_density_def}, $$\mathcal S := \{s \in \mathbb R^{n-k} : \psi'_s := \Psi'(x,s) \pitchfork \Sigma\}$$ is dense in $\mathbb R^{n-k}$. We claim that $\forall s \in \mathcal S$, $\psi_s \pitchfork \Sigma$. Indeed, at $x = 0$ we have $D_0\psi_s (\mathbb R^k) = H$ which is transverse to $\Sigma$ at $\psi_s(0)=0$, so $\psi_s \pitchfork_0 \Sigma$. For $x \neq 0$, we have $\psi_s = \psi_s'$ which is transverse to $\Sigma$. This proves the claim. Next, we fix an $s \in \mathcal S$ and show that there exists a definable smooth map $h : M \to \mathbb R^k$ such that $h(x) = 0$ for some $x \in M$ which is a submersion at $x$ and $h \pitchfork \psi_s^{-1}(\Sigma)$. Take a point $x \in M$ and a definable smooth coordinate chart $\xi : V \to \mathbb R^m$ around $x$ such that $\xi(x) = 0$. Now, define $L : \mathbb R^m \to \mathbb R^k$ by $$L(x_1,\ldots,x_m) = (x_1,\ldots,x_k).$$ Since $m \geq k$, $L$ is a linear submersion at $0 \in \mathbb R^m$. Put $h' := L\circ \xi : V \to \mathbb R^k$. Since $\xi^{-1}$ is a diffeomorphism, $h'$ is a submersion at $x$. We can then smoothly extend $h'$ to $M$ by means of a smooth definable bump function. Let us call this extension $\tilde{h}$ and notice that it is a submersion at $x$. Consider the set $\mathscr S$ of definable smooth maps between $M$ and $\mathbb R^k$ that are submersion at $x \in M$. Obviously, $\mathscr S$ is non-empty since $\tilde{h}\in \mathscr S$. Moreover, it is easy to see that $\mathscr S$ is an open set in $\mathcal D^{\infty}(M,\mathbb R^n)$ with the $\mathcal D^\infty$-topology. Since the set of smooth definabe maps transverse to $\psi_s^{-1}(\Sigma)$, say $\mathscr T$, is dense in $\mathcal D^{\infty}(M,\mathbb R^n)$ with the $\mathcal D^\infty$-topology (by Corollary \ref{cor_non_transverse}) the intersection of $\mathscr S$ with $\mathscr T$ is non-empty. Take $h \in \mathscr S \cap \mathscr T$. Then $h$ is a smooth definable map transverse to $\psi_s^{-1}(\Sigma)$ which is also a submersion at $x \in M$. Now, put $f := \psi_s \circ h : M \to \mathbb R^n$. We claim that $f$ is the required map, i.e. $D_xf(T_xM) = H$ and $f$ is transverse to $\Sigma$. Just following the construction of $f$ it is easy to see that $D_xf(T_xM) = H$. We need to verify that $f$ is transverse to $\Sigma$. This fact immediately follows from Lemma \ref{lem_composetrans} and the proof is complete. \end{proof} We can now prove a definable version of the main result of Trotman \cite{trotman1} (see also \cite{trotman2}). The result is stated as follows. \begin{thm}\label{thm_openness_a_regularity} Let $N$ be a definable smooth manifold. Let $\Sigma$ be a definable stratification of some definable closed subset of $N$. Then the following statements are equivalent: (1) $\Sigma$ is Whitney $(a)$-regular. (2) For any definable smooth manifold $M$, the set $\{f \in \mathcal D^{\infty}(M,N) : f \pitchfork \Sigma\}$ is open in $\mathcal D^{\infty}(M,N)$ with the $\mathcal D^\infty$-topology. (3) For any definable smooth manifold $M$ of dimension greater than or equal to the codimension (in $N$) of the smallest stratum in $\Sigma$, the set $$\{f \in \mathcal D^\infty(M, N): f \pitchfork \Sigma\}$$ is open in $\mathcal D^\infty(M, N)$ with the $\mathcal D^\infty$-topology. \end{thm} \begin{proof} That (1) implies (2) follows from \ref{lem_a_open}. That (2) implies (3) is trivial. We show $(3)$ implies $(1)$. Suppose $\Sigma$ is not Whitney $(a)$-regular. Then there exist a pair of adjacent strata $(X, Y)$ of $\Sigma$ and a point $y\in Y$ such that $X$ is not $(a)$-regular over $Y$ at $y$. Let $(U, \phi)$ be a definable smooth chart of $N$ around the point $y$ such that $\phi(U) = \mathbb R^n$ and $\phi(y) = 0$ where $n = \dim N$. Set $\Sigma' := \phi (\Sigma' \cap U)$. Since diffeomorphisms preserve\footnote{In fact, $(a)$-regularity is a $C^1$-invariant; see Proposition 2.1.2 in the Ph.D. thesis \cite{Nhan2} of the first author.} Whitney $(a)$-regularity, $\Sigma'$ is not $(a)$-regular. More precisely, if $X':= \phi (X \cap U)$ and $Y' := \phi(Y \cap U)$, then $X'$ is not $(a)$-regular over $Y'$ at $0$. This means that there is a sequence $\{x_i\}$ in $X'$ tending to $0$ such that $T_{x_i} X'$ tends to $\tau \in \mathbb G_n^{l}$ ($l =\dim X'$) and $T_0 Y' \not\subset \tau$. Let $r = \dim Y'$. By the classical arguments as in \cite{trotman1}, there is a linear subspace $H$ of $\mathbb R^n$ with $\dim H = n -r$ such that $H + T_0 Y' = \mathbb R^n$ and $H +\tau \neq \mathbb R^n$. Suppose that $\dim (H +\tau) = b$ and $\dim (H \cap \tau) = c$. Note that $c \leq n-r\leq b < n$. Let $\{v_1, \ldots,v_n\}$ be basis of $\mathbb R^n$ such that $H$ is spanned by $\{v_1, \ldots, v_{n-r}\}$ and $\tau$ is spanned by $\{v_{n-r-c}, \ldots, v_{b}\}$. Now, at each point $x_i \in X'$, choose $\{v^i_1, \ldots v^i_n\}$ to be a basis of $T_{x_i} \mathbb R^n$ such that $T_{x_i}X'$ spanned by $\{v^i_{n-r-c}, \ldots, v^i_{b}\}$ and $\lim_{i\to \infty} v^i_j= v_j$ for $j = 1, \ldots, n$. Denote by $H_i$ the linear subspace spanned by $\{v^i_1, \ldots, v^i_{n-r}\}$. We have \begin{equation*}\label{eq_linear_spaces} \lim_{i\to \infty}H_i = H \hspace{1cm} \text{and}\hspace{1cm} H_i + T_{x_i}X' \neq T_{x_i} \mathbb R^n =\mathbb R^n. \tag{$\dagger$} \end{equation*} Next, for each $i$ we denote by $A_i: \mathbb R^n \to \mathbb R^n$ the linear isomorphism such that $A_i (v_j) = v_j^i, \forall j = 1\ldots, n$. Note that $A_i$ is uniquely determined. Since $\lim_{i \to \infty} v_j^i = v_j, \forall j = 1\ldots, n$, $\lim_{i \to \infty} A_i = {Id}_n$. Let $M$ be a definable smooth manifold as in the hypothesis. Note that $\dim M \geq \dim H$ since $\dim M \geq n- \min_{S\in \Sigma}\{\dim S\}$. By Lemma \ref{givendf}, there exists a definable smooth map $f: M \to \mathbb R^n$ such that $f(x) =0$ (for some $x\in M$), $D_x f (T_xM) = H$ and $f\pitchfork \Sigma'$. Let $f_i := A_i\circ f + x_i$. We have $f_i(x) = x_i$, $D_x f_i (T_x M) = H_i$ but $f_i$ does not converge to $f$ in the $\mathcal D^\infty$-topology since it is not `close' to $f$ at infinity. Now let $\lambda: M \to [0, 1]$ be a definable smooth function whose value equals $1$ in a small definable neighborhood of $x$ and equals $0$ outside a bigger definable compact neighborhood of $x$ ($\lambda$ can be constructed by using exponential bump functions). Define $g_i := \lambda f_i + (1-\lambda) f$. It is easy to check that $g_i(x) = x_i$, $D_x g_i (T_x M) = H_i$ and $\{g_i\}$ converges to $f$ in $\mathcal D^\infty$-topology. It follows from (\ref{eq_linear_spaces}) that $f_i$ is not transverse to $\Sigma'$ at $x$ while $f$ is transverse to $\Sigma$ at every point on $M$. Set $h:= \phi^{-1}\circ f$ and $h_i =\phi^{-1} \circ g_i$. By Lemma \ref{lem_composetrans}, $h_i \not \pitchfork \Sigma$ and $ h\pitchfork \Sigma$. On the other hand, since $\phi$ is a definable smooth diffeomorphism and $\{f_i\}$ tends to $f$, $\{h_i\}$ tends to $h$ (in the $\mathcal D^\infty$-topology). This shows that the set $$\{f \in \mathcal D^\infty(M, N): f \pitchfork \Sigma\}$$ is not open, a contradiction. \end{proof} \begin{rem} (i) The results of Theorem \ref{cor_strong_transversality} and Theorem \ref{thm_openness_a_regularity} hold for the $\mathcal D^k$-topology ($0<k < \infty$) without any restriction on the o-minimal structure. (ii) The method used to prove the above theorem works also in the smooth case. Thus we get a new more constructive proof of the theorem of Trotman (see \cite{trotman2}, \cite{trotman1}). \end{rem}

\section{Introduction} \label{} Consider the integral domain \[ \z[\sqrt{-5}] = \{ a+b\sqrt{-5} \mid a,b \in \mathbb{Z}\}. \] Your undergraduate abstract algebra text probably used it as the base example of an integral domain that is not a unique factorization domain (or UFD). The Fundamental Theorem of Arithmetic fails in $\z[\sqrt{-5}]$ as this domain contains elements with multiple factorizations into irreducibles; for example, \begin{equation}\label{2ndzroot5} 6=2\cdot 3 = (1 - \sqrt{-5})(1 + \sqrt{-5}) \end{equation} even though $2,3, 1 - \sqrt{-5}$, and $1+ \sqrt{-5}$ are pairwise non-associate irreducible elements in $\z[\sqrt{-5}]$. To argue this, the norm on $\z[\sqrt{-5}]$, i.e., \begin{align} N(a+b\sqrt{-5})=a^2+5b^2, \label{ norm} \end{align} plays an important role, as it is a multiplicative function satisfying the following properties: \begin{itemize} \item $N(\alpha)=0$ if and only if $\alpha=0$; \vspace{2pt} \item $N(\alpha\beta)=N(\alpha)N(\beta)$ for all $\alpha ,\beta\in \z[\sqrt{-5}]$; \vspace{2pt} \item $\alpha$ is a unit if and only if $N(\alpha) = 1$ (i.e., $\pm 1$ are the only units of $\z[\sqrt{-5}]$); \vspace{2pt} \item if $N(\alpha)$ is prime, then $\alpha$ is irreducible. \end{itemize} However, introductory abstract algebra books seldom dig deeper than what Equation \eqref{2ndzroot5} does. The goal of this paper is to use ideal theory to describe exactly how elements in $\z[\sqrt{-5}]$ factor into products of irreducibles. In doing so, we will show that $\z[\sqrt{-5}]$ satisfies a nice factorization property, which is known as \textit{half-factoriality}. Thus, we say that $\z[\sqrt{-5}]$ is a \emph{half-factorial domain} (or HFD). Our journey will require nothing more than elementary algebra, but will give the reader a glimpse of how The Fundamental Theorem of Ideal Theory resolves the non-unique factorizations of $\z[\sqrt{-5}]$. The notion that unique factorization in rings of integers could be recovered via ideals was important in the late 1800's in attempts to prove Fermat's Last Theorem (see \cite[Chapter~11]{PD}). Our presentation is somewhat interactive, as many steps that follow from standard techniques of basic algebra are left to the reader as exercises. The only background we expect from the reader are introductory courses in linear algebra and abstract algebra. Assuming such prerequisites, we have tried to present here a self-contained and friendly approach to the phenomenon of non-uniqueness of factorizations occurring in $\z[\sqrt{-5}]$. More advanced and general arguments (which apply to any ring of integers) can be found in~\cite{Ma} and~\cite{PD}. \medskip \section{Integral Bases and Discriminants} \label{ bases} Although in this paper we are primarily concerned with the phenomenon of non-unique factorizations in the particular ring of integers $\z[\sqrt{-5}]$, it is more enlightening from an algebraic perspective to introduce our needed concepts for arbitrary commutative rings with identity, rings of integers, or quadratic rings of integers, depending on the most appropriate context for each concept being introduced. In what follows, we shall proceed in this manner while trying, by all means, to keep the exposition as elementary as possible. An element $\alpha \in {\mathbb{C}}$ is said to be \emph{algebraic} provided that it is a root of a nonzero polynomial with rational coefficients, while $\alpha$ is said to be an \emph{algebraic integer} provided that it is a root of a monic polynomial with integer coefficients. It is not hard to argue that every subfield of ${\mathbb{C}}$ contains $\q$ and is a $\q$-vector space. \begin{definition} A subfield $K$ of ${\mathbb{C}}$ is called an \emph{algebraic number field} provided that it has finite dimension as a vector space over $\q$. The subset \[ \mathcal{O}_K := \{\alpha \in K \mid \alpha \ \text{ is an algebraic integer}\} \] of $K$ is called the \emph{ring of integers} of $K$. \end{definition} The ring of integers of any algebraic number field is, indeed, a ring. The reader is invited to verify this observation. If $\alpha$ is a complex number, then $\q(\alpha)$ denotes the smallest subfield of ${\mathbb{C}}$ containing $\alpha$. It is well known that a subfield $K$ of ${\mathbb{C}}$ is an algebraic number field if and only if there exists an algebraic number $\alpha \in {\mathbb{C}}$ such that $K = \q(\alpha)$ (see, for example, \cite[Theorem~2.17]{Ja}). Among all algebraic number fields, we are primarily interested in those that are two-dimensional vector spaces over~$\q$. \begin{definition} An algebraic number field that is a two-dimensional vector space over $\q$ is called a \emph{quadratic number field}. If $K$ is a quadratic number field, then $\mathcal{O}_K$ is called a \emph{quadratic ring of integers}. \end{definition} For $\alpha \in {\mathbb{C}}$, let $\mathbb{Z}[\alpha]$ denote the set of all polynomial expressions in $\alpha$ having integer coefficients. Clearly, $\mathbb{Z}[\alpha]$ is a subring of $\q(\alpha)$. It is also clear that, for $d \in \mathbb{Z}$, the field $\q(\sqrt{d})$ has dimension at most two as a $\q$-vector space and, therefore, it is an algebraic number field. Moreover, if $d \notin \{0,1\}$ and $d$ is squarefree (i.e., $d$ is not divisible by the square of any prime), then it immediately follows that $\q(\sqrt{d})$ is a two-dimensional vector space over $\q$ and, as a result, a quadratic number field. As we are mainly interested in the case when $d = -5$, we propose the following exercise. \begin{exercise} \label{ of integers of quadratic fields} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer such that $d \equiv 2, 3 \pmod{4}$. Prove that $\mathbb{Z}[\sqrt{d}]$ is the ring of integers of the quadratic number field $\q(\sqrt{d})$. \end{exercise} \begin{remark} When $d \in \mathbb{Z} \setminus \{0,1\}$ is a squarefree integer satisfying $d \equiv 1 \pmod{4}$, it is not hard to argue that the ring of integers of $\q(\sqrt{d})$ is $\mathbb{Z}[\frac{1 + \sqrt{d}}{2}]$. However, we will not be concerned with this case as our case of interest is $d = -5$. \end{remark} For $d$ as specified in Exercise~\ref{ of integers of quadratic fields}, the elements of $\mathbb{Z}[\sqrt{d}]$ can be written in the form $a + b \sqrt{d}$ for $a,b \in \mathbb{Z}$. The \emph{norm} $N$ on $\mathbb{Z}[\sqrt{d}]$ is defined by \[ N(a + b\sqrt{d}) = a^2 - d b^2 \] (cf. Equation~(\ref{ norm})). The norm $N$ on $\mathbb{Z}[\sqrt{d}]$ also satisfies the four properties listed in the introduction. Let us now take a look at the structure of an algebraic number field $K$ with linear algebra in mind. For $\alpha \in K$ consider the function $m_\alpha \colon K \to K$ defined via multiplication by $\alpha$, i.e., $m_\alpha(x) = \alpha x$ for all $x \in K$. One can easily see that $m_\alpha$ is a linear transformation of $\q$-vector spaces. Therefore, after fixing a basis for the $\q$-vector space $K$, we can represent $m_\alpha$ by a matrix $M$. The \emph{trace} of $\alpha$, which is denoted by $\text{Tr}(\alpha)$, is defined to be the trace of the matrix $M$. It is worth noting that $\text{Tr}(\alpha)$ does not depend on the chosen basis for $K$. Also, notice that $\text{Tr}(\alpha) \in \q$. Furthermore, if $\alpha \in \mathcal{O}_K$, then $\text{Tr}(\alpha) \in \mathbb{Z}$ (see \cite[Lemma~4.1.1]{RE}, or Exercise~\ref{ and discriminant} for the case when $K = \q(\sqrt{d})$). \begin{definition} Let $K$ be an algebraic number field that has dimension $n$ as a $\q$-vector space. The \emph{discriminant} of a subset $\{\omega_1, \dots, \omega_n\}$ of $K$, which is denoted by $\Delta[\omega_1, \dots, \omega_n]$, is $\det \, T$, where $T$ is the $n \times n$ matrix $\big(\text{Tr}(\omega_i \omega_j)\big)_{1 \le i,j \le n}$. \end{definition} With $K$ as introduced above, if $\{\omega_1, \dots, \omega_n\}$ is a subset of $\mathcal{O}_K$, then it follows that $\Delta[\omega_1, \dots, \omega_n] \in \mathbb{Z}$ (see Exercise~\ref{ and discriminant} for the case when $K = \q(\sqrt{d})$). In addition, the discriminant of any basis for the $\q$-vector space $K$ is nonzero; we will prove this for $K = \q(\sqrt{d})$ in Proposition~\ref{}. \begin{exercise} \label{ and discriminant} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer such that $d \equiv 2,3 \pmod{4}$. \begin{enumerate} \item If $\alpha= a_1 + a_2\sqrt{d} \in \mathbb{Z}[\sqrt{d}]$, then $\text{Tr}(\alpha) = 2a_1$. \vspace{3pt} \item If, in addition, $\beta = b_1 + b_2 \sqrt{d} \in \mathbb{Z}[\sqrt{d}]$, then \[ \Delta[\alpha,\beta] = \left( \det \left[ \begin{matrix} \alpha \, & \sigma(\alpha) \\ \beta \, & \sigma(\beta) \end{matrix}\right] \right)^2 = 4d(a_1b_2 - a_2b_1)^2, \] where $\sigma(x + y\sqrt{d}) = x - y \sqrt{d}$ for all $x,y \in \mathbb{Z}$. \end{enumerate} \end{exercise} \medskip \begin{example} \label{ of the canonical integral basis} Let $d \notin \{0,1\}$ be a squarefree integer such that $d \equiv 2,3 \pmod{4}$. It follows from Exercise~\ref{ and discriminant} that the subset $\{1, \sqrt{d}\}$ of the ring of integers $\mathbb{Z}[\sqrt{d}]$ satisfies that $\Delta[1, \sqrt{d}] =4d$. \end{example} \medskip We proceed to introduce the concept of integral basis. \begin{definition} Let $K$ be an algebraic number field of dimension $n$ as a vector space over $\q$. The elements $\omega_1, \dots , \omega_n \in \mathcal{O}_K$ form an \emph{integral basis} for $\mathcal{O}_K$ if for each $\beta \in \mathcal{O}_K$ there are unique $z_1,\dots ,z_n \in \mathbb{Z}$ satisfying $\beta = z_1 \omega_1 + \dots + z_n \omega_n$. \end{definition} \begin{example} \label{ integral basis} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer such that $d \equiv 2,3 \pmod{4}$. Clearly, every element in $\mathbb{Z}[\sqrt{d}]$ is an integral linear combination of $1$ and $\sqrt{d}$. Suppose, on the other hand, that $a_1 + a_2 \sqrt{d} = b_1 + b_2 \sqrt{d}$ for some $a_1, a_2, b_1, b_2 \in \mathbb{Z}$. Note that $a_2 = b_2$; otherwise $\sqrt{d} = \frac{a_1 - b_1}{b_2 - a_2}$ would be a rational number. As a result, $a_1 = b_1$. Thus, we have verified that every element of $\mathbb{Z}[\sqrt{d}]$ can be uniquely written as an integral linear combination of $1$ and $\sqrt{d}$. Hence, the set $\{1, \sqrt{d}\}$ is an integral basis for $\mathbb{Z}[\sqrt{d}]$. \end{example} In general, the ring of integers of any algebraic number field has an integral basis (see~\cite[Theorem~3.27]{Ja}). On the other hand, although integral bases are not unique, any two integral bases for the same ring of integers have the same discriminant. We shall prove this for $\mathbb{Z}[\sqrt{d}]$ in Theorem~\ref{}. \\ \noindent {\bf Notation:} If $S$ is a subset of the complex numbers, then we let $S^\bullet$ denote $S \! \setminus \! \{0\}$. \medskip \begin{lemma}\label{intbasis} Let $K$ be an algebraic number field of dimension $n$ as a $\q$-vector space. An integral basis for $\mathcal{O}_K$ is a basis for $K$ as a vector space over $\q$. \end{lemma} \begin{proof} Suppose that $\{\omega_1, \dots, \omega_n\}$ is an integral basis for $\mathcal{O}_K$, and take rational coefficients $q_1, \dots, q_n$ such that \[ q_1 \omega_1 + \dots + q_n \omega_n = 0. \] Multiplying the above equality by the common denominator of the nonzero $q_i$'s and using the fact that $\{\omega_1, \dots, \omega_n\}$ is an integral basis for the ring of integers $\mathcal{O}_K$, we obtain that $q_1 = \dots = q_n = 0$. Hence, $\{\omega_1, \dots, \omega_n\}$ is a linearly independent set of the $\q$-vector space $K$. As $K$ has dimension $n$ over $\q$, the set $\{\omega_1, \dots, \omega_n\}$ is a basis for the vector space $K$ over $\q$. \end{proof} \medskip \begin{proposition} \label{} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer with $d \equiv 2,3 \pmod{4}$. If $\{\alpha_1,\alpha_2\}$ is a vector space basis for $\q(\sqrt{d})$ contained in $\mathbb{Z}[\sqrt{d}]$, then $\Delta[\alpha_1,\alpha_2] \in \mathbb{Z}^\bullet$. \end{proposition} \begin{proof} From the fact that $\{\alpha_1, \alpha_2\} \subseteq \mathbb{Z}[\sqrt{d}]$, it follows that $\Delta[\alpha_1, \alpha_2] \in \mathbb{Z}$. So suppose, by way of contradiction, that $\Delta[\alpha_1, \alpha_2] = 0$. Taking $\{\omega_1, \omega_2\}$ to be an integral basis for $\mathbb{Z}[\sqrt{d}]$, one has that \[ \begin{array} {lclcl} \alpha_1 & = & z_{1,1}\omega_1 & + & z_{1,2}\omega_2 \\ \alpha_2 & = & z_{2,1}\omega_1 & + & z_{2,2}\omega_2, \end{array} \] for some $z_{i,j} \in \mathbb{Z}$. Using Exercise~\ref{ and discriminant}, we obtain \begin{multline}\label{discrim} \Delta[\alpha_1,\alpha_2] = \left(\det \left[ \begin{matrix} \alpha_1 & \sigma(\alpha_1) \\ \alpha_2 & \sigma(\alpha_2) \end{matrix}\right]\right)^2 = \left(\det \left( \left[\begin{matrix} z_{1,1} & z_{1,2} \\ z_{2,1} & z_{2,2} \end{matrix}\right] \left[ \begin{matrix} \omega_1 & \sigma(\omega_1) \\ \omega_2 & \sigma(\omega_2) \end{matrix}\right]\right)\right)^2 \\ = \left(\det \left[\begin{matrix} z_{1,1} & z_{1,2} \\ z_{2,1} & z_{2,2} \end{matrix}\right]\right)^2 \left(\det \left[ \begin{matrix} \omega_1 & \sigma(\omega_1) \\ \omega_2 & \sigma(\omega_2) \end{matrix}\right]\right)^2 = \left(\det \left[\begin{matrix} z_{1,1} & z_{1,2} \\ z_{2,1} & z_{2,2} \end{matrix}\right]\right)^2\Delta[\omega_1,\omega_2], \end{multline} where $\sigma(x + y\sqrt{d}) = x - y \sqrt{d}$ for all $x,y \in \mathbb{Z}$. If $\omega_1 = 1$ and $\omega_2 = \sqrt{d}$, then \[ \det \left[\begin{matrix} z_{1,1} & z_{2,1} \\ z_{1,2} & z_{2,2} \end{matrix}\right]= \det \left[\begin{matrix} z_{1,1} & z_{1,2} \\ z_{2,1} & z_{2,2} \end{matrix}\right]=0, \] and so there are elements $q_1, q_2\in \q$ not both zero with \[ \left[\begin{matrix} z_{1,1} & z_{2,1} \\ z_{1,2} & z_{2,2} \end{matrix}\right] \left[\begin{matrix} q_1 \\ q_2 \end{matrix}\right] = \left[\begin{matrix} 0 \\ 0 \end{matrix}\right]. \] Hence, \begin{align*} 0 &= \omega_1(q_1 z_{1,1}+q_2 z_{2,1})+\omega_2(q_1 z_{1,2}+q_2 z_{2,2}) \\ &= q_1(z_{1,1}\omega_1 + z_{1,2}\omega_2) + q_2(z_{2,1}\omega_1 + z_{2,2}\omega_2) \\ &= q_1\alpha_1 + q_2\alpha_2, \end{align*} which is a contradiction because the set $\{\alpha_1, \alpha_2\}$ is linearly independent in the vector space $\q(\sqrt{d})$. Thus, $\Delta[\alpha_1,\alpha_2]\neq 0$, as desired. \end{proof} \medskip \begin{exercise} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer with $d \equiv 2,3 \pmod{4}$. Show that $\Delta[\alpha_1,\alpha_2] \neq 0$ whenever $\{\alpha_1,\alpha_2\}$ is a basis for the $\q$-vector space $\q(\sqrt{d})$. \end{exercise} Using Lemma~\ref{intbasis} and Exercise~\ref{ and discriminant}, we obtain the following important result. \begin{corollary} \label{} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer with $d \equiv 2,3 \pmod{4}$. The discriminant of each integral basis for $\mathbb{Z}[\sqrt{d}]$ is in $\mathbb{Z}^\bullet$. \end{corollary} \medskip \noindent {\bf Notation:} Let $\n$ denote the set of positive integers, and set $\n_0 = \{0\} \cup \n$. \medskip \begin{theorem} \label{} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer with $d \equiv 2,3 \pmod{4}$. Any two integral bases for $\mathbb{Z}[\sqrt{d}]$ have the same discriminant. \end{theorem} \begin{proof} Let $\{\alpha_1, \alpha_2\}$ and $\{\omega_1, \omega_2\}$ be integral bases for $\mathbb{Z}[\sqrt{d}]$, and let $z_{i,j}$ be defined as in the proof of Proposition~\ref{}. Since $\Delta[\alpha_1, \alpha_2]$ and $\Delta[\omega_1, \omega_2]$ are both integers, Equation~\eqref{discrim} in the proof of Proposition~\ref{}, along with the fact that $\left(\det \left[\begin{matrix} z_{1,1} & z_{1,2} \\ z_{2,1} & z_{2,2} \end{matrix}\right]\right)^2$ belongs to $\n$, implies that $\Delta[\omega_1,\omega_2]$ divides $\Delta[\alpha_1,\alpha_2]$. Using a similar argument, we can show that $\Delta[\alpha_1,\alpha_2]$ divides $\Delta[\omega_1,\omega_2]$. As both discriminants have the same sign, $\Delta[\alpha_1,\alpha_2]= \Delta[\omega_1,\omega_2]$. \end{proof} \medskip Using Example~\ref{ of the canonical integral basis} and Example~\ref{ integral basis}, we obtain the following corollary. \begin{corollary} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer with $d \equiv 2,3 \pmod{4}$. Every integral basis for $\mathbb{Z}[\sqrt{d}]$ has discriminant $4d$. \end{corollary} \medskip \section{General Properties of Ideals} \label{ properties of ideals} Let $R$ be a commutative ring with identity. In most beginning algebra classes, the units, irreducibles, and associate elements in $R$ are standard concepts of interest. Recall that the units of $R$ are precisely the invertible elements, while nonunit elements $x, y \in R$ are associates if $a=ub$ for a unit $u$ of $R$. A nonunit $x \in R^\bullet := R \setminus \{0\}$ is irreducible if whenever $x = uv$ in $R$, then either $u$ or $v$ is a unit. To truly understand factorizations in $\z[\sqrt{-5}]$, we will need to know first how ideals of $\z[\sqrt{-5}]$ are generated. Recall that a subset $I$ of a commutative ring $R$ with identity is called an ideal of $R$ provided that $I$ is a subring with the property that $rI \subseteq I$ for all $r \in R$. It follows immediately that if $x_1, \dots, x_k \in R$, then the set \[ I = \langle x_1, \dots, x_k \rangle=\{r_1 x_1 + \dots + r_k x_k \mid \mbox{ each } r_i \in R\} \] is an ideal of $R$, that is, the ideal generated by $x_1, \dots, x_k$. Recall that $I$ is said to be principal if $I = \langle x \rangle$ for some $x \in R$, and $R$ is said to be a principal ideal domain (or a~PID) if each ideal of $R$ is principal. The zero ideal $\langle 0 \rangle$ and the entire ring $R = \langle 1 \rangle$ are principal ideals. May it be that all the ideals of $\z[\sqrt{-5}]$ are principal? It turns out that the answer is ``no'' as we shall see in the next example. \begin{example}\label{ex:a nonprincipal ideal} The ring of integers $\z[\sqrt{-5}]$ is not a PID. We argue that the ideal \[ I = \langle 2, 1+\sqrt{-5}\rangle \] is not principal. If $I=\langle \alpha \rangle$, then $\alpha$ divides both $2$ and $1+\sqrt{-5}$. The reader will verify in Exercise~\ref{ and non-associate} below that both of these elements are irreducible and non-associates. Hence, $\alpha=\pm 1$ and $I=\langle \pm 1\rangle = \z[\sqrt{-5}]$. Now we show that $3 \notin I$. Suppose there exist $a, b, c, d \in \mathbb{Z}$ so that \[ (a+b\sqrt{-5})2 + (c+d\sqrt{-5})(1+\sqrt{-5}) = 3. \] Expanding the previous equality, we obtain \begin{equation}\label{idealarray} \begin{array}{lclclcl} 2a & + & c & - & 5d & = & 3 \\ 2b & + & c & + & d & = & 0. \end{array} \end{equation} After subtracting, we are left with $2(a-b)-6d=3$, which implies that $2$ divides $3$ in $\mathbb{Z}$, a contradiction. \end{example} \medskip \begin{exercise} \label{ and non-associate} Show that the elements $2$ and $1+\sqrt{-5}$ are irreducible and non-associates in $\z[\sqrt{-5}]$. (Hint: use the norm function.) \end{exercise} \medskip Let us recall that a proper ideal $I$ of a commutative ring $R$ with identity is said to be prime if whenever $xy\in I$ for $x,y \in R$, then either $x\in I$ or $y\in I$. In addition, we know that an element $p \in R \setminus \{0\}$ is said to be prime provided that the principal ideal $\langle p \rangle$ is prime. It follows immediately that, in any integral domain, every prime element is irreducible. \begin{exercise} \label{ of ideals inside a prime ideal} Let $P$ be an ideal of a commutative ring $R$ with identity. Show that $P$ is prime if and only if the containment $IJ \subseteq P$ for ideals $I$ and $J$ of $R$ implies that either $I \subseteq P$ or $J \subseteq P$. \end{exercise} \medskip \begin{example} \label{ex:a non-prime ideal} We argue that the ideal $I = \langle 2 \rangle$ is not prime in $\z[\sqrt{-5}]$ and will in fact use Equation~\eqref{2ndzroot5}. Since $(1 - \sqrt{-5})(1 + \sqrt{-5}) = 2 \cdot 3$, it follows that \[ (1 - \sqrt{-5})(1 + \sqrt{-5}) \in \langle 2 \rangle. \] Now if $1 - \sqrt{-5} \in \langle 2 \rangle$, then there is an element $\alpha \in \z[\sqrt{-5}]$ with $1 - \sqrt{-5} = 2\alpha$. But then $\alpha = \frac{1}{2} - \frac{\sqrt{-5}}{2} \notin \z[\sqrt{-5}]$, a contradiction. A similar argument works with $1 + \sqrt{-5}$. Hence, $\langle 2 \rangle$ is not a prime ideal in $\z[\sqrt{-5}]$. \end{example} \medskip We remind the reader that a proper ideal $I$ of a commutative ring $R$ with identity is called maximal if for each ideal $J$ the containment $I \subseteq J \subseteq R$ implies that either $J = I$ or $J = R$. What we ask the reader to verify in the next exercise is a well-known result from basic abstract algebra. \begin{exercise} \label{ by prime/maximal ideals are domain/field} Let $I$ be a proper ideal of a commutative ring $R$ with identity, and let $R/I = \{r + I \mid r \in R\}$ be the quotient ring of $R$ by $I$. \begin{enumerate} \item Show that $I$ is prime if and only if $R/I$ is an integral domain. \label{ by prime} \vspace{3pt} \item Show that $I$ is maximal if and only if $R/I$ is a field. Deduce that maximal ideals are prime. \label{ by maximal} \end{enumerate} \end{exercise} \medskip \begin{example} \label{ ideal dividing <2>} Now we shall expand our analysis of $I = \langle 2, 1+\sqrt{-5}\rangle$ in Example~\ref{ex:a nonprincipal ideal} by showing that $I$ is a prime ideal in $\z[\sqrt{-5}]$. To do this, we first argue that an element $\alpha = z_1 + z_2\sqrt{-5} \in \z[\sqrt{-5}]$ is contained in $I$ if and only if $z_1$ and $z_2$ have the same parity. If $\alpha \in I$, then there are integers $a, b, c$, and $d$ so that \[ z_1 + z_2\sqrt{-5} = (a+b\sqrt{-5})2 + (c+d\sqrt{-5})(1 + \sqrt{-5}). \] Adjusting the equations from \eqref{idealarray} yields \begin{equation}\label{newidealarray} \begin{array}{lclclcl} 2a & + & c & - & 5d & = & z_1 \\ 2b & + & c & + & d & = & z_2. \end{array} \end{equation} Notice that if $c \equiv d \pmod{2}$, then both $z_1$ and $z_2$ are even, while $c \not\equiv d \pmod{2}$ implies that both $z_1$ and $z_2$ are odd. Hence, $z_1$ and $z_2$ must have the same parity. Conversely, suppose that $z_1$ and $z_2$ have the same parity. As, clearly, every element of the form $2k_1+2k_2\sqrt{-5}=2(k_1+k_2\sqrt{-5})$ is in $I$, let us assume that $z_1$ and $z_2$ are both odd. The equations in \eqref{newidealarray} form a linear system that obviously has solutions over $\q$ for any choice of $z_1$ and $z_2$ in $\mathbb{Z}$. By solving this system, we find that $a$ and $b$ are dependent variables and \[ a = \frac{z_1 -c +5d}{2} \quad \mbox{ and } \quad b=\frac{z_2-c -d}{2}. \] Letting $c$ be any even integer and $d$ any odd integer now yields a solution with both $a$ and $b$ integers. Thus, $z_1+z_2\sqrt{-5}\in I$. Now consider $\z[\sqrt{-5}] /I$. As $I$ is not principal (Example~\ref{ex:a nonprincipal ideal}), $1 \notin I$. Therefore $1 + I \neq 0 + I$. If $c_1+c_2\sqrt{-5}\not\in I$, then $c_1$ and $c_2$ have opposite parity. If $c_1$ is odd and $c_2$ even, then $((c_1-1) + c_2\sqrt{-5}) + I = 0+I$ implies that $(c_1+c_2\sqrt{-5})+I = 1+ I$. If $c_1$ is even and $c_2$ odd, then $((c_1-1) + c_2\sqrt{-5}) + I = 0+I$ again implies that $(c_1+c_2\sqrt{-5})+I = 1+ I$. Hence, $\z[\sqrt{-5}] /I \cong \{0+I, 1+I\} \cong \mathbb{Z}_2$. Since $\mathbb{Z}_2$ is a field, $I$ is a maximal ideal and thus prime (by Exercise~\ref{ by prime/maximal ideals are domain/field}). \end{example} \medskip \begin{exercise} \label{ ideals dividing <3>} Show that $\langle 3, 1 - 2 \sqrt{-5} \rangle$ and $\langle 3, 1 + 2 \sqrt{-5} \rangle$ are prime ideals in the ring of integers $\z[\sqrt{-5}]$. \end{exercise} \medskip Let $R$ be a commutative ring with identity. If every ideal of $R$ is finitely generated, then $R$ is called a \emph{Noetherian ring}. In addition, $R$ satisfies the \emph{ascending chain condition on ideals} (ACC) if every increasing (under inclusion) sequence of ideals of $R$ eventually stabilizes. \begin{exercise} Let $R$ be a commutative ring with identity. Show that $R$ is Noetherian if and only if it satisfies the ACC. \end{exercise} \medskip We shall see in Theorem~\ref{2-gen} that the rings of integers $\mathbb{Z}[\sqrt{d}]$ are Noetherian and, therefore, satisfy the ACC. \medskip \section{Ideals in $\z[\sqrt{-5}]$} \label{ inside the ring of integers} In this section we explore the algebraic structure of all ideals of $\z[\sqrt{-5}]$ under ideal multiplication, encapsulating the basic properties of multiplication of ideals. Let us begin by generalizing the notion of an integral basis, which also plays an important role in ideal theory. \begin{definition} Let $K$ be an algebraic number field of dimension $n$ as a vector space over $\q$, and let $I$ be a proper ideal of the ring of integers $\mathcal{O}_K$. We say that the elements $\omega_1, \dots, \omega_n \in I$ form an \emph{integral basis} for $I$ provided that for each $\beta \in I$ there exist unique $z_1,\dots ,z_n \in \mathbb{Z}$ satisfying that $\beta = z_1 \omega_1 + \dots + z_n \omega_n$. \end{definition} \medskip With notation as in the above definition, notice that if $\{\omega_1, \dots, \omega_n\}$ is an integral basis for $I$, then $I = \langle \omega_1, \dots, \omega_n \rangle$. Care is needed here as the converse is not necessarily true. For instance, $\{3\}$ is not an integral basis for the ideal $I=\langle 3 \rangle$ of $\z[\sqrt{-5}]$ (note that $3\sqrt{-5} \in I$). \begin{exercise} Argue that $\{3, 3\sqrt{-5}\}$ is an integral basis for the ideal $I = \langle 3 \rangle$ of the ring of integers $\z[\sqrt{-5}]$. \end{exercise} \medskip We now show that every proper ideal of $\z[\sqrt{-5}]$ has an integral basis. \begin{theorem} \label{2-gen} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer with $d \equiv 2,3 \pmod{4}$. Every nonzero proper ideal of $\mathbb{Z}[\sqrt{d}]$ has an integral basis. Hence, every ideal of $\mathbb{Z}[\sqrt{d}]$ is finitely generated. \end{theorem} \begin{proof} Let $I$ be a nonzero proper ideal of $\mathbb{Z}[\sqrt{d}]$. To find an integral basis for $I$ consider the collection $\mathcal{B}$ of all subsets of $I$ which form a vector space basis for $\q(\sqrt{d})$. Note that if $\{\omega_1, \omega_2\}$ is an integral basis for $\mathbb{Z}[\sqrt{d}]$ and $\alpha \in I^\bullet$, then the subset $\{\alpha \omega_1, \alpha \omega_2\}$ of $I$ is also a linearly independent subset inside the vector space $\q(\sqrt{d})$. As a result, the collection $\mathcal{B}$ is nonempty. As $I \subseteq \mathbb{Z}[\sqrt{d}]$, Proposition~\ref{} ensures that $\Delta[\delta_1,\delta_2] \in \mathbb{Z}^\bullet$ for every member $\{\delta_1, \delta_2\}$ of $\mathcal{B}$. Then we can take a pair $\{\delta_1, \delta_2\}$ in $\mathcal{B}$ and assume that the absolute value of its discriminant, i.e., $|\Delta[\delta_1,\delta_2]|$, is as small as possible. We argue now that $\{\delta_1, \delta_2\}$ is an integral basis for $I$. Assume, by way of contradiction, that $\{\delta_1, \delta_2\}$ is not an integral basis for $I$. Since $\{\delta_1, \delta_2\}$ is a basis for $\q(\sqrt{d})$ as a vector space over $\q$, there must exist $\beta \in I$ and $q_1, q_2 \in \q$ such that $\beta = q_1 \delta_1 + q_2 \delta_2$, where not both $q_1$ and $q_2$ are in $\mathbb{Z}$. Without loss of generality, we can assume that $q_1 \in \q \setminus \mathbb{Z}$. Write $q_1 = z + r$, where $z \in \mathbb{Z}$ and $0 < r < 1$. Let \begin{align*} \delta_1^\ast &= \beta - z \delta_1 = (q_1 - z) \delta_1 + q_2 \delta_2 \\ \delta_2^\ast &=\delta_2. \end{align*} It is easy to verify that $\{\delta_1^\ast, \delta_2^\ast\}$ is linearly independent and thus is another vector space basis for $\q(\sqrt{d})$ which consists of elements of $I$, that is, $\{\delta_1^\ast, \delta_2^\ast\}$ is a member of $\mathcal{B}$. Proceeding as we did in the proof of Proposition~\ref{}, we find that \[ \Delta[\delta_1^\ast, \delta_2^\ast] = r^2\Delta [\delta_1,\delta_2]; \] this is because $\left(\det\left[ \begin{matrix} q_1 - z\;\;\; & q_2 \\ 0 & 1 \end{matrix}\right]\right)^2 = r^2$. It immediately follows from $0<r<~1$ that $|\Delta[\delta_1^\ast, \delta_2^\ast]| < |\Delta[\delta_1,\delta_2]|$, contradicting the minimality of $| \Delta[\delta_1,\delta_2] |$. Hence, $\{\delta_1,\delta_2\}$ is an integral basis for~$I$, which completes the proof. \end{proof} \medskip Theorem~\ref{2-gen} yields the next important corollary. \begin{corollary} \label{ generating ideals} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer with $d \equiv 2,3 \pmod{4}$. If $I$ is a proper ideal of the ring of integers $\mathbb{Z}[\sqrt{d}]$, then there exist elements $\alpha_1, \alpha_2 \in I$ such that $I = \langle \alpha_1, \alpha_2 \rangle$. Thus, $\mathbb{Z}[\sqrt{d}]$ is a Noetherian ring. \end{corollary} \medskip \begin{remark} One can actually say much more. For $d$ as in Corollary~\ref{ generating ideals}, the following stronger statement is true: if $I$ is a nonzero proper ideal of $\mathbb{Z}[\sqrt{d}]$ and $\alpha_1 \in I^\bullet$, then there exists $\alpha_2 \in I$ satisfying that $I = \langle \alpha_1, \alpha_2 \rangle$. This condition is known as the $1 \frac{1}{2}$-\emph{generator property}. The interested reader can find a proof of this result in \cite[Theorem~9.3]{PD}. \end{remark} \medskip \begin{definition} A pair $(M, *)$, where $M$ is a set and $*$ is a binary operation on $M$, is called a \emph{monoid} if $*$ is associative and there exists $e \in M$ satisfying that $e*x = x*e = x$ for all $x \in M$. The element $e$ is called the \emph{identity element}. The monoid $M$ is called \emph{commutative} if the operation $*$ is commutative. \end{definition} Let $R$ be a commutative ring with identity. Recall that we have a natural multiplication on the collection consisting of all ideals of $R$, that is, for any two ideals $I$ and $J$ of $R$, the product \begin{align} \label{ multiplication} IJ = \big\{\sum_{i=1}^k a_ib_i \ \big{|} \ k \in \n, \ a_1, \dots, a_k \in I, \mbox{ and } b_1, \dots, b_k \in J \big\} \end{align} is again an ideal. It is not hard to check that ideal multiplication is both associative and commutative, and satisfies that $R I = I$ for each ideal $I$ of $R$. This amounts to arguing the following exercise. \begin{exercise} \label{ of ideal multiplication} Let $R$ be a commutative ring with identity. Show that the set of all ideals of $R$ is a commutative monoid under ideal multiplication. \end{exercise} \medskip \begin{example} \label{ of <2> into prime ideals} To give the reader a notion of how ideal multiplication works, we show that \[ \langle 2, 1+\sqrt{-5}\rangle^2 = \langle 2\rangle. \] It follows by~(\ref{ multiplication}) that ideal multiplication can be achieved by merely multiplying generators. For instance, \begin{align*} \langle 2, 1+\sqrt{-5}\rangle^2 &= \langle 2, 1+\sqrt{-5}\rangle \langle 2, 1+\sqrt{-5}\rangle\\ &=\langle 4, 2(1+\sqrt{-5}), 2(1+\sqrt{-5}), -2(2 - \sqrt{-5}) \rangle. \end{align*} Since $2$ divides each of the generators of $\langle 2, 1+\sqrt{-5}\rangle^2$ in $\z[\sqrt{-5}]$, we clearly have that $\langle 2, 1+\sqrt{-5}\rangle^2 \subseteq \langle 2\rangle$. To verify the reverse inclusion, let us first observe that \[ 2\sqrt{-5} = 4 - 2(2 - \sqrt{-5}) \in \langle 2, 1 + \sqrt{-5} \rangle^2. \] As $2\sqrt{-5} \in \langle 2, 1 + \sqrt{-5} \rangle^2$, one immediately sees that \[ 2 = 2 (1 + \sqrt{-5}) - 2\sqrt{-5} \in \langle 2, 1 + \sqrt{-5} \rangle^2. \] Hence, the inclusion $\langle 2\rangle\subseteq \langle 2, 1+\sqrt{-5}\rangle^2$ holds, and equality follows. \end{example} \medskip \begin{exercise} \label{ of <3> into prime ideals} Verify that the next equalities hold: \begin{align*} \langle 3 \rangle &= \langle 3, 1 - 2\sqrt{-5} \rangle \langle 3, 1 + 2\sqrt{-5} \rangle , \\ \langle 1 - \sqrt{-5} \rangle &= \langle 2, 1 + \sqrt{-5} \rangle \langle 3, 1 + 2\sqrt{-5} \rangle ,\\% \label{ of ideals 2} \\ \langle 1 + \sqrt{-5} \rangle &= \langle 2, 1 + \sqrt{-5} \rangle \langle 3, 1 - 2\sqrt{-5} \rangle . \end{align*} \end{exercise} \medskip Example~\ref{ of <3> into prime ideals} is no accident. Indeed, every nonprincipal ideal of $\z[\sqrt{-5}]$ has a multiple which is a principal ideal as it is established in the following theorem. \begin{theorem} \label{inverse} Let $I$ be an ideal of $\z[\sqrt{-5}]$. Then there exists a nonzero ideal $J$ of $\z[\sqrt{-5}]$ such that $IJ$ is principal. \end{theorem} \begin{proof} If $I$ is a principal ideal, then the result follows by letting $J=\langle 1\rangle$. So suppose $I=\langle \alpha, \beta\rangle$ is not a principal ideal of $\z[\sqrt{-5}]$, where $\alpha= a+b\sqrt{-5}$ and $\beta=c+d\sqrt{-5}$. Notice that it is enough to verify the existence of such an ideal $J$ when $\gcd(a,b,c,d) = 1$, and we make this assumption. It is easy to check that $\alpha\overline{\beta}+\overline{\alpha}\beta= 2ac+10bd \in \mathbb{Z}$. Hence, $\alpha\overline{\alpha}$, $\alpha \overline{\beta}+\overline{\alpha}\beta$, and $\beta\overline{\beta}$ are all integers. Let \begin{align*} f &= \gcd(\alpha\overline{\alpha}, \alpha \overline{\beta}+\overline{\alpha}\beta, \beta \bar{\beta}) \\ &= \gcd(a^2 + 5b^2, 2ac + 10bd, c^2 + 5d^2). \end{align*} Take $J = \langle \overline{\alpha}, \overline{\beta}\rangle$. We claim that $IJ = \langle f \rangle$. Since $f = \gcd(\alpha\overline{\alpha}, \alpha \overline{\beta}+\overline{\alpha}\beta, \beta \overline{\beta})$, there are integers $z_1, z_2$, and $z_3$ so that \[ f = z_1 \alpha\overline{\alpha} + z_2 \beta\overline{\beta} + z_3 (\alpha\overline{\beta}+\overline{\alpha}\beta). \] Because $IJ = \langle \alpha\overline{\alpha}, \alpha\overline{\beta}, \beta \overline{\alpha}, \beta\overline{\beta}\rangle$, we have that $f$ is a linear combination of the generating elements. Thus, $f \in IJ$ and, therefore, $\langle f \rangle \subseteq IJ$. To prove the reverse containment, we first show that $f$ divides $bc - ad$. Suppose, by way of contradiction, that this is not the case. Notice that $25 \nmid f$; otherwise $25 \mid a^2 + 5b^2$ and $25 \mid c^2 + 5d^2$ would imply that $5 \mid \gcd(a,b,c,d)$. On the other hand, $4 \mid f$ would imply $4 \mid a^2 + 5b^2$ and $4 \mid c^2 + 5d^2$, forcing $a$, $b$, $c$, and $d$ to be even, which is not possible as $\gcd(a,b,c,d) = 1$. Hence, $4 \nmid f$ and $25 \nmid f$. Because \begin{align*} 2c (a^2 + 5b^2) - a(2ac + 10bd) &= 10b(bc - ad) \\ 2a (c^2 + 5d^2) - c(2ac + 10bd) &= 10d(ad - bc), \end{align*} $f$ must divide both $10b(bc - ad)$ and $10d (bc - ad)$. As, by assumption, $f \nmid bc - ad$, there must be a prime $p$ and a natural $n$ such that $p^n \mid f$ but $p^n \nmid bc - ad$. If $p = 2$, then $4 \nmid f$ forces $n=1$. In this case, both $a^2 + 5b^2$ and $c^2 + 5d^2$ would be even, and so $2 \mid a-b$ and $2 \mid c-d$, which implies that $2 \mid bc - ad$, a contradiction. Thus, $p \neq 2$. On the other hand, if $p=5$, then again $n=1$. In this case, $5 \mid a^2 + 5b^2$ and $5 \mid c^2 + 5d^2$ and so $5$ would divide both $a$ and $c$, contradicting that $5 \nmid bc - ad$. Then, we can assume that $p \notin \{2,5\}$. As $p^n \mid 10b(bc - ad)$ but $p^n \nmid bc - ad$, we have that $p \mid 10b$. Similarly, $p \mid 10d$. Since $p \notin \{2,5\}$, it follows that $p \mid b$ and $p \mid d$. Now the fact that $p$ divides both $a^2 + 5b^2$ and $c^2 + 5d^2$ yields that $p \mid a$ and $p \mid c$, contradicting that $\gcd(a,b,c,d) = 1$. Hence, $f \mid bc - ad$. Let us verify now that $f \mid ac + 5bd$. If $f$ is odd, then $f \mid ac + 5bd$. Assume, therefore, that $f = 2f_1$, where $f_1 \in \mathbb{Z}$. As $4 \nmid f$, the integer $f_1$ is odd. Now, $f \mid a^2 + 5b^2$ implies that $a$ and $b$ have the same parity. Similarly, one sees that $c$ and $d$ have the same parity. As a consequence, $ac + 5bd$ is even. Since $f_1$ is odd, it must divide $(ac + 5bd)/2$, which means that $f$ divides $ac + 5bd$, as desired. Because $f$ divides both $\alpha \overline{\alpha}$ and $\beta \overline{\beta}$ in $\mathbb{Z}$, proving that $IJ \subseteq \langle f \rangle$ amounts to verifying that $f$ divides both $\alpha \overline{\beta}$ and $\overline{\alpha} \beta$ in $\z[\sqrt{-5}]$. Since $f$ divides both $ac + 5 bd$ and $bc - ad$ in $\mathbb{Z}$, one has that \[ x = \frac{ac + 5bd}{f} \in \mathbb{Z} \quad \text{ and } \quad y = \frac{bc - ad}{f} \in \mathbb{Z}. \] Therefore \begin{align*} \alpha \overline{\beta} &= ac + 5bd + (bc - ad)\sqrt{-5} = (x + y \sqrt{-5})f \in \langle f \rangle. \end{align*} Also, $\overline{\alpha} \beta = \overline{\alpha \overline{\beta}} = (x - y \sqrt{-5})f \in \langle f \rangle$. Hence, the reverse inclusion $IJ \subseteq \langle f \rangle$ also holds, which completes the proof. \end{proof} \medskip A commutative monoid $(M,*)$ is said to be \emph{cancellative} if for all $a,b,c \in M$, the equality $a*b = a*c$ implies that $b = c$. By Exercise~\ref{ of ideal multiplication}, the set \[ \mathcal{I} := \{I \mid I \ \text{is an ideal of } \ \z[\sqrt{-5}]\} \] is a commutative monoid. As the next corollary states, the set $\mathcal{I}^\bullet := \mathcal{I} \setminus \{\langle 0 \rangle \}$ is indeed a commutative cancellative monoid. \begin{corollary} The set $\mathcal{I}^\bullet$ under ideal multiplication is a commutative cancellative monoid. \end{corollary} \begin{proof} Because $\mathcal{I}$ is a commutative monoid under ideal multiplication, it immediately follows that $\mathcal{I}^\bullet$ is also a commutative monoid. To prove that $\mathcal{I}^\bullet$ is cancellative, take $I,J,K \in \mathcal{I}^\bullet$ such that $IJ = IK$. By Theorem~\ref{inverse}, there exists an ideal $I'$ of $\z[\sqrt{-5}]$ and $x \in \z[\sqrt{-5}]^\bullet$ with $I'I = \langle x \rangle$. Then \[ \langle x \rangle J = I'IJ = I'IK = \langle x \rangle K. \] As $x \neq 0$ and the product in $\z[\sqrt{-5}]^\bullet$ is cancellative, $J = K$ \end{proof} \medskip \section{The Fundamental Theorem of Ideal Theory} \label{ theorem} We devote this section to prove a version of the Fundamental Theorem of Ideal Theory for the ring of integers $\z[\sqrt{-5}]$. To do this, we need to develop a few tools. In particular, we introduce the concept of a fractional ideal of $\z[\sqrt{-5}]$ and show that the set of such fractional ideals is an abelian group. Let us begin by exploring the relationship between the concepts of prime and maximal ideals. We recall that every proper ideal of a commutative ring $R$ with identity is contained in a maximal ideal, which implies, in particular, that maximal ideals always exist. \begin{exercise} \label{ ideals are prime} Show that every maximal ideal of a commutative ring with identity is prime. \end{exercise} \medskip Prime ideals, however, are not necessarily maximal. The following example sheds some light upon this observation. \begin{example} Let $\mathbb{Z}[X]$ denote the ring of polynomials with integer coefficients. Clearly, $\mathbb{Z}[X]$ is an integral domain. It is not hard to verify that the ideal $\langle X \rangle$ of $\mathbb{Z}[X]$ is prime. Because $2 \notin \langle X \rangle$, one obtains that $\langle X \rangle \subsetneq \langle 2, X \rangle$. It is left to the reader to argue that $\langle 2, X \rangle$ is a proper ideal of $\mathbb{Z}[X]$. Since $\langle X \rangle \subsetneq \langle 2, X \rangle \subsetneq \mathbb{Z}[X]$, it follows that $\langle X \rangle$ is not a maximal ideal of $\mathbb{Z}[X]$. (An alternate argument can easily be given using Exercise~\ref{ by prime/maximal ideals are domain/field}.) \end{example} \medskip In the ring of integers $\mathcal{O}_K$ of any algebraic number field $K$, every nonzero prime ideal is maximal (see, for instance, \cite[Proposition~5.21]{Ja}). Let us establish this result here for our case of interest. \begin{proposition} \label{ ideals are maximal} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer with $d \equiv 2,3 \pmod{4}$. Then every nonzero prime ideal of $\mathbb{Z}[\sqrt{d}]$ is maximal. \end{proposition} \begin{proof} Let $P$ be a nonzero prime ideal in $\mathbb{Z}[\sqrt{d}]$, and let $\{\omega_1, \omega_2\}$ be an integral basis for $\mathbb{Z}[\sqrt{d}]$. Fix $\beta \in P^\bullet$. Note that $n := N(\beta) = \beta \bar{\beta} \in P \cap \n$. Consider the finite subset \[ S = \big\{ n_1 \omega_1 + n_2 \omega_2 + P \mid n_1, n_2 \in \{0,1, \dots, n-1\} \big\} \] of $\mathbb{Z}[\sqrt{d}]/P$. Take $x \in \mathbb{Z}[\sqrt{d}]$. As $\{\omega_1, \omega_2\}$ is an integral basis, there exist $z_1, z_2 \in \mathbb{Z}$ such that $x = z_1 \omega_1 + z_2 \omega_2$ and, therefore, $x + P = n_1 \omega_1 + n_2 \omega_2 + P \in S$, where $n_i \in \{0, \dots, n-1\}$ and $n_i \equiv z_i \pmod{n}$. Hence, $\mathbb{Z}[\sqrt{d}]/P = S$, which implies that $\mathbb{Z}[\sqrt{d}]/P$ is finite. It follows by Exercise~\ref{ by prime/maximal ideals are domain/field}(\ref{ by prime}) that $\mathbb{Z}[\sqrt{d}]/P$ is an integral domain. As a result, $\mathbb{Z}[\sqrt{d}]/P$ is a field (see Exercise~\ref{ domains are fields} below). Thus, Exercise~\ref{ by prime/maximal ideals are domain/field}(\ref{ by maximal}) guarantees that $P$ is a maximal ideal. \end{proof} \medskip \begin{exercise} \label{ domains are fields} Let $R$ be a finite integral domain. Show that $R$ is a field. \end{exercise} \medskip Although the concepts of (nonzero) prime and maximal ideals coincide in $\mathbb{Z}[\sqrt{d}]$, we will use both terms depending on the ideal property we are willing to apply. \begin{lemma} \label{ product of prime ideals is contained in a given ideal} If $I$ is a nonzero ideal of a Noetherian ring $R$, then there exist nonzero prime ideals $P_1, \dots, P_n$ of $R$ such that $P_1 \cdots P_n \subseteq I$. \end{lemma} \begin{proof} Assume, by way of contradiction, that the statement of the lemma does not hold. Because $R$ is a Noetherian ring and, therefore, satisfies the ACC, there exists an ideal $I$ of $R$ that is maximal among all the ideals failing to satisfy the statement of the lemma. Clearly, $I$ cannot be prime. By Exercise~\ref{ of ideals inside a prime ideal}, there exist ideals $J$ and $K$ of $R$ such that $JK \subseteq I$ but neither $J \subseteq I$ nor $K \subseteq I$. Now notice that the ideals $J' = I + J$ and $K' = I + K$ both strictly contain $I$. The maximality of $I$ implies that both $J'$ and $K'$ contain products of nonzero prime ideals. Now the fact that $J'K' \subseteq I$ would also imply that $I$ contains a product of nonzero prime ideals, a contradiction. \end{proof} \medskip Recall that if $R$ is an integral domain contained in a field $F$, then the field of fractions of $R$ is the smallest subfield of $F$ containing $R$. If $K$ is an algebraic number field, then it is not hard to argue that the field of fractions of $\mathcal{O}_K$ is precisely $K$. \begin{definition} Let $R$ be an integral domain with field of fractions $F$. A \emph{fractional ideal} of $R$ is a subset of $F$ of the form $\alpha^{-1}I$, where $\alpha \in R^\bullet$ and $I$ is an ideal of $R$. \end{definition} \medskip With notation as in the previous definition, it is clear that every ideal of $R$ is a fractional ideal. However, fractional ideals are not necessarily ideals. The product of fractional ideals is defined similarly to the product of standard ideals. Therefore it is easily seen that the product of two fractional ideals is again a fractional ideal. Indeed, for elements $\alpha$ and $\beta$ of $R^\bullet$ and for ideals $I$ and $J$ of $R$, we only need to observe that $(\alpha^{-1}I)(\beta^{-1}J) = (\alpha \beta)^{-1} IJ$. \medskip \noindent {\bf Notation:} Let $\mathcal{F}$ denote the set of all fractional ideals of $\z[\sqrt{-5}]$, and let $\mathcal{F}^\bullet$ denote the set $\mathcal{F} \setminus \{\langle 0 \rangle \}$. \medskip \begin{definition} Let $R$ be an integral domain with field of fractions $F$. For a fractional ideal $I$ of $R$, the set \[ I^{-1} := \{\alpha \in F \mid \alpha I \subseteq R\} \] is called the \emph{inverse} of $I$. \end{definition} \medskip \begin{exercise} \label{ ideals are fractional} Show that the inverse of a fractional ideal is again a fractional ideal. \end{exercise} \medskip \begin{lemma} \label{ of proper ideals strictly contain the ring} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer with $d \equiv 2,3 \pmod{4}$. If $I$ is a proper ideal of the ring of integers $\mathbb{Z}[\sqrt{d}]$, then $\mathbb{Z}[\sqrt{d}]$ is strictly contained in the fractional ideal $I^{-1}$. \end{lemma} \begin{proof} Since $I$ is a proper ideal of $\mathbb{Z}[\sqrt{d}]$, then there exists a maximal ideal $M$ of $\mathbb{Z}[\sqrt{d}]$ containing $I$. Fix $\alpha \in M^\bullet$. By the definition of the inverse of an ideal, $\mathbb{Z}[\sqrt{d}] \subseteq M^{-1}$. Since $\mathbb{Z}[\sqrt{d}]$ is a Noetherian ring, Lemma~\ref{ product of prime ideals is contained in a given ideal} ensures the existence of $m \in \n$ and prime ideals $P_1, \dots, P_m$ in $\mathbb{Z}[\sqrt{d}]$ such that $P_1 \cdots P_m \subseteq \langle \alpha \rangle \subseteq M$. Assume that $m$ is the minimum natural number satisfying this property. Since $M$ is a prime ideal (Exercise~\ref{ ideals are prime}), by Exercise~\ref{ of ideals inside a prime ideal} there exists $P \in \{P_1, \dots, P_m\}$ such that $P \subseteq M$. There is no loss of generality in assuming that $P = P_1$. Now it follows by Proposition~\ref{ ideals are maximal}, the ideal $P_1$ is maximal, which implies that $P_1 = M$. By the minimality of $m$, there exists $\alpha' \in P_2 \cdots P_m \setminus \langle \alpha \rangle$. Thus, $\alpha^{-1} \alpha' \notin \mathbb{Z}[\sqrt{d}]$ and $\alpha' M = \alpha' P_1 \subseteq P_1 \cdots P_m \subseteq \langle \alpha \rangle$, that is $\alpha^{-1}\alpha' M \subseteq \langle 1 \rangle = \mathbb{Z}[\sqrt{d}]$. As a result, $\alpha^{-1} \alpha' \in M^{-1} \setminus \mathbb{Z}[\sqrt{d}]$. Hence, we find that $\mathbb{Z}[\sqrt{d}] \subsetneq M^{-1} \subseteq I^{-1}$, and the proof follows. \end{proof} \medskip We focus throughout the remainder of our work on the ring of integers $\z[\sqrt{-5}]$. This, via Theorem~\ref{inverse}, will substantially simplify our remaining arguments. \begin{lemma} \label{ ideal by element of Q(sqrt{-5})} If $I \in \mathcal{I}^\bullet$ and $\alpha \in \q(\sqrt{-5})$, then $\alpha I \subseteq I$ implies $\alpha \in \z[\sqrt{-5}]$. \end{lemma} \begin{proof} Let $I$ and $\alpha$ be as in the statement of the lemma. By Theorem~\ref{inverse}, there exists a nonzero ideal $J$ of $\z[\sqrt{-5}]$ such that $IJ = \langle \beta \rangle$ for some $\beta \in \z[\sqrt{-5}]$. Then $\alpha \langle \beta \rangle = \alpha IJ \subseteq IJ = \langle \beta \rangle$, which means that $\alpha \beta = \sigma \beta$ for some $\sigma \in \z[\sqrt{-5}]$. As $\beta \neq 0$, it follows that $\alpha = \sigma \in \z[\sqrt{-5}]$. \end{proof} \medskip \begin{theorem} \label{ set of fractional ideals forms a group} The set $\mathcal{F}^\bullet$ is an abelian group under multiplication of fractional ideals. \end{theorem} \begin{proof} Clearly, multiplication of fractional ideals is associative. In addition, it immediately follows that the fractional ideal $\z[\sqrt{-5}] = 1^{-1} \langle 1 \rangle$ is the identity. The most involved part of the proof consists in arguing that each fractional ideal is invertible. Let $M \in \mathcal{I}^\bullet$ be a maximal ideal of $\z[\sqrt{-5}]$. By definition of $M^{-1}$, we have that $M M^{-1} \subseteq \z[\sqrt{-5}]$, which implies that $M M^{-1} \in \mathcal{I}^\bullet$. As $M = M \z[\sqrt{-5}] \subseteq M M^{-1}$ and $M$ is maximal, $M M^{-1} = M$ or $M M^{-1} = \z[\sqrt{-5}]$. As $M$ is proper, Lemma~\ref{ of proper ideals strictly contain the ring} ensures that $M^{-1}$ strictly contains $\z[\sqrt{-5}]$, which implies, by Lemma~\ref{ ideal by element of Q(sqrt{-5})}, that $M M^{-1} \neq M$. So $M M^{-1} = \z[\sqrt{-5}]$. As a result, each maximal ideal of $\z[\sqrt{-5}]$ is invertible. Now suppose, by way of contradiction, that not every ideal in $\mathcal{I}^\bullet$ is invertible. Among all the nonzero non-invertible ideals take one, say $J$, maximal under inclusion (this is possible because $\z[\sqrt{-5}]$ satisfies the ACC). Because $\z[\sqrt{-5}]$ is an invertible fractional ideal, $J \subsetneq \z[\sqrt{-5}]$. Let $M$ be a maximal ideal containing $J$. By Lemma~\ref{ of proper ideals strictly contain the ring}, one has that $\z[\sqrt{-5}] \subsetneq M^{-1} \subseteq J^{-1}$. This, along with Lemma~\ref{ ideal by element of Q(sqrt{-5})}, yields $J \subsetneq J M^{-1} \subseteq J J^{-1} \subseteq \z[\sqrt{-5}]$. Thus, $J M^{-1}$ is an ideal of $\z[\sqrt{-5}]$ strictly containing $J$. The maximality of $J$ now implies that $J M^{-1}(J M^{-1})^{-1} = \z[\sqrt{-5}]$ and, therefore, $M^{-1}(J M^{-1})^{-1} \subseteq J^{-1}$. Then \[ \z[\sqrt{-5}] = J M^{-1}(J M^{-1})^{-1} \subseteq J J^{-1} \subseteq \z[\sqrt{-5}], \] which forces $J J^{-1} = \z[\sqrt{-5}]$, a contradiction. Finally, take $F \in \mathcal{F}^\bullet$. Then there exist an ideal $I \in \mathcal{I}^\bullet$ and $\alpha \in \z[\sqrt{-5}]^\bullet$ such that $F = \alpha^{-1} I$. So one obtains that \[ (\alpha I^{-1}) F = (\alpha I^{-1})(\alpha^{-1} I) = I^{-1} I = \z[\sqrt{-5}]. \] As a consequence, the fractional ideal $\alpha I^{-1}$ is the inverse of $F$ in $\mathcal{F}^\bullet$. Because each nonzero fractional ideal of $\z[\sqrt{-5}]$ is invertible, $\mathcal{F}^\bullet$ is a group. Since the multiplication of fractional ideals is commutative, $\mathcal{F}^\bullet$ is abelian. \end{proof} \medskip \begin{corollary} \label{cor:an ideal divides any principal ideal it contains} If $I \in \mathcal{I}^\bullet$ and $\alpha \in I^\bullet$, then $IJ = \langle \alpha \rangle$ for some $J \in \mathcal{I}^\bullet$. \end{corollary} \begin{proof} Let $I$ and $\alpha$ be as in the statement of the corollary. As $\alpha^{-1} I$ is a nonzero fractional ideal, there exists a nonzero fractional ideal $J$ such that $\alpha^{-1}I J = \z[\sqrt{-5}]$, that is $I J = \langle \alpha \rangle$. Since $\beta I \subseteq JI = \langle \alpha \rangle \subseteq I$ for all $\beta \in J$, Lemma~\ref{ ideal by element of Q(sqrt{-5})} guarantees that $J \subseteq \z[\sqrt{-5}]$. Hence, $J$ is a nonzero ideal of $\z[\sqrt{-5}]$. \end{proof} \medskip \begin{theorem}\label{FTIT} [The Fundamental Theorem of Ideal Theory] Let $I$ be a nonzero proper ideal of $\z[\sqrt{-5}]$. There exists a unique (up to order) list of prime ideals $P_1, \dots ,P_k$ of $\z[\sqrt{-5}]$ such that $I = P_1 \cdots P_k$. \end{theorem} \begin{proof} Suppose, by way of contradiction, that not every ideal in $\mathcal{I}^\bullet$ can be written as the product of prime ideals. From the set of ideals of $\z[\sqrt{-5}]$ which are not the product of primes ideals, take one, say $I$, maximal under inclusion. Clearly, $I$ is not prime. Therefore $I$ is contained in a maximal ideal $P_1$, and such containment must be strict by Exercise~\ref{ ideals are prime}. By Lemma~\ref{ of proper ideals strictly contain the ring}, one has that $\z[\sqrt{-5}] \subsetneq P_1^{-1}$ and so $I \subseteq I P_1^{-1}$. Now Lemma~\ref{ ideal by element of Q(sqrt{-5})} ensures that the latter inclusion is strict. The maximality of $I$ now implies that $I P_1^{-1} = P_2 \cdots P_k$ for some prime ideals $P_2, \dots, P_k$. This, along with Theorem~\ref{ set of fractional ideals forms a group}, ensures that $I = P_1 \cdots P_k$, a contradiction. To argue uniqueness, let us assume, by contradiction, that there exists an ideal having two distinct prime factorizations. Let $m$ be the minimum natural number such that there exists $I \in \mathcal{I}$ with two distinct factorizations into prime ideals, one of them containing $m$ factors. Suppose that \begin{align} \label{ prime factorizations} I = P_1 \cdots P_m = Q_1 \cdots Q_n. \end{align} Because $Q_1 \cdots Q_n \subseteq P_m$, there exists $Q \in \{Q_1, \dots, Q_n\}$ such that $Q \subseteq P_m$ (Exercise~\ref{ of ideals inside a prime ideal}). By Proposition~\ref{ ideals are maximal}, both $Q$ and $P_m$ are maximal ideals, which implies that $P_m = Q$. As $IQ^{-1} \subseteq I I^{-1} \subseteq \z[\sqrt{-5}]$, it follows that $I Q^{-1} \in \mathcal{I}$. Multiplying the equality (\ref{ prime factorizations}) by the fractional ideal $Q^{-1}$, we obtain that $I Q^{-1}$ is an ideal of $\z[\sqrt{-5}]$ with two distinct factorizations into prime ideals such that one of them, namely $P_1 \cdots P_{m-1}$, contains less than $m$ factors. As this contradicts the minimality of $m$, uniqueness follows. \end{proof} \medskip An element $a$ of a commutative monoid $M$ is said to be an \emph{atom} if for all $x,y \in M$ such that $a = xy$, either $x$ is a unit or $y$ is a unit (i.e., has an inverse). A commutative cancellative monoid is called \emph{atomic} if every nonzero nonunit element can be factored into atoms. \begin{corollary} The monoid $\mathcal{I}^\bullet$ is atomic. \end{corollary} \medskip \section{The Class Group} \label{ class group} To understand the phenomenon of non-unique factorization in $\z[\sqrt{-5}]$, we first need to understand certain classes of ideals of $\z[\sqrt{-5}]$. Let \[ \mathcal{P} := \{ I \in \mathcal{I} \mid I \mbox{ is a principal ideal of } \z[\sqrt{-5}]\}. \] Two ideals $I,J \in \mathcal{I}$ are \emph{equivalent} if $\langle \alpha \rangle I = \langle \beta \rangle J$ for some $\alpha, \beta \in \z[\sqrt{-5}]^\bullet$. In this case, we write $I \sim J$. It is clear that $\sim$ defines an equivalence relation on $\z[\sqrt{-5}]$. The equivalence classes of $\sim$ are called \emph{ideal classes}. Let $I\mathcal{P}$ denote the ideal class of $I$, and we also let $\mathcal{C}(\z[\sqrt{-5}])$ denote the set of all nonzero ideal classes. Now define a binary operation $\ast$ on $\mathcal{C}(\z[\sqrt{-5}])$ by \[ I \mathcal{P} \ast J\mathcal{P} = (IJ)\mathcal{P}. \] It turns out that $\mathcal{C}(\z[\sqrt{-5}])$ is, indeed, a group under the $\ast$ operation. \begin{theorem} The set of ideal classes $\mathcal{C}(\z[\sqrt{-5}])$ is an abelian group under $\ast$. \end{theorem} \begin{proof} Because the product of ideals is associative and commutative, so is $\ast$. Also, it follows immediately that $\langle 1 \rangle \mathcal P \ast I \mathcal P = (\langle 1 \rangle I) \mathcal P = I \mathcal P$ for each $I \in \mathcal{I}^\bullet$, which means that $\mathcal P = \langle 1 \rangle \mathcal{P}$ is the identity element of $\mathcal{C}(\z[\sqrt{-5}])$. In addition, as any two nonzero principal ideals are in the same ideal class, Theorem~\ref{inverse} guarantees that, for any $I \mathcal P \in \mathcal C (\z[\sqrt{-5}])$, there exists $J \in \mathcal{I}^\bullet$ such that $I\mathcal{P} \ast J\mathcal{P} = IJ \in \mathcal P = \langle 1 \rangle \mathcal{P}$. So $J \mathcal P$ is the inverse of $I \mathcal P$ in $\mathcal{C}(\z[\sqrt{-5}])$. Hence, $\mathcal C (\z[\sqrt{-5}])$ is an abelian group. \end{proof} \medskip \begin{definition} The group $\mathcal{C}(\z[\sqrt{-5}])$ is called the \emph{class group} of $\z[\sqrt{-5}]$, and the order of $\mathcal{C}(\z[\sqrt{-5}])$ is called the \emph{class number} of $\z[\sqrt{-5}]$. \end{definition} \medskip Recall that if $\theta \colon R \to S$ is a ring homomorphism, then $\ker \theta = \{r \in R \mid \theta(r) = 0\}$ is an ideal of $R$. Moreover, the First Isomorphism Theorem for rings states that $R/\ker \theta \cong \theta(R)$. \begin{definition}\label{idealnorm} Let $K$ be an algebraic number field. For any nonzero ideal $I$ of $\mathcal{O}_K$, the cardinality $|\mathcal{O}_K/I|$ is called the \emph{norm} of $I$ and is denoted by $N(I)$. \end{definition} \medskip \begin{proposition} \label{ are only finitely many residue classes} Let $d \in \mathbb{Z} \setminus \{0,1\}$ be a squarefree integer with $d \equiv 2,3 \pmod{4}$. Then $N(I)$ is finite for all nonzero ideals $I$ of $\mathbb{Z}[\sqrt{d}]$. \end{proposition} \begin{proof} Take $n = \alpha \bar{\alpha}$ for any nonzero $\alpha \in I$. Then $n \in I \cap \n$. As $\langle n \rangle \subseteq I$, it follows that $|\mathbb{Z}[\sqrt{d}]/I| \le |\mathbb{Z}[\sqrt{d}]/\langle n \rangle|$. In addition, each element of $\mathbb{Z}[\sqrt{d}]/\langle n \rangle$ has a representative $n_1 + n_2\sqrt{d}$ with $n_1, n_2 \in \{0,1,\dots, n-1\}$. Hence, $\mathbb{Z}[\sqrt{d}]/\langle n \rangle$ is finite and, therefore, $N(I) = |\mathbb{Z}[\sqrt{d}]/I| < \infty$. \end{proof} \medskip As ideal norms generalize the notion of standard norms given in (\ref{ norm}), we expect they satisfy some similar properties. Indeed, this is the case. \begin{exercise} \label{ of quotients} Let $I$ and $P$ be a nonzero ideal and a nonzero prime ideal of $\z[\sqrt{-5}]$, respectively. Show that $|\z[\sqrt{-5}]/P| = |I/IP|$. \end{exercise} \medskip \begin{proposition} \label{ is multiplicative} $N(IJ) = N(I)N(J)$ for all $I,J \in \mathcal{I}^\bullet$. \end{proposition} \begin{proof} By factoring $J$ as the product of prime ideals (Theorem~\ref{FTIT}) and applying induction on the number of factors, we can assume that $J$ is a prime ideal. Consider the ring homomorphism $\theta \colon \z[\sqrt{-5}]/IJ \to \z[\sqrt{-5}]/I$ defined by $\theta(\alpha + IJ) = \alpha + I$. It follows immediately that $\theta$ is surjective and $\ker \theta = \{\alpha + IJ \mid \alpha \in I\}$. Therefore \[ \frac{\z[\sqrt{-5}]/IJ}{I/IJ} \cong \z[\sqrt{-5}]/I \] by the First Isomorphism Theorem. As $IJ$ is nonzero, $|\z[\sqrt{-5}]/IJ| = N(IJ)$ is finite and so $|\z[\sqrt{-5}]/IJ| = |\z[\sqrt{-5}]/I| \cdot |I/IJ|$. Since $J$ is prime, we can use Exercise~\ref{ of quotients} to conclude that \begin{align*} N(IJ) &= |\z[\sqrt{-5}]/IJ| = |\z[\sqrt{-5}]/I| \cdot |I/IJ|\\ &= |\z[\sqrt{-5}]/I| \cdot |\z[\sqrt{-5}]/J| = N(I)N(J). \end{align*} \end{proof} \medskip \begin{corollary} \label{ norm implies prime ideal} If $N(I)$ is prime for some $I \in \mathcal{I}^\bullet$, then $I$ is a prime ideal. \end{corollary} \medskip Let us verify now that the ideal norm is consistent with the standard norm on principal ideals. \begin{proposition} \label{ ideal of a principal ideal} $N(\langle \alpha \rangle) = N(\alpha)$ for all $\alpha \in \z[\sqrt{-5}]^\bullet$. \end{proposition} \begin{proof} Set $S = \big\{a + b\sqrt{-5} \mid a,b \in \{0,1, \dots, n-1\} \big\}$. Clearly, $|S| = n^2$. In addition, \[ \z[\sqrt{-5}]/ \langle n \rangle = \{s + \langle n \rangle \mid s \in S\}. \] Note that if $s + \langle n \rangle = s' + \langle n \rangle$ for $s,s' \in S$, then we have $s = s'$. As a consequence, $N(\langle n \rangle) = n^2 = N(n)$ for each $n \in \n$. It can also be readily verified that the map $\theta \colon \z[\sqrt{-5}] \to \z[\sqrt{-5}]/ \langle \bar{\alpha} \rangle$ defined by $\theta(x) = \bar{x} + \langle \bar{\alpha} \rangle$ is a surjective ring homomorphism with $\ker \theta = \langle \alpha \rangle$. Therefore the rings $\z[\sqrt{-5}]/ \langle \alpha \rangle$ and $\z[\sqrt{-5}]/ \langle \bar{\alpha} \rangle$ are isomorphic by the First Isomorphism Theorem. This implies that $N(\langle \alpha \rangle) = N(\langle \bar{\alpha} \rangle)$. Because $\alpha \bar{\alpha} \in \n$, using Proposition~\ref{ is multiplicative}, one obtains \[ N(\langle \alpha \rangle) = \sqrt{N(\langle \alpha \rangle)N(\langle\bar{\alpha} \rangle)} = \sqrt{N(\langle \alpha \bar{\alpha} \rangle)} = \alpha \bar{\alpha} = N(\alpha). \] \end{proof} \medskip \begin{lemma} \label{ ideal as a factor of principal integer ideal} If $P$ is a nonzero prime ideal in $\z[\sqrt{-5}]$, then $P$ divides exactly one ideal $\langle p \rangle$, where $p$ is a prime number. \end{lemma} \begin{proof} For $\alpha \in P^\bullet$, it follows that $z = \alpha \bar{\alpha} \in P \cap \n$. Then, writing $z = p_1 \cdots p_k$ for some prime numbers $p_1, \dots, p_k$, we get $\langle z \rangle = \langle p_1 \rangle \cdots \langle p_k \rangle$. As $\langle p_1 \rangle \cdots \langle p_k \rangle \subseteq P$, we have that $\langle p_i \rangle \subseteq P$ for some $i \in \{1, \dots, k\}$ (Exercise~\ref{ of ideals inside a prime ideal}). As $p_i \in P^\bullet$, Corollary~\ref{cor:an ideal divides any principal ideal it contains} ensures that $P$ divides $\langle p_i \rangle$. For the uniqueness, note that if $P$ divides $\langle p \rangle$ and $\langle p' \rangle$ for distinct primes $p$ and $p'$, then the fact that $mp + np' = 1$ for some $m,n \in \mathbb{Z}$ would imply that $P$ divides the full ideal $\langle 1 \rangle = \z[\sqrt{-5}]$, a contradiction. \\ \end{proof} \medskip \begin{theorem} The class group of $\z[\sqrt{-5}]$ is $\mathbb{Z}_2$. \end{theorem} \begin{proof} First, we verify that every nonzero ideal $I$ of $\z[\sqrt{-5}]$ contains a nonzero element $\alpha$ with $N(\alpha) \le 6 N(I)$. For $I \in \mathcal{I}^\bullet$, take $B = \lfloor \sqrt{N(I)} \rfloor$ and define \[ S_I := \big\{ a + b \sqrt{-5} \mid a,b \in \{0,1, \dots, B\} \big\} \subsetneq \z[\sqrt{-5}]. \] Observe that $|S_I| = (B+1)^2 > N(I)$. Thus, there exist $\alpha_1 = a_1 + b_1 \sqrt{-5} \in S_I$ and $\alpha_2 = a_2 + b_2 \sqrt{-5} \in S_I$ such that $\alpha = \alpha_1 - \alpha_2 \in I \setminus \{0\}$ and \[ N(\alpha) = (a_1 - a_2)^2 + 5(b_1 - b_2)^2 \le 6B^2 \le 6 N(I). \] Now, let $I \mathcal{P}$ be a nonzero ideal class of $\z[\sqrt{-5}]$. Take $J \in \mathcal{I}^\bullet$ satisfying $IJ \mathcal{P} = \mathcal{P}$. By the argument given in the previous paragraph, there exists $\beta \in J^\bullet$ such that $N(\beta) \le 6N(J)$. By Corollary~\ref{cor:an ideal divides any principal ideal it contains}, there exists an ideal $K \in \mathcal{I}^\bullet$ such that $JK = \langle \beta \rangle$. Using Proposition~\ref{ is multiplicative} and Proposition~\ref{ ideal of a principal ideal}, one obtains \[ N(J)N(K) = N(\langle \beta \rangle) = N(\beta) \le 6 N(J), \] which implies that $N(K) \le 6$. Because $KJ \sim IJ$ (they are both principal), it follows that $K \in I\mathcal{P}$. Hence, every nonzero ideal class of $\z[\sqrt{-5}]$ contains an ideal whose norm is at most $6$. To show that the class group of $\z[\sqrt{-5}]$ is $\mathbb{Z}_2$, let us first determine the congruence relations among ideals of norm at most $6$. Every ideal $P$ of norm $p \in \{2,3,5\}$ must be prime by Corollary~\ref{ norm implies prime ideal}. Moreover, by Lemma~\ref{ ideal as a factor of principal integer ideal}, Theorem~\ref{FTIT}, and Proposition~\ref{ is multiplicative}, the ideal $P$ must show in the prime factorization \begin{align} \label{factorization of (p)} \langle p \rangle = P_1^{n_1} \cdots P_k^{n_k} \end{align} of the ideal $\langle p \rangle$. The following ideal factorizations have been already verified in Example~\ref{ of <2> into prime ideals} and Exercise~\ref{ of <3> into prime ideals}: \begin{align} \label{ of ideals 1} \langle 2 \rangle &= \langle 2, 1 + \sqrt{-5} \rangle^2, \nonumber \\ \langle 3 \rangle &= \langle 3, 1 - 2\sqrt{-5} \rangle \langle 3, 1 + 2\sqrt{-5} \rangle, \\ \langle 5 \rangle &= \langle \sqrt{-5} \rangle^2. \nonumber \end{align} In addition, we have proved in Example~\ref{ ideal dividing <2>} and Exercise~\ref{ ideals dividing <3>} that the ideals on the right-hand side of the first two equalities in~(\ref{ of ideals 1}) are prime. Also, $N(\langle \sqrt{-5} \rangle) = N(\sqrt{-5}) = 5$ implies that the ideal $\langle \sqrt{-5} \rangle$ is prime. It follows now by the uniqueness of Theorem~\ref{FTIT} that the ideals on the right-hand side of the equalities~(\ref{ of ideals 1}) are the only ideals of $\z[\sqrt{-5}]$ having norm in the set $\{2,3,5\}$. Once again, combining Lemma~\ref{ ideal as a factor of principal integer ideal}, Theorem~\ref{FTIT}, and Proposition~\ref{ is multiplicative}, we obtain that any ideal $I$ whose norm is $4$ must be a product of prime ideals dividing $\langle 2 \rangle$, which forces $I = \langle 2 \rangle$. Similarly, any ideal $J$ with norm $6$ must be the product of ideals dividing the ideals $\langle 2 \rangle$ and $\langle 3 \rangle$. The reader can readily verify that, \begin{align} \langle 1 - \sqrt{-5} \rangle &= \langle 2, 1 + \sqrt{-5} \rangle \langle 3, 1 + 2\sqrt{-5} \rangle \label{ of ideals 2} \\ \langle 1 + \sqrt{-5} \rangle &= \langle 2, 1 + \sqrt{-5} \rangle \langle 3, 1 - 2\sqrt{-5} \rangle. \label{ of ideals 3} \end{align} Therefore $\langle 1 - \sqrt{-5} \rangle$ and $\langle 1 + \sqrt{-5} \rangle$ are the only two ideals having norm $6$. Now since we know all ideals of $\z[\sqrt{-5}]$ with norm at most 6, it is not difficult to check that $|\mathcal{C}(\z[\sqrt{-5}])| = 2$. Because each principal ideal of $\z[\sqrt{-5}]$ represents the identity ideal class $\mathcal{P}$, we find that \[ \langle 1 \rangle \mathcal{P} = \langle 2 \rangle \mathcal{P} = \langle \sqrt{-5} \rangle \mathcal{P} = \langle 1 - \sqrt{-5} \rangle \mathcal{P}. \] On the other hand, we have seen that the product of $\langle 2, 1 + \sqrt{-5} \rangle$ and each of the three nonprincipal ideals with norm at most $6$ is a principal ideal. Thus, \[ \langle 2, 1 + \sqrt{-5} \rangle \mathcal{P} = \langle 3, 1 + 2 \sqrt{-5}\rangle \mathcal{P} = \langle 3, 1 - 2\sqrt{-5} \rangle \mathcal{P}. \] Since there are only two ideal classes, $\mathcal{C}(\z[\sqrt{-5}]) = \mathbb{Z}_2$. \end{proof} \medskip \begin{exercise} Verify the equalities (\ref{ of ideals 2}), and (\ref{ of ideals 3}). \end{exercise} \medskip From this observation, we deduce an important property of the ideals of $\z[\sqrt{-5}]$. \begin{corollary}\label{multcorollary} If $I, J \in \mathcal{I}^\bullet$ are not principal, then $IJ$ is principal. \end{corollary} \medskip \section{Half-factoriality} \label{} The class group, in tandem with The Fundamental Theorem of Ideal Theory, will allow us to determine exactly what elements of $\z[\sqrt{-5}]$ are irreducible. \begin{proposition}\label{irreducibles} Let $\alpha$ be a nonzero nonunit element in $\z[\sqrt{-5}]$. Then $\alpha$ is irreducible in $\z[\sqrt{-5}]$ if and only if \begin{enumerate} \vspace{1pt} \item $\langle \alpha \rangle$ is a prime ideal in $\z[\sqrt{-5}]$ (and hence $\alpha$ is a prime element), or \vspace{3pt} \item $\langle \alpha \rangle = P_1 P_2$ where $P_1$ and $P_2$ are nonprincipal prime ideals of $\z[\sqrt{-5}]$. \end{enumerate} \end{proposition} \begin{proof} ($\Rightarrow$) Suppose $\alpha$ is irreducible in $\z[\sqrt{-5}]$. If $\langle \alpha \rangle$ is a prime ideal, then we are done. Assume $\langle \alpha \rangle$ is not a prime ideal. Then by Theorem~\ref{FTIT} there are prime ideals $P_1, \dots, P_k$ of $\z[\sqrt{-5}]$ with $\langle \alpha \rangle = P_1 \cdots P_k$ for some $k \geq 2$. Suppose that one of the $P_i$'s is a principal ideal. Without loss of generality, assume that $P_1 = \langle \beta \rangle$ for some prime $\beta$ in $\z[\sqrt{-5}]$. Using the class group, $P_2 \cdots P_k = \langle \gamma \rangle$, where $\gamma$ is a nonzero nonunit of $\z[\sqrt{-5}]$. Thus, $\langle \alpha \rangle = \langle \beta \rangle \langle \gamma \rangle$ implies that $\alpha = (u \beta) \gamma$ for some unit $u$ of $\z[\sqrt{-5}]$. This contradicts the irreducibility of $\alpha$ in $\z[\sqrt{-5}]$. Therefore all the $P_i$'s are nonprincipal. Since the class group of $\z[\sqrt{-5}]$ is $\mathbb{Z}_2$, it follows that $k$ is even. Now suppose that $k > 2$. Using Corollary~\ref{multcorollary} and proceeding in a manner similar to the previous argument, $P_1 P_2 = \langle \beta \rangle$ and $P_3 \cdots P_k = \langle \gamma \rangle$, and again $\alpha = u \beta \gamma$ for some unit $u$, which contradicts the irreducibility of $\alpha$. Hence, either $k=1$ and $\alpha$ is a prime element, or $k=2$. ($\Leftarrow$) If $\langle \alpha \rangle$ is a prime ideal, then $\alpha$ is prime and so irreducible. Then suppose that $\langle \alpha \rangle = P_1 P_2$, where $P_1$ and $P_2$ are nonprincipal prime ideals of $\z[\sqrt{-5}]$. Let $\alpha = \beta \gamma$ for some $\beta, \gamma \in \z[\sqrt{-5}]$, and assume, without loss of generality, that $\beta$ is a nonzero nonunit of $\z[\sqrt{-5}]$. Notice that $\langle \beta \gamma \rangle = \langle \beta \rangle \langle \gamma \rangle = P_1 P_2$. Because $P_1$ and $P_2$ are nonprincipal ideals, $\langle \beta \rangle \notin \{P_1, P_2\}$. As a consequence of Theorem~\ref{FTIT}, we have that $\langle \beta \rangle = P_1 P_2$. This forces $\langle \gamma \rangle = \langle 1 \rangle$, which implies that $\gamma \in \{\pm 1\}$. Thus, $\alpha$ is irreducible. \end{proof} \medskip Let us use Proposition~\ref{irreducibles} to analyze the factorizations presented in (\ref{2ndzroot5}) at the beginning of the exposition. As the product of any two nonprincipal ideals of $\z[\sqrt{-5}]$ is a principal ideal, the decompositions \begin{align*} \langle 6 \rangle &= \langle 2 \rangle \langle 3 \rangle = \langle 2,\; 1 + \sqrt{-5} \rangle^2 \langle 3,\; 1 - \sqrt{-5} \rangle \langle 3,\; 1 + \sqrt{-5}\rangle \\ &= \langle 2,\; 1+\sqrt{-5} \rangle \langle 3,\; 1 + \sqrt{-5} \rangle \langle 2,\; 1 + \sqrt{-5} \rangle \langle 3,\; 1 - \sqrt{-5} \rangle \\ &= \langle 1 + \sqrt{-5} \rangle \langle 1 - \sqrt{-5} \rangle \end{align*} yield that $2 \cdot 3$ and $(1 + \sqrt{-5})(1 - \sqrt{-5})$ are the only two irreducible factorizations of $6$ in $\z[\sqrt{-5}]$. Thus, any two irreducible factorizations of $6$ in $\z[\sqrt{-5}]$ have the same factorization length. We can take this observation a step further. \begin{theorem} \label{zroot5hf} If $\alpha$ is a nonzero nonunit of $\z[\sqrt{-5}]$ and $\beta_1, \dots, \beta_s, \gamma_1, \dots, \gamma_t$ are irreducible elements of $\z[\sqrt{-5}]$ with $\alpha = \beta_1 \cdots \beta_s = \gamma_1 \cdots \gamma_t$, then $s=t$. \end{theorem} \begin{proof} Let $\alpha = \omega_1 \cdots \omega_m$ be a factorization of $\alpha$ in $\z[\sqrt{-5}]$ into irreducible elements. By Theorem~\ref{FTIT}, there are unique prime ideals $P_1, \dots, P_k$ in $\z[\sqrt{-5}]$ satisfying that $\langle \alpha \rangle = P_1 \cdots P_k$. Suppose that exactly $d$ of these prime ideals are principal and assume, without loss, that $P_i = \langle \alpha_i \rangle$ for all $i \in \{1,\dots,d\}$, where each $\alpha_i$ is prime in $\z[\sqrt{-5}]$. Since the class group of $\z[\sqrt{-5}]$ is $\mathbb{Z}_2$, there exists $n \in \n$ such that $k - d = 2n$. Hence, \[ \langle \alpha \rangle = \left( P_1 \cdots P_d \right) \left(P_{d+1} \cdots P_k \right) = \langle \alpha_1 \cdots \alpha_d \rangle \left(P_{d+1} \cdots P_k \right), \] and any factorization into irreducibles of $\alpha$ will be of the form $u \alpha_1 \cdots \alpha_d \cdot \beta_1 \cdots \beta_n$, where each ideal $\langle \beta_j \rangle$ is the product of two ideals chosen from $P_{d+1}, \dots, P_k$. As a result, $m = d+n$ and, clearly, $s = t = m$, completing the proof. \end{proof} \medskip Thus, while some elements of $\z[\sqrt{-5}]$ admit many factorizations into irreducibles, the number of irreducible factors in any two factorizations of a given element is the same. As we mentioned in the introduction, this phenomenon is called half-factoriality. Since the concept of half-factoriality does not involve the addition of $\z[\sqrt{-5}]$, it can also be defined for commutative monoids. \begin{definition} An atomic monoid $M$ is called \emph{half-factorial} if any two factorizations of each nonzero nonunit element of $M$ have the same number of irreducible factors. \end{definition} \medskip Half-factorial domains and monoids have been systematically studied since the 1950's, when Carlitz gave a characterization theorem of half-factorial rings of integers, which generalizes the case of $\z[\sqrt{-5}]$ considered in this exposition. \begin{theorem}[Carlitz \cite{Ca}] Let $R$ be the ring of integers in a finite extension field of $\q$. Then $R$ is half-factorial if and only if $R$ has class number less than or equal to two. \end{theorem} \medskip A list of factorization inspired characterizations of class number two can be found in \cite{Ch2}. In addition, a few families of half-factorial domains in a more general setting are presented in \cite{Ki}. We will conclude this paper by exhibiting two simple examples of half-factorial monoids, using the second one to illustrate how to compute the number of factorizations in $\z[\sqrt{-5}]$ of a given element. \begin{example}[Hilbert monoid] \label{ monoid} It is easily seen that \[ H = \{1 + 4k \mid k \in \n_0\} \] is a multiplicative submonoid of $\n$. The monoid $H$ is called \emph{Hilbert monoid}. It is not hard to verify (Exercise~\ref{ of Hilbert monoid}) that the irreducible elements of $H$ are \begin{enumerate} \item the prime numbers $p$ satisfying $p \equiv 1 \pmod{4}$ and \vspace{3pt} \item $p_1p_2$, where $p_1$ and $p_2$ are prime numbers satisfying $p_i \equiv 3 \pmod{4}$. \end{enumerate} Therefore every element of $H$ is a product of irreducibles. Also, in the factorization of any element of $H$ into primes, there must be an even number of prime factors congruent to $3$ modulo $4$. Hence, any factorization of an element $x \in H$ comes from pairing the prime factors of $x$ that are congruent to $3$ modulo $4$. This implies that $H$ is half-factorial. For instance, $x = 5^2 \cdot 3^2 \cdot 11 \cdot 13 \cdot 19$ has exactly two factorizations into irreducibles, each of them contains five factors: \begin{align*} x = 5^2 \cdot 13 \cdot (3^2) \cdot (11 \cdot 19) = 5^2 \cdot 13 \cdot (3 \cdot 11) \cdot (3 \cdot 19). \end{align*} \end{example} \medskip \begin{exercise} \label{ of Hilbert monoid} Argue that the irreducible elements of the Hilbert monoid are precisely those described in Example~\ref{ monoid}. \end{exercise} \medskip \begin{definition} Let $p$ be a prime number. \begin{enumerate} \item We say that $p$ is \emph{inert} if $\langle p \rangle$ is a prime ideal in $\z[\sqrt{-5}]$. \vspace{3pt} \item We say that $p$ is \emph{ramified} if $\langle p \rangle = P^2$ for some prime ideal $P$ in $\z[\sqrt{-5}]$. \vspace{3pt} \item We say that $p$ \emph{splits} if $\langle p \rangle = PP'$ for two distinct prime ideals in $\z[\sqrt{-5}]$. \end{enumerate} \end{definition} \medskip Prime numbers $p$ can be classified according to the above definition. Indeed, we have seen that $p$ is ramified when $p \in \{2,5\}$. On the other hand, it is also known that $p$ splits if $p \equiv 1, 3, 7, 9 \pmod{20}$ and is inert if $p \not\equiv 1, 3, 7, 9 \pmod{20}$ (except $2$ and $5$). A proof of this result is given in \cite{Ma}. \begin{example} When $n \ge 2$, the submonoid $\mathbb{X}_n$ of the additive monoid $\n_0^{n+1}$ given by \[ \mathbb{X}_n = \{(x_1, \dots, x_{n+1}) \mid x_i \in \mathbb{N}_0 \ \text{ and } \ x_1 + \dots + x_n = x_{n+1}\} \] is a half-factorial Krull monoid with divisor class group $\mathbb{Z}_2$ (see \cite[Section~2]{CKO} for more details). Following \cite{CHR}, we will use $\mathbb{X}_n$ to count the number of distinct factorizations into irreducibles of a given nonzero nonunit $\alpha \in \z[\sqrt{-5}]$. Let \[ \langle \alpha \rangle = P_1^{n_1} \cdots P_k^{n_k}Q_1^{m_1} \cdots Q_t^{m_t}, \] where the $P_i$'s are distinct prime ideals in the trivial class ideal of $\z[\sqrt{-5}]$, the $Q_j$'s are distinct prime ideals in the nontrivial class ideal of $\z[\sqrt{-5}]$, and $m_1 \le \dots \le m_t$. Then the desired number of factorizations $\eta(\alpha)$ of $\alpha$ in $\z[\sqrt{-5}]$ is given by \[ \eta(\alpha) = \eta_{\mathbb{X}_t}\bigg( m_1, \dots, m_t, \frac{m_1 + \dots + m_t}2 \bigg), \] which, when $t = 3$, can be computed by the formula \[ \eta_{\mathbb{X}_3}(x_1, x_2, x_3, x_4) = \sum_{j=0}^{\lfloor x_1/2 \rfloor} \sum_{k=0}^{x_1 - 2j} \bigg(\bigg\lfloor \frac{\min\{x_2 - k, x_3 - x_1 + 2j + k\}}{2} \bigg\rfloor + 1\bigg). \] For instance, let us find how many factorizations $1980 = 2^2 \cdot 3^2 \cdot 5 \cdot 11$ has in $\z[\sqrt{-5}]$. We have seen that $5$ ramifies as $\langle 5 \rangle = P_1^2$, where $P_1$ is principal. As $11$ is inert, $P_2 = \langle 11 \rangle$ is prime. In addition, $3$ splits as $\langle 3 \rangle = Q_1 Q_2$, where $Q_1$ and $Q_2$ are nonprincipal. Finally, $2$ ramifies as $\langle 2 \rangle = Q_3^2$, where $Q_3$ is nonprincipal. Therefore one has that $\langle 1980 \rangle = P_1^2 P_2 Q_1^2Q_2^2Q_3^4$, and so \[ \eta(1980) = \eta_{\mathbb{X}_3}(2,2,4,4) = \sum_{j=0}^1 \sum_{k=0}^{2-2j} \bigg(\bigg\lfloor \frac{\min\{2 - k, 2 + 2j + k\}}{2} \bigg\rfloor + 1\bigg) = 6. \] \end{example} \medskip \section*{Acknowledgements} It is a pleasure for the authors to thank the referee, whose helpful suggestions vastly improved the final version of this paper.